Linux-Fsdevel Archive on lore.kernel.org
 help / color / Atom feed
* [PATCH v3 0/5] kvm "virtio pmem" device
@ 2019-01-09 14:47 Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 1/5] libnvdimm: nd_region flush callback support Pankaj Gupta
                   ` (6 more replies)
  0 siblings, 7 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

 This patch series has implementation for "virtio pmem". 
 "virtio pmem" is fake persistent memory(nvdimm) in guest 
 which allows to bypass the guest page cache. This also
 implements a VIRTIO based asynchronous flush mechanism.  
 
 Sharing guest kernel driver in this patchset with the 
 changes suggested in v2. Tested with Qemu side device 
 emulation for virtio-pmem [6]. 
 
 Details of project idea for 'virtio pmem' flushing interface 
 is shared [3] & [4].

 Implementation is divided into two parts:
 New virtio pmem guest driver and qemu code changes for new 
 virtio pmem paravirtualized device.

1. Guest virtio-pmem kernel driver
---------------------------------
   - Reads persistent memory range from paravirt device and 
     registers with 'nvdimm_bus'.  
   - 'nvdimm/pmem' driver uses this information to allocate 
     persistent memory region and setup filesystem operations 
     to the allocated memory. 
   - virtio pmem driver implements asynchronous flushing 
     interface to flush from guest to host.

2. Qemu virtio-pmem device
---------------------------------
   - Creates virtio pmem device and exposes a memory range to 
     KVM guest. 
   - At host side this is file backed memory which acts as 
     persistent memory. 
   - Qemu side flush uses aio thread pool API's and virtio 
     for asynchronous guest multi request handling. 

   David Hildenbrand CCed also posted a modified version[6] of 
   qemu virtio-pmem code based on updated Qemu memory device API. 

 Virtio-pmem errors handling:
 ----------------------------------------
  Checked behaviour of virtio-pmem for below types of errors
  Need suggestions on expected behaviour for handling these errors?

  - Hardware Errors: Uncorrectable recoverable Errors: 
  a] virtio-pmem: 
    - As per current logic if error page belongs to Qemu process, 
      host MCE handler isolates(hwpoison) that page and send SIGBUS. 
      Qemu SIGBUS handler injects exception to KVM guest. 
    - KVM guest then isolates the page and send SIGBUS to guest 
      userspace process which has mapped the page. 
  
  b] Existing implementation for ACPI pmem driver: 
    - Handles such errors with MCE notifier and creates a list 
      of bad blocks. Read/direct access DAX operation return EIO 
      if accessed memory page fall in bad block list.
    - It also starts backgound scrubbing.  
    - Similar functionality can be reused in virtio-pmem with MCE 
      notifier but without scrubbing(no ACPI/ARS)? Need inputs to 
      confirm if this behaviour is ok or needs any change?

Changes from PATCH v2: [1]
- Disable MAP_SYNC for ext4 & XFS filesystems - [Dan] 
- Use name 'virtio pmem' in place of 'fake dax' 

Changes from PATCH v1: [2]
- 0-day build test for build dependency on libnvdimm 

 Changes suggested by - [Dan Williams]
- Split the driver into two parts virtio & pmem  
- Move queuing of async block request to block layer
- Add "sync" parameter in nvdimm_flush function
- Use indirect call for nvdimm_flush
- Don’t move declarations to common global header e.g nd.h
- nvdimm_flush() return 0 or -EIO if it fails
- Teach nsio_rw_bytes() that the flush can fail
- Rename nvdimm_flush() to generic_nvdimm_flush()
- Use 'nd_region->provider_data' for long dereferencing
- Remove virtio_pmem_freeze/restore functions
- Remove BSD license text with SPDX license text

- Add might_sleep() in virtio_pmem_flush - [Luiz]
- Make spin_lock_irqsave() narrow

Changes from RFC v3
- Rebase to latest upstream - Luiz
- Call ndregion->flush in place of nvdimm_flush- Luiz
- kmalloc return check - Luiz
- virtqueue full handling - Stefan
- Don't map entire virtio_pmem_req to device - Stefan
- request leak, correct sizeof req- Stefan
- Move declaration to virtio_pmem.c

Changes from RFC v2:
- Add flush function in the nd_region in place of switching
  on a flag - Dan & Stefan
- Add flush completion function with proper locking and wait
  for host side flush completion - Stefan & Dan
- Keep userspace API in uapi header file - Stefan, MST
- Use LE fields & New device id - MST
- Indentation & spacing suggestions - MST & Eric
- Remove extra header files & add licensing - Stefan

Changes from RFC v1:
- Reuse existing 'pmem' code for registering persistent 
  memory and other operations instead of creating an entirely 
  new block driver.
- Use VIRTIO driver to register memory information with 
  nvdimm_bus and create region_type accordingly. 
- Call VIRTIO flush from existing pmem driver.

Pankaj Gupta (5):
   libnvdimm: nd_region flush callback support
   virtio-pmem: Add virtio-pmem guest driver
   libnvdimm: add nd_region buffered dax_dev flag
   ext4: disable map_sync for virtio pmem
   xfs: disable map_sync for virtio pmem

[2] https://lkml.org/lkml/2018/8/31/407
[3] https://www.spinics.net/lists/kvm/msg149761.html
[4] https://www.spinics.net/lists/kvm/msg153095.html  
[5] https://lkml.org/lkml/2018/8/31/413
[6] https://marc.info/?l=qemu-devel&m=153555721901824&w=2

 drivers/acpi/nfit/core.c         |    4 -
 drivers/dax/super.c              |   17 +++++
 drivers/nvdimm/claim.c           |    6 +
 drivers/nvdimm/nd.h              |    1 
 drivers/nvdimm/pmem.c            |   15 +++-
 drivers/nvdimm/region_devs.c     |   45 +++++++++++++-
 drivers/nvdimm/virtio_pmem.c     |   84 ++++++++++++++++++++++++++
 drivers/virtio/Kconfig           |   10 +++
 drivers/virtio/Makefile          |    1 
 drivers/virtio/pmem.c            |  125 +++++++++++++++++++++++++++++++++++++++
 fs/ext4/file.c                   |   11 +++
 fs/xfs/xfs_file.c                |    8 ++
 include/linux/dax.h              |    9 ++
 include/linux/libnvdimm.h        |   11 +++
 include/linux/virtio_pmem.h      |   60 ++++++++++++++++++
 include/uapi/linux/virtio_ids.h  |    1 
 include/uapi/linux/virtio_pmem.h |   10 +++
 17 files changed, 406 insertions(+), 12 deletions(-)

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 1/5] libnvdimm: nd_region flush callback support
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
@ 2019-01-09 14:47 ` Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 2/5] virtio-pmem: Add virtio pmem driver Pankaj Gupta
                   ` (5 subsequent siblings)
  6 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

This patch adds functionality to perform flush from guest
to host over VIRTIO. We are registering a callback based
on 'nd_region' type. virtio_pmem driver requires this special
flush function. For rest of the region types we are registering
existing flush function. Report error returned by host fsync
failure to userspace.

This also handles asynchronous flush requests from the block layer
by creating a child bio and chaining it with parent bio.

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 drivers/acpi/nfit/core.c     |  4 ++--
 drivers/nvdimm/claim.c       |  6 ++++--
 drivers/nvdimm/nd.h          |  1 +
 drivers/nvdimm/pmem.c        | 12 ++++++++----
 drivers/nvdimm/region_devs.c | 38 ++++++++++++++++++++++++++++++++++++--
 include/linux/libnvdimm.h    |  5 ++++-
 6 files changed, 55 insertions(+), 11 deletions(-)

diff --git a/drivers/acpi/nfit/core.c b/drivers/acpi/nfit/core.c
index b072cfc..f154852 100644
--- a/drivers/acpi/nfit/core.c
+++ b/drivers/acpi/nfit/core.c
@@ -2234,7 +2234,7 @@ static void write_blk_ctl(struct nfit_blk *nfit_blk, unsigned int bw,
 		offset = to_interleave_offset(offset, mmio);
 
 	writeq(cmd, mmio->addr.base + offset);
-	nvdimm_flush(nfit_blk->nd_region);
+	nvdimm_flush(nfit_blk->nd_region, NULL, false);
 
 	if (nfit_blk->dimm_flags & NFIT_BLK_DCR_LATCH)
 		readq(mmio->addr.base + offset);
@@ -2283,7 +2283,7 @@ static int acpi_nfit_blk_single_io(struct nfit_blk *nfit_blk,
 	}
 
 	if (rw)
-		nvdimm_flush(nfit_blk->nd_region);
+		nvdimm_flush(nfit_blk->nd_region, NULL, false);
 
 	rc = read_blk_stat(nfit_blk, lane) ? -EIO : 0;
 	return rc;
diff --git a/drivers/nvdimm/claim.c b/drivers/nvdimm/claim.c
index fb667bf..a1dfa06 100644
--- a/drivers/nvdimm/claim.c
+++ b/drivers/nvdimm/claim.c
@@ -263,7 +263,7 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,
 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 	unsigned int sz_align = ALIGN(size + (offset & (512 - 1)), 512);
 	sector_t sector = offset >> 9;
-	int rc = 0;
+	int rc = 0, ret = 0;
 
 	if (unlikely(!size))
 		return 0;
@@ -301,7 +301,9 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,
 	}
 
 	memcpy_flushcache(nsio->addr + offset, buf, size);
-	nvdimm_flush(to_nd_region(ndns->dev.parent));
+	ret = nvdimm_flush(to_nd_region(ndns->dev.parent), NULL, false);
+	if (ret)
+		rc = ret;
 
 	return rc;
 }
diff --git a/drivers/nvdimm/nd.h b/drivers/nvdimm/nd.h
index 98317e7..d53a2d1 100644
--- a/drivers/nvdimm/nd.h
+++ b/drivers/nvdimm/nd.h
@@ -160,6 +160,7 @@ struct nd_region {
 	struct nd_interleave_set *nd_set;
 	struct nd_percpu_lane __percpu *lane;
 	struct nd_mapping mapping[0];
+	int (*flush)(struct nd_region *nd_region);
 };
 
 struct nd_blk_region {
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 6071e29..5d6a4a1 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -192,6 +192,7 @@ static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
 
 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 {
+	int ret = 0;
 	blk_status_t rc = 0;
 	bool do_acct;
 	unsigned long start;
@@ -201,7 +202,7 @@ static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 	struct nd_region *nd_region = to_region(pmem);
 
 	if (bio->bi_opf & REQ_PREFLUSH)
-		nvdimm_flush(nd_region);
+		ret = nvdimm_flush(nd_region, bio, true);
 
 	do_acct = nd_iostat_start(bio, &start);
 	bio_for_each_segment(bvec, bio, iter) {
@@ -216,7 +217,10 @@ static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 		nd_iostat_end(bio, start);
 
 	if (bio->bi_opf & REQ_FUA)
-		nvdimm_flush(nd_region);
+		ret = nvdimm_flush(nd_region, bio, true);
+
+	if (ret)
+		bio->bi_status = errno_to_blk_status(ret);
 
 	bio_endio(bio);
 	return BLK_QC_T_NONE;
@@ -528,14 +532,14 @@ static int nd_pmem_remove(struct device *dev)
 		sysfs_put(pmem->bb_state);
 		pmem->bb_state = NULL;
 	}
-	nvdimm_flush(to_nd_region(dev->parent));
+	nvdimm_flush(to_nd_region(dev->parent), NULL, false);
 
 	return 0;
 }
 
 static void nd_pmem_shutdown(struct device *dev)
 {
-	nvdimm_flush(to_nd_region(dev->parent));
+	nvdimm_flush(to_nd_region(dev->parent), NULL, false);
 }
 
 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
index fa37afc..5508727 100644
--- a/drivers/nvdimm/region_devs.c
+++ b/drivers/nvdimm/region_devs.c
@@ -290,7 +290,9 @@ static ssize_t deep_flush_store(struct device *dev, struct device_attribute *att
 		return rc;
 	if (!flush)
 		return -EINVAL;
-	nvdimm_flush(nd_region);
+	rc = nvdimm_flush(nd_region, NULL, false);
+	if (rc)
+		return rc;
 
 	return len;
 }
@@ -1065,6 +1067,11 @@ static struct nd_region *nd_region_create(struct nvdimm_bus *nvdimm_bus,
 	dev->of_node = ndr_desc->of_node;
 	nd_region->ndr_size = resource_size(ndr_desc->res);
 	nd_region->ndr_start = ndr_desc->res->start;
+	if (ndr_desc->flush)
+		nd_region->flush = ndr_desc->flush;
+	else
+		nd_region->flush = generic_nvdimm_flush;
+
 	nd_device_register(dev);
 
 	return nd_region;
@@ -1105,11 +1112,36 @@ struct nd_region *nvdimm_volatile_region_create(struct nvdimm_bus *nvdimm_bus,
 }
 EXPORT_SYMBOL_GPL(nvdimm_volatile_region_create);
 
+int nvdimm_flush(struct nd_region *nd_region, struct bio *bio, bool async)
+{
+	int rc = 0;
+
+	/* Create child bio for asynchronous flush and chain with
+	 * parent bio. Otherwise directly call nd_region flush.
+	 */
+	if (async && bio->bi_iter.bi_sector != -1) {
+
+		struct bio *child = bio_alloc(GFP_ATOMIC, 0);
+
+		if (!child)
+			return -ENOMEM;
+		bio_copy_dev(child, bio);
+		child->bi_opf = REQ_PREFLUSH;
+		child->bi_iter.bi_sector = -1;
+		bio_chain(child, bio);
+		submit_bio(child);
+	} else {
+		if (nd_region->flush(nd_region))
+			rc = -EIO;
+	}
+
+	return rc;
+}
 /**
  * nvdimm_flush - flush any posted write queues between the cpu and pmem media
  * @nd_region: blk or interleaved pmem region
  */
-void nvdimm_flush(struct nd_region *nd_region)
+int generic_nvdimm_flush(struct nd_region *nd_region)
 {
 	struct nd_region_data *ndrd = dev_get_drvdata(&nd_region->dev);
 	int i, idx;
@@ -1133,6 +1165,8 @@ void nvdimm_flush(struct nd_region *nd_region)
 		if (ndrd_get_flush_wpq(ndrd, i, 0))
 			writeq(1, ndrd_get_flush_wpq(ndrd, i, idx));
 	wmb();
+
+	return 0;
 }
 EXPORT_SYMBOL_GPL(nvdimm_flush);
 
diff --git a/include/linux/libnvdimm.h b/include/linux/libnvdimm.h
index 097072c..b49632c 100644
--- a/include/linux/libnvdimm.h
+++ b/include/linux/libnvdimm.h
@@ -115,6 +115,7 @@ struct nd_mapping_desc {
 	int position;
 };
 
+struct nd_region;
 struct nd_region_desc {
 	struct resource *res;
 	struct nd_mapping_desc *mapping;
@@ -126,6 +127,7 @@ struct nd_region_desc {
 	int numa_node;
 	unsigned long flags;
 	struct device_node *of_node;
+	int (*flush)(struct nd_region *nd_region);
 };
 
 struct device;
@@ -201,7 +203,8 @@ unsigned long nd_blk_memremap_flags(struct nd_blk_region *ndbr);
 unsigned int nd_region_acquire_lane(struct nd_region *nd_region);
 void nd_region_release_lane(struct nd_region *nd_region, unsigned int lane);
 u64 nd_fletcher64(void *addr, size_t len, bool le);
-void nvdimm_flush(struct nd_region *nd_region);
+int nvdimm_flush(struct nd_region *nd_region, struct bio *bio, bool async);
+int generic_nvdimm_flush(struct nd_region *nd_region);
 int nvdimm_has_flush(struct nd_region *nd_region);
 int nvdimm_has_cache(struct nd_region *nd_region);
 
-- 
2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 2/5]  virtio-pmem: Add virtio pmem driver
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 1/5] libnvdimm: nd_region flush callback support Pankaj Gupta
@ 2019-01-09 14:47 ` Pankaj Gupta
  2019-01-14 15:54   ` Michael S. Tsirkin
  2019-01-09 14:47 ` [PATCH v3 3/5] libnvdimm: add nd_region buffered dax_dev flag Pankaj Gupta
                   ` (4 subsequent siblings)
  6 siblings, 1 reply; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

This patch adds virtio-pmem driver for KVM guest.

Guest reads the persistent memory range information from
Qemu over VIRTIO and registers it on nvdimm_bus. It also
creates a nd_region object with the persistent memory
range information so that existing 'nvdimm/pmem' driver
can reserve this into system memory map. This way
'virtio-pmem' driver uses existing functionality of pmem
driver to register persistent memory compatible for DAX
capable filesystems.

This also provides function to perform guest flush over
VIRTIO from 'pmem' driver when userspace performs flush
on DAX memory range.

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 drivers/nvdimm/virtio_pmem.c     |  84 ++++++++++++++++++++++++++
 drivers/virtio/Kconfig           |  10 ++++
 drivers/virtio/Makefile          |   1 +
 drivers/virtio/pmem.c            | 124 +++++++++++++++++++++++++++++++++++++++
 include/linux/virtio_pmem.h      |  60 +++++++++++++++++++
 include/uapi/linux/virtio_ids.h  |   1 +
 include/uapi/linux/virtio_pmem.h |  10 ++++
 7 files changed, 290 insertions(+)
 create mode 100644 drivers/nvdimm/virtio_pmem.c
 create mode 100644 drivers/virtio/pmem.c
 create mode 100644 include/linux/virtio_pmem.h
 create mode 100644 include/uapi/linux/virtio_pmem.h

diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
new file mode 100644
index 0000000..2a1b1ba
--- /dev/null
+++ b/drivers/nvdimm/virtio_pmem.c
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * virtio_pmem.c: Virtio pmem Driver
+ *
+ * Discovers persistent memory range information
+ * from host and provides a virtio based flushing
+ * interface.
+ */
+#include <linux/virtio_pmem.h>
+#include "nd.h"
+
+ /* The interrupt handler */
+void host_ack(struct virtqueue *vq)
+{
+	unsigned int len;
+	unsigned long flags;
+	struct virtio_pmem_request *req, *req_buf;
+	struct virtio_pmem *vpmem = vq->vdev->priv;
+
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
+		req->done = true;
+		wake_up(&req->host_acked);
+
+		if (!list_empty(&vpmem->req_list)) {
+			req_buf = list_first_entry(&vpmem->req_list,
+					struct virtio_pmem_request, list);
+			list_del(&vpmem->req_list);
+			req_buf->wq_buf_avail = true;
+			wake_up(&req_buf->wq_buf);
+		}
+	}
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+}
+EXPORT_SYMBOL_GPL(host_ack);
+
+ /* The request submission function */
+int virtio_pmem_flush(struct nd_region *nd_region)
+{
+	int err;
+	unsigned long flags;
+	struct scatterlist *sgs[2], sg, ret;
+	struct virtio_device *vdev = nd_region->provider_data;
+	struct virtio_pmem *vpmem = vdev->priv;
+	struct virtio_pmem_request *req;
+
+	might_sleep();
+	req = kmalloc(sizeof(*req), GFP_KERNEL);
+	if (!req)
+		return -ENOMEM;
+
+	req->done = req->wq_buf_avail = false;
+	strcpy(req->name, "FLUSH");
+	init_waitqueue_head(&req->host_acked);
+	init_waitqueue_head(&req->wq_buf);
+	sg_init_one(&sg, req->name, strlen(req->name));
+	sgs[0] = &sg;
+	sg_init_one(&ret, &req->ret, sizeof(req->ret));
+	sgs[1] = &ret;
+
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
+	if (err) {
+		dev_err(&vdev->dev, "failed to send command to virtio pmem device\n");
+
+		list_add_tail(&vpmem->req_list, &req->list);
+		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+
+		/* When host has read buffer, this completes via host_ack */
+		wait_event(req->wq_buf, req->wq_buf_avail);
+		spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	}
+	virtqueue_kick(vpmem->req_vq);
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+
+	/* When host has read buffer, this completes via host_ack */
+	wait_event(req->host_acked, req->done);
+	err = req->ret;
+	kfree(req);
+
+	return err;
+};
+EXPORT_SYMBOL_GPL(virtio_pmem_flush);
+MODULE_LICENSE("GPL");
diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
index 3589764..9f634a2 100644
--- a/drivers/virtio/Kconfig
+++ b/drivers/virtio/Kconfig
@@ -42,6 +42,16 @@ config VIRTIO_PCI_LEGACY
 
 	  If unsure, say Y.
 
+config VIRTIO_PMEM
+	tristate "Support for virtio pmem driver"
+	depends on VIRTIO
+	depends on LIBNVDIMM
+	help
+	This driver provides support for virtio based flushing interface
+	for persistent memory range.
+
+	If unsure, say M.
+
 config VIRTIO_BALLOON
 	tristate "Virtio balloon driver"
 	depends on VIRTIO
diff --git a/drivers/virtio/Makefile b/drivers/virtio/Makefile
index 3a2b5c5..143ce91 100644
--- a/drivers/virtio/Makefile
+++ b/drivers/virtio/Makefile
@@ -6,3 +6,4 @@ virtio_pci-y := virtio_pci_modern.o virtio_pci_common.o
 virtio_pci-$(CONFIG_VIRTIO_PCI_LEGACY) += virtio_pci_legacy.o
 obj-$(CONFIG_VIRTIO_BALLOON) += virtio_balloon.o
 obj-$(CONFIG_VIRTIO_INPUT) += virtio_input.o
+obj-$(CONFIG_VIRTIO_PMEM) += pmem.o ../nvdimm/virtio_pmem.o
diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
new file mode 100644
index 0000000..51f5349
--- /dev/null
+++ b/drivers/virtio/pmem.c
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * virtio_pmem.c: Virtio pmem Driver
+ *
+ * Discovers persistent memory range information
+ * from host and registers the virtual pmem device
+ * with libnvdimm core.
+ */
+#include <linux/virtio_pmem.h>
+#include <../../drivers/nvdimm/nd.h>
+
+static struct virtio_device_id id_table[] = {
+	{ VIRTIO_ID_PMEM, VIRTIO_DEV_ANY_ID },
+	{ 0 },
+};
+
+ /* Initialize virt queue */
+static int init_vq(struct virtio_pmem *vpmem)
+{
+	struct virtqueue *vq;
+
+	/* single vq */
+	vpmem->req_vq = vq = virtio_find_single_vq(vpmem->vdev,
+				host_ack, "flush_queue");
+	if (IS_ERR(vq))
+		return PTR_ERR(vq);
+
+	spin_lock_init(&vpmem->pmem_lock);
+	INIT_LIST_HEAD(&vpmem->req_list);
+
+	return 0;
+};
+
+static int virtio_pmem_probe(struct virtio_device *vdev)
+{
+	int err = 0;
+	struct resource res;
+	struct virtio_pmem *vpmem;
+	struct nvdimm_bus *nvdimm_bus;
+	struct nd_region_desc ndr_desc;
+	int nid = dev_to_node(&vdev->dev);
+	struct nd_region *nd_region;
+
+	if (!vdev->config->get) {
+		dev_err(&vdev->dev, "%s failure: config disabled\n",
+			__func__);
+		return -EINVAL;
+	}
+
+	vdev->priv = vpmem = devm_kzalloc(&vdev->dev, sizeof(*vpmem),
+			GFP_KERNEL);
+	if (!vpmem) {
+		err = -ENOMEM;
+		goto out_err;
+	}
+
+	vpmem->vdev = vdev;
+	err = init_vq(vpmem);
+	if (err)
+		goto out_err;
+
+	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
+			start, &vpmem->start);
+	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
+			size, &vpmem->size);
+
+	res.start = vpmem->start;
+	res.end   = vpmem->start + vpmem->size-1;
+	vpmem->nd_desc.provider_name = "virtio-pmem";
+	vpmem->nd_desc.module = THIS_MODULE;
+
+	vpmem->nvdimm_bus = nvdimm_bus = nvdimm_bus_register(&vdev->dev,
+						&vpmem->nd_desc);
+	if (!nvdimm_bus)
+		goto out_vq;
+
+	dev_set_drvdata(&vdev->dev, nvdimm_bus);
+	memset(&ndr_desc, 0, sizeof(ndr_desc));
+
+	ndr_desc.res = &res;
+	ndr_desc.numa_node = nid;
+	ndr_desc.flush = virtio_pmem_flush;
+	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
+	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
+	nd_region->provider_data =  dev_to_virtio
+					(nd_region->dev.parent->parent);
+
+	if (!nd_region)
+		goto out_nd;
+
+	//virtio_device_ready(vdev);
+	return 0;
+out_nd:
+	err = -ENXIO;
+	nvdimm_bus_unregister(nvdimm_bus);
+out_vq:
+	vdev->config->del_vqs(vdev);
+out_err:
+	dev_err(&vdev->dev, "failed to register virtio pmem memory\n");
+	return err;
+}
+
+static void virtio_pmem_remove(struct virtio_device *vdev)
+{
+	struct virtio_pmem *vpmem = vdev->priv;
+	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
+
+	nvdimm_bus_unregister(nvdimm_bus);
+	vdev->config->del_vqs(vdev);
+	kfree(vpmem);
+}
+
+static struct virtio_driver virtio_pmem_driver = {
+	.driver.name		= KBUILD_MODNAME,
+	.driver.owner		= THIS_MODULE,
+	.id_table		= id_table,
+	.probe			= virtio_pmem_probe,
+	.remove			= virtio_pmem_remove,
+};
+
+module_virtio_driver(virtio_pmem_driver);
+MODULE_DEVICE_TABLE(virtio, id_table);
+MODULE_DESCRIPTION("Virtio pmem driver");
+MODULE_LICENSE("GPL");
diff --git a/include/linux/virtio_pmem.h b/include/linux/virtio_pmem.h
new file mode 100644
index 0000000..224f9d9
--- /dev/null
+++ b/include/linux/virtio_pmem.h
@@ -0,0 +1,60 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * virtio_pmem.h: virtio pmem Driver
+ *
+ * Discovers persistent memory range information
+ * from host and provides a virtio based flushing
+ * interface.
+ **/
+
+#ifndef _LINUX_VIRTIO_PMEM_H
+#define _LINUX_VIRTIO_PMEM_H
+
+#include <linux/virtio_ids.h>
+#include <linux/module.h>
+#include <linux/virtio_config.h>
+#include <uapi/linux/virtio_pmem.h>
+#include <linux/libnvdimm.h>
+#include <linux/spinlock.h>
+
+struct virtio_pmem_request {
+	/* Host return status corresponding to flush request */
+	int ret;
+
+	/* command name*/
+	char name[16];
+
+	/* Wait queue to process deferred work after ack from host */
+	wait_queue_head_t host_acked;
+	bool done;
+
+	/* Wait queue to process deferred work after virt queue buffer avail */
+	wait_queue_head_t wq_buf;
+	bool wq_buf_avail;
+	struct list_head list;
+};
+
+struct virtio_pmem {
+	struct virtio_device *vdev;
+
+	/* Virtio pmem request queue */
+	struct virtqueue *req_vq;
+
+	/* nvdimm bus registers virtio pmem device */
+	struct nvdimm_bus *nvdimm_bus;
+	struct nvdimm_bus_descriptor nd_desc;
+
+	/* List to store deferred work if virtqueue is full */
+	struct list_head req_list;
+
+	/* Synchronize virtqueue data */
+	spinlock_t pmem_lock;
+
+	/* Memory region information */
+	uint64_t start;
+	uint64_t size;
+};
+
+void host_ack(struct virtqueue *vq);
+int virtio_pmem_flush(struct nd_region *nd_region);
+#endif
diff --git a/include/uapi/linux/virtio_ids.h b/include/uapi/linux/virtio_ids.h
index 6d5c3b2..3463895 100644
--- a/include/uapi/linux/virtio_ids.h
+++ b/include/uapi/linux/virtio_ids.h
@@ -43,5 +43,6 @@
 #define VIRTIO_ID_INPUT        18 /* virtio input */
 #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
 #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
+#define VIRTIO_ID_PMEM         25 /* virtio pmem */
 
 #endif /* _LINUX_VIRTIO_IDS_H */
diff --git a/include/uapi/linux/virtio_pmem.h b/include/uapi/linux/virtio_pmem.h
new file mode 100644
index 0000000..fa3f7d5
--- /dev/null
+++ b/include/uapi/linux/virtio_pmem.h
@@ -0,0 +1,10 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef _UAPI_LINUX_VIRTIO_PMEM_H
+#define _UAPI_LINUX_VIRTIO_PMEM_H
+
+struct virtio_pmem_config {
+	__le64 start;
+	__le64 size;
+};
+#endif
-- 
2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 3/5] libnvdimm: add nd_region buffered dax_dev flag
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 1/5] libnvdimm: nd_region flush callback support Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 2/5] virtio-pmem: Add virtio pmem driver Pankaj Gupta
@ 2019-01-09 14:47 ` Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 4/5] ext4: disable map_sync for virtio pmem Pankaj Gupta
                   ` (3 subsequent siblings)
  6 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

This patch adds 'DAXDEV_BUFFERED' flag which is set
for virtio pmem corresponding nd_region. This later
is used to disable MAP_SYNC functionality for ext4
& xfs filesystem.

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 drivers/dax/super.c          | 17 +++++++++++++++++
 drivers/nvdimm/pmem.c        |  3 +++
 drivers/nvdimm/region_devs.c |  7 +++++++
 drivers/virtio/pmem.c        |  1 +
 include/linux/dax.h          |  9 +++++++++
 include/linux/libnvdimm.h    |  6 ++++++
 6 files changed, 43 insertions(+)

diff --git a/drivers/dax/super.c b/drivers/dax/super.c
index 6e928f3..9128740 100644
--- a/drivers/dax/super.c
+++ b/drivers/dax/super.c
@@ -167,6 +167,8 @@ enum dax_device_flags {
 	DAXDEV_ALIVE,
 	/* gate whether dax_flush() calls the low level flush routine */
 	DAXDEV_WRITE_CACHE,
+	/* flag to disable MAP_SYNC for virtio based host page cache flush */
+	DAXDEV_BUFFERED,
 };
 
 /**
@@ -335,6 +337,21 @@ bool dax_write_cache_enabled(struct dax_device *dax_dev)
 }
 EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
 
+void virtio_pmem_host_cache(struct dax_device *dax_dev, bool wc)
+{
+	if (wc)
+		set_bit(DAXDEV_BUFFERED, &dax_dev->flags);
+	else
+		clear_bit(DAXDEV_BUFFERED, &dax_dev->flags);
+}
+EXPORT_SYMBOL_GPL(virtio_pmem_host_cache);
+
+bool virtio_pmem_host_cache_enabled(struct dax_device *dax_dev)
+{
+	return test_bit(DAXDEV_BUFFERED, &dax_dev->flags);
+}
+EXPORT_SYMBOL_GPL(virtio_pmem_host_cache_enabled);
+
 bool dax_alive(struct dax_device *dax_dev)
 {
 	lockdep_assert_held(&dax_srcu);
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index fe1217b..8d190a3 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -472,6 +472,9 @@ static int pmem_attach_disk(struct device *dev,
 		return -ENOMEM;
 	}
 	dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
+
+	/* Set buffered bit in 'dax_dev' for virtio pmem */
+	virtio_pmem_host_cache(dax_dev, nvdimm_is_buffered(nd_region));
 	pmem->dax_dev = dax_dev;
 
 	gendev = disk_to_dev(disk);
diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
index f8218b4..1f8b2be 100644
--- a/drivers/nvdimm/region_devs.c
+++ b/drivers/nvdimm/region_devs.c
@@ -1264,6 +1264,13 @@ int nd_region_conflict(struct nd_region *nd_region, resource_size_t start,
 	return device_for_each_child(&nvdimm_bus->dev, &ctx, region_conflict);
 }
 
+int nvdimm_is_buffered(struct nd_region *nd_region)
+{
+	return is_nd_pmem(&nd_region->dev) &&
+		test_bit(ND_REGION_BUFFERED, &nd_region->flags);
+}
+EXPORT_SYMBOL_GPL(nvdimm_is_buffered);
+
 void __exit nd_region_devs_exit(void)
 {
 	ida_destroy(&region_ida);
diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
index 51f5349..901767b 100644
--- a/drivers/virtio/pmem.c
+++ b/drivers/virtio/pmem.c
@@ -81,6 +81,7 @@ static int virtio_pmem_probe(struct virtio_device *vdev)
 	ndr_desc.numa_node = nid;
 	ndr_desc.flush = virtio_pmem_flush;
 	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
+	set_bit(ND_REGION_BUFFERED, &ndr_desc.flags);
 	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
 	nd_region->provider_data =  dev_to_virtio
 					(nd_region->dev.parent->parent);
diff --git a/include/linux/dax.h b/include/linux/dax.h
index 0dd316a..d16e03e 100644
--- a/include/linux/dax.h
+++ b/include/linux/dax.h
@@ -37,6 +37,8 @@ void put_dax(struct dax_device *dax_dev);
 void kill_dax(struct dax_device *dax_dev);
 void dax_write_cache(struct dax_device *dax_dev, bool wc);
 bool dax_write_cache_enabled(struct dax_device *dax_dev);
+void virtio_pmem_host_cache(struct dax_device *dax_dev, bool wc);
+bool virtio_pmem_host_cache_enabled(struct dax_device *dax_dev);
 #else
 static inline struct dax_device *dax_get_by_host(const char *host)
 {
@@ -64,6 +66,13 @@ static inline bool dax_write_cache_enabled(struct dax_device *dax_dev)
 {
 	return false;
 }
+static inline void virtio_pmem_host_cache(struct dax_device *dax_dev, bool wc)
+{
+}
+static inline bool virtio_pmem_host_cache_enabled(struct dax_device *dax_dev)
+{
+	return false;
+}
 #endif
 
 struct writeback_control;
diff --git a/include/linux/libnvdimm.h b/include/linux/libnvdimm.h
index ca8bc07..94616f1 100644
--- a/include/linux/libnvdimm.h
+++ b/include/linux/libnvdimm.h
@@ -64,6 +64,11 @@ enum {
 	 */
 	ND_REGION_PERSIST_MEMCTRL = 2,
 
+	/* provides virtio based asynchronous flush mechanism for buffered
+	 * host page cache.
+	 */
+	ND_REGION_BUFFERED = 3,
+
 	/* mark newly adjusted resources as requiring a label update */
 	DPA_RESOURCE_ADJUSTED = 1 << 0,
 };
@@ -265,6 +270,7 @@ int generic_nvdimm_flush(struct nd_region *nd_region);
 int nvdimm_has_flush(struct nd_region *nd_region);
 int nvdimm_has_cache(struct nd_region *nd_region);
 int nvdimm_in_overwrite(struct nvdimm *nvdimm);
+int nvdimm_is_buffered(struct nd_region *nd_region);
 
 static inline int nvdimm_ctl(struct nvdimm *nvdimm, unsigned int cmd, void *buf,
 		unsigned int buf_len, int *cmd_rc)
-- 
2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 4/5] ext4: disable map_sync for virtio pmem
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
                   ` (2 preceding siblings ...)
  2019-01-09 14:47 ` [PATCH v3 3/5] libnvdimm: add nd_region buffered dax_dev flag Pankaj Gupta
@ 2019-01-09 14:47 ` Pankaj Gupta
  2019-01-09 14:47 ` [PATCH v3 5/5] xfs: " Pankaj Gupta
                   ` (2 subsequent siblings)
  6 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

Virtio pmem provides asynchronous host page cache flush
mechanism. We don't support 'MAP_SYNC' with virtio pmem 
and ext4. 

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 fs/ext4/file.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/fs/ext4/file.c b/fs/ext4/file.c
index 69d65d4..e54f48b 100644
--- a/fs/ext4/file.c
+++ b/fs/ext4/file.c
@@ -360,8 +360,10 @@ static const struct vm_operations_struct ext4_file_vm_ops = {
 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 {
 	struct inode *inode = file->f_mapping->host;
+	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+	struct dax_device *dax_dev = sbi->s_daxdev;
 
-	if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
+	if (unlikely(ext4_forced_shutdown(sbi)))
 		return -EIO;
 
 	/*
@@ -371,6 +373,13 @@ static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 	if (!IS_DAX(file_inode(file)) && (vma->vm_flags & VM_SYNC))
 		return -EOPNOTSUPP;
 
+	/* We don't support synchronous mappings with guest direct access
+	 * and virtio based host page cache flush mechanism.
+	 */
+	if (IS_DAX(file_inode(file)) && virtio_pmem_host_cache_enabled(dax_dev)
+						&& (vma->vm_flags & VM_SYNC))
+		return -EOPNOTSUPP;
+
 	file_accessed(file);
 	if (IS_DAX(file_inode(file))) {
 		vma->vm_ops = &ext4_dax_vm_ops;
-- 
2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 5/5] xfs: disable map_sync for virtio pmem
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
                   ` (3 preceding siblings ...)
  2019-01-09 14:47 ` [PATCH v3 4/5] ext4: disable map_sync for virtio pmem Pankaj Gupta
@ 2019-01-09 14:47 ` " Pankaj Gupta
  2019-01-09 14:47   ` Pankaj Gupta
  2019-01-09 16:26   ` Darrick J. Wong
       [not found] ` <20190109144736.17452-1-pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  2019-02-04 22:56 ` security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device) Michael S. Tsirkin
  6 siblings, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

Virtio pmem provides asynchronous host page cache flush
mechanism. we don't support 'MAP_SYNC' with virtio pmem 
and xfs.

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 fs/xfs/xfs_file.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index e474250..eae4aa4 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -1190,6 +1190,14 @@ xfs_file_mmap(
 	if (!IS_DAX(file_inode(filp)) && (vma->vm_flags & VM_SYNC))
 		return -EOPNOTSUPP;
 
+	/* We don't support synchronous mappings with guest direct access
+	 * and virtio based host page cache mechanism.
+	 */
+	if (IS_DAX(file_inode(filp)) && virtio_pmem_host_cache_enabled(
+		xfs_find_daxdev_for_inode(file_inode(filp))) &&
+					(vma->vm_flags & VM_SYNC))
+		return -EOPNOTSUPP;
+
 	file_accessed(filp);
 	vma->vm_ops = &xfs_file_vm_ops;
 	if (IS_DAX(file_inode(filp)))
-- 
2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* [PATCH v3 5/5] xfs: disable map_sync for virtio pmem
  2019-01-09 14:47 ` [PATCH v3 5/5] xfs: " Pankaj Gupta
@ 2019-01-09 14:47   ` Pankaj Gupta
  2019-01-09 16:26   ` Darrick J. Wong
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 14:47 UTC (permalink / raw)
  To: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs
  Cc: jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw, pagupta

Virtio pmem provides asynchronous host page cache flush
mechanism. we don't support 'MAP_SYNC' with virtio pmem 
and xfs.

Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
---
 fs/xfs/xfs_file.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index e474250..eae4aa4 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -1190,6 +1190,14 @@ xfs_file_mmap(
 	if (!IS_DAX(file_inode(filp)) && (vma->vm_flags & VM_SYNC))
 		return -EOPNOTSUPP;
 
+	/* We don't support synchronous mappings with guest direct access
+	 * and virtio based host page cache mechanism.
+	 */
+	if (IS_DAX(file_inode(filp)) && virtio_pmem_host_cache_enabled(
+		xfs_find_daxdev_for_inode(file_inode(filp))) &&
+					(vma->vm_flags & VM_SYNC))
+		return -EOPNOTSUPP;
+
 	file_accessed(filp);
 	vma->vm_ops = &xfs_file_vm_ops;
 	if (IS_DAX(file_inode(filp)))
-- 
2.9.3


^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 5/5] xfs: disable map_sync for virtio pmem
  2019-01-09 14:47 ` [PATCH v3 5/5] xfs: " Pankaj Gupta
  2019-01-09 14:47   ` Pankaj Gupta
@ 2019-01-09 16:26   ` Darrick J. Wong
  2019-01-09 18:08     ` Pankaj Gupta
  1 sibling, 1 reply; 61+ messages in thread
From: Darrick J. Wong @ 2019-01-09 16:26 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, rjw

On Wed, Jan 09, 2019 at 08:17:36PM +0530, Pankaj Gupta wrote:
> Virtio pmem provides asynchronous host page cache flush
> mechanism. we don't support 'MAP_SYNC' with virtio pmem 
> and xfs.
> 
> Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
> ---
>  fs/xfs/xfs_file.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index e474250..eae4aa4 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -1190,6 +1190,14 @@ xfs_file_mmap(
>  	if (!IS_DAX(file_inode(filp)) && (vma->vm_flags & VM_SYNC))
>  		return -EOPNOTSUPP;
>  
> +	/* We don't support synchronous mappings with guest direct access
> +	 * and virtio based host page cache mechanism.
> +	 */
> +	if (IS_DAX(file_inode(filp)) && virtio_pmem_host_cache_enabled(

Echoing what Jan said, this ought to be some sort of generic function
that tells us whether or not memory mapped from the dax device will
always still be accessible even after a crash (i.e. supports MAP_SYNC).

What if the underlying file on the host is itself on pmem and can be
MAP_SYNC'd?  Shouldn't the guest be able to use MAP_SYNC as well?

--D

> +		xfs_find_daxdev_for_inode(file_inode(filp))) &&
> +					(vma->vm_flags & VM_SYNC))
> +		return -EOPNOTSUPP;
> +
>  	file_accessed(filp);
>  	vma->vm_ops = &xfs_file_vm_ops;
>  	if (IS_DAX(file_inode(filp)))
> -- 
> 2.9.3
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 5/5] xfs: disable map_sync for virtio pmem
  2019-01-09 16:26   ` Darrick J. Wong
@ 2019-01-09 18:08     ` Pankaj Gupta
  0 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-09 18:08 UTC (permalink / raw)
  To: Darrick J. Wong
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel, rjw


> > Virtio pmem provides asynchronous host page cache flush
> > mechanism. we don't support 'MAP_SYNC' with virtio pmem
> > and xfs.
> > 
> > Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
> > ---
> >  fs/xfs/xfs_file.c | 8 ++++++++
> >  1 file changed, 8 insertions(+)
> > 
> > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > index e474250..eae4aa4 100644
> > --- a/fs/xfs/xfs_file.c
> > +++ b/fs/xfs/xfs_file.c
> > @@ -1190,6 +1190,14 @@ xfs_file_mmap(
> >  	if (!IS_DAX(file_inode(filp)) && (vma->vm_flags & VM_SYNC))
> >  		return -EOPNOTSUPP;
> >  
> > +	/* We don't support synchronous mappings with guest direct access
> > +	 * and virtio based host page cache mechanism.
> > +	 */
> > +	if (IS_DAX(file_inode(filp)) && virtio_pmem_host_cache_enabled(
> 
> Echoing what Jan said, this ought to be some sort of generic function
> that tells us whether or not memory mapped from the dax device will
> always still be accessible even after a crash (i.e. supports MAP_SYNC).

o.k
> 
> What if the underlying file on the host is itself on pmem and can be
> MAP_SYNC'd?  Shouldn't the guest be able to use MAP_SYNC as well?

Guest MAP_SYNC on actual host pmem will sync guest metadata, as guest 
writes are persistent on actual pmem. Host side Qemu MAP_SYNC enabling
work for pmem is in progress. It will make sure metadata at host side
is in consistent state after a crash or any other metadata corruption 
operation.

For virtio-pmem, we are emulating pmem device over regular storage on
host side. Guest need to call fsync followed by write to make sure
guest metadata is in consistent state(or journalled). Host backing file
data & metadata will also be persistent after guest to host fsync.

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found] ` <20190109144736.17452-1-pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-10  1:26   ` Dave Chinner
  2019-01-10  1:26     ` Dave Chinner
                       ` (3 more replies)
  0 siblings, 4 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-10  1:26 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger.kernel-m1MBpc4rdrD3fQ9qLvQP4Q,
	zwisler-DgEjT+Ai2ygdnm+yROfE0A, eblake-H+wXaHxf7aLQT0dZR+AlfA,
	darrick.wong-QHcLZuEGTsvQT0dZR+AlfA, mst-H+wXaHxf7aLQT0dZR+AlfA,
	willy-wEGCiKHe2LqWVfeAwA7xHQ, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA, nilal-H+wXaHxf7aLQT0dZR+AlfA,
	riel-ebMLmSuQjDVBDgjK7y7TUQ, stefanha-H+wXaHxf7aLQT0dZR+AlfA,
	imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, tytso-3s7WtUTddSA,
	xiaoguangrong.eric-Re5JQEeQqe8AvxtiuMwx3w,
	rjw-LthD3rsA81gm4RdzfppkhA, linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
>  This patch series has implementation for "virtio pmem". 
>  "virtio pmem" is fake persistent memory(nvdimm) in guest 
>  which allows to bypass the guest page cache. This also
>  implements a VIRTIO based asynchronous flush mechanism.  

Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
like a good idea, but.....

This means the guest VM can now run timing attacks to observe host
side page cache residency, and depending on the implementation I'm
guessing that the guest will be able to control host side page
cache eviction, too (e.g. via discard or hole punch operations).

Which means this functionality looks to me like a new vector for
information leakage into and out of the guest VM via guest
controlled host page cache manipulation.

https://arxiv.org/pdf/1901.01161

I might be wrong, but if I'm not we're going to have to be very
careful about how guest VMs can access and manipulate host side
resources like the page cache.....

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10  1:26   ` [PATCH v3 0/5] kvm "virtio pmem" device Dave Chinner
@ 2019-01-10  1:26     ` Dave Chinner
  2019-01-10  2:40     ` Rik van Riel
                       ` (2 subsequent siblings)
  3 siblings, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-10  1:26 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw

On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
>  This patch series has implementation for "virtio pmem". 
>  "virtio pmem" is fake persistent memory(nvdimm) in guest 
>  which allows to bypass the guest page cache. This also
>  implements a VIRTIO based asynchronous flush mechanism.  

Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
like a good idea, but.....

This means the guest VM can now run timing attacks to observe host
side page cache residency, and depending on the implementation I'm
guessing that the guest will be able to control host side page
cache eviction, too (e.g. via discard or hole punch operations).

Which means this functionality looks to me like a new vector for
information leakage into and out of the guest VM via guest
controlled host page cache manipulation.

https://arxiv.org/pdf/1901.01161

I might be wrong, but if I'm not we're going to have to be very
careful about how guest VMs can access and manipulate host side
resources like the page cache.....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10  1:26   ` [PATCH v3 0/5] kvm "virtio pmem" device Dave Chinner
  2019-01-10  1:26     ` Dave Chinner
@ 2019-01-10  2:40     ` Rik van Riel
  2019-01-10  2:40       ` Rik van Riel
  2019-01-10 10:17     ` Jan Kara
  2019-01-11  7:45     ` Pankaj Gupta
  3 siblings, 1 reply; 61+ messages in thread
From: Rik van Riel @ 2019-01-10  2:40 UTC (permalink / raw)
  To: Dave Chinner, Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw

[-- Attachment #1: Type: text/plain, Size: 1243 bytes --]

On Thu, 2019-01-10 at 12:26 +1100, Dave Chinner wrote:
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem". 
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest 
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.  
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).

Even if the guest can only access its own disk image,
using its own page cache pages, Pankaj's patches could
still be a win because they allow the host kernel to
easily evict page cache pages without doing IO.

Evicting page cache pages that live inside a guest
involves something like swapping, ballooning, or some
sort of callback into the guest to figure out whether
a page is clean.

Evicting page cache pages that live in the host, OTOH...

-- 
All Rights Reversed.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10  2:40     ` Rik van Riel
@ 2019-01-10  2:40       ` Rik van Riel
  0 siblings, 0 replies; 61+ messages in thread
From: Rik van Riel @ 2019-01-10  2:40 UTC (permalink / raw)
  To: Dave Chinner, Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw

[-- Attachment #1: Type: text/plain, Size: 1243 bytes --]

On Thu, 2019-01-10 at 12:26 +1100, Dave Chinner wrote:
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem". 
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest 
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.  
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).

Even if the guest can only access its own disk image,
using its own page cache pages, Pankaj's patches could
still be a win because they allow the host kernel to
easily evict page cache pages without doing IO.

Evicting page cache pages that live inside a guest
involves something like swapping, ballooning, or some
sort of callback into the guest to figure out whether
a page is clean.

Evicting page cache pages that live in the host, OTOH...

-- 
All Rights Reversed.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10  1:26   ` [PATCH v3 0/5] kvm "virtio pmem" device Dave Chinner
  2019-01-10  1:26     ` Dave Chinner
  2019-01-10  2:40     ` Rik van Riel
@ 2019-01-10 10:17     ` Jan Kara
  2019-01-10 10:17       ` Jan Kara
       [not found]       ` <20190110101757.GC15790-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
  2019-01-11  7:45     ` Pankaj Gupta
  3 siblings, 2 replies; 61+ messages in thread
From: Jan Kara @ 2019-01-10 10:17 UTC (permalink / raw)
  To: Dave Chinner
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger.kernel-m1MBpc4rdrD3fQ9qLvQP4Q,
	zwisler-DgEjT+Ai2ygdnm+yROfE0A, eblake-H+wXaHxf7aLQT0dZR+AlfA,
	darrick.wong-QHcLZuEGTsvQT0dZR+AlfA, mst-H+wXaHxf7aLQT0dZR+AlfA,
	willy-wEGCiKHe2LqWVfeAwA7xHQ, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong.eric-Re5JQEeQqe8AvxtiuMwx3w,
	rjw-LthD3rsA81gm4RdzfppkhA, linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem". 
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest 
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.  
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).
> 
> Which means this functionality looks to me like a new vector for
> information leakage into and out of the guest VM via guest
> controlled host page cache manipulation.
> 
> https://arxiv.org/pdf/1901.01161
> 
> I might be wrong, but if I'm not we're going to have to be very
> careful about how guest VMs can access and manipulate host side
> resources like the page cache.....

Right. Thinking about this I would be more concerned about the fact that
guest can effectively pin amount of host's page cache upto size of the
device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
magic that avoids this?

								Honza
-- 
Jan Kara <jack-IBi9RG/b67k@public.gmane.org>
SUSE Labs, CR

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10 10:17     ` Jan Kara
@ 2019-01-10 10:17       ` Jan Kara
       [not found]       ` <20190110101757.GC15790-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Jan Kara @ 2019-01-10 10:17 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Pankaj Gupta, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	jack, stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal.l.verma, dave.jiang, david, jmoyer,
	xiaoguangrong.eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, willy, tytso, adilger.kernel, darrick.wong, rjw

On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem". 
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest 
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.  
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).
> 
> Which means this functionality looks to me like a new vector for
> information leakage into and out of the guest VM via guest
> controlled host page cache manipulation.
> 
> https://arxiv.org/pdf/1901.01161
> 
> I might be wrong, but if I'm not we're going to have to be very
> careful about how guest VMs can access and manipulate host side
> resources like the page cache.....

Right. Thinking about this I would be more concerned about the fact that
guest can effectively pin amount of host's page cache upto size of the
device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
magic that avoids this?

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-10  1:26   ` [PATCH v3 0/5] kvm "virtio pmem" device Dave Chinner
                       ` (2 preceding siblings ...)
  2019-01-10 10:17     ` Jan Kara
@ 2019-01-11  7:45     ` Pankaj Gupta
  2019-01-11  7:45       ` Pankaj Gupta
       [not found]       ` <1326478078.61913951.1547192704870.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  3 siblings, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-11  7:45 UTC (permalink / raw)
  To: Dave Chinner
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel


> 
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem".
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).

Not sure how? this is similar to mmapping virtual memory by any userspace 
process. Any host userspace process can do such attack on host page cache
using mincore & mmap shared file. 

But i don't think guest can do this alone. For virtio-pmem usecase guest 
won't be using page cache so timing attack from only guest side is not 
possible unless host userspace can run checks on page cache eviction state
using mincore etc. 

As rightly described by Rik, guest will only access its own page cache pages 
and if guest page cache is managed directly by host, this saves alot of 
effort for guest in transferring guest state of page cache.  

> 
> Which means this functionality looks to me like a new vector for
> information leakage into and out of the guest VM via guest
> controlled host page cache manipulation.
> 
> https://arxiv.org/pdf/1901.01161
> 
> I might be wrong, but if I'm not we're going to have to be very
> careful about how guest VMs can access and manipulate host side
> resources like the page cache.....

If I am following correctly the discussions in MM thread. 
Important steps to mitigate this:

* Avoid running mincore in privilege mode: to safeguard page evict state of any 
  page cache page.
* tweaking RWF_NOWAIT 

I think if we secure ways to find current state(cached/evicted) of a page in host, 
we should be able to mitigate the impact for any page cache page access attack 
including virtio-pmem.

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-11  7:45     ` Pankaj Gupta
@ 2019-01-11  7:45       ` Pankaj Gupta
       [not found]       ` <1326478078.61913951.1547192704870.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-11  7:45 UTC (permalink / raw)
  To: Dave Chinner
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel, darrick wong, rjw


> 
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >  This patch series has implementation for "virtio pmem".
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.
> 
> Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> like a good idea, but.....
> 
> This means the guest VM can now run timing attacks to observe host
> side page cache residency, and depending on the implementation I'm
> guessing that the guest will be able to control host side page
> cache eviction, too (e.g. via discard or hole punch operations).

Not sure how? this is similar to mmapping virtual memory by any userspace 
process. Any host userspace process can do such attack on host page cache
using mincore & mmap shared file. 

But i don't think guest can do this alone. For virtio-pmem usecase guest 
won't be using page cache so timing attack from only guest side is not 
possible unless host userspace can run checks on page cache eviction state
using mincore etc. 

As rightly described by Rik, guest will only access its own page cache pages 
and if guest page cache is managed directly by host, this saves alot of 
effort for guest in transferring guest state of page cache.  

> 
> Which means this functionality looks to me like a new vector for
> information leakage into and out of the guest VM via guest
> controlled host page cache manipulation.
> 
> https://arxiv.org/pdf/1901.01161
> 
> I might be wrong, but if I'm not we're going to have to be very
> careful about how guest VMs can access and manipulate host side
> resources like the page cache.....

If I am following correctly the discussions in MM thread. 
Important steps to mitigate this:

* Avoid running mincore in privilege mode: to safeguard page evict state of any 
  page cache page.
* tweaking RWF_NOWAIT 

I think if we secure ways to find current state(cached/evicted) of a page in host, 
we should be able to mitigate the impact for any page cache page access attack 
including virtio-pmem.

Thanks,
Pankaj



^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]       ` <20190110101757.GC15790-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
@ 2019-01-13  1:38         ` Pankaj Gupta
  2019-01-13  1:38           ` Pankaj Gupta
       [not found]           ` <1354249849.63357171.1547343519970.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  0 siblings, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-13  1:38 UTC (permalink / raw)
  To: Jan Kara
  Cc: kvm-u79uwXL29TY76Z2rM5mHXA, david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA, Dave Chinner,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, willy-wEGCiKHe2LqWVfeAwA7xHQ,
	hch-wEGCiKHe2LqWVfeAwA7xHQ, linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA



> 
> On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > >  This patch series has implementation for "virtio pmem".
> > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > >  which allows to bypass the guest page cache. This also
> > >  implements a VIRTIO based asynchronous flush mechanism.
> > 
> > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > like a good idea, but.....
> > 
> > This means the guest VM can now run timing attacks to observe host
> > side page cache residency, and depending on the implementation I'm
> > guessing that the guest will be able to control host side page
> > cache eviction, too (e.g. via discard or hole punch operations).
> > 
> > Which means this functionality looks to me like a new vector for
> > information leakage into and out of the guest VM via guest
> > controlled host page cache manipulation.
> > 
> > https://arxiv.org/pdf/1901.01161
> > 
> > I might be wrong, but if I'm not we're going to have to be very
> > careful about how guest VMs can access and manipulate host side
> > resources like the page cache.....
> 
> Right. Thinking about this I would be more concerned about the fact that
> guest can effectively pin amount of host's page cache upto size of the
> device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
> magic that avoids this?

Yes, guest will pin these host page cache pages using 'get_user_pages' by
elevating the page reference count. But these pages can be reclaimed by host
at any time when there is memory pressure.

KVM does not permanently pin pages. vfio does that but we are not using
it here.

Could you please elaborate what you are thinking?

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13  1:38         ` Pankaj Gupta
@ 2019-01-13  1:38           ` Pankaj Gupta
       [not found]           ` <1354249849.63357171.1547343519970.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-13  1:38 UTC (permalink / raw)
  To: Jan Kara
  Cc: Dave Chinner, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel, darrick wong, rjw



> 
> On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > >  This patch series has implementation for "virtio pmem".
> > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > >  which allows to bypass the guest page cache. This also
> > >  implements a VIRTIO based asynchronous flush mechanism.
> > 
> > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > like a good idea, but.....
> > 
> > This means the guest VM can now run timing attacks to observe host
> > side page cache residency, and depending on the implementation I'm
> > guessing that the guest will be able to control host side page
> > cache eviction, too (e.g. via discard or hole punch operations).
> > 
> > Which means this functionality looks to me like a new vector for
> > information leakage into and out of the guest VM via guest
> > controlled host page cache manipulation.
> > 
> > https://arxiv.org/pdf/1901.01161
> > 
> > I might be wrong, but if I'm not we're going to have to be very
> > careful about how guest VMs can access and manipulate host side
> > resources like the page cache.....
> 
> Right. Thinking about this I would be more concerned about the fact that
> guest can effectively pin amount of host's page cache upto size of the
> device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
> magic that avoids this?

Yes, guest will pin these host page cache pages using 'get_user_pages' by
elevating the page reference count. But these pages can be reclaimed by host
at any time when there is memory pressure.

KVM does not permanently pin pages. vfio does that but we are not using
it here.

Could you please elaborate what you are thinking?

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]           ` <1354249849.63357171.1547343519970.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-13  1:43             ` Dan Williams
  2019-01-13  1:43               ` Dan Williams
       [not found]               ` <CAPcyv4hwcgTUpgNCefCGu4DvgkYBp5b=f+hJ+FC=s5APYKoycg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  0 siblings, 2 replies; 61+ messages in thread
From: Dan Williams @ 2019-01-13  1:43 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, Eric Blake, darrick wong,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangro

On Sat, Jan 12, 2019 at 5:38 PM Pankaj Gupta <pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:
>
>
>
> >
> > On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > > >  This patch series has implementation for "virtio pmem".
> > > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > > >  which allows to bypass the guest page cache. This also
> > > >  implements a VIRTIO based asynchronous flush mechanism.
> > >
> > > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > > like a good idea, but.....
> > >
> > > This means the guest VM can now run timing attacks to observe host
> > > side page cache residency, and depending on the implementation I'm
> > > guessing that the guest will be able to control host side page
> > > cache eviction, too (e.g. via discard or hole punch operations).
> > >
> > > Which means this functionality looks to me like a new vector for
> > > information leakage into and out of the guest VM via guest
> > > controlled host page cache manipulation.
> > >
> > > https://arxiv.org/pdf/1901.01161
> > >
> > > I might be wrong, but if I'm not we're going to have to be very
> > > careful about how guest VMs can access and manipulate host side
> > > resources like the page cache.....
> >
> > Right. Thinking about this I would be more concerned about the fact that
> > guest can effectively pin amount of host's page cache upto size of the
> > device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
> > magic that avoids this?
>
> Yes, guest will pin these host page cache pages using 'get_user_pages' by
> elevating the page reference count. But these pages can be reclaimed by host
> at any time when there is memory pressure.

Wait, how can the guest pin the host pages? I would expect this to
happen only when using vfio and device assignment. Otherwise, no the
host can't reclaim a pinned page, that's the whole point of a pin to
prevent the mm from reclaiming ownership.

> KVM does not permanently pin pages. vfio does that but we are not using
> it here.

Right, so I'm confused by your pin assertion above.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13  1:43             ` Dan Williams
@ 2019-01-13  1:43               ` Dan Williams
       [not found]               ` <CAPcyv4hwcgTUpgNCefCGu4DvgkYBp5b=f+hJ+FC=s5APYKoycg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Dan Williams @ 2019-01-13  1:43 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Jan Kara, Dave Chinner, Linux Kernel Mailing List, KVM list,
	Qemu Developers, linux-nvdimm, linux-fsdevel, virtualization,
	Linux ACPI, linux-ext4, linux-xfs, Stefan Hajnoczi, Rik van Riel,
	Nitesh Narayan Lal, Kevin Wolf, Paolo Bonzini, Ross Zwisler,
	vishal l verma, dave jiang, David Hildenbrand, jmoyer,
	xiaoguangrong eric, Christoph Hellwig, Michael S. Tsirkin,
	Jason Wang, lcapitulino, Igor Mammedov, Eric Blake,
	Matthew Wilcox, Theodore Ts'o, adilger kernel, darrick wong,
	Rafael J. Wysocki

On Sat, Jan 12, 2019 at 5:38 PM Pankaj Gupta <pagupta@redhat.com> wrote:
>
>
>
> >
> > On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > > >  This patch series has implementation for "virtio pmem".
> > > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > > >  which allows to bypass the guest page cache. This also
> > > >  implements a VIRTIO based asynchronous flush mechanism.
> > >
> > > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > > like a good idea, but.....
> > >
> > > This means the guest VM can now run timing attacks to observe host
> > > side page cache residency, and depending on the implementation I'm
> > > guessing that the guest will be able to control host side page
> > > cache eviction, too (e.g. via discard or hole punch operations).
> > >
> > > Which means this functionality looks to me like a new vector for
> > > information leakage into and out of the guest VM via guest
> > > controlled host page cache manipulation.
> > >
> > > https://arxiv.org/pdf/1901.01161
> > >
> > > I might be wrong, but if I'm not we're going to have to be very
> > > careful about how guest VMs can access and manipulate host side
> > > resources like the page cache.....
> >
> > Right. Thinking about this I would be more concerned about the fact that
> > guest can effectively pin amount of host's page cache upto size of the
> > device/file passed to guest as PMEM, can't it Pankaj? Or is there some QEMU
> > magic that avoids this?
>
> Yes, guest will pin these host page cache pages using 'get_user_pages' by
> elevating the page reference count. But these pages can be reclaimed by host
> at any time when there is memory pressure.

Wait, how can the guest pin the host pages? I would expect this to
happen only when using vfio and device assignment. Otherwise, no the
host can't reclaim a pinned page, that's the whole point of a pin to
prevent the mm from reclaiming ownership.

> KVM does not permanently pin pages. vfio does that but we are not using
> it here.

Right, so I'm confused by your pin assertion above.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]               ` <CAPcyv4hwcgTUpgNCefCGu4DvgkYBp5b=f+hJ+FC=s5APYKoycg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
@ 2019-01-13  2:17                 ` " Pankaj Gupta
  2019-01-13  2:17                   ` Pankaj Gupta
       [not found]                   ` <540171952.63371441.1547345866585.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  0 siblings, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-13  2:17 UTC (permalink / raw)
  To: Dan Williams
  Cc: Jan Kara, KVM list, Michael S. Tsirkin, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, darrick wong, David Hildenbrand,
	Matthew Wilcox, Christoph Hellwig, Linux ACPI, linux-ext4,
	Rik van Riel, Stefan Hajnoczi, Paolo Bonzini,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric


> >
> >
> >
> > >
> > > On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > > > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > > > >  This patch series has implementation for "virtio pmem".
> > > > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > > > >  which allows to bypass the guest page cache. This also
> > > > >  implements a VIRTIO based asynchronous flush mechanism.
> > > >
> > > > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > > > like a good idea, but.....
> > > >
> > > > This means the guest VM can now run timing attacks to observe host
> > > > side page cache residency, and depending on the implementation I'm
> > > > guessing that the guest will be able to control host side page
> > > > cache eviction, too (e.g. via discard or hole punch operations).
> > > >
> > > > Which means this functionality looks to me like a new vector for
> > > > information leakage into and out of the guest VM via guest
> > > > controlled host page cache manipulation.
> > > >
> > > > https://arxiv.org/pdf/1901.01161
> > > >
> > > > I might be wrong, but if I'm not we're going to have to be very
> > > > careful about how guest VMs can access and manipulate host side
> > > > resources like the page cache.....
> > >
> > > Right. Thinking about this I would be more concerned about the fact that
> > > guest can effectively pin amount of host's page cache upto size of the
> > > device/file passed to guest as PMEM, can't it Pankaj? Or is there some
> > > QEMU
> > > magic that avoids this?
> >
> > Yes, guest will pin these host page cache pages using 'get_user_pages' by
> > elevating the page reference count. But these pages can be reclaimed by
> > host
> > at any time when there is memory pressure.
> 
> Wait, how can the guest pin the host pages? I would expect this to
> happen only when using vfio and device assignment. Otherwise, no the
> host can't reclaim a pinned page, that's the whole point of a pin to
> prevent the mm from reclaiming ownership.

yes. You are right I just used the pin word but it does not actually pin pages 
permanently. I had gone through the discussion on existing problems with 
get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP 
pin pages so I also used the word 'pin'. But guest does not permanently pin 
these pages and these pages can be reclaimed by host.

> 
> > KVM does not permanently pin pages. vfio does that but we are not using
> > it here.
> 
> Right, so I'm confused by your pin assertion above.

Sorry! for the confusion. 

[1] https://lwn.net/Articles/753027/

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13  2:17                 ` [Qemu-devel] " Pankaj Gupta
@ 2019-01-13  2:17                   ` Pankaj Gupta
       [not found]                   ` <540171952.63371441.1547345866585.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-13  2:17 UTC (permalink / raw)
  To: Dan Williams
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers, virtualization, adilger kernel,
	Ross Zwisler, dave jiang, darrick wong, vishal l verma,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, jmoyer, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino, Kevin Wolf, Nitesh Narayan Lal,
	Theodore Ts'o, xiaoguangrong eric, Rafael J. Wysocki,
	Linux Kernel Mailing List, linux-xfs, linux-fsdevel,
	Paolo Bonzini


> >
> >
> >
> > >
> > > On Thu 10-01-19 12:26:17, Dave Chinner wrote:
> > > > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > > > >  This patch series has implementation for "virtio pmem".
> > > > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > > > >  which allows to bypass the guest page cache. This also
> > > > >  implements a VIRTIO based asynchronous flush mechanism.
> > > >
> > > > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > > > like a good idea, but.....
> > > >
> > > > This means the guest VM can now run timing attacks to observe host
> > > > side page cache residency, and depending on the implementation I'm
> > > > guessing that the guest will be able to control host side page
> > > > cache eviction, too (e.g. via discard or hole punch operations).
> > > >
> > > > Which means this functionality looks to me like a new vector for
> > > > information leakage into and out of the guest VM via guest
> > > > controlled host page cache manipulation.
> > > >
> > > > https://arxiv.org/pdf/1901.01161
> > > >
> > > > I might be wrong, but if I'm not we're going to have to be very
> > > > careful about how guest VMs can access and manipulate host side
> > > > resources like the page cache.....
> > >
> > > Right. Thinking about this I would be more concerned about the fact that
> > > guest can effectively pin amount of host's page cache upto size of the
> > > device/file passed to guest as PMEM, can't it Pankaj? Or is there some
> > > QEMU
> > > magic that avoids this?
> >
> > Yes, guest will pin these host page cache pages using 'get_user_pages' by
> > elevating the page reference count. But these pages can be reclaimed by
> > host
> > at any time when there is memory pressure.
> 
> Wait, how can the guest pin the host pages? I would expect this to
> happen only when using vfio and device assignment. Otherwise, no the
> host can't reclaim a pinned page, that's the whole point of a pin to
> prevent the mm from reclaiming ownership.

yes. You are right I just used the pin word but it does not actually pin pages 
permanently. I had gone through the discussion on existing problems with 
get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP 
pin pages so I also used the word 'pin'. But guest does not permanently pin 
these pages and these pages can be reclaimed by host.

> 
> > KVM does not permanently pin pages. vfio does that but we are not using
> > it here.
> 
> Right, so I'm confused by your pin assertion above.

Sorry! for the confusion. 

[1] https://lwn.net/Articles/753027/

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]       ` <1326478078.61913951.1547192704870.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-13 23:29         ` Dave Chinner
  2019-01-13 23:29           ` Dave Chinner
  2019-01-13 23:38           ` Matthew Wilcox
  0 siblings, 2 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-13 23:29 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, willy-wEGCiKHe2LqWVfeAwA7xHQ,
	hch-wEGCiKHe2LqWVfeAwA7xHQ, linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Fri, Jan 11, 2019 at 02:45:04AM -0500, Pankaj Gupta wrote:
> 
> > 
> > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > >  This patch series has implementation for "virtio pmem".
> > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > >  which allows to bypass the guest page cache. This also
> > >  implements a VIRTIO based asynchronous flush mechanism.
> > 
> > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > like a good idea, but.....
> > 
> > This means the guest VM can now run timing attacks to observe host
> > side page cache residency, and depending on the implementation I'm
> > guessing that the guest will be able to control host side page
> > cache eviction, too (e.g. via discard or hole punch operations).
> 
> Not sure how? this is similar to mmapping virtual memory by any userspace 
> process. Any host userspace process can do such attack on host page cache
> using mincore & mmap shared file. 

Mincore is for monitoring, not cached eviction. And it's not
required to observe cache residency, either. That's a wide open
field containing an uncountable number of moles...

> But i don't think guest can do this alone. For virtio-pmem usecase
> guest won't be using page cache so timing attack from only guest
> side is not possible unless host userspace can run checks on page
> cache eviction state using mincore etc.  As rightly described by
> Rik, guest will only access its own page cache pages and if guest
> page cache is managed directly by host, this saves alot of effort
> for guest in transferring guest state of page cache.  

Until you have images (and hence host page cache) shared between
multiple guests. People will want to do this, because it means they
only need a single set of pages in host memory for executable
binaries rather than a set of pages per guest. Then you have
multiple guests being able to detect residency of the same set of
pages. If the guests can then, in any way, control eviction of the
pages from the host cache, then we have a guest-to-guest information
leak channel.

i.e. it's something we need to be aware of and really careful about
enabling infrastructure that /will/ be abused if guests can find a
way to influence the host side cache residency.

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13 23:29         ` Dave Chinner
@ 2019-01-13 23:29           ` Dave Chinner
  2019-01-13 23:38           ` Matthew Wilcox
  1 sibling, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-13 23:29 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, mst, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel, darrick wong, rjw

On Fri, Jan 11, 2019 at 02:45:04AM -0500, Pankaj Gupta wrote:
> 
> > 
> > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> > >  This patch series has implementation for "virtio pmem".
> > >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> > >  which allows to bypass the guest page cache. This also
> > >  implements a VIRTIO based asynchronous flush mechanism.
> > 
> > Hmmmm. Sharing the host page cache direct into the guest VM. Sounds
> > like a good idea, but.....
> > 
> > This means the guest VM can now run timing attacks to observe host
> > side page cache residency, and depending on the implementation I'm
> > guessing that the guest will be able to control host side page
> > cache eviction, too (e.g. via discard or hole punch operations).
> 
> Not sure how? this is similar to mmapping virtual memory by any userspace 
> process. Any host userspace process can do such attack on host page cache
> using mincore & mmap shared file. 

Mincore is for monitoring, not cached eviction. And it's not
required to observe cache residency, either. That's a wide open
field containing an uncountable number of moles...

> But i don't think guest can do this alone. For virtio-pmem usecase
> guest won't be using page cache so timing attack from only guest
> side is not possible unless host userspace can run checks on page
> cache eviction state using mincore etc.  As rightly described by
> Rik, guest will only access its own page cache pages and if guest
> page cache is managed directly by host, this saves alot of effort
> for guest in transferring guest state of page cache.  

Until you have images (and hence host page cache) shared between
multiple guests. People will want to do this, because it means they
only need a single set of pages in host memory for executable
binaries rather than a set of pages per guest. Then you have
multiple guests being able to detect residency of the same set of
pages. If the guests can then, in any way, control eviction of the
pages from the host cache, then we have a guest-to-guest information
leak channel.

i.e. it's something we need to be aware of and really careful about
enabling infrastructure that /will/ be abused if guests can find a
way to influence the host side cache residency.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13 23:29         ` Dave Chinner
  2019-01-13 23:29           ` Dave Chinner
@ 2019-01-13 23:38           ` Matthew Wilcox
  2019-01-13 23:38             ` Matthew Wilcox
       [not found]             ` <20190113233820.GX6310-PfSpb0PWhxZc2C7mugBRk2EX/6BAtgUQ@public.gmane.org>
  1 sibling, 2 replies; 61+ messages in thread
From: Matthew Wilcox @ 2019-01-13 23:38 UTC (permalink / raw)
  To: Dave Chinner
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Mon, Jan 14, 2019 at 10:29:02AM +1100, Dave Chinner wrote:
> Until you have images (and hence host page cache) shared between
> multiple guests. People will want to do this, because it means they
> only need a single set of pages in host memory for executable
> binaries rather than a set of pages per guest. Then you have
> multiple guests being able to detect residency of the same set of
> pages. If the guests can then, in any way, control eviction of the
> pages from the host cache, then we have a guest-to-guest information
> leak channel.

I don't think we should ever be considering something that would allow a
guest to evict page's from the host's pagecache [1].  The guest should
be able to kick its own references to the host's pagecache out of its
own pagecache, but not be able to influence whether the host or another
guest has a read-only mapping cached.

[1] Unless the guest is allowed to modify the host's file; obviously
truncation, holepunching, etc are going to evict pages from the host's
page cache.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-13 23:38           ` Matthew Wilcox
@ 2019-01-13 23:38             ` Matthew Wilcox
       [not found]             ` <20190113233820.GX6310-PfSpb0PWhxZc2C7mugBRk2EX/6BAtgUQ@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Matthew Wilcox @ 2019-01-13 23:38 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Pankaj Gupta, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	jack, stefanha, dan j williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal l verma, dave jiang, david, jmoyer,
	xiaoguangrong eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, tytso, adilger kernel, darrick wong, rjw

On Mon, Jan 14, 2019 at 10:29:02AM +1100, Dave Chinner wrote:
> Until you have images (and hence host page cache) shared between
> multiple guests. People will want to do this, because it means they
> only need a single set of pages in host memory for executable
> binaries rather than a set of pages per guest. Then you have
> multiple guests being able to detect residency of the same set of
> pages. If the guests can then, in any way, control eviction of the
> pages from the host cache, then we have a guest-to-guest information
> leak channel.

I don't think we should ever be considering something that would allow a
guest to evict page's from the host's pagecache [1].  The guest should
be able to kick its own references to the host's pagecache out of its
own pagecache, but not be able to influence whether the host or another
guest has a read-only mapping cached.

[1] Unless the guest is allowed to modify the host's file; obviously
truncation, holepunching, etc are going to evict pages from the host's
page cache.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]             ` <20190113233820.GX6310-PfSpb0PWhxZc2C7mugBRk2EX/6BAtgUQ@public.gmane.org>
@ 2019-01-14  2:50               ` Dave Chinner
  2019-01-14  2:50                 ` Dave Chinner
  2019-01-14  7:15               ` Pankaj Gupta
  1 sibling, 1 reply; 61+ messages in thread
From: Dave Chinner @ 2019-01-14  2:50 UTC (permalink / raw)
  To: Matthew Wilcox
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Sun, Jan 13, 2019 at 03:38:21PM -0800, Matthew Wilcox wrote:
> On Mon, Jan 14, 2019 at 10:29:02AM +1100, Dave Chinner wrote:
> > Until you have images (and hence host page cache) shared between
> > multiple guests. People will want to do this, because it means they
> > only need a single set of pages in host memory for executable
> > binaries rather than a set of pages per guest. Then you have
> > multiple guests being able to detect residency of the same set of
> > pages. If the guests can then, in any way, control eviction of the
> > pages from the host cache, then we have a guest-to-guest information
> > leak channel.
> 
> I don't think we should ever be considering something that would allow a
> guest to evict page's from the host's pagecache [1].  The guest should
> be able to kick its own references to the host's pagecache out of its
> own pagecache, but not be able to influence whether the host or another
> guest has a read-only mapping cached.
> 
> [1] Unless the guest is allowed to modify the host's file; obviously
> truncation, holepunching, etc are going to evict pages from the host's
> page cache.

Right, and that's exactly what I mean by "we need to be real careful
with functionality like this".

To be honest, I really don't think I've even touched the surface
here.

e.g. Filesystems and storage can share logical and physical extents.
Which means that image files that share storage (e.g.  because they
are all cloned from the same master image and/or there's in-line
deduplication running on the storage) and can be directly accessed
by guests may very well be susceptible to detection of host side
deduplication and subsequent copy-on-write operations.

This really doesn't seem much different to me from the guest being
able to infer host side KSM page deduplication and COW operation in
the guest side page cache.  The only difference is that DAX is being
used to probe the host side page cache and storage rather than the
guest side.

IOWs, I suspect there's a world of pain waiting for us if we punch
huge holes through the virtual machine abstractions like this.

Improving performance is a laudible goal, but at what price?

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14  2:50               ` Dave Chinner
@ 2019-01-14  2:50                 ` Dave Chinner
  0 siblings, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-14  2:50 UTC (permalink / raw)
  To: Matthew Wilcox
  Cc: Pankaj Gupta, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	jack, stefanha, dan j williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal l verma, dave jiang, david, jmoyer,
	xiaoguangrong eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, tytso, adilger kernel, darrick wong, rjw

On Sun, Jan 13, 2019 at 03:38:21PM -0800, Matthew Wilcox wrote:
> On Mon, Jan 14, 2019 at 10:29:02AM +1100, Dave Chinner wrote:
> > Until you have images (and hence host page cache) shared between
> > multiple guests. People will want to do this, because it means they
> > only need a single set of pages in host memory for executable
> > binaries rather than a set of pages per guest. Then you have
> > multiple guests being able to detect residency of the same set of
> > pages. If the guests can then, in any way, control eviction of the
> > pages from the host cache, then we have a guest-to-guest information
> > leak channel.
> 
> I don't think we should ever be considering something that would allow a
> guest to evict page's from the host's pagecache [1].  The guest should
> be able to kick its own references to the host's pagecache out of its
> own pagecache, but not be able to influence whether the host or another
> guest has a read-only mapping cached.
> 
> [1] Unless the guest is allowed to modify the host's file; obviously
> truncation, holepunching, etc are going to evict pages from the host's
> page cache.

Right, and that's exactly what I mean by "we need to be real careful
with functionality like this".

To be honest, I really don't think I've even touched the surface
here.

e.g. Filesystems and storage can share logical and physical extents.
Which means that image files that share storage (e.g.  because they
are all cloned from the same master image and/or there's in-line
deduplication running on the storage) and can be directly accessed
by guests may very well be susceptible to detection of host side
deduplication and subsequent copy-on-write operations.

This really doesn't seem much different to me from the guest being
able to infer host side KSM page deduplication and COW operation in
the guest side page cache.  The only difference is that DAX is being
used to probe the host side page cache and storage rather than the
guest side.

IOWs, I suspect there's a world of pain waiting for us if we punch
huge holes through the virtual machine abstractions like this.

Improving performance is a laudible goal, but at what price?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]             ` <20190113233820.GX6310-PfSpb0PWhxZc2C7mugBRk2EX/6BAtgUQ@public.gmane.org>
  2019-01-14  2:50               ` Dave Chinner
@ 2019-01-14  7:15               ` Pankaj Gupta
  2019-01-14  7:15                 ` Pankaj Gupta
       [not found]                 ` <942065073.64011540.1547450140670.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-14  7:15 UTC (permalink / raw)
  To: Matthew Wilcox
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA, Dave Chinner,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA


> > Until you have images (and hence host page cache) shared between
> > multiple guests. People will want to do this, because it means they
> > only need a single set of pages in host memory for executable
> > binaries rather than a set of pages per guest. Then you have
> > multiple guests being able to detect residency of the same set of
> > pages. If the guests can then, in any way, control eviction of the
> > pages from the host cache, then we have a guest-to-guest information
> > leak channel.
> 
> I don't think we should ever be considering something that would allow a
> guest to evict page's from the host's pagecache [1].  The guest should
> be able to kick its own references to the host's pagecache out of its
> own pagecache, but not be able to influence whether the host or another
> guest has a read-only mapping cached.
> 
> [1] Unless the guest is allowed to modify the host's file; obviously
> truncation, holepunching, etc are going to evict pages from the host's
> page cache.

This is so correct. Guest does not not evict host page cache pages directly. 
In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
Its solely decision of host to take action on the host page cache pages.

In case of virtio-pmem, guest does not modify host file directly i.e don't
perform hole punch & truncation operation directly on host file.  

Thanks,
Pankaj 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14  7:15               ` Pankaj Gupta
@ 2019-01-14  7:15                 ` Pankaj Gupta
       [not found]                 ` <942065073.64011540.1547450140670.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-14  7:15 UTC (permalink / raw)
  To: Matthew Wilcox
  Cc: Dave Chinner, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	jack, stefanha, dan j williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal l verma, dave jiang, david, jmoyer,
	xiaoguangrong eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, tytso, adilger kernel, darrick wong, rjw


> > Until you have images (and hence host page cache) shared between
> > multiple guests. People will want to do this, because it means they
> > only need a single set of pages in host memory for executable
> > binaries rather than a set of pages per guest. Then you have
> > multiple guests being able to detect residency of the same set of
> > pages. If the guests can then, in any way, control eviction of the
> > pages from the host cache, then we have a guest-to-guest information
> > leak channel.
> 
> I don't think we should ever be considering something that would allow a
> guest to evict page's from the host's pagecache [1].  The guest should
> be able to kick its own references to the host's pagecache out of its
> own pagecache, but not be able to influence whether the host or another
> guest has a read-only mapping cached.
> 
> [1] Unless the guest is allowed to modify the host's file; obviously
> truncation, holepunching, etc are going to evict pages from the host's
> page cache.

This is so correct. Guest does not not evict host page cache pages directly. 
In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
Its solely decision of host to take action on the host page cache pages.

In case of virtio-pmem, guest does not modify host file directly i.e don't
perform hole punch & truncation operation directly on host file.  

Thanks,
Pankaj 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                   ` <540171952.63371441.1547345866585.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-14  9:55                     ` Jan Kara
  2019-01-14  9:55                       ` Jan Kara
       [not found]                       ` <20190114095520.GC13316-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
  0 siblings, 2 replies; 61+ messages in thread
From: Jan Kara @ 2019-01-14  9:55 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Jan Kara, KVM list, Michael S. Tsirkin, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, darrick wong, David Hildenbrand,
	Matthew Wilcox, Christoph Hellwig, Linux ACPI, linux-ext4,
	Rik van Riel, Stefan Hajnoczi, Paolo Bonzini,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric

On Sat 12-01-19 21:17:46, Pankaj Gupta wrote:
> > > > Right. Thinking about this I would be more concerned about the fact that
> > > > guest can effectively pin amount of host's page cache upto size of the
> > > > device/file passed to guest as PMEM, can't it Pankaj? Or is there some
> > > > QEMU
> > > > magic that avoids this?
> > >
> > > Yes, guest will pin these host page cache pages using 'get_user_pages' by
> > > elevating the page reference count. But these pages can be reclaimed by
> > > host
> > > at any time when there is memory pressure.
> > 
> > Wait, how can the guest pin the host pages? I would expect this to
> > happen only when using vfio and device assignment. Otherwise, no the
> > host can't reclaim a pinned page, that's the whole point of a pin to
> > prevent the mm from reclaiming ownership.
> 
> yes. You are right I just used the pin word but it does not actually pin pages 
> permanently. I had gone through the discussion on existing problems with 
> get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP 
> pin pages so I also used the word 'pin'. But guest does not permanently pin 
> these pages and these pages can be reclaimed by host.

OK, then I was just confused how virtio-pmem is going to work. Thanks for
explanation! So can I imagine this as guest mmaping the host file and
providing the mapped range as "NVDIMM pages" to the kernel inside the
guest? Or is it more complex?

								Honza

-- 
Jan Kara <jack-IBi9RG/b67k@public.gmane.org>
SUSE Labs, CR

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14  9:55                     ` Jan Kara
@ 2019-01-14  9:55                       ` Jan Kara
       [not found]                       ` <20190114095520.GC13316-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Jan Kara @ 2019-01-14  9:55 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Dan Williams, Jan Kara, KVM list, David Hildenbrand,
	linux-nvdimm, Jason Wang, Dave Chinner, Qemu Developers,
	virtualization, adilger kernel, Ross Zwisler, dave jiang,
	darrick wong, vishal l verma, Michael S. Tsirkin, Matthew Wilcox,
	Christoph Hellwig, Linux ACPI, jmoyer, linux-ext4, Rik van Riel,
	Stefan Hajnoczi, Igor Mammedov, lcapitulino, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric,
	Rafael J. Wysocki, Linux Kernel Mailing List, linux-xfs,
	linux-fsdevel, Paolo Bonzini

On Sat 12-01-19 21:17:46, Pankaj Gupta wrote:
> > > > Right. Thinking about this I would be more concerned about the fact that
> > > > guest can effectively pin amount of host's page cache upto size of the
> > > > device/file passed to guest as PMEM, can't it Pankaj? Or is there some
> > > > QEMU
> > > > magic that avoids this?
> > >
> > > Yes, guest will pin these host page cache pages using 'get_user_pages' by
> > > elevating the page reference count. But these pages can be reclaimed by
> > > host
> > > at any time when there is memory pressure.
> > 
> > Wait, how can the guest pin the host pages? I would expect this to
> > happen only when using vfio and device assignment. Otherwise, no the
> > host can't reclaim a pinned page, that's the whole point of a pin to
> > prevent the mm from reclaiming ownership.
> 
> yes. You are right I just used the pin word but it does not actually pin pages 
> permanently. I had gone through the discussion on existing problems with 
> get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP 
> pin pages so I also used the word 'pin'. But guest does not permanently pin 
> these pages and these pages can be reclaimed by host.

OK, then I was just confused how virtio-pmem is going to work. Thanks for
explanation! So can I imagine this as guest mmaping the host file and
providing the mapped range as "NVDIMM pages" to the kernel inside the
guest? Or is it more complex?

								Honza

-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                       ` <20190114095520.GC13316-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
@ 2019-01-14 10:16                         ` Pankaj Gupta
  2019-01-14 10:16                           ` Pankaj Gupta
  0 siblings, 1 reply; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-14 10:16 UTC (permalink / raw)
  To: Jan Kara
  Cc: KVM list, Michael S. Tsirkin, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, darrick wong, David Hildenbrand,
	Matthew Wilcox, Christoph Hellwig, Linux ACPI, linux-ext4,
	Rik van Riel, Stefan Hajnoczi, Paolo Bonzini,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric,
	Rafael J. Wysock


> > > > > Right. Thinking about this I would be more concerned about the fact
> > > > > that
> > > > > guest can effectively pin amount of host's page cache upto size of
> > > > > the
> > > > > device/file passed to guest as PMEM, can't it Pankaj? Or is there
> > > > > some
> > > > > QEMU
> > > > > magic that avoids this?
> > > >
> > > > Yes, guest will pin these host page cache pages using 'get_user_pages'
> > > > by
> > > > elevating the page reference count. But these pages can be reclaimed by
> > > > host
> > > > at any time when there is memory pressure.
> > > 
> > > Wait, how can the guest pin the host pages? I would expect this to
> > > happen only when using vfio and device assignment. Otherwise, no the
> > > host can't reclaim a pinned page, that's the whole point of a pin to
> > > prevent the mm from reclaiming ownership.
> > 
> > yes. You are right I just used the pin word but it does not actually pin
> > pages
> > permanently. I had gone through the discussion on existing problems with
> > get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP
> > pin pages so I also used the word 'pin'. But guest does not permanently pin
> > these pages and these pages can be reclaimed by host.
> 
> OK, then I was just confused how virtio-pmem is going to work. Thanks for
> explanation! So can I imagine this as guest mmaping the host file and
> providing the mapped range as "NVDIMM pages" to the kernel inside the
> guest? Or is it more complex?

yes, that's correct. Host's Qemu process virtual address range is used as guest physical 
address and a direct mapping(EPT/NPT) is established. At guest side, this physical memory
range is plugged into guest system memory map and DAX mapping is setup using nvdimm calls.

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 10:16                         ` Pankaj Gupta
@ 2019-01-14 10:16                           ` Pankaj Gupta
  0 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-14 10:16 UTC (permalink / raw)
  To: Jan Kara
  Cc: Dan Williams, KVM list, David Hildenbrand, linux-nvdimm,
	Jason Wang, Dave Chinner, Qemu Developers, virtualization,
	adilger kernel, Ross Zwisler, dave jiang, darrick wong,
	vishal l verma, Michael S. Tsirkin, Matthew Wilcox,
	Christoph Hellwig, Linux ACPI, jmoyer, linux-ext4, Rik van Riel,
	Stefan Hajnoczi, Igor Mammedov, lcapitulino, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric,
	Rafael J. Wysocki, Linux Kernel Mailing List, linux-xfs,
	linux-fsdevel, Paolo Bonzini


> > > > > Right. Thinking about this I would be more concerned about the fact
> > > > > that
> > > > > guest can effectively pin amount of host's page cache upto size of
> > > > > the
> > > > > device/file passed to guest as PMEM, can't it Pankaj? Or is there
> > > > > some
> > > > > QEMU
> > > > > magic that avoids this?
> > > >
> > > > Yes, guest will pin these host page cache pages using 'get_user_pages'
> > > > by
> > > > elevating the page reference count. But these pages can be reclaimed by
> > > > host
> > > > at any time when there is memory pressure.
> > > 
> > > Wait, how can the guest pin the host pages? I would expect this to
> > > happen only when using vfio and device assignment. Otherwise, no the
> > > host can't reclaim a pinned page, that's the whole point of a pin to
> > > prevent the mm from reclaiming ownership.
> > 
> > yes. You are right I just used the pin word but it does not actually pin
> > pages
> > permanently. I had gone through the discussion on existing problems with
> > get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP
> > pin pages so I also used the word 'pin'. But guest does not permanently pin
> > these pages and these pages can be reclaimed by host.
> 
> OK, then I was just confused how virtio-pmem is going to work. Thanks for
> explanation! So can I imagine this as guest mmaping the host file and
> providing the mapped range as "NVDIMM pages" to the kernel inside the
> guest? Or is it more complex?

yes, that's correct. Host's Qemu process virtual address range is used as guest physical 
address and a direct mapping(EPT/NPT) is established. At guest side, this physical memory
range is plugged into guest system memory map and DAX mapping is setup using nvdimm calls.

Thanks,
Pankaj

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 2/5]  virtio-pmem: Add virtio pmem driver
  2019-01-09 14:47 ` [PATCH v3 2/5] virtio-pmem: Add virtio pmem driver Pankaj Gupta
@ 2019-01-14 15:54   ` Michael S. Tsirkin
  2019-01-14 15:54     ` Michael S. Tsirkin
       [not found]     ` <20190114105314-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
  0 siblings, 2 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-01-14 15:54 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw

On Wed, Jan 09, 2019 at 08:17:33PM +0530, Pankaj Gupta wrote:
> This patch adds virtio-pmem driver for KVM guest.
> 
> Guest reads the persistent memory range information from
> Qemu over VIRTIO and registers it on nvdimm_bus. It also
> creates a nd_region object with the persistent memory
> range information so that existing 'nvdimm/pmem' driver
> can reserve this into system memory map. This way
> 'virtio-pmem' driver uses existing functionality of pmem
> driver to register persistent memory compatible for DAX
> capable filesystems.
> 
> This also provides function to perform guest flush over
> VIRTIO from 'pmem' driver when userspace performs flush
> on DAX memory range.
> 
> Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
> ---
>  drivers/nvdimm/virtio_pmem.c     |  84 ++++++++++++++++++++++++++
>  drivers/virtio/Kconfig           |  10 ++++
>  drivers/virtio/Makefile          |   1 +
>  drivers/virtio/pmem.c            | 124 +++++++++++++++++++++++++++++++++++++++
>  include/linux/virtio_pmem.h      |  60 +++++++++++++++++++
>  include/uapi/linux/virtio_ids.h  |   1 +
>  include/uapi/linux/virtio_pmem.h |  10 ++++

As with any uapi change, you need to CC the virtio dev
mailing list (subscribers only, sorry about that).


>  7 files changed, 290 insertions(+)
>  create mode 100644 drivers/nvdimm/virtio_pmem.c
>  create mode 100644 drivers/virtio/pmem.c
>  create mode 100644 include/linux/virtio_pmem.h
>  create mode 100644 include/uapi/linux/virtio_pmem.h
> 
> diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> new file mode 100644
> index 0000000..2a1b1ba
> --- /dev/null
> +++ b/drivers/nvdimm/virtio_pmem.c
> @@ -0,0 +1,84 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * virtio_pmem.c: Virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and provides a virtio based flushing
> + * interface.
> + */
> +#include <linux/virtio_pmem.h>
> +#include "nd.h"
> +
> + /* The interrupt handler */
> +void host_ack(struct virtqueue *vq)
> +{
> +	unsigned int len;
> +	unsigned long flags;
> +	struct virtio_pmem_request *req, *req_buf;
> +	struct virtio_pmem *vpmem = vq->vdev->priv;
> +
> +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
> +		req->done = true;
> +		wake_up(&req->host_acked);
> +
> +		if (!list_empty(&vpmem->req_list)) {
> +			req_buf = list_first_entry(&vpmem->req_list,
> +					struct virtio_pmem_request, list);
> +			list_del(&vpmem->req_list);
> +			req_buf->wq_buf_avail = true;
> +			wake_up(&req_buf->wq_buf);
> +		}
> +	}
> +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +}
> +EXPORT_SYMBOL_GPL(host_ack);
> +
> + /* The request submission function */
> +int virtio_pmem_flush(struct nd_region *nd_region)
> +{
> +	int err;
> +	unsigned long flags;
> +	struct scatterlist *sgs[2], sg, ret;
> +	struct virtio_device *vdev = nd_region->provider_data;
> +	struct virtio_pmem *vpmem = vdev->priv;
> +	struct virtio_pmem_request *req;
> +
> +	might_sleep();
> +	req = kmalloc(sizeof(*req), GFP_KERNEL);
> +	if (!req)
> +		return -ENOMEM;
> +
> +	req->done = req->wq_buf_avail = false;
> +	strcpy(req->name, "FLUSH");
> +	init_waitqueue_head(&req->host_acked);
> +	init_waitqueue_head(&req->wq_buf);
> +	sg_init_one(&sg, req->name, strlen(req->name));
> +	sgs[0] = &sg;
> +	sg_init_one(&ret, &req->ret, sizeof(req->ret));
> +	sgs[1] = &ret;
> +
> +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> +	if (err) {
> +		dev_err(&vdev->dev, "failed to send command to virtio pmem device\n");
> +
> +		list_add_tail(&vpmem->req_list, &req->list);
> +		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +
> +		/* When host has read buffer, this completes via host_ack */
> +		wait_event(req->wq_buf, req->wq_buf_avail);
> +		spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	}
> +	virtqueue_kick(vpmem->req_vq);
> +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +
> +	/* When host has read buffer, this completes via host_ack */
> +	wait_event(req->host_acked, req->done);
> +	err = req->ret;
> +	kfree(req);
> +
> +	return err;
> +};
> +EXPORT_SYMBOL_GPL(virtio_pmem_flush);
> +MODULE_LICENSE("GPL");
> diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
> index 3589764..9f634a2 100644
> --- a/drivers/virtio/Kconfig
> +++ b/drivers/virtio/Kconfig
> @@ -42,6 +42,16 @@ config VIRTIO_PCI_LEGACY
>  
>  	  If unsure, say Y.
>  
> +config VIRTIO_PMEM
> +	tristate "Support for virtio pmem driver"
> +	depends on VIRTIO
> +	depends on LIBNVDIMM
> +	help
> +	This driver provides support for virtio based flushing interface
> +	for persistent memory range.
> +
> +	If unsure, say M.
> +
>  config VIRTIO_BALLOON
>  	tristate "Virtio balloon driver"
>  	depends on VIRTIO
> diff --git a/drivers/virtio/Makefile b/drivers/virtio/Makefile
> index 3a2b5c5..143ce91 100644
> --- a/drivers/virtio/Makefile
> +++ b/drivers/virtio/Makefile
> @@ -6,3 +6,4 @@ virtio_pci-y := virtio_pci_modern.o virtio_pci_common.o
>  virtio_pci-$(CONFIG_VIRTIO_PCI_LEGACY) += virtio_pci_legacy.o
>  obj-$(CONFIG_VIRTIO_BALLOON) += virtio_balloon.o
>  obj-$(CONFIG_VIRTIO_INPUT) += virtio_input.o
> +obj-$(CONFIG_VIRTIO_PMEM) += pmem.o ../nvdimm/virtio_pmem.o
> diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
> new file mode 100644
> index 0000000..51f5349
> --- /dev/null
> +++ b/drivers/virtio/pmem.c
> @@ -0,0 +1,124 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * virtio_pmem.c: Virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and registers the virtual pmem device
> + * with libnvdimm core.
> + */
> +#include <linux/virtio_pmem.h>
> +#include <../../drivers/nvdimm/nd.h>
> +
> +static struct virtio_device_id id_table[] = {
> +	{ VIRTIO_ID_PMEM, VIRTIO_DEV_ANY_ID },
> +	{ 0 },
> +};
> +
> + /* Initialize virt queue */
> +static int init_vq(struct virtio_pmem *vpmem)
> +{
> +	struct virtqueue *vq;
> +
> +	/* single vq */
> +	vpmem->req_vq = vq = virtio_find_single_vq(vpmem->vdev,
> +				host_ack, "flush_queue");
> +	if (IS_ERR(vq))
> +		return PTR_ERR(vq);
> +
> +	spin_lock_init(&vpmem->pmem_lock);
> +	INIT_LIST_HEAD(&vpmem->req_list);
> +
> +	return 0;
> +};
> +
> +static int virtio_pmem_probe(struct virtio_device *vdev)
> +{
> +	int err = 0;
> +	struct resource res;
> +	struct virtio_pmem *vpmem;
> +	struct nvdimm_bus *nvdimm_bus;
> +	struct nd_region_desc ndr_desc;
> +	int nid = dev_to_node(&vdev->dev);
> +	struct nd_region *nd_region;
> +
> +	if (!vdev->config->get) {
> +		dev_err(&vdev->dev, "%s failure: config disabled\n",
> +			__func__);
> +		return -EINVAL;
> +	}
> +
> +	vdev->priv = vpmem = devm_kzalloc(&vdev->dev, sizeof(*vpmem),
> +			GFP_KERNEL);
> +	if (!vpmem) {
> +		err = -ENOMEM;
> +		goto out_err;
> +	}
> +
> +	vpmem->vdev = vdev;
> +	err = init_vq(vpmem);
> +	if (err)
> +		goto out_err;
> +
> +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> +			start, &vpmem->start);
> +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> +			size, &vpmem->size);
> +
> +	res.start = vpmem->start;
> +	res.end   = vpmem->start + vpmem->size-1;
> +	vpmem->nd_desc.provider_name = "virtio-pmem";
> +	vpmem->nd_desc.module = THIS_MODULE;
> +
> +	vpmem->nvdimm_bus = nvdimm_bus = nvdimm_bus_register(&vdev->dev,
> +						&vpmem->nd_desc);
> +	if (!nvdimm_bus)
> +		goto out_vq;
> +
> +	dev_set_drvdata(&vdev->dev, nvdimm_bus);
> +	memset(&ndr_desc, 0, sizeof(ndr_desc));
> +
> +	ndr_desc.res = &res;
> +	ndr_desc.numa_node = nid;
> +	ndr_desc.flush = virtio_pmem_flush;
> +	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
> +	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
> +	nd_region->provider_data =  dev_to_virtio
> +					(nd_region->dev.parent->parent);
> +
> +	if (!nd_region)
> +		goto out_nd;
> +
> +	//virtio_device_ready(vdev);
> +	return 0;
> +out_nd:
> +	err = -ENXIO;
> +	nvdimm_bus_unregister(nvdimm_bus);
> +out_vq:
> +	vdev->config->del_vqs(vdev);
> +out_err:
> +	dev_err(&vdev->dev, "failed to register virtio pmem memory\n");
> +	return err;
> +}
> +
> +static void virtio_pmem_remove(struct virtio_device *vdev)
> +{
> +	struct virtio_pmem *vpmem = vdev->priv;
> +	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
> +
> +	nvdimm_bus_unregister(nvdimm_bus);
> +	vdev->config->del_vqs(vdev);
> +	kfree(vpmem);
> +}
> +
> +static struct virtio_driver virtio_pmem_driver = {
> +	.driver.name		= KBUILD_MODNAME,
> +	.driver.owner		= THIS_MODULE,
> +	.id_table		= id_table,
> +	.probe			= virtio_pmem_probe,
> +	.remove			= virtio_pmem_remove,
> +};
> +
> +module_virtio_driver(virtio_pmem_driver);
> +MODULE_DEVICE_TABLE(virtio, id_table);
> +MODULE_DESCRIPTION("Virtio pmem driver");
> +MODULE_LICENSE("GPL");
> diff --git a/include/linux/virtio_pmem.h b/include/linux/virtio_pmem.h
> new file mode 100644
> index 0000000..224f9d9
> --- /dev/null
> +++ b/include/linux/virtio_pmem.h
> @@ -0,0 +1,60 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +/*
> + * virtio_pmem.h: virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and provides a virtio based flushing
> + * interface.
> + **/
> +
> +#ifndef _LINUX_VIRTIO_PMEM_H
> +#define _LINUX_VIRTIO_PMEM_H
> +
> +#include <linux/virtio_ids.h>
> +#include <linux/module.h>
> +#include <linux/virtio_config.h>
> +#include <uapi/linux/virtio_pmem.h>
> +#include <linux/libnvdimm.h>
> +#include <linux/spinlock.h>
> +
> +struct virtio_pmem_request {
> +	/* Host return status corresponding to flush request */
> +	int ret;
> +
> +	/* command name*/
> +	char name[16];
> +
> +	/* Wait queue to process deferred work after ack from host */
> +	wait_queue_head_t host_acked;
> +	bool done;
> +
> +	/* Wait queue to process deferred work after virt queue buffer avail */
> +	wait_queue_head_t wq_buf;
> +	bool wq_buf_avail;
> +	struct list_head list;
> +};
> +
> +struct virtio_pmem {
> +	struct virtio_device *vdev;
> +
> +	/* Virtio pmem request queue */
> +	struct virtqueue *req_vq;
> +
> +	/* nvdimm bus registers virtio pmem device */
> +	struct nvdimm_bus *nvdimm_bus;
> +	struct nvdimm_bus_descriptor nd_desc;
> +
> +	/* List to store deferred work if virtqueue is full */
> +	struct list_head req_list;
> +
> +	/* Synchronize virtqueue data */
> +	spinlock_t pmem_lock;
> +
> +	/* Memory region information */
> +	uint64_t start;
> +	uint64_t size;
> +};
> +
> +void host_ack(struct virtqueue *vq);
> +int virtio_pmem_flush(struct nd_region *nd_region);
> +#endif
> diff --git a/include/uapi/linux/virtio_ids.h b/include/uapi/linux/virtio_ids.h
> index 6d5c3b2..3463895 100644
> --- a/include/uapi/linux/virtio_ids.h
> +++ b/include/uapi/linux/virtio_ids.h
> @@ -43,5 +43,6 @@
>  #define VIRTIO_ID_INPUT        18 /* virtio input */
>  #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
>  #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
> +#define VIRTIO_ID_PMEM         25 /* virtio pmem */
>  
>  #endif /* _LINUX_VIRTIO_IDS_H */
> diff --git a/include/uapi/linux/virtio_pmem.h b/include/uapi/linux/virtio_pmem.h
> new file mode 100644
> index 0000000..fa3f7d5
> --- /dev/null
> +++ b/include/uapi/linux/virtio_pmem.h
> @@ -0,0 +1,10 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +
> +#ifndef _UAPI_LINUX_VIRTIO_PMEM_H
> +#define _UAPI_LINUX_VIRTIO_PMEM_H
> +
> +struct virtio_pmem_config {
> +	__le64 start;
> +	__le64 size;
> +};
> +#endif
> -- 
> 2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 2/5]  virtio-pmem: Add virtio pmem driver
  2019-01-14 15:54   ` Michael S. Tsirkin
@ 2019-01-14 15:54     ` Michael S. Tsirkin
       [not found]     ` <20190114105314-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-01-14 15:54 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw

On Wed, Jan 09, 2019 at 08:17:33PM +0530, Pankaj Gupta wrote:
> This patch adds virtio-pmem driver for KVM guest.
> 
> Guest reads the persistent memory range information from
> Qemu over VIRTIO and registers it on nvdimm_bus. It also
> creates a nd_region object with the persistent memory
> range information so that existing 'nvdimm/pmem' driver
> can reserve this into system memory map. This way
> 'virtio-pmem' driver uses existing functionality of pmem
> driver to register persistent memory compatible for DAX
> capable filesystems.
> 
> This also provides function to perform guest flush over
> VIRTIO from 'pmem' driver when userspace performs flush
> on DAX memory range.
> 
> Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
> ---
>  drivers/nvdimm/virtio_pmem.c     |  84 ++++++++++++++++++++++++++
>  drivers/virtio/Kconfig           |  10 ++++
>  drivers/virtio/Makefile          |   1 +
>  drivers/virtio/pmem.c            | 124 +++++++++++++++++++++++++++++++++++++++
>  include/linux/virtio_pmem.h      |  60 +++++++++++++++++++
>  include/uapi/linux/virtio_ids.h  |   1 +
>  include/uapi/linux/virtio_pmem.h |  10 ++++

As with any uapi change, you need to CC the virtio dev
mailing list (subscribers only, sorry about that).


>  7 files changed, 290 insertions(+)
>  create mode 100644 drivers/nvdimm/virtio_pmem.c
>  create mode 100644 drivers/virtio/pmem.c
>  create mode 100644 include/linux/virtio_pmem.h
>  create mode 100644 include/uapi/linux/virtio_pmem.h
> 
> diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> new file mode 100644
> index 0000000..2a1b1ba
> --- /dev/null
> +++ b/drivers/nvdimm/virtio_pmem.c
> @@ -0,0 +1,84 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * virtio_pmem.c: Virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and provides a virtio based flushing
> + * interface.
> + */
> +#include <linux/virtio_pmem.h>
> +#include "nd.h"
> +
> + /* The interrupt handler */
> +void host_ack(struct virtqueue *vq)
> +{
> +	unsigned int len;
> +	unsigned long flags;
> +	struct virtio_pmem_request *req, *req_buf;
> +	struct virtio_pmem *vpmem = vq->vdev->priv;
> +
> +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
> +		req->done = true;
> +		wake_up(&req->host_acked);
> +
> +		if (!list_empty(&vpmem->req_list)) {
> +			req_buf = list_first_entry(&vpmem->req_list,
> +					struct virtio_pmem_request, list);
> +			list_del(&vpmem->req_list);
> +			req_buf->wq_buf_avail = true;
> +			wake_up(&req_buf->wq_buf);
> +		}
> +	}
> +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +}
> +EXPORT_SYMBOL_GPL(host_ack);
> +
> + /* The request submission function */
> +int virtio_pmem_flush(struct nd_region *nd_region)
> +{
> +	int err;
> +	unsigned long flags;
> +	struct scatterlist *sgs[2], sg, ret;
> +	struct virtio_device *vdev = nd_region->provider_data;
> +	struct virtio_pmem *vpmem = vdev->priv;
> +	struct virtio_pmem_request *req;
> +
> +	might_sleep();
> +	req = kmalloc(sizeof(*req), GFP_KERNEL);
> +	if (!req)
> +		return -ENOMEM;
> +
> +	req->done = req->wq_buf_avail = false;
> +	strcpy(req->name, "FLUSH");
> +	init_waitqueue_head(&req->host_acked);
> +	init_waitqueue_head(&req->wq_buf);
> +	sg_init_one(&sg, req->name, strlen(req->name));
> +	sgs[0] = &sg;
> +	sg_init_one(&ret, &req->ret, sizeof(req->ret));
> +	sgs[1] = &ret;
> +
> +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> +	if (err) {
> +		dev_err(&vdev->dev, "failed to send command to virtio pmem device\n");
> +
> +		list_add_tail(&vpmem->req_list, &req->list);
> +		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +
> +		/* When host has read buffer, this completes via host_ack */
> +		wait_event(req->wq_buf, req->wq_buf_avail);
> +		spin_lock_irqsave(&vpmem->pmem_lock, flags);
> +	}
> +	virtqueue_kick(vpmem->req_vq);
> +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +
> +	/* When host has read buffer, this completes via host_ack */
> +	wait_event(req->host_acked, req->done);
> +	err = req->ret;
> +	kfree(req);
> +
> +	return err;
> +};
> +EXPORT_SYMBOL_GPL(virtio_pmem_flush);
> +MODULE_LICENSE("GPL");
> diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
> index 3589764..9f634a2 100644
> --- a/drivers/virtio/Kconfig
> +++ b/drivers/virtio/Kconfig
> @@ -42,6 +42,16 @@ config VIRTIO_PCI_LEGACY
>  
>  	  If unsure, say Y.
>  
> +config VIRTIO_PMEM
> +	tristate "Support for virtio pmem driver"
> +	depends on VIRTIO
> +	depends on LIBNVDIMM
> +	help
> +	This driver provides support for virtio based flushing interface
> +	for persistent memory range.
> +
> +	If unsure, say M.
> +
>  config VIRTIO_BALLOON
>  	tristate "Virtio balloon driver"
>  	depends on VIRTIO
> diff --git a/drivers/virtio/Makefile b/drivers/virtio/Makefile
> index 3a2b5c5..143ce91 100644
> --- a/drivers/virtio/Makefile
> +++ b/drivers/virtio/Makefile
> @@ -6,3 +6,4 @@ virtio_pci-y := virtio_pci_modern.o virtio_pci_common.o
>  virtio_pci-$(CONFIG_VIRTIO_PCI_LEGACY) += virtio_pci_legacy.o
>  obj-$(CONFIG_VIRTIO_BALLOON) += virtio_balloon.o
>  obj-$(CONFIG_VIRTIO_INPUT) += virtio_input.o
> +obj-$(CONFIG_VIRTIO_PMEM) += pmem.o ../nvdimm/virtio_pmem.o
> diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
> new file mode 100644
> index 0000000..51f5349
> --- /dev/null
> +++ b/drivers/virtio/pmem.c
> @@ -0,0 +1,124 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * virtio_pmem.c: Virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and registers the virtual pmem device
> + * with libnvdimm core.
> + */
> +#include <linux/virtio_pmem.h>
> +#include <../../drivers/nvdimm/nd.h>
> +
> +static struct virtio_device_id id_table[] = {
> +	{ VIRTIO_ID_PMEM, VIRTIO_DEV_ANY_ID },
> +	{ 0 },
> +};
> +
> + /* Initialize virt queue */
> +static int init_vq(struct virtio_pmem *vpmem)
> +{
> +	struct virtqueue *vq;
> +
> +	/* single vq */
> +	vpmem->req_vq = vq = virtio_find_single_vq(vpmem->vdev,
> +				host_ack, "flush_queue");
> +	if (IS_ERR(vq))
> +		return PTR_ERR(vq);
> +
> +	spin_lock_init(&vpmem->pmem_lock);
> +	INIT_LIST_HEAD(&vpmem->req_list);
> +
> +	return 0;
> +};
> +
> +static int virtio_pmem_probe(struct virtio_device *vdev)
> +{
> +	int err = 0;
> +	struct resource res;
> +	struct virtio_pmem *vpmem;
> +	struct nvdimm_bus *nvdimm_bus;
> +	struct nd_region_desc ndr_desc;
> +	int nid = dev_to_node(&vdev->dev);
> +	struct nd_region *nd_region;
> +
> +	if (!vdev->config->get) {
> +		dev_err(&vdev->dev, "%s failure: config disabled\n",
> +			__func__);
> +		return -EINVAL;
> +	}
> +
> +	vdev->priv = vpmem = devm_kzalloc(&vdev->dev, sizeof(*vpmem),
> +			GFP_KERNEL);
> +	if (!vpmem) {
> +		err = -ENOMEM;
> +		goto out_err;
> +	}
> +
> +	vpmem->vdev = vdev;
> +	err = init_vq(vpmem);
> +	if (err)
> +		goto out_err;
> +
> +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> +			start, &vpmem->start);
> +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> +			size, &vpmem->size);
> +
> +	res.start = vpmem->start;
> +	res.end   = vpmem->start + vpmem->size-1;
> +	vpmem->nd_desc.provider_name = "virtio-pmem";
> +	vpmem->nd_desc.module = THIS_MODULE;
> +
> +	vpmem->nvdimm_bus = nvdimm_bus = nvdimm_bus_register(&vdev->dev,
> +						&vpmem->nd_desc);
> +	if (!nvdimm_bus)
> +		goto out_vq;
> +
> +	dev_set_drvdata(&vdev->dev, nvdimm_bus);
> +	memset(&ndr_desc, 0, sizeof(ndr_desc));
> +
> +	ndr_desc.res = &res;
> +	ndr_desc.numa_node = nid;
> +	ndr_desc.flush = virtio_pmem_flush;
> +	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
> +	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
> +	nd_region->provider_data =  dev_to_virtio
> +					(nd_region->dev.parent->parent);
> +
> +	if (!nd_region)
> +		goto out_nd;
> +
> +	//virtio_device_ready(vdev);
> +	return 0;
> +out_nd:
> +	err = -ENXIO;
> +	nvdimm_bus_unregister(nvdimm_bus);
> +out_vq:
> +	vdev->config->del_vqs(vdev);
> +out_err:
> +	dev_err(&vdev->dev, "failed to register virtio pmem memory\n");
> +	return err;
> +}
> +
> +static void virtio_pmem_remove(struct virtio_device *vdev)
> +{
> +	struct virtio_pmem *vpmem = vdev->priv;
> +	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
> +
> +	nvdimm_bus_unregister(nvdimm_bus);
> +	vdev->config->del_vqs(vdev);
> +	kfree(vpmem);
> +}
> +
> +static struct virtio_driver virtio_pmem_driver = {
> +	.driver.name		= KBUILD_MODNAME,
> +	.driver.owner		= THIS_MODULE,
> +	.id_table		= id_table,
> +	.probe			= virtio_pmem_probe,
> +	.remove			= virtio_pmem_remove,
> +};
> +
> +module_virtio_driver(virtio_pmem_driver);
> +MODULE_DEVICE_TABLE(virtio, id_table);
> +MODULE_DESCRIPTION("Virtio pmem driver");
> +MODULE_LICENSE("GPL");
> diff --git a/include/linux/virtio_pmem.h b/include/linux/virtio_pmem.h
> new file mode 100644
> index 0000000..224f9d9
> --- /dev/null
> +++ b/include/linux/virtio_pmem.h
> @@ -0,0 +1,60 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +/*
> + * virtio_pmem.h: virtio pmem Driver
> + *
> + * Discovers persistent memory range information
> + * from host and provides a virtio based flushing
> + * interface.
> + **/
> +
> +#ifndef _LINUX_VIRTIO_PMEM_H
> +#define _LINUX_VIRTIO_PMEM_H
> +
> +#include <linux/virtio_ids.h>
> +#include <linux/module.h>
> +#include <linux/virtio_config.h>
> +#include <uapi/linux/virtio_pmem.h>
> +#include <linux/libnvdimm.h>
> +#include <linux/spinlock.h>
> +
> +struct virtio_pmem_request {
> +	/* Host return status corresponding to flush request */
> +	int ret;
> +
> +	/* command name*/
> +	char name[16];
> +
> +	/* Wait queue to process deferred work after ack from host */
> +	wait_queue_head_t host_acked;
> +	bool done;
> +
> +	/* Wait queue to process deferred work after virt queue buffer avail */
> +	wait_queue_head_t wq_buf;
> +	bool wq_buf_avail;
> +	struct list_head list;
> +};
> +
> +struct virtio_pmem {
> +	struct virtio_device *vdev;
> +
> +	/* Virtio pmem request queue */
> +	struct virtqueue *req_vq;
> +
> +	/* nvdimm bus registers virtio pmem device */
> +	struct nvdimm_bus *nvdimm_bus;
> +	struct nvdimm_bus_descriptor nd_desc;
> +
> +	/* List to store deferred work if virtqueue is full */
> +	struct list_head req_list;
> +
> +	/* Synchronize virtqueue data */
> +	spinlock_t pmem_lock;
> +
> +	/* Memory region information */
> +	uint64_t start;
> +	uint64_t size;
> +};
> +
> +void host_ack(struct virtqueue *vq);
> +int virtio_pmem_flush(struct nd_region *nd_region);
> +#endif
> diff --git a/include/uapi/linux/virtio_ids.h b/include/uapi/linux/virtio_ids.h
> index 6d5c3b2..3463895 100644
> --- a/include/uapi/linux/virtio_ids.h
> +++ b/include/uapi/linux/virtio_ids.h
> @@ -43,5 +43,6 @@
>  #define VIRTIO_ID_INPUT        18 /* virtio input */
>  #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
>  #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
> +#define VIRTIO_ID_PMEM         25 /* virtio pmem */
>  
>  #endif /* _LINUX_VIRTIO_IDS_H */
> diff --git a/include/uapi/linux/virtio_pmem.h b/include/uapi/linux/virtio_pmem.h
> new file mode 100644
> index 0000000..fa3f7d5
> --- /dev/null
> +++ b/include/uapi/linux/virtio_pmem.h
> @@ -0,0 +1,10 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +
> +#ifndef _UAPI_LINUX_VIRTIO_PMEM_H
> +#define _UAPI_LINUX_VIRTIO_PMEM_H
> +
> +struct virtio_pmem_config {
> +	__le64 start;
> +	__le64 size;
> +};
> +#endif
> -- 
> 2.9.3

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                 ` <942065073.64011540.1547450140670.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-14 21:25                   ` Dave Chinner
  2019-01-14 21:25                     ` Dave Chinner
  2019-01-14 21:35                     ` Dan Williams
  0 siblings, 2 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-14 21:25 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	mst-H+wXaHxf7aLQT0dZR+AlfA, Matthew Wilcox,
	hch-wEGCiKHe2LqWVfeAwA7xHQ, linux-acpi-u79uwXL29TY76Z2rM5mHXA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, riel-ebMLmSuQjDVBDgjK7y7TUQ,
	stefanha-H+wXaHxf7aLQT0dZR+AlfA, imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	nilal-H+wXaHxf7aLQT0dZR+AlfA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA

On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> 
> > > Until you have images (and hence host page cache) shared between
> > > multiple guests. People will want to do this, because it means they
> > > only need a single set of pages in host memory for executable
> > > binaries rather than a set of pages per guest. Then you have
> > > multiple guests being able to detect residency of the same set of
> > > pages. If the guests can then, in any way, control eviction of the
> > > pages from the host cache, then we have a guest-to-guest information
> > > leak channel.
> > 
> > I don't think we should ever be considering something that would allow a
> > guest to evict page's from the host's pagecache [1].  The guest should
> > be able to kick its own references to the host's pagecache out of its
> > own pagecache, but not be able to influence whether the host or another
> > guest has a read-only mapping cached.
> > 
> > [1] Unless the guest is allowed to modify the host's file; obviously
> > truncation, holepunching, etc are going to evict pages from the host's
> > page cache.
> 
> This is so correct. Guest does not not evict host page cache pages directly. 

They don't right now.

But someone is going to end up asking for discard to work so that
the guest can free unused space in the underlying spares image (i.e.
make use of fstrim or mount -o discard) because they have workloads
that have bursts of space usage and they need to trim the image
files afterwards to keep their overall space usage under control.

And then....

> In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> Its solely decision of host to take action on the host page cache pages.
> 
> In case of virtio-pmem, guest does not modify host file directly i.e don't
> perform hole punch & truncation operation directly on host file.  

... this will no longer be true, and the nuclear landmine in this
driver interface will have been armed....

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 21:25                   ` Dave Chinner
@ 2019-01-14 21:25                     ` Dave Chinner
  2019-01-14 21:35                     ` Dan Williams
  1 sibling, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-14 21:25 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Matthew Wilcox, linux-kernel, kvm, qemu-devel, linux-nvdimm,
	linux-fsdevel, virtualization, linux-acpi, linux-ext4, linux-xfs,
	jack, stefanha, dan j williams, riel, nilal, kwolf, pbonzini,
	zwisler, vishal l verma, dave jiang, david, jmoyer,
	xiaoguangrong eric, hch, mst, jasowang, lcapitulino, imammedo,
	eblake, tytso, adilger kernel, darrick wong, rjw

On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> 
> > > Until you have images (and hence host page cache) shared between
> > > multiple guests. People will want to do this, because it means they
> > > only need a single set of pages in host memory for executable
> > > binaries rather than a set of pages per guest. Then you have
> > > multiple guests being able to detect residency of the same set of
> > > pages. If the guests can then, in any way, control eviction of the
> > > pages from the host cache, then we have a guest-to-guest information
> > > leak channel.
> > 
> > I don't think we should ever be considering something that would allow a
> > guest to evict page's from the host's pagecache [1].  The guest should
> > be able to kick its own references to the host's pagecache out of its
> > own pagecache, but not be able to influence whether the host or another
> > guest has a read-only mapping cached.
> > 
> > [1] Unless the guest is allowed to modify the host's file; obviously
> > truncation, holepunching, etc are going to evict pages from the host's
> > page cache.
> 
> This is so correct. Guest does not not evict host page cache pages directly. 

They don't right now.

But someone is going to end up asking for discard to work so that
the guest can free unused space in the underlying spares image (i.e.
make use of fstrim or mount -o discard) because they have workloads
that have bursts of space usage and they need to trim the image
files afterwards to keep their overall space usage under control.

And then....

> In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> Its solely decision of host to take action on the host page cache pages.
> 
> In case of virtio-pmem, guest does not modify host file directly i.e don't
> perform hole punch & truncation operation directly on host file.  

... this will no longer be true, and the nuclear landmine in this
driver interface will have been armed....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 21:25                   ` Dave Chinner
  2019-01-14 21:25                     ` Dave Chinner
@ 2019-01-14 21:35                     ` Dan Williams
  2019-01-14 21:35                       ` Dan Williams
       [not found]                       ` <CAPcyv4jtPcLV-s0sKNHwwk0ug7GLBV6699dpm1h3r2xSo879dg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  1 sibling, 2 replies; 61+ messages in thread
From: Dan Williams @ 2019-01-14 21:35 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, Eric Blake, darrick wong,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric

On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org> wrote:
>
> On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> >
> > > > Until you have images (and hence host page cache) shared between
> > > > multiple guests. People will want to do this, because it means they
> > > > only need a single set of pages in host memory for executable
> > > > binaries rather than a set of pages per guest. Then you have
> > > > multiple guests being able to detect residency of the same set of
> > > > pages. If the guests can then, in any way, control eviction of the
> > > > pages from the host cache, then we have a guest-to-guest information
> > > > leak channel.
> > >
> > > I don't think we should ever be considering something that would allow a
> > > guest to evict page's from the host's pagecache [1].  The guest should
> > > be able to kick its own references to the host's pagecache out of its
> > > own pagecache, but not be able to influence whether the host or another
> > > guest has a read-only mapping cached.
> > >
> > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > truncation, holepunching, etc are going to evict pages from the host's
> > > page cache.
> >
> > This is so correct. Guest does not not evict host page cache pages directly.
>
> They don't right now.
>
> But someone is going to end up asking for discard to work so that
> the guest can free unused space in the underlying spares image (i.e.
> make use of fstrim or mount -o discard) because they have workloads
> that have bursts of space usage and they need to trim the image
> files afterwards to keep their overall space usage under control.
>
> And then....

...we reject / push back on that patch citing the above concern.

> > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > Its solely decision of host to take action on the host page cache pages.
> >
> > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > perform hole punch & truncation operation directly on host file.
>
> ... this will no longer be true, and the nuclear landmine in this
> driver interface will have been armed....

I agree with the need to be careful when / if explicit cache control
is added, but that's not the case today.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 21:35                     ` Dan Williams
@ 2019-01-14 21:35                       ` Dan Williams
       [not found]                       ` <CAPcyv4jtPcLV-s0sKNHwwk0ug7GLBV6699dpm1h3r2xSo879dg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Dan Williams @ 2019-01-14 21:35 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Pankaj Gupta, Matthew Wilcox, Linux Kernel Mailing List,
	KVM list, Qemu Developers, linux-nvdimm, linux-fsdevel,
	virtualization, Linux ACPI, linux-ext4, linux-xfs, Jan Kara,
	Stefan Hajnoczi, Rik van Riel, Nitesh Narayan Lal, Kevin Wolf,
	Paolo Bonzini, Ross Zwisler, vishal l verma, dave jiang,
	David Hildenbrand, jmoyer, xiaoguangrong eric, Christoph Hellwig,
	Michael S. Tsirkin, Jason Wang, lcapitulino, Igor Mammedov,
	Eric Blake, Theodore Ts'o, adilger kernel, darrick wong,
	Rafael J. Wysocki

On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david@fromorbit.com> wrote:
>
> On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> >
> > > > Until you have images (and hence host page cache) shared between
> > > > multiple guests. People will want to do this, because it means they
> > > > only need a single set of pages in host memory for executable
> > > > binaries rather than a set of pages per guest. Then you have
> > > > multiple guests being able to detect residency of the same set of
> > > > pages. If the guests can then, in any way, control eviction of the
> > > > pages from the host cache, then we have a guest-to-guest information
> > > > leak channel.
> > >
> > > I don't think we should ever be considering something that would allow a
> > > guest to evict page's from the host's pagecache [1].  The guest should
> > > be able to kick its own references to the host's pagecache out of its
> > > own pagecache, but not be able to influence whether the host or another
> > > guest has a read-only mapping cached.
> > >
> > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > truncation, holepunching, etc are going to evict pages from the host's
> > > page cache.
> >
> > This is so correct. Guest does not not evict host page cache pages directly.
>
> They don't right now.
>
> But someone is going to end up asking for discard to work so that
> the guest can free unused space in the underlying spares image (i.e.
> make use of fstrim or mount -o discard) because they have workloads
> that have bursts of space usage and they need to trim the image
> files afterwards to keep their overall space usage under control.
>
> And then....

...we reject / push back on that patch citing the above concern.

> > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > Its solely decision of host to take action on the host page cache pages.
> >
> > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > perform hole punch & truncation operation directly on host file.
>
> ... this will no longer be true, and the nuclear landmine in this
> driver interface will have been armed....

I agree with the need to be careful when / if explicit cache control
is added, but that's not the case today.

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                       ` <CAPcyv4jtPcLV-s0sKNHwwk0ug7GLBV6699dpm1h3r2xSo879dg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
@ 2019-01-14 22:21                         ` Dave Chinner
  2019-01-14 22:21                           ` Dave Chinner
                                             ` (2 more replies)
  0 siblings, 3 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-14 22:21 UTC (permalink / raw)
  To: Dan Williams
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, Eric Blake, darrick wong,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric

On Mon, Jan 14, 2019 at 01:35:57PM -0800, Dan Williams wrote:
> On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org> wrote:
> >
> > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > >
> > > > > Until you have images (and hence host page cache) shared between
> > > > > multiple guests. People will want to do this, because it means they
> > > > > only need a single set of pages in host memory for executable
> > > > > binaries rather than a set of pages per guest. Then you have
> > > > > multiple guests being able to detect residency of the same set of
> > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > pages from the host cache, then we have a guest-to-guest information
> > > > > leak channel.
> > > >
> > > > I don't think we should ever be considering something that would allow a
> > > > guest to evict page's from the host's pagecache [1].  The guest should
> > > > be able to kick its own references to the host's pagecache out of its
> > > > own pagecache, but not be able to influence whether the host or another
> > > > guest has a read-only mapping cached.
> > > >
> > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > truncation, holepunching, etc are going to evict pages from the host's
> > > > page cache.
> > >
> > > This is so correct. Guest does not not evict host page cache pages directly.
> >
> > They don't right now.
> >
> > But someone is going to end up asking for discard to work so that
> > the guest can free unused space in the underlying spares image (i.e.
> > make use of fstrim or mount -o discard) because they have workloads
> > that have bursts of space usage and they need to trim the image
> > files afterwards to keep their overall space usage under control.
> >
> > And then....
> 
> ...we reject / push back on that patch citing the above concern.

So at what point do we draw the line?

We're allowing writable DAX mappings, but as I've pointed out that
means we are going to be allowing  a potential information leak via
files with shared extents to be directly mapped and written to.

But we won't allow useful admin operations that allow better
management of host side storage space similar to how normal image
files are used by guests because it's an information leak vector?

That's splitting some really fine hairs there...

> > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > > Its solely decision of host to take action on the host page cache pages.
> > >
> > > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > > perform hole punch & truncation operation directly on host file.
> >
> > ... this will no longer be true, and the nuclear landmine in this
> > driver interface will have been armed....
> 
> I agree with the need to be careful when / if explicit cache control
> is added, but that's not the case today.

"if"?

I expect it to be "when", not if. Expect the worst, plan for it now.

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 22:21                         ` Dave Chinner
@ 2019-01-14 22:21                           ` Dave Chinner
  2019-01-15  2:19                           ` Michael S. Tsirkin
  2019-01-15  5:35                           ` Pankaj Gupta
  2 siblings, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-14 22:21 UTC (permalink / raw)
  To: Dan Williams
  Cc: Pankaj Gupta, Matthew Wilcox, Linux Kernel Mailing List,
	KVM list, Qemu Developers, linux-nvdimm, linux-fsdevel,
	virtualization, Linux ACPI, linux-ext4, linux-xfs, Jan Kara,
	Stefan Hajnoczi, Rik van Riel, Nitesh Narayan Lal, Kevin Wolf,
	Paolo Bonzini, Ross Zwisler, vishal l verma, dave jiang,
	David Hildenbrand, jmoyer, xiaoguangrong eric, Christoph Hellwig,
	Michael S. Tsirkin, Jason Wang, lcapitulino, Igor Mammedov,
	Eric Blake, Theodore Ts'o, adilger kernel, darrick wong,
	Rafael J. Wysocki

On Mon, Jan 14, 2019 at 01:35:57PM -0800, Dan Williams wrote:
> On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david@fromorbit.com> wrote:
> >
> > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > >
> > > > > Until you have images (and hence host page cache) shared between
> > > > > multiple guests. People will want to do this, because it means they
> > > > > only need a single set of pages in host memory for executable
> > > > > binaries rather than a set of pages per guest. Then you have
> > > > > multiple guests being able to detect residency of the same set of
> > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > pages from the host cache, then we have a guest-to-guest information
> > > > > leak channel.
> > > >
> > > > I don't think we should ever be considering something that would allow a
> > > > guest to evict page's from the host's pagecache [1].  The guest should
> > > > be able to kick its own references to the host's pagecache out of its
> > > > own pagecache, but not be able to influence whether the host or another
> > > > guest has a read-only mapping cached.
> > > >
> > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > truncation, holepunching, etc are going to evict pages from the host's
> > > > page cache.
> > >
> > > This is so correct. Guest does not not evict host page cache pages directly.
> >
> > They don't right now.
> >
> > But someone is going to end up asking for discard to work so that
> > the guest can free unused space in the underlying spares image (i.e.
> > make use of fstrim or mount -o discard) because they have workloads
> > that have bursts of space usage and they need to trim the image
> > files afterwards to keep their overall space usage under control.
> >
> > And then....
> 
> ...we reject / push back on that patch citing the above concern.

So at what point do we draw the line?

We're allowing writable DAX mappings, but as I've pointed out that
means we are going to be allowing  a potential information leak via
files with shared extents to be directly mapped and written to.

But we won't allow useful admin operations that allow better
management of host side storage space similar to how normal image
files are used by guests because it's an information leak vector?

That's splitting some really fine hairs there...

> > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > > Its solely decision of host to take action on the host page cache pages.
> > >
> > > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > > perform hole punch & truncation operation directly on host file.
> >
> > ... this will no longer be true, and the nuclear landmine in this
> > driver interface will have been armed....
> 
> I agree with the need to be careful when / if explicit cache control
> is added, but that's not the case today.

"if"?

I expect it to be "when", not if. Expect the worst, plan for it now.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 22:21                         ` Dave Chinner
  2019-01-14 22:21                           ` Dave Chinner
@ 2019-01-15  2:19                           ` Michael S. Tsirkin
  2019-01-15  2:19                             ` Michael S. Tsirkin
       [not found]                             ` <20190114205031-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
  2019-01-15  5:35                           ` Pankaj Gupta
  2 siblings, 2 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-01-15  2:19 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Dan Williams, Pankaj Gupta, Matthew Wilcox,
	Linux Kernel Mailing List, KVM list, Qemu Developers,
	linux-nvdimm, linux-fsdevel, virtualization, Linux ACPI,
	linux-ext4, linux-xfs, Jan Kara, Stefan Hajnoczi, Rik van Riel,
	Nitesh Narayan Lal, Kevin Wolf, Paolo Bonzini, Ross Zwisler,
	vishal l verma <

On Tue, Jan 15, 2019 at 09:21:32AM +1100, Dave Chinner wrote:
> On Mon, Jan 14, 2019 at 01:35:57PM -0800, Dan Williams wrote:
> > On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david@fromorbit.com> wrote:
> > >
> > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > >
> > > > > > Until you have images (and hence host page cache) shared between
> > > > > > multiple guests. People will want to do this, because it means they
> > > > > > only need a single set of pages in host memory for executable
> > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > multiple guests being able to detect residency of the same set of
> > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > pages from the host cache, then we have a guest-to-guest information
> > > > > > leak channel.
> > > > >
> > > > > I don't think we should ever be considering something that would allow a
> > > > > guest to evict page's from the host's pagecache [1].  The guest should
> > > > > be able to kick its own references to the host's pagecache out of its
> > > > > own pagecache, but not be able to influence whether the host or another
> > > > > guest has a read-only mapping cached.
> > > > >
> > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > truncation, holepunching, etc are going to evict pages from the host's
> > > > > page cache.
> > > >
> > > > This is so correct. Guest does not not evict host page cache pages directly.
> > >
> > > They don't right now.
> > >
> > > But someone is going to end up asking for discard to work so that
> > > the guest can free unused space in the underlying spares image (i.e.
> > > make use of fstrim or mount -o discard) because they have workloads
> > > that have bursts of space usage and they need to trim the image
> > > files afterwards to keep their overall space usage under control.
> > >
> > > And then....
> > 
> > ...we reject / push back on that patch citing the above concern.
> 
> So at what point do we draw the line?
> 
> We're allowing writable DAX mappings, but as I've pointed out that
> means we are going to be allowing  a potential information leak via
> files with shared extents to be directly mapped and written to.
> 
> But we won't allow useful admin operations that allow better
> management of host side storage space similar to how normal image
> files are used by guests because it's an information leak vector?
> 
> That's splitting some really fine hairs there...

May I summarize that th security implications need to
be documented?

In fact that would make a fine security implications section
in the device specification.





> > > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > > > Its solely decision of host to take action on the host page cache pages.
> > > >
> > > > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > > > perform hole punch & truncation operation directly on host file.
> > >
> > > ... this will no longer be true, and the nuclear landmine in this
> > > driver interface will have been armed....
> > 
> > I agree with the need to be careful when / if explicit cache control
> > is added, but that's not the case today.
> 
> "if"?
> 
> I expect it to be "when", not if. Expect the worst, plan for it now.
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-15  2:19                           ` Michael S. Tsirkin
@ 2019-01-15  2:19                             ` Michael S. Tsirkin
       [not found]                             ` <20190114205031-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-01-15  2:19 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Dan Williams, Pankaj Gupta, Matthew Wilcox,
	Linux Kernel Mailing List, KVM list, Qemu Developers,
	linux-nvdimm, linux-fsdevel, virtualization, Linux ACPI,
	linux-ext4, linux-xfs, Jan Kara, Stefan Hajnoczi, Rik van Riel,
	Nitesh Narayan Lal, Kevin Wolf, Paolo Bonzini, Ross Zwisler,
	vishal l verma, dave jiang, David Hildenbrand, jmoyer,
	xiaoguangrong eric, Christoph Hellwig, Jason Wang, lcapitulino,
	Igor Mammedov, Eric Blake, Theodore Ts'o, adilger kernel,
	darrick wong, Rafael J. Wysocki

On Tue, Jan 15, 2019 at 09:21:32AM +1100, Dave Chinner wrote:
> On Mon, Jan 14, 2019 at 01:35:57PM -0800, Dan Williams wrote:
> > On Mon, Jan 14, 2019 at 1:25 PM Dave Chinner <david@fromorbit.com> wrote:
> > >
> > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > >
> > > > > > Until you have images (and hence host page cache) shared between
> > > > > > multiple guests. People will want to do this, because it means they
> > > > > > only need a single set of pages in host memory for executable
> > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > multiple guests being able to detect residency of the same set of
> > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > pages from the host cache, then we have a guest-to-guest information
> > > > > > leak channel.
> > > > >
> > > > > I don't think we should ever be considering something that would allow a
> > > > > guest to evict page's from the host's pagecache [1].  The guest should
> > > > > be able to kick its own references to the host's pagecache out of its
> > > > > own pagecache, but not be able to influence whether the host or another
> > > > > guest has a read-only mapping cached.
> > > > >
> > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > truncation, holepunching, etc are going to evict pages from the host's
> > > > > page cache.
> > > >
> > > > This is so correct. Guest does not not evict host page cache pages directly.
> > >
> > > They don't right now.
> > >
> > > But someone is going to end up asking for discard to work so that
> > > the guest can free unused space in the underlying spares image (i.e.
> > > make use of fstrim or mount -o discard) because they have workloads
> > > that have bursts of space usage and they need to trim the image
> > > files afterwards to keep their overall space usage under control.
> > >
> > > And then....
> > 
> > ...we reject / push back on that patch citing the above concern.
> 
> So at what point do we draw the line?
> 
> We're allowing writable DAX mappings, but as I've pointed out that
> means we are going to be allowing  a potential information leak via
> files with shared extents to be directly mapped and written to.
> 
> But we won't allow useful admin operations that allow better
> management of host side storage space similar to how normal image
> files are used by guests because it's an information leak vector?
> 
> That's splitting some really fine hairs there...

May I summarize that th security implications need to
be documented?

In fact that would make a fine security implications section
in the device specification.





> > > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional entries.
> > > > Its solely decision of host to take action on the host page cache pages.
> > > >
> > > > In case of virtio-pmem, guest does not modify host file directly i.e don't
> > > > perform hole punch & truncation operation directly on host file.
> > >
> > > ... this will no longer be true, and the nuclear landmine in this
> > > driver interface will have been armed....
> > 
> > I agree with the need to be careful when / if explicit cache control
> > is added, but that's not the case today.
> 
> "if"?
> 
> I expect it to be "when", not if. Expect the worst, plan for it now.
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-14 22:21                         ` Dave Chinner
  2019-01-14 22:21                           ` Dave Chinner
  2019-01-15  2:19                           ` Michael S. Tsirkin
@ 2019-01-15  5:35                           ` Pankaj Gupta
  2019-01-15  5:35                             ` Pankaj Gupta
       [not found]                             ` <1684638419.64320214.1547530506805.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  2 siblings, 2 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  5:35 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, Eric Blake, darrick wong,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric


> > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > >
> > > > > > Until you have images (and hence host page cache) shared between
> > > > > > multiple guests. People will want to do this, because it means they
> > > > > > only need a single set of pages in host memory for executable
> > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > multiple guests being able to detect residency of the same set of
> > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > information
> > > > > > leak channel.
> > > > >
> > > > > I don't think we should ever be considering something that would
> > > > > allow a
> > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > should
> > > > > be able to kick its own references to the host's pagecache out of its
> > > > > own pagecache, but not be able to influence whether the host or
> > > > > another
> > > > > guest has a read-only mapping cached.
> > > > >
> > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > host's
> > > > > page cache.
> > > >
> > > > This is so correct. Guest does not not evict host page cache pages
> > > > directly.
> > >
> > > They don't right now.
> > >
> > > But someone is going to end up asking for discard to work so that
> > > the guest can free unused space in the underlying spares image (i.e.
> > > make use of fstrim or mount -o discard) because they have workloads
> > > that have bursts of space usage and they need to trim the image
> > > files afterwards to keep their overall space usage under control.
> > >
> > > And then....
> > 
> > ...we reject / push back on that patch citing the above concern.
> 
> So at what point do we draw the line?
> 
> We're allowing writable DAX mappings, but as I've pointed out that
> means we are going to be allowing  a potential information leak via
> files with shared extents to be directly mapped and written to.
> 
> But we won't allow useful admin operations that allow better
> management of host side storage space similar to how normal image
> files are used by guests because it's an information leak vector?

First of all Thank you for all the useful discussions. 
I am summarizing here:

- We have to live with the limitation to not support fstrim and 
  mount -o discard options with virtio-pmem as they will evict 
  host page cache pages. We cannot allow this for virtio-pmem
  for security reasons. These filesystem commands will just zero out 
  unused pages currently.

- If alot of space is unused and not freed guest can request host 
  Administrator for truncating the host backing image. 
  We are also planning to support qcow2 sparse image format at 
  host side with virtio-pmem.

- There is no existing solution for Qemu persistent memory 
  emulation with write support currently. This solution provides 
  us the paravartualized way of emulating persistent memory. It 
  does not emulate of ACPI structures instead it just uses VIRTIO 
  for communication between guest & host. It is fast because of its
  asynchronous nature and it works well. This makes use of at guest 
  side libnvdimm API's 
  
- If disk size freeing problem with guest files trim truncate is 
  very important for users, they can still use real hardware which 
  will provide them both (advance disk features & page cache by pass).

Considering all the above reasons I think this feature is useful
from virtualization point of view. As Dave rightly said we should
be careful and I think now we are careful with the security implications
of this device. 

Thanks again for all the inputs.

Best regards,
Pankaj  


> 
> That's splitting some really fine hairs there...
> 
> > > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional
> > > > entries.
> > > > Its solely decision of host to take action on the host page cache
> > > > pages.
> > > >
> > > > In case of virtio-pmem, guest does not modify host file directly i.e
> > > > don't
> > > > perform hole punch & truncation operation directly on host file.
> > >
> > > ... this will no longer be true, and the nuclear landmine in this
> > > driver interface will have been armed....
> > 
> > I agree with the need to be careful when / if explicit cache control
> > is added, but that's not the case today.
> 
> "if"?
> 
> I expect it to be "when", not if. Expect the worst, plan for it now.
> 
> Cheers,
> 
> Dave.
> --
> Dave Chinner
> david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-15  5:35                           ` Pankaj Gupta
@ 2019-01-15  5:35                             ` Pankaj Gupta
       [not found]                             ` <1684638419.64320214.1547530506805.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
  1 sibling, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  5:35 UTC (permalink / raw)
  To: Dave Chinner
  Cc: Dan Williams, Matthew Wilcox, Linux Kernel Mailing List,
	KVM list, Qemu Developers, linux-nvdimm, linux-fsdevel,
	virtualization, Linux ACPI, linux-ext4, linux-xfs, Jan Kara,
	Stefan Hajnoczi, Rik van Riel, Nitesh Narayan Lal, Kevin Wolf,
	Paolo Bonzini, Ross Zwisler, vishal l verma, dave jiang,
	David Hildenbrand, jmoyer, xiaoguangrong eric, Christoph Hellwig,
	Michael S. Tsirkin, Jason Wang, lcapitulino, Igor Mammedov,
	Eric Blake, Theodore Ts'o, adilger kernel, darrick wong,
	Rafael J. Wysocki


> > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > >
> > > > > > Until you have images (and hence host page cache) shared between
> > > > > > multiple guests. People will want to do this, because it means they
> > > > > > only need a single set of pages in host memory for executable
> > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > multiple guests being able to detect residency of the same set of
> > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > information
> > > > > > leak channel.
> > > > >
> > > > > I don't think we should ever be considering something that would
> > > > > allow a
> > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > should
> > > > > be able to kick its own references to the host's pagecache out of its
> > > > > own pagecache, but not be able to influence whether the host or
> > > > > another
> > > > > guest has a read-only mapping cached.
> > > > >
> > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > host's
> > > > > page cache.
> > > >
> > > > This is so correct. Guest does not not evict host page cache pages
> > > > directly.
> > >
> > > They don't right now.
> > >
> > > But someone is going to end up asking for discard to work so that
> > > the guest can free unused space in the underlying spares image (i.e.
> > > make use of fstrim or mount -o discard) because they have workloads
> > > that have bursts of space usage and they need to trim the image
> > > files afterwards to keep their overall space usage under control.
> > >
> > > And then....
> > 
> > ...we reject / push back on that patch citing the above concern.
> 
> So at what point do we draw the line?
> 
> We're allowing writable DAX mappings, but as I've pointed out that
> means we are going to be allowing  a potential information leak via
> files with shared extents to be directly mapped and written to.
> 
> But we won't allow useful admin operations that allow better
> management of host side storage space similar to how normal image
> files are used by guests because it's an information leak vector?

First of all Thank you for all the useful discussions. 
I am summarizing here:

- We have to live with the limitation to not support fstrim and 
  mount -o discard options with virtio-pmem as they will evict 
  host page cache pages. We cannot allow this for virtio-pmem
  for security reasons. These filesystem commands will just zero out 
  unused pages currently.

- If alot of space is unused and not freed guest can request host 
  Administrator for truncating the host backing image. 
  We are also planning to support qcow2 sparse image format at 
  host side with virtio-pmem.

- There is no existing solution for Qemu persistent memory 
  emulation with write support currently. This solution provides 
  us the paravartualized way of emulating persistent memory. It 
  does not emulate of ACPI structures instead it just uses VIRTIO 
  for communication between guest & host. It is fast because of its
  asynchronous nature and it works well. This makes use of at guest 
  side libnvdimm API's 
  
- If disk size freeing problem with guest files trim truncate is 
  very important for users, they can still use real hardware which 
  will provide them both (advance disk features & page cache by pass).

Considering all the above reasons I think this feature is useful
from virtualization point of view. As Dave rightly said we should
be careful and I think now we are careful with the security implications
of this device. 

Thanks again for all the inputs.

Best regards,
Pankaj  


> 
> That's splitting some really fine hairs there...
> 
> > > > In case of virtio-pmem & DAX, guest clears guest page cache exceptional
> > > > entries.
> > > > Its solely decision of host to take action on the host page cache
> > > > pages.
> > > >
> > > > In case of virtio-pmem, guest does not modify host file directly i.e
> > > > don't
> > > > perform hole punch & truncation operation directly on host file.
> > >
> > > ... this will no longer be true, and the nuclear landmine in this
> > > driver interface will have been armed....
> > 
> > I agree with the need to be careful when / if explicit cache control
> > is added, but that's not the case today.
> 
> "if"?
> 
> I expect it to be "when", not if. Expect the worst, plan for it now.
> 
> Cheers,
> 
> Dave.
> --
> Dave Chinner
> david@fromorbit.com
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                             ` <20190114205031-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
@ 2019-01-15  5:37                               ` " Pankaj Gupta
  2019-01-15  5:37                                 ` Pankaj Gupta
  0 siblings, 1 reply; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  5:37 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Dave Chinner, Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, darrick wong, Matthew Wilcox,
	Christoph Hellwig, Linux ACPI, linux-ext4, Rik van Riel,
	Stefan Hajnoczi, Paolo Bonzini,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric,
	Rafael J. Wysocki


> > > >
> > > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > > >
> > > > > > > Until you have images (and hence host page cache) shared between
> > > > > > > multiple guests. People will want to do this, because it means
> > > > > > > they
> > > > > > > only need a single set of pages in host memory for executable
> > > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > > multiple guests being able to detect residency of the same set of
> > > > > > > pages. If the guests can then, in any way, control eviction of
> > > > > > > the
> > > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > > information
> > > > > > > leak channel.
> > > > > >
> > > > > > I don't think we should ever be considering something that would
> > > > > > allow a
> > > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > > should
> > > > > > be able to kick its own references to the host's pagecache out of
> > > > > > its
> > > > > > own pagecache, but not be able to influence whether the host or
> > > > > > another
> > > > > > guest has a read-only mapping cached.
> > > > > >
> > > > > > [1] Unless the guest is allowed to modify the host's file;
> > > > > > obviously
> > > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > > host's
> > > > > > page cache.
> > > > >
> > > > > This is so correct. Guest does not not evict host page cache pages
> > > > > directly.
> > > >
> > > > They don't right now.
> > > >
> > > > But someone is going to end up asking for discard to work so that
> > > > the guest can free unused space in the underlying spares image (i.e.
> > > > make use of fstrim or mount -o discard) because they have workloads
> > > > that have bursts of space usage and they need to trim the image
> > > > files afterwards to keep their overall space usage under control.
> > > >
> > > > And then....
> > > 
> > > ...we reject / push back on that patch citing the above concern.
> > 
> > So at what point do we draw the line?
> > 
> > We're allowing writable DAX mappings, but as I've pointed out that
> > means we are going to be allowing  a potential information leak via
> > files with shared extents to be directly mapped and written to.
> > 
> > But we won't allow useful admin operations that allow better
> > management of host side storage space similar to how normal image
> > files are used by guests because it's an information leak vector?
> > 
> > That's splitting some really fine hairs there...
> 
> May I summarize that th security implications need to
> be documented?
> 
> In fact that would make a fine security implications section
> in the device specification.

This is a very good suggestion. 

I will document the security implications in details in device specification
with details of what all filesystem features we don't support and why.

Best regards,
Pankaj

> 
> 
> 
> 
> 
> > > > > In case of virtio-pmem & DAX, guest clears guest page cache
> > > > > exceptional entries.
> > > > > Its solely decision of host to take action on the host page cache
> > > > > pages.
> > > > >
> > > > > In case of virtio-pmem, guest does not modify host file directly i.e
> > > > > don't
> > > > > perform hole punch & truncation operation directly on host file.
> > > >
> > > > ... this will no longer be true, and the nuclear landmine in this
> > > > driver interface will have been armed....
> > > 
> > > I agree with the need to be careful when / if explicit cache control
> > > is added, but that's not the case today.
> > 
> > "if"?
> > 
> > I expect it to be "when", not if. Expect the worst, plan for it now.
> > 
> > Cheers,
> > 
> > Dave.
> > --
> > Dave Chinner
> > david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org
> 
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-15  5:37                               ` [Qemu-devel] " Pankaj Gupta
@ 2019-01-15  5:37                                 ` Pankaj Gupta
  0 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  5:37 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: Dave Chinner, Jan Kara, KVM list, David Hildenbrand,
	linux-nvdimm, Jason Wang, Qemu Developers, virtualization,
	adilger kernel, Ross Zwisler, dave jiang, darrick wong,
	vishal l verma, Matthew Wilcox, Christoph Hellwig, Linux ACPI,
	jmoyer, linux-ext4, Rik van Riel, Stefan Hajnoczi, Igor Mammedov,
	Dan Williams, lcapitulino, Kevin Wolf, Nitesh Narayan Lal,
	Theodore Ts'o, xiaoguangrong eric, Rafael J. Wysocki,
	Linux Kernel Mailing List, linux-xfs, linux-fsdevel,
	Paolo Bonzini


> > > >
> > > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > > >
> > > > > > > Until you have images (and hence host page cache) shared between
> > > > > > > multiple guests. People will want to do this, because it means
> > > > > > > they
> > > > > > > only need a single set of pages in host memory for executable
> > > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > > multiple guests being able to detect residency of the same set of
> > > > > > > pages. If the guests can then, in any way, control eviction of
> > > > > > > the
> > > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > > information
> > > > > > > leak channel.
> > > > > >
> > > > > > I don't think we should ever be considering something that would
> > > > > > allow a
> > > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > > should
> > > > > > be able to kick its own references to the host's pagecache out of
> > > > > > its
> > > > > > own pagecache, but not be able to influence whether the host or
> > > > > > another
> > > > > > guest has a read-only mapping cached.
> > > > > >
> > > > > > [1] Unless the guest is allowed to modify the host's file;
> > > > > > obviously
> > > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > > host's
> > > > > > page cache.
> > > > >
> > > > > This is so correct. Guest does not not evict host page cache pages
> > > > > directly.
> > > >
> > > > They don't right now.
> > > >
> > > > But someone is going to end up asking for discard to work so that
> > > > the guest can free unused space in the underlying spares image (i.e.
> > > > make use of fstrim or mount -o discard) because they have workloads
> > > > that have bursts of space usage and they need to trim the image
> > > > files afterwards to keep their overall space usage under control.
> > > >
> > > > And then....
> > > 
> > > ...we reject / push back on that patch citing the above concern.
> > 
> > So at what point do we draw the line?
> > 
> > We're allowing writable DAX mappings, but as I've pointed out that
> > means we are going to be allowing  a potential information leak via
> > files with shared extents to be directly mapped and written to.
> > 
> > But we won't allow useful admin operations that allow better
> > management of host side storage space similar to how normal image
> > files are used by guests because it's an information leak vector?
> > 
> > That's splitting some really fine hairs there...
> 
> May I summarize that th security implications need to
> be documented?
> 
> In fact that would make a fine security implications section
> in the device specification.

This is a very good suggestion. 

I will document the security implications in details in device specification
with details of what all filesystem features we don't support and why.

Best regards,
Pankaj

> 
> 
> 
> 
> 
> > > > > In case of virtio-pmem & DAX, guest clears guest page cache
> > > > > exceptional entries.
> > > > > Its solely decision of host to take action on the host page cache
> > > > > pages.
> > > > >
> > > > > In case of virtio-pmem, guest does not modify host file directly i.e
> > > > > don't
> > > > > perform hole punch & truncation operation directly on host file.
> > > >
> > > > ... this will no longer be true, and the nuclear landmine in this
> > > > driver interface will have been armed....
> > > 
> > > I agree with the need to be careful when / if explicit cache control
> > > is added, but that's not the case today.
> > 
> > "if"?
> > 
> > I expect it to be "when", not if. Expect the worst, plan for it now.
> > 
> > Cheers,
> > 
> > Dave.
> > --
> > Dave Chinner
> > david@fromorbit.com
> 
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 2/5]  virtio-pmem: Add virtio pmem driver
       [not found]     ` <20190114105314-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
@ 2019-01-15  6:33       ` Pankaj Gupta
  2019-01-15  6:33         ` Pankaj Gupta
  0 siblings, 1 reply; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  6:33 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: jack-AlSwsSmVLrQ, kvm-u79uwXL29TY76Z2rM5mHXA,
	david-H+wXaHxf7aLQT0dZR+AlfA,
	linux-nvdimm-y27Ovi1pjclAfugRpC6u6w,
	jasowang-H+wXaHxf7aLQT0dZR+AlfA,
	qemu-devel-qX2TKyscuCcdnm+yROfE0A,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, zwisler-DgEjT+Ai2ygdnm+yROfE0A,
	eblake-H+wXaHxf7aLQT0dZR+AlfA, darrick wong,
	willy-wEGCiKHe2LqWVfeAwA7xHQ, hch-wEGCiKHe2LqWVfeAwA7xHQ,
	linux-acpi-u79uwXL29TY76Z2rM5mHXA, nilal-H+wXaHxf7aLQT0dZR+AlfA,
	riel-ebMLmSuQjDVBDgjK7y7TUQ, stefanha-H+wXaHxf7aLQT0dZR+AlfA,
	imammedo-H+wXaHxf7aLQT0dZR+AlfA,
	lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, kwolf-H+wXaHxf7aLQT0dZR+AlfA,
	linux-ext4-u79uwXL29TY76Z2rM5mHXA, tytso-3s7WtUTddSA,
	xiaoguangrong eric, rjw-LthD3rsA81gm4RdzfppkhA,
	linux-kernel-u79uwXL29TY76Z2rM5mHXA,
	linux-xfs-u79uwXL29TY76Z2rM5mHXA,
	linux-fsdevel-u79uwXL29TY76Z2rM5mHXA,
	pbonzini-H+wXaHxf7aLQT0dZR+AlfA


> > This patch adds virtio-pmem driver for KVM guest.
> > 
> > Guest reads the persistent memory range information from
> > Qemu over VIRTIO and registers it on nvdimm_bus. It also
> > creates a nd_region object with the persistent memory
> > range information so that existing 'nvdimm/pmem' driver
> > can reserve this into system memory map. This way
> > 'virtio-pmem' driver uses existing functionality of pmem
> > driver to register persistent memory compatible for DAX
> > capable filesystems.
> > 
> > This also provides function to perform guest flush over
> > VIRTIO from 'pmem' driver when userspace performs flush
> > on DAX memory range.
> > 
> > Signed-off-by: Pankaj Gupta <pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> > ---
> >  drivers/nvdimm/virtio_pmem.c     |  84 ++++++++++++++++++++++++++
> >  drivers/virtio/Kconfig           |  10 ++++
> >  drivers/virtio/Makefile          |   1 +
> >  drivers/virtio/pmem.c            | 124
> >  +++++++++++++++++++++++++++++++++++++++
> >  include/linux/virtio_pmem.h      |  60 +++++++++++++++++++
> >  include/uapi/linux/virtio_ids.h  |   1 +
> >  include/uapi/linux/virtio_pmem.h |  10 ++++
> 
> As with any uapi change, you need to CC the virtio dev
> mailing list (subscribers only, sorry about that).

Sure, will add virtio dev mailing list while sending v4.

Thanks,
Pankaj

> 
> 
> >  7 files changed, 290 insertions(+)
> >  create mode 100644 drivers/nvdimm/virtio_pmem.c
> >  create mode 100644 drivers/virtio/pmem.c
> >  create mode 100644 include/linux/virtio_pmem.h
> >  create mode 100644 include/uapi/linux/virtio_pmem.h
> > 
> > diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> > new file mode 100644
> > index 0000000..2a1b1ba
> > --- /dev/null
> > +++ b/drivers/nvdimm/virtio_pmem.c
> > @@ -0,0 +1,84 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * virtio_pmem.c: Virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and provides a virtio based flushing
> > + * interface.
> > + */
> > +#include <linux/virtio_pmem.h>
> > +#include "nd.h"
> > +
> > + /* The interrupt handler */
> > +void host_ack(struct virtqueue *vq)
> > +{
> > +	unsigned int len;
> > +	unsigned long flags;
> > +	struct virtio_pmem_request *req, *req_buf;
> > +	struct virtio_pmem *vpmem = vq->vdev->priv;
> > +
> > +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
> > +		req->done = true;
> > +		wake_up(&req->host_acked);
> > +
> > +		if (!list_empty(&vpmem->req_list)) {
> > +			req_buf = list_first_entry(&vpmem->req_list,
> > +					struct virtio_pmem_request, list);
> > +			list_del(&vpmem->req_list);
> > +			req_buf->wq_buf_avail = true;
> > +			wake_up(&req_buf->wq_buf);
> > +		}
> > +	}
> > +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +}
> > +EXPORT_SYMBOL_GPL(host_ack);
> > +
> > + /* The request submission function */
> > +int virtio_pmem_flush(struct nd_region *nd_region)
> > +{
> > +	int err;
> > +	unsigned long flags;
> > +	struct scatterlist *sgs[2], sg, ret;
> > +	struct virtio_device *vdev = nd_region->provider_data;
> > +	struct virtio_pmem *vpmem = vdev->priv;
> > +	struct virtio_pmem_request *req;
> > +
> > +	might_sleep();
> > +	req = kmalloc(sizeof(*req), GFP_KERNEL);
> > +	if (!req)
> > +		return -ENOMEM;
> > +
> > +	req->done = req->wq_buf_avail = false;
> > +	strcpy(req->name, "FLUSH");
> > +	init_waitqueue_head(&req->host_acked);
> > +	init_waitqueue_head(&req->wq_buf);
> > +	sg_init_one(&sg, req->name, strlen(req->name));
> > +	sgs[0] = &sg;
> > +	sg_init_one(&ret, &req->ret, sizeof(req->ret));
> > +	sgs[1] = &ret;
> > +
> > +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> > +	if (err) {
> > +		dev_err(&vdev->dev, "failed to send command to virtio pmem device\n");
> > +
> > +		list_add_tail(&vpmem->req_list, &req->list);
> > +		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +
> > +		/* When host has read buffer, this completes via host_ack */
> > +		wait_event(req->wq_buf, req->wq_buf_avail);
> > +		spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	}
> > +	virtqueue_kick(vpmem->req_vq);
> > +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +
> > +	/* When host has read buffer, this completes via host_ack */
> > +	wait_event(req->host_acked, req->done);
> > +	err = req->ret;
> > +	kfree(req);
> > +
> > +	return err;
> > +};
> > +EXPORT_SYMBOL_GPL(virtio_pmem_flush);
> > +MODULE_LICENSE("GPL");
> > diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
> > index 3589764..9f634a2 100644
> > --- a/drivers/virtio/Kconfig
> > +++ b/drivers/virtio/Kconfig
> > @@ -42,6 +42,16 @@ config VIRTIO_PCI_LEGACY
> >  
> >  	  If unsure, say Y.
> >  
> > +config VIRTIO_PMEM
> > +	tristate "Support for virtio pmem driver"
> > +	depends on VIRTIO
> > +	depends on LIBNVDIMM
> > +	help
> > +	This driver provides support for virtio based flushing interface
> > +	for persistent memory range.
> > +
> > +	If unsure, say M.
> > +
> >  config VIRTIO_BALLOON
> >  	tristate "Virtio balloon driver"
> >  	depends on VIRTIO
> > diff --git a/drivers/virtio/Makefile b/drivers/virtio/Makefile
> > index 3a2b5c5..143ce91 100644
> > --- a/drivers/virtio/Makefile
> > +++ b/drivers/virtio/Makefile
> > @@ -6,3 +6,4 @@ virtio_pci-y := virtio_pci_modern.o virtio_pci_common.o
> >  virtio_pci-$(CONFIG_VIRTIO_PCI_LEGACY) += virtio_pci_legacy.o
> >  obj-$(CONFIG_VIRTIO_BALLOON) += virtio_balloon.o
> >  obj-$(CONFIG_VIRTIO_INPUT) += virtio_input.o
> > +obj-$(CONFIG_VIRTIO_PMEM) += pmem.o ../nvdimm/virtio_pmem.o
> > diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
> > new file mode 100644
> > index 0000000..51f5349
> > --- /dev/null
> > +++ b/drivers/virtio/pmem.c
> > @@ -0,0 +1,124 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * virtio_pmem.c: Virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and registers the virtual pmem device
> > + * with libnvdimm core.
> > + */
> > +#include <linux/virtio_pmem.h>
> > +#include <../../drivers/nvdimm/nd.h>
> > +
> > +static struct virtio_device_id id_table[] = {
> > +	{ VIRTIO_ID_PMEM, VIRTIO_DEV_ANY_ID },
> > +	{ 0 },
> > +};
> > +
> > + /* Initialize virt queue */
> > +static int init_vq(struct virtio_pmem *vpmem)
> > +{
> > +	struct virtqueue *vq;
> > +
> > +	/* single vq */
> > +	vpmem->req_vq = vq = virtio_find_single_vq(vpmem->vdev,
> > +				host_ack, "flush_queue");
> > +	if (IS_ERR(vq))
> > +		return PTR_ERR(vq);
> > +
> > +	spin_lock_init(&vpmem->pmem_lock);
> > +	INIT_LIST_HEAD(&vpmem->req_list);
> > +
> > +	return 0;
> > +};
> > +
> > +static int virtio_pmem_probe(struct virtio_device *vdev)
> > +{
> > +	int err = 0;
> > +	struct resource res;
> > +	struct virtio_pmem *vpmem;
> > +	struct nvdimm_bus *nvdimm_bus;
> > +	struct nd_region_desc ndr_desc;
> > +	int nid = dev_to_node(&vdev->dev);
> > +	struct nd_region *nd_region;
> > +
> > +	if (!vdev->config->get) {
> > +		dev_err(&vdev->dev, "%s failure: config disabled\n",
> > +			__func__);
> > +		return -EINVAL;
> > +	}
> > +
> > +	vdev->priv = vpmem = devm_kzalloc(&vdev->dev, sizeof(*vpmem),
> > +			GFP_KERNEL);
> > +	if (!vpmem) {
> > +		err = -ENOMEM;
> > +		goto out_err;
> > +	}
> > +
> > +	vpmem->vdev = vdev;
> > +	err = init_vq(vpmem);
> > +	if (err)
> > +		goto out_err;
> > +
> > +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> > +			start, &vpmem->start);
> > +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> > +			size, &vpmem->size);
> > +
> > +	res.start = vpmem->start;
> > +	res.end   = vpmem->start + vpmem->size-1;
> > +	vpmem->nd_desc.provider_name = "virtio-pmem";
> > +	vpmem->nd_desc.module = THIS_MODULE;
> > +
> > +	vpmem->nvdimm_bus = nvdimm_bus = nvdimm_bus_register(&vdev->dev,
> > +						&vpmem->nd_desc);
> > +	if (!nvdimm_bus)
> > +		goto out_vq;
> > +
> > +	dev_set_drvdata(&vdev->dev, nvdimm_bus);
> > +	memset(&ndr_desc, 0, sizeof(ndr_desc));
> > +
> > +	ndr_desc.res = &res;
> > +	ndr_desc.numa_node = nid;
> > +	ndr_desc.flush = virtio_pmem_flush;
> > +	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
> > +	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
> > +	nd_region->provider_data =  dev_to_virtio
> > +					(nd_region->dev.parent->parent);
> > +
> > +	if (!nd_region)
> > +		goto out_nd;
> > +
> > +	//virtio_device_ready(vdev);
> > +	return 0;
> > +out_nd:
> > +	err = -ENXIO;
> > +	nvdimm_bus_unregister(nvdimm_bus);
> > +out_vq:
> > +	vdev->config->del_vqs(vdev);
> > +out_err:
> > +	dev_err(&vdev->dev, "failed to register virtio pmem memory\n");
> > +	return err;
> > +}
> > +
> > +static void virtio_pmem_remove(struct virtio_device *vdev)
> > +{
> > +	struct virtio_pmem *vpmem = vdev->priv;
> > +	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
> > +
> > +	nvdimm_bus_unregister(nvdimm_bus);
> > +	vdev->config->del_vqs(vdev);
> > +	kfree(vpmem);
> > +}
> > +
> > +static struct virtio_driver virtio_pmem_driver = {
> > +	.driver.name		= KBUILD_MODNAME,
> > +	.driver.owner		= THIS_MODULE,
> > +	.id_table		= id_table,
> > +	.probe			= virtio_pmem_probe,
> > +	.remove			= virtio_pmem_remove,
> > +};
> > +
> > +module_virtio_driver(virtio_pmem_driver);
> > +MODULE_DEVICE_TABLE(virtio, id_table);
> > +MODULE_DESCRIPTION("Virtio pmem driver");
> > +MODULE_LICENSE("GPL");
> > diff --git a/include/linux/virtio_pmem.h b/include/linux/virtio_pmem.h
> > new file mode 100644
> > index 0000000..224f9d9
> > --- /dev/null
> > +++ b/include/linux/virtio_pmem.h
> > @@ -0,0 +1,60 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +/*
> > + * virtio_pmem.h: virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and provides a virtio based flushing
> > + * interface.
> > + **/
> > +
> > +#ifndef _LINUX_VIRTIO_PMEM_H
> > +#define _LINUX_VIRTIO_PMEM_H
> > +
> > +#include <linux/virtio_ids.h>
> > +#include <linux/module.h>
> > +#include <linux/virtio_config.h>
> > +#include <uapi/linux/virtio_pmem.h>
> > +#include <linux/libnvdimm.h>
> > +#include <linux/spinlock.h>
> > +
> > +struct virtio_pmem_request {
> > +	/* Host return status corresponding to flush request */
> > +	int ret;
> > +
> > +	/* command name*/
> > +	char name[16];
> > +
> > +	/* Wait queue to process deferred work after ack from host */
> > +	wait_queue_head_t host_acked;
> > +	bool done;
> > +
> > +	/* Wait queue to process deferred work after virt queue buffer avail */
> > +	wait_queue_head_t wq_buf;
> > +	bool wq_buf_avail;
> > +	struct list_head list;
> > +};
> > +
> > +struct virtio_pmem {
> > +	struct virtio_device *vdev;
> > +
> > +	/* Virtio pmem request queue */
> > +	struct virtqueue *req_vq;
> > +
> > +	/* nvdimm bus registers virtio pmem device */
> > +	struct nvdimm_bus *nvdimm_bus;
> > +	struct nvdimm_bus_descriptor nd_desc;
> > +
> > +	/* List to store deferred work if virtqueue is full */
> > +	struct list_head req_list;
> > +
> > +	/* Synchronize virtqueue data */
> > +	spinlock_t pmem_lock;
> > +
> > +	/* Memory region information */
> > +	uint64_t start;
> > +	uint64_t size;
> > +};
> > +
> > +void host_ack(struct virtqueue *vq);
> > +int virtio_pmem_flush(struct nd_region *nd_region);
> > +#endif
> > diff --git a/include/uapi/linux/virtio_ids.h
> > b/include/uapi/linux/virtio_ids.h
> > index 6d5c3b2..3463895 100644
> > --- a/include/uapi/linux/virtio_ids.h
> > +++ b/include/uapi/linux/virtio_ids.h
> > @@ -43,5 +43,6 @@
> >  #define VIRTIO_ID_INPUT        18 /* virtio input */
> >  #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
> >  #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
> > +#define VIRTIO_ID_PMEM         25 /* virtio pmem */
> >  
> >  #endif /* _LINUX_VIRTIO_IDS_H */
> > diff --git a/include/uapi/linux/virtio_pmem.h
> > b/include/uapi/linux/virtio_pmem.h
> > new file mode 100644
> > index 0000000..fa3f7d5
> > --- /dev/null
> > +++ b/include/uapi/linux/virtio_pmem.h
> > @@ -0,0 +1,10 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +
> > +#ifndef _UAPI_LINUX_VIRTIO_PMEM_H
> > +#define _UAPI_LINUX_VIRTIO_PMEM_H
> > +
> > +struct virtio_pmem_config {
> > +	__le64 start;
> > +	__le64 size;
> > +};
> > +#endif
> > --
> > 2.9.3
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 2/5]  virtio-pmem: Add virtio pmem driver
  2019-01-15  6:33       ` Pankaj Gupta
@ 2019-01-15  6:33         ` Pankaj Gupta
  0 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-01-15  6:33 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan j williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal l verma, dave jiang, david, jmoyer, xiaoguangrong eric,
	hch, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger kernel, darrick wong, rjw


> > This patch adds virtio-pmem driver for KVM guest.
> > 
> > Guest reads the persistent memory range information from
> > Qemu over VIRTIO and registers it on nvdimm_bus. It also
> > creates a nd_region object with the persistent memory
> > range information so that existing 'nvdimm/pmem' driver
> > can reserve this into system memory map. This way
> > 'virtio-pmem' driver uses existing functionality of pmem
> > driver to register persistent memory compatible for DAX
> > capable filesystems.
> > 
> > This also provides function to perform guest flush over
> > VIRTIO from 'pmem' driver when userspace performs flush
> > on DAX memory range.
> > 
> > Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
> > ---
> >  drivers/nvdimm/virtio_pmem.c     |  84 ++++++++++++++++++++++++++
> >  drivers/virtio/Kconfig           |  10 ++++
> >  drivers/virtio/Makefile          |   1 +
> >  drivers/virtio/pmem.c            | 124
> >  +++++++++++++++++++++++++++++++++++++++
> >  include/linux/virtio_pmem.h      |  60 +++++++++++++++++++
> >  include/uapi/linux/virtio_ids.h  |   1 +
> >  include/uapi/linux/virtio_pmem.h |  10 ++++
> 
> As with any uapi change, you need to CC the virtio dev
> mailing list (subscribers only, sorry about that).

Sure, will add virtio dev mailing list while sending v4.

Thanks,
Pankaj

> 
> 
> >  7 files changed, 290 insertions(+)
> >  create mode 100644 drivers/nvdimm/virtio_pmem.c
> >  create mode 100644 drivers/virtio/pmem.c
> >  create mode 100644 include/linux/virtio_pmem.h
> >  create mode 100644 include/uapi/linux/virtio_pmem.h
> > 
> > diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> > new file mode 100644
> > index 0000000..2a1b1ba
> > --- /dev/null
> > +++ b/drivers/nvdimm/virtio_pmem.c
> > @@ -0,0 +1,84 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * virtio_pmem.c: Virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and provides a virtio based flushing
> > + * interface.
> > + */
> > +#include <linux/virtio_pmem.h>
> > +#include "nd.h"
> > +
> > + /* The interrupt handler */
> > +void host_ack(struct virtqueue *vq)
> > +{
> > +	unsigned int len;
> > +	unsigned long flags;
> > +	struct virtio_pmem_request *req, *req_buf;
> > +	struct virtio_pmem *vpmem = vq->vdev->priv;
> > +
> > +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
> > +		req->done = true;
> > +		wake_up(&req->host_acked);
> > +
> > +		if (!list_empty(&vpmem->req_list)) {
> > +			req_buf = list_first_entry(&vpmem->req_list,
> > +					struct virtio_pmem_request, list);
> > +			list_del(&vpmem->req_list);
> > +			req_buf->wq_buf_avail = true;
> > +			wake_up(&req_buf->wq_buf);
> > +		}
> > +	}
> > +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +}
> > +EXPORT_SYMBOL_GPL(host_ack);
> > +
> > + /* The request submission function */
> > +int virtio_pmem_flush(struct nd_region *nd_region)
> > +{
> > +	int err;
> > +	unsigned long flags;
> > +	struct scatterlist *sgs[2], sg, ret;
> > +	struct virtio_device *vdev = nd_region->provider_data;
> > +	struct virtio_pmem *vpmem = vdev->priv;
> > +	struct virtio_pmem_request *req;
> > +
> > +	might_sleep();
> > +	req = kmalloc(sizeof(*req), GFP_KERNEL);
> > +	if (!req)
> > +		return -ENOMEM;
> > +
> > +	req->done = req->wq_buf_avail = false;
> > +	strcpy(req->name, "FLUSH");
> > +	init_waitqueue_head(&req->host_acked);
> > +	init_waitqueue_head(&req->wq_buf);
> > +	sg_init_one(&sg, req->name, strlen(req->name));
> > +	sgs[0] = &sg;
> > +	sg_init_one(&ret, &req->ret, sizeof(req->ret));
> > +	sgs[1] = &ret;
> > +
> > +	spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> > +	if (err) {
> > +		dev_err(&vdev->dev, "failed to send command to virtio pmem device\n");
> > +
> > +		list_add_tail(&vpmem->req_list, &req->list);
> > +		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +
> > +		/* When host has read buffer, this completes via host_ack */
> > +		wait_event(req->wq_buf, req->wq_buf_avail);
> > +		spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +	}
> > +	virtqueue_kick(vpmem->req_vq);
> > +	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +
> > +	/* When host has read buffer, this completes via host_ack */
> > +	wait_event(req->host_acked, req->done);
> > +	err = req->ret;
> > +	kfree(req);
> > +
> > +	return err;
> > +};
> > +EXPORT_SYMBOL_GPL(virtio_pmem_flush);
> > +MODULE_LICENSE("GPL");
> > diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
> > index 3589764..9f634a2 100644
> > --- a/drivers/virtio/Kconfig
> > +++ b/drivers/virtio/Kconfig
> > @@ -42,6 +42,16 @@ config VIRTIO_PCI_LEGACY
> >  
> >  	  If unsure, say Y.
> >  
> > +config VIRTIO_PMEM
> > +	tristate "Support for virtio pmem driver"
> > +	depends on VIRTIO
> > +	depends on LIBNVDIMM
> > +	help
> > +	This driver provides support for virtio based flushing interface
> > +	for persistent memory range.
> > +
> > +	If unsure, say M.
> > +
> >  config VIRTIO_BALLOON
> >  	tristate "Virtio balloon driver"
> >  	depends on VIRTIO
> > diff --git a/drivers/virtio/Makefile b/drivers/virtio/Makefile
> > index 3a2b5c5..143ce91 100644
> > --- a/drivers/virtio/Makefile
> > +++ b/drivers/virtio/Makefile
> > @@ -6,3 +6,4 @@ virtio_pci-y := virtio_pci_modern.o virtio_pci_common.o
> >  virtio_pci-$(CONFIG_VIRTIO_PCI_LEGACY) += virtio_pci_legacy.o
> >  obj-$(CONFIG_VIRTIO_BALLOON) += virtio_balloon.o
> >  obj-$(CONFIG_VIRTIO_INPUT) += virtio_input.o
> > +obj-$(CONFIG_VIRTIO_PMEM) += pmem.o ../nvdimm/virtio_pmem.o
> > diff --git a/drivers/virtio/pmem.c b/drivers/virtio/pmem.c
> > new file mode 100644
> > index 0000000..51f5349
> > --- /dev/null
> > +++ b/drivers/virtio/pmem.c
> > @@ -0,0 +1,124 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * virtio_pmem.c: Virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and registers the virtual pmem device
> > + * with libnvdimm core.
> > + */
> > +#include <linux/virtio_pmem.h>
> > +#include <../../drivers/nvdimm/nd.h>
> > +
> > +static struct virtio_device_id id_table[] = {
> > +	{ VIRTIO_ID_PMEM, VIRTIO_DEV_ANY_ID },
> > +	{ 0 },
> > +};
> > +
> > + /* Initialize virt queue */
> > +static int init_vq(struct virtio_pmem *vpmem)
> > +{
> > +	struct virtqueue *vq;
> > +
> > +	/* single vq */
> > +	vpmem->req_vq = vq = virtio_find_single_vq(vpmem->vdev,
> > +				host_ack, "flush_queue");
> > +	if (IS_ERR(vq))
> > +		return PTR_ERR(vq);
> > +
> > +	spin_lock_init(&vpmem->pmem_lock);
> > +	INIT_LIST_HEAD(&vpmem->req_list);
> > +
> > +	return 0;
> > +};
> > +
> > +static int virtio_pmem_probe(struct virtio_device *vdev)
> > +{
> > +	int err = 0;
> > +	struct resource res;
> > +	struct virtio_pmem *vpmem;
> > +	struct nvdimm_bus *nvdimm_bus;
> > +	struct nd_region_desc ndr_desc;
> > +	int nid = dev_to_node(&vdev->dev);
> > +	struct nd_region *nd_region;
> > +
> > +	if (!vdev->config->get) {
> > +		dev_err(&vdev->dev, "%s failure: config disabled\n",
> > +			__func__);
> > +		return -EINVAL;
> > +	}
> > +
> > +	vdev->priv = vpmem = devm_kzalloc(&vdev->dev, sizeof(*vpmem),
> > +			GFP_KERNEL);
> > +	if (!vpmem) {
> > +		err = -ENOMEM;
> > +		goto out_err;
> > +	}
> > +
> > +	vpmem->vdev = vdev;
> > +	err = init_vq(vpmem);
> > +	if (err)
> > +		goto out_err;
> > +
> > +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> > +			start, &vpmem->start);
> > +	virtio_cread(vpmem->vdev, struct virtio_pmem_config,
> > +			size, &vpmem->size);
> > +
> > +	res.start = vpmem->start;
> > +	res.end   = vpmem->start + vpmem->size-1;
> > +	vpmem->nd_desc.provider_name = "virtio-pmem";
> > +	vpmem->nd_desc.module = THIS_MODULE;
> > +
> > +	vpmem->nvdimm_bus = nvdimm_bus = nvdimm_bus_register(&vdev->dev,
> > +						&vpmem->nd_desc);
> > +	if (!nvdimm_bus)
> > +		goto out_vq;
> > +
> > +	dev_set_drvdata(&vdev->dev, nvdimm_bus);
> > +	memset(&ndr_desc, 0, sizeof(ndr_desc));
> > +
> > +	ndr_desc.res = &res;
> > +	ndr_desc.numa_node = nid;
> > +	ndr_desc.flush = virtio_pmem_flush;
> > +	set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
> > +	nd_region = nvdimm_pmem_region_create(nvdimm_bus, &ndr_desc);
> > +	nd_region->provider_data =  dev_to_virtio
> > +					(nd_region->dev.parent->parent);
> > +
> > +	if (!nd_region)
> > +		goto out_nd;
> > +
> > +	//virtio_device_ready(vdev);
> > +	return 0;
> > +out_nd:
> > +	err = -ENXIO;
> > +	nvdimm_bus_unregister(nvdimm_bus);
> > +out_vq:
> > +	vdev->config->del_vqs(vdev);
> > +out_err:
> > +	dev_err(&vdev->dev, "failed to register virtio pmem memory\n");
> > +	return err;
> > +}
> > +
> > +static void virtio_pmem_remove(struct virtio_device *vdev)
> > +{
> > +	struct virtio_pmem *vpmem = vdev->priv;
> > +	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
> > +
> > +	nvdimm_bus_unregister(nvdimm_bus);
> > +	vdev->config->del_vqs(vdev);
> > +	kfree(vpmem);
> > +}
> > +
> > +static struct virtio_driver virtio_pmem_driver = {
> > +	.driver.name		= KBUILD_MODNAME,
> > +	.driver.owner		= THIS_MODULE,
> > +	.id_table		= id_table,
> > +	.probe			= virtio_pmem_probe,
> > +	.remove			= virtio_pmem_remove,
> > +};
> > +
> > +module_virtio_driver(virtio_pmem_driver);
> > +MODULE_DEVICE_TABLE(virtio, id_table);
> > +MODULE_DESCRIPTION("Virtio pmem driver");
> > +MODULE_LICENSE("GPL");
> > diff --git a/include/linux/virtio_pmem.h b/include/linux/virtio_pmem.h
> > new file mode 100644
> > index 0000000..224f9d9
> > --- /dev/null
> > +++ b/include/linux/virtio_pmem.h
> > @@ -0,0 +1,60 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +/*
> > + * virtio_pmem.h: virtio pmem Driver
> > + *
> > + * Discovers persistent memory range information
> > + * from host and provides a virtio based flushing
> > + * interface.
> > + **/
> > +
> > +#ifndef _LINUX_VIRTIO_PMEM_H
> > +#define _LINUX_VIRTIO_PMEM_H
> > +
> > +#include <linux/virtio_ids.h>
> > +#include <linux/module.h>
> > +#include <linux/virtio_config.h>
> > +#include <uapi/linux/virtio_pmem.h>
> > +#include <linux/libnvdimm.h>
> > +#include <linux/spinlock.h>
> > +
> > +struct virtio_pmem_request {
> > +	/* Host return status corresponding to flush request */
> > +	int ret;
> > +
> > +	/* command name*/
> > +	char name[16];
> > +
> > +	/* Wait queue to process deferred work after ack from host */
> > +	wait_queue_head_t host_acked;
> > +	bool done;
> > +
> > +	/* Wait queue to process deferred work after virt queue buffer avail */
> > +	wait_queue_head_t wq_buf;
> > +	bool wq_buf_avail;
> > +	struct list_head list;
> > +};
> > +
> > +struct virtio_pmem {
> > +	struct virtio_device *vdev;
> > +
> > +	/* Virtio pmem request queue */
> > +	struct virtqueue *req_vq;
> > +
> > +	/* nvdimm bus registers virtio pmem device */
> > +	struct nvdimm_bus *nvdimm_bus;
> > +	struct nvdimm_bus_descriptor nd_desc;
> > +
> > +	/* List to store deferred work if virtqueue is full */
> > +	struct list_head req_list;
> > +
> > +	/* Synchronize virtqueue data */
> > +	spinlock_t pmem_lock;
> > +
> > +	/* Memory region information */
> > +	uint64_t start;
> > +	uint64_t size;
> > +};
> > +
> > +void host_ack(struct virtqueue *vq);
> > +int virtio_pmem_flush(struct nd_region *nd_region);
> > +#endif
> > diff --git a/include/uapi/linux/virtio_ids.h
> > b/include/uapi/linux/virtio_ids.h
> > index 6d5c3b2..3463895 100644
> > --- a/include/uapi/linux/virtio_ids.h
> > +++ b/include/uapi/linux/virtio_ids.h
> > @@ -43,5 +43,6 @@
> >  #define VIRTIO_ID_INPUT        18 /* virtio input */
> >  #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
> >  #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
> > +#define VIRTIO_ID_PMEM         25 /* virtio pmem */
> >  
> >  #endif /* _LINUX_VIRTIO_IDS_H */
> > diff --git a/include/uapi/linux/virtio_pmem.h
> > b/include/uapi/linux/virtio_pmem.h
> > new file mode 100644
> > index 0000000..fa3f7d5
> > --- /dev/null
> > +++ b/include/uapi/linux/virtio_pmem.h
> > @@ -0,0 +1,10 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +
> > +#ifndef _UAPI_LINUX_VIRTIO_PMEM_H
> > +#define _UAPI_LINUX_VIRTIO_PMEM_H
> > +
> > +struct virtio_pmem_config {
> > +	__le64 start;
> > +	__le64 size;
> > +};
> > +#endif
> > --
> > 2.9.3
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
       [not found]                             ` <1684638419.64320214.1547530506805.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-01-15 20:42                               ` Dave Chinner
  2019-01-15 20:42                                 ` Dave Chinner
  0 siblings, 1 reply; 61+ messages in thread
From: Dave Chinner @ 2019-01-15 20:42 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Jan Kara, KVM list, David Hildenbrand, linux-nvdimm, Jason Wang,
	Qemu Developers,
	virtualization-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA,
	adilger kernel, Ross Zwisler, Eric Blake, darrick wong,
	Michael S. Tsirkin, Matthew Wilcox, Christoph Hellwig,
	Linux ACPI, linux-ext4, Rik van Riel, Stefan Hajnoczi,
	Igor Mammedov, lcapitulino-H+wXaHxf7aLQT0dZR+AlfA, Kevin Wolf,
	Nitesh Narayan Lal, Theodore Ts'o, xiaoguangrong eric

On Tue, Jan 15, 2019 at 12:35:06AM -0500, Pankaj Gupta wrote:
> 
> > > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > > >
> > > > > > > Until you have images (and hence host page cache) shared between
> > > > > > > multiple guests. People will want to do this, because it means they
> > > > > > > only need a single set of pages in host memory for executable
> > > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > > multiple guests being able to detect residency of the same set of
> > > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > > information
> > > > > > > leak channel.
> > > > > >
> > > > > > I don't think we should ever be considering something that would
> > > > > > allow a
> > > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > > should
> > > > > > be able to kick its own references to the host's pagecache out of its
> > > > > > own pagecache, but not be able to influence whether the host or
> > > > > > another
> > > > > > guest has a read-only mapping cached.
> > > > > >
> > > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > > host's
> > > > > > page cache.
> > > > >
> > > > > This is so correct. Guest does not not evict host page cache pages
> > > > > directly.
> > > >
> > > > They don't right now.
> > > >
> > > > But someone is going to end up asking for discard to work so that
> > > > the guest can free unused space in the underlying spares image (i.e.
> > > > make use of fstrim or mount -o discard) because they have workloads
> > > > that have bursts of space usage and they need to trim the image
> > > > files afterwards to keep their overall space usage under control.
> > > >
> > > > And then....
> > > 
> > > ...we reject / push back on that patch citing the above concern.
> > 
> > So at what point do we draw the line?
> > 
> > We're allowing writable DAX mappings, but as I've pointed out that
> > means we are going to be allowing  a potential information leak via
> > files with shared extents to be directly mapped and written to.
> > 
> > But we won't allow useful admin operations that allow better
> > management of host side storage space similar to how normal image
> > files are used by guests because it's an information leak vector?
> 
> First of all Thank you for all the useful discussions. 
> I am summarizing here:
> 
> - We have to live with the limitation to not support fstrim and 
>   mount -o discard options with virtio-pmem as they will evict 
>   host page cache pages. We cannot allow this for virtio-pmem
>   for security reasons. These filesystem commands will just zero out 
>   unused pages currently.

Not sure I follow you here - what pages are going to be zeroed and
when will they be zeroed? If discard is not allowed, filesystems
just don't issue such commands and the underlying device will never
seen them.

> - If alot of space is unused and not freed guest can request host 
>   Administrator for truncating the host backing image. 

You can't use truncate to free space in a disk image file. The only
way to do it safely in a generic, filesystem agnositic way is to
mount the disk image (e.g. on loopback) and run fstrim on it. The
loopback device will punches holes in the file where all the free
space is reported by the filesystem via discard requests.

Which is kinda my point - this could only be done if the guest is
shut down, which makes it very difficult for admins to manage. 

>   We are also planning to support qcow2 sparse image format at 
>   host side with virtio-pmem.

So you're going to be remapping a huge number of disjoint regions
into a linear pmem mapping? ISTR discussions about similar things
for virtio+fuse+dax that came up against "large numbers of mapped
regions don't scale" and so it wasn't a practical solution compared
to a just using raw sparse files....

> - There is no existing solution for Qemu persistent memory 
>   emulation with write support currently. This solution provides 
>   us the paravartualized way of emulating persistent memory.

Sure, but the question is why do you need to create an emulation
that doesn't actually perform like pmem? The whole point of pmem is
performance, and emulating pmem by mmap() of a file on spinning
disks is going to be horrible for performance. Even on SSDs it's
going to be orders of magnitudes slower than real pmem.

So exactly what problem are you trying to solve with this driver?

Cheers,

Dave.
-- 
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [PATCH v3 0/5] kvm "virtio pmem" device
  2019-01-15 20:42                               ` Dave Chinner
@ 2019-01-15 20:42                                 ` Dave Chinner
  0 siblings, 0 replies; 61+ messages in thread
From: Dave Chinner @ 2019-01-15 20:42 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Dan Williams, Matthew Wilcox, Linux Kernel Mailing List,
	KVM list, Qemu Developers, linux-nvdimm, linux-fsdevel,
	virtualization, Linux ACPI, linux-ext4, linux-xfs, Jan Kara,
	Stefan Hajnoczi, Rik van Riel, Nitesh Narayan Lal, Kevin Wolf,
	Paolo Bonzini, Ross Zwisler, vishal l verma, dave jiang,
	David Hildenbrand, jmoyer, xiaoguangrong eric, Christoph Hellwig,
	Michael S. Tsirkin, Jason Wang, lcapitulino, Igor Mammedov,
	Eric Blake, Theodore Ts'o, adilger kernel, darrick wong,
	Rafael J. Wysocki

On Tue, Jan 15, 2019 at 12:35:06AM -0500, Pankaj Gupta wrote:
> 
> > > > On Mon, Jan 14, 2019 at 02:15:40AM -0500, Pankaj Gupta wrote:
> > > > >
> > > > > > > Until you have images (and hence host page cache) shared between
> > > > > > > multiple guests. People will want to do this, because it means they
> > > > > > > only need a single set of pages in host memory for executable
> > > > > > > binaries rather than a set of pages per guest. Then you have
> > > > > > > multiple guests being able to detect residency of the same set of
> > > > > > > pages. If the guests can then, in any way, control eviction of the
> > > > > > > pages from the host cache, then we have a guest-to-guest
> > > > > > > information
> > > > > > > leak channel.
> > > > > >
> > > > > > I don't think we should ever be considering something that would
> > > > > > allow a
> > > > > > guest to evict page's from the host's pagecache [1].  The guest
> > > > > > should
> > > > > > be able to kick its own references to the host's pagecache out of its
> > > > > > own pagecache, but not be able to influence whether the host or
> > > > > > another
> > > > > > guest has a read-only mapping cached.
> > > > > >
> > > > > > [1] Unless the guest is allowed to modify the host's file; obviously
> > > > > > truncation, holepunching, etc are going to evict pages from the
> > > > > > host's
> > > > > > page cache.
> > > > >
> > > > > This is so correct. Guest does not not evict host page cache pages
> > > > > directly.
> > > >
> > > > They don't right now.
> > > >
> > > > But someone is going to end up asking for discard to work so that
> > > > the guest can free unused space in the underlying spares image (i.e.
> > > > make use of fstrim or mount -o discard) because they have workloads
> > > > that have bursts of space usage and they need to trim the image
> > > > files afterwards to keep their overall space usage under control.
> > > >
> > > > And then....
> > > 
> > > ...we reject / push back on that patch citing the above concern.
> > 
> > So at what point do we draw the line?
> > 
> > We're allowing writable DAX mappings, but as I've pointed out that
> > means we are going to be allowing  a potential information leak via
> > files with shared extents to be directly mapped and written to.
> > 
> > But we won't allow useful admin operations that allow better
> > management of host side storage space similar to how normal image
> > files are used by guests because it's an information leak vector?
> 
> First of all Thank you for all the useful discussions. 
> I am summarizing here:
> 
> - We have to live with the limitation to not support fstrim and 
>   mount -o discard options with virtio-pmem as they will evict 
>   host page cache pages. We cannot allow this for virtio-pmem
>   for security reasons. These filesystem commands will just zero out 
>   unused pages currently.

Not sure I follow you here - what pages are going to be zeroed and
when will they be zeroed? If discard is not allowed, filesystems
just don't issue such commands and the underlying device will never
seen them.

> - If alot of space is unused and not freed guest can request host 
>   Administrator for truncating the host backing image. 

You can't use truncate to free space in a disk image file. The only
way to do it safely in a generic, filesystem agnositic way is to
mount the disk image (e.g. on loopback) and run fstrim on it. The
loopback device will punches holes in the file where all the free
space is reported by the filesystem via discard requests.

Which is kinda my point - this could only be done if the guest is
shut down, which makes it very difficult for admins to manage. 

>   We are also planning to support qcow2 sparse image format at 
>   host side with virtio-pmem.

So you're going to be remapping a huge number of disjoint regions
into a linear pmem mapping? ISTR discussions about similar things
for virtio+fuse+dax that came up against "large numbers of mapped
regions don't scale" and so it wasn't a practical solution compared
to a just using raw sparse files....

> - There is no existing solution for Qemu persistent memory 
>   emulation with write support currently. This solution provides 
>   us the paravartualized way of emulating persistent memory.

Sure, but the question is why do you need to create an emulation
that doesn't actually perform like pmem? The whole point of pmem is
performance, and emulating pmem by mmap() of a file on spinning
disks is going to be horrible for performance. Even on SSDs it's
going to be orders of magnitudes slower than real pmem.

So exactly what problem are you trying to solve with this driver?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
                   ` (5 preceding siblings ...)
       [not found] ` <20190109144736.17452-1-pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2019-02-04 22:56 ` Michael S. Tsirkin
  2019-02-05  7:29   ` [Qemu-devel] " Pankaj Gupta
                     ` (2 more replies)
  6 siblings, 3 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-02-04 22:56 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, david, jmoyer, xiaoguangrong.eric,
	hch, jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw, Andrea Arcangeli


On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
>  This patch series has implementation for "virtio pmem". 
>  "virtio pmem" is fake persistent memory(nvdimm) in guest 
>  which allows to bypass the guest page cache. This also
>  implements a VIRTIO based asynchronous flush mechanism.  


At Pankaj's request I looked at information leak implications of virtio
pmem in light of the recent page cache side channels paper
(https://arxiv.org/pdf/1901.01161.pdf) - to see what
kind of side channels it might create if any.  TLDR - I think that
depending on the host side implementation there could be some, but this
might be addressable by better documentation in both code and spec.
The fake dax approach backing the guest memory by a host page cache
does seem to have potential issues.

For clarity: we are talking about leaking information either to a VM, or
within a VM (I did not look into leaks to hypervisor in configurations
such as SEV) through host page cache.

Leaks into a VM: It seems clear that while pmem allows memory accesses
versus read/write with e.g. a block device, from host page cache point
of view this doesn't matter much: reads populate cache in the same way
as memory faults.  Thus ignoring presence of information leaks (which is
an interesting question e.g. in light of recent discard support) pmem
doesn't seem to be any better or worse for leaking information into a
VM.

Leaks within VM: Right now pmem seems to bypass the guest page cache
completely.  Whether pmem memory is then resident in a page cache would
be up to the device/host. Assuming that it is, the "Preventing
Efficient Eviction while Increasing the System Performance"
countermeasure for the page cache side channel attack would appear to
become ineffective with pmem. What is suggested is a per-process
management of the page cache, and host does not have visibility of
processes within a VM. Another possible countermeasure - not discussed
in the paper - could be modify the applications to lock the security
relevant pages in memory.  Again this becomes impractical with pmem as
host does not have visibility into that. However note that as long
as the only countermeasure linux uses is "Privileged Access"
(i.e. blocking mincore) nothing can be done as guest page cache
remains as vulnerable as host page cache.


Countermeasures: which host-side countermeasures can be designed would
depend on which countermeasures are used guest-side - we would need to
make sure they are not broken by pmem.  For "Preventing Efficient
Eviction while Increasing the System Performance" modifying the host
implementation to ensure that pmem device bypasses the host page cache
would seem to address the security problem.Similarly, ensuring that a
real memory device (e.g. DAX, RAM such as hugetlbfs, pmem for nested
virt) is used for pmem would make the memory locking countermeasure
work.  Whether with such limitations the device is still useful
performance wise is an open question.  These questions probably should
be addressed in the documentation, spec and possible qemu code.



Severity of the security implications: some people argue that the
security implications of the page cache leaks are minor.  I do not have
an opinion on this: the severity would seem to depend on the specific
configuration.


Other security implications: recent discussion seems to suggest there
are other concerns around e.g. resource management and thus DOS
potential. If that's so, it's a matter for a separate discussion
as I didn't look into that in depth.

Some or all of the above might be based on a misunderstanding of the
current pmem code, the whitepaper and linux page cache in general.
If so I apologise, do not hesitate to call out any mistakes.

Thanks!

-- 
MST

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-04 22:56 ` security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device) Michael S. Tsirkin
@ 2019-02-05  7:29   ` " Pankaj Gupta
  2019-02-06 14:00   ` David Hildenbrand
  2019-02-11  7:29   ` [Qemu-devel] " Pankaj Gupta
  2 siblings, 0 replies; 61+ messages in thread
From: Pankaj Gupta @ 2019-02-05  7:29 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: jack, kvm, david, linux-nvdimm, jasowang, qemu-devel,
	virtualization, adilger kernel, zwisler, Andrea Arcangeli,
	dave jiang, darrick wong, vishal l verma, willy, hch, linux-acpi,
	jmoyer, nilal, riel, stefanha, imammedo, dan j williams,
	lcapitulino, kwolf, linux-ext4, tytso, xiaoguangrong eric, rjw,
	linux-kernel, linux-xfs, linux-fsdevel, pbonzini, david


+CC [Dave Chinner], to maintain updated CC list  

> >  This patch series has implementation for "virtio pmem".
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.
> 
> 
> At Pankaj's request I looked at information leak implications of virtio
> pmem in light of the recent page cache side channels paper
> (https://arxiv.org/pdf/1901.01161.pdf) - to see what
> kind of side channels it might create if any.  TLDR - I think that
> depending on the host side implementation there could be some, but this
> might be addressable by better documentation in both code and spec.
> The fake dax approach backing the guest memory by a host page cache
> does seem to have potential issues.
> 
> For clarity: we are talking about leaking information either to a VM, or
> within a VM (I did not look into leaks to hypervisor in configurations
> such as SEV) through host page cache.
> 
> Leaks into a VM: It seems clear that while pmem allows memory accesses
> versus read/write with e.g. a block device, from host page cache point
> of view this doesn't matter much: reads populate cache in the same way
> as memory faults.  Thus ignoring presence of information leaks (which is
> an interesting question e.g. in light of recent discard support) pmem
> doesn't seem to be any better or worse for leaking information into a
> VM.
> 
> Leaks within VM: Right now pmem seems to bypass the guest page cache
> completely.  Whether pmem memory is then resident in a page cache would
> be up to the device/host. Assuming that it is, the "Preventing
> Efficient Eviction while Increasing the System Performance"
> countermeasure for the page cache side channel attack would appear to
> become ineffective with pmem. What is suggested is a per-process
> management of the page cache, and host does not have visibility of
> processes within a VM. Another possible countermeasure - not discussed
> in the paper - could be modify the applications to lock the security
> relevant pages in memory.  Again this becomes impractical with pmem as
> host does not have visibility into that. However note that as long
> as the only countermeasure linux uses is "Privileged Access"
> (i.e. blocking mincore) nothing can be done as guest page cache
> remains as vulnerable as host page cache.
> 
> 
> Countermeasures: which host-side countermeasures can be designed would
> depend on which countermeasures are used guest-side - we would need to
> make sure they are not broken by pmem.  For "Preventing Efficient
> Eviction while Increasing the System Performance" modifying the host
> implementation to ensure that pmem device bypasses the host page cache
> would seem to address the security problem.Similarly, ensuring that a
> real memory device (e.g. DAX, RAM such as hugetlbfs, pmem for nested
> virt) is used for pmem would make the memory locking countermeasure
> work.  Whether with such limitations the device is still useful
> performance wise is an open question.  These questions probably should
> be addressed in the documentation, spec and possible qemu code.
> 
> 
> 
> Severity of the security implications: some people argue that the
> security implications of the page cache leaks are minor.  I do not have
> an opinion on this: the severity would seem to depend on the specific
> configuration.
> 
> 
> Other security implications: recent discussion seems to suggest there
> are other concerns around e.g. resource management and thus DOS
> potential. If that's so, it's a matter for a separate discussion
> as I didn't look into that in depth.
> 
> Some or all of the above might be based on a misunderstanding of the
> current pmem code, the whitepaper and linux page cache in general.
> If so I apologise, do not hesitate to call out any mistakes.
> 
> Thanks!
> 
> --
> MST
> 
> 

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-04 22:56 ` security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device) Michael S. Tsirkin
  2019-02-05  7:29   ` [Qemu-devel] " Pankaj Gupta
@ 2019-02-06 14:00   ` David Hildenbrand
  2019-02-06 18:01     ` Michael S. Tsirkin
  2019-02-11  7:29   ` [Qemu-devel] " Pankaj Gupta
  2 siblings, 1 reply; 61+ messages in thread
From: David Hildenbrand @ 2019-02-06 14:00 UTC (permalink / raw)
  To: Michael S. Tsirkin, Pankaj Gupta
  Cc: linux-kernel, kvm, qemu-devel, linux-nvdimm, linux-fsdevel,
	virtualization, linux-acpi, linux-ext4, linux-xfs, jack,
	stefanha, dan.j.williams, riel, nilal, kwolf, pbonzini, zwisler,
	vishal.l.verma, dave.jiang, jmoyer, xiaoguangrong.eric, hch,
	jasowang, lcapitulino, imammedo, eblake, willy, tytso,
	adilger.kernel, darrick.wong, rjw, Andrea Arcangeli

On 04.02.19 23:56, Michael S. Tsirkin wrote:
> 
> On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
>>  This patch series has implementation for "virtio pmem". 
>>  "virtio pmem" is fake persistent memory(nvdimm) in guest 
>>  which allows to bypass the guest page cache. This also
>>  implements a VIRTIO based asynchronous flush mechanism.  
> 
> 
> At Pankaj's request I looked at information leak implications of virtio
> pmem in light of the recent page cache side channels paper
> (https://arxiv.org/pdf/1901.01161.pdf) - to see what
> kind of side channels it might create if any.  TLDR - I think that
> depending on the host side implementation there could be some, but this
> might be addressable by better documentation in both code and spec.
> The fake dax approach backing the guest memory by a host page cache
> does seem to have potential issues.
> 
> For clarity: we are talking about leaking information either to a VM, or
> within a VM (I did not look into leaks to hypervisor in configurations
> such as SEV) through host page cache.
> 
> Leaks into a VM: It seems clear that while pmem allows memory accesses
> versus read/write with e.g. a block device, from host page cache point
> of view this doesn't matter much: reads populate cache in the same way
> as memory faults.  Thus ignoring presence of information leaks (which is
> an interesting question e.g. in light of recent discard support) pmem
> doesn't seem to be any better or worse for leaking information into a
> VM.

+1, just a different way to access that cache.

Conceptually a virtio-pmem devices is from the guest view a "device with
a managed buffer". Some accesses might be faster than others. There are
no guarantees on how fast a certain access is. And yes, actions on other
guests can result in accesses being slower but not faster.

Also other storage devices have caches like that (well, the caches size
depends on the device) - thinking especially about storage systems -
which would in my opinion, also allow similar leaks. How are such
security concerns handled there? Are they different (besides eventually
access speed)?

> 
> Leaks within VM: Right now pmem seems to bypass the guest page cache
> completely.  Whether pmem memory is then resident in a page cache would
> be up to the device/host. Assuming that it is, the "Preventing
> Efficient Eviction while Increasing the System Performance"
> countermeasure for the page cache side channel attack would appear to
> become ineffective with pmem. What is suggested is a per-process
> management of the page cache, and host does not have visibility of
> processes within a VM. Another possible countermeasure - not discussed
> in the paper - could be modify the applications to lock the security
> relevant pages in memory.  Again this becomes impractical with pmem as
> host does not have visibility into that. However note that as long
> as the only countermeasure linux uses is "Privileged Access"
> (i.e. blocking mincore) nothing can be done as guest page cache
> remains as vulnerable as host page cache.

This sounds very use-case specific. If I run a VM only with a very
specific workload (say, a container running one application), I usually
don't care about leaks within the VM. At least not leaks between
applications ;)

In contrast, to running different applications (e.g. containers from
different customers) on one system, I really care about leaks within a VM.

> 
> 
> Countermeasures: which host-side countermeasures can be designed would
> depend on which countermeasures are used guest-side - we would need to
> make sure they are not broken by pmem.  For "Preventing Efficient
> Eviction while Increasing the System Performance" modifying the host
> implementation to ensure that pmem device bypasses the host page cache
> would seem to address the security problem.Similarly, ensuring that a
> real memory device (e.g. DAX, RAM such as hugetlbfs, pmem for nested
> virt) is used for pmem would make the memory locking countermeasure
> work.  Whether with such limitations the device is still useful
> performance wise is an open question.  These questions probably should
> be addressed in the documentation, spec and possible qemu code.
> 
I also want to note that using a disk/file as memory backend with
NVDIMMs in QEMU essentially results in the exact same questions we have
with virtio-pmem.

E.g. kata-containers use nvdimms for the rootfile system (read-only) as
far as I am aware.

Conceptually, a virtio-pmem device is just an emulated nvdimm device
with a flush interface. And the nice thing is, that it is designed to
also work on architectures that don't speak "nvdimm".

> 
> Severity of the security implications: some people argue that the
> security implications of the page cache leaks are minor.  I do not have
> an opinion on this: the severity would seem to depend on the specific
> configuration.

I guess configuration and use case.

Nice summary, thanks for looking into this Michael!


-- 

Thanks,

David / dhildenb

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-06 14:00   ` David Hildenbrand
@ 2019-02-06 18:01     ` Michael S. Tsirkin
  0 siblings, 0 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-02-06 18:01 UTC (permalink / raw)
  To: David Hildenbrand
  Cc: Pankaj Gupta, jack, kvm, linux-nvdimm, qemu-devel,
	virtualization, adilger.kernel, zwisler, eblake,
	Andrea Arcangeli, dave.jiang, darrick.wong, vishal.l.verma,
	willy, hch, linux-acpi, jmoyer, nilal, riel, stefanha, imammedo,
	dan.j.williams, lcapitulino, linux-ext4, tytso,
	xiaoguangrong.eric, rjw, linux-kernel, linux-xfs, linux-fsdevel,
	pbonzini

On Wed, Feb 06, 2019 at 03:00:26PM +0100, David Hildenbrand wrote:
> On 04.02.19 23:56, Michael S. Tsirkin wrote:
> > 
> > On Wed, Jan 09, 2019 at 08:17:31PM +0530, Pankaj Gupta wrote:
> >>  This patch series has implementation for "virtio pmem". 
> >>  "virtio pmem" is fake persistent memory(nvdimm) in guest 
> >>  which allows to bypass the guest page cache. This also
> >>  implements a VIRTIO based asynchronous flush mechanism.  
> > 
> > 
> > At Pankaj's request I looked at information leak implications of virtio
> > pmem in light of the recent page cache side channels paper
> > (https://arxiv.org/pdf/1901.01161.pdf) - to see what
> > kind of side channels it might create if any.  TLDR - I think that
> > depending on the host side implementation there could be some, but this
> > might be addressable by better documentation in both code and spec.
> > The fake dax approach backing the guest memory by a host page cache
> > does seem to have potential issues.
> > 
> > For clarity: we are talking about leaking information either to a VM, or
> > within a VM (I did not look into leaks to hypervisor in configurations
> > such as SEV) through host page cache.
> > 
> > Leaks into a VM: It seems clear that while pmem allows memory accesses
> > versus read/write with e.g. a block device, from host page cache point
> > of view this doesn't matter much: reads populate cache in the same way
> > as memory faults.  Thus ignoring presence of information leaks (which is
> > an interesting question e.g. in light of recent discard support) pmem
> > doesn't seem to be any better or worse for leaking information into a
> > VM.
> 
> +1, just a different way to access that cache.
> 
> Conceptually a virtio-pmem devices is from the guest view a "device with
> a managed buffer". Some accesses might be faster than others. There are
> no guarantees on how fast a certain access is. And yes, actions on other
> guests can result in accesses being slower but not faster.
> 
> Also other storage devices have caches like that (well, the caches size
> depends on the device) - thinking especially about storage systems -
> which would in my opinion, also allow similar leaks. How are such
> security concerns handled there? Are they different (besides eventually
> access speed)?
> 
> > 
> > Leaks within VM: Right now pmem seems to bypass the guest page cache
> > completely.  Whether pmem memory is then resident in a page cache would
> > be up to the device/host. Assuming that it is, the "Preventing
> > Efficient Eviction while Increasing the System Performance"
> > countermeasure for the page cache side channel attack would appear to
> > become ineffective with pmem. What is suggested is a per-process
> > management of the page cache, and host does not have visibility of
> > processes within a VM. Another possible countermeasure - not discussed
> > in the paper - could be modify the applications to lock the security
> > relevant pages in memory.  Again this becomes impractical with pmem as
> > host does not have visibility into that. However note that as long
> > as the only countermeasure linux uses is "Privileged Access"
> > (i.e. blocking mincore) nothing can be done as guest page cache
> > remains as vulnerable as host page cache.
> 
> This sounds very use-case specific. If I run a VM only with a very
> specific workload (say, a container running one application), I usually
> don't care about leaks within the VM. At least not leaks between
> applications ;)
> 
> In contrast, to running different applications (e.g. containers from
> different customers) on one system, I really care about leaks within a VM.

Clearly, not everyone cares about closing off information leaks.

> > 
> > 
> > Countermeasures: which host-side countermeasures can be designed would
> > depend on which countermeasures are used guest-side - we would need to
> > make sure they are not broken by pmem.  For "Preventing Efficient
> > Eviction while Increasing the System Performance" modifying the host
> > implementation to ensure that pmem device bypasses the host page cache
> > would seem to address the security problem.Similarly, ensuring that a
> > real memory device (e.g. DAX, RAM such as hugetlbfs, pmem for nested
> > virt) is used for pmem would make the memory locking countermeasure
> > work.  Whether with such limitations the device is still useful
> > performance wise is an open question.  These questions probably should
> > be addressed in the documentation, spec and possible qemu code.
> > 
> I also want to note that using a disk/file as memory backend with
> NVDIMMs in QEMU essentially results in the exact same questions we have
> with virtio-pmem.
> 
> E.g. kata-containers use nvdimms for the rootfile system (read-only) as
> far as I am aware.
> 
> Conceptually, a virtio-pmem device is just an emulated nvdimm device
> with a flush interface. And the nice thing is, that it is designed to
> also work on architectures that don't speak "nvdimm".
> 
> > 
> > Severity of the security implications: some people argue that the
> > security implications of the page cache leaks are minor.  I do not have
> > an opinion on this: the severity would seem to depend on the specific
> > configuration.
> 
> I guess configuration and use case.

Good point.

> Nice summary, thanks for looking into this Michael!
> 
> 
> -- 
> 
> Thanks,
> 
> David / dhildenb
> _______________________________________________
> Virtualization mailing list
> Virtualization@lists.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/virtualization

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-04 22:56 ` security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device) Michael S. Tsirkin
  2019-02-05  7:29   ` [Qemu-devel] " Pankaj Gupta
  2019-02-06 14:00   ` David Hildenbrand
@ 2019-02-11  7:29   ` " Pankaj Gupta
  2019-02-11 22:29     ` Dave Chinner
  2 siblings, 1 reply; 61+ messages in thread
From: Pankaj Gupta @ 2019-02-11  7:29 UTC (permalink / raw)
  To: Michael S. Tsirkin, dchinner
  Cc: jack, kvm, david, linux-nvdimm, jasowang, qemu-devel,
	virtualization, adilger kernel, zwisler, Andrea Arcangeli,
	dave jiang, darrick wong, vishal l verma, willy, hch, linux-acpi,
	jmoyer, nilal, riel, stefanha, imammedo, dan j williams,
	lcapitulino, kwolf, linux-ext4, tytso, xiaoguangrong eric, rjw,
	linux-kernel, linux-xfs, linux-fsdevel, pbonzini


Hi Michael, 

Thanks for looking into this and summarizing in detail. 

> >  This patch series has implementation for "virtio pmem".
> >  "virtio pmem" is fake persistent memory(nvdimm) in guest
> >  which allows to bypass the guest page cache. This also
> >  implements a VIRTIO based asynchronous flush mechanism.
> 
> 
> At Pankaj's request I looked at information leak implications of virtio
> pmem in light of the recent page cache side channels paper
> (https://arxiv.org/pdf/1901.01161.pdf) - to see what
> kind of side channels it might create if any.  TLDR - I think that
> depending on the host side implementation there could be some, but this
> might be addressable by better documentation in both code and spec.
> The fake dax approach backing the guest memory by a host page cache
> does seem to have potential issues.
> 
> For clarity: we are talking about leaking information either to a VM, or
> within a VM (I did not look into leaks to hypervisor in configurations
> such as SEV) through host page cache.
> 
> Leaks into a VM: It seems clear that while pmem allows memory accesses
> versus read/write with e.g. a block device, from host page cache point
> of view this doesn't matter much: reads populate cache in the same way
> as memory faults.  Thus ignoring presence of information leaks (which is
> an interesting question e.g. in light of recent discard support) pmem
> doesn't seem to be any better or worse for leaking information into a
> VM.
> 
> Leaks within VM: Right now pmem seems to bypass the guest page cache
> completely.  Whether pmem memory is then resident in a page cache would
> be up to the device/host. Assuming that it is, the "Preventing
> Efficient Eviction while Increasing the System Performance"
> countermeasure for the page cache side channel attack would appear to
> become ineffective with pmem. What is suggested is a per-process
> management of the page cache, and host does not have visibility of
> processes within a VM. Another possible countermeasure - not discussed
> in the paper - could be modify the applications to lock the security
> relevant pages in memory.  Again this becomes impractical with pmem as
> host does not have visibility into that. However note that as long
> as the only countermeasure linux uses is "Privileged Access"
> (i.e. blocking mincore) nothing can be done as guest page cache
> remains as vulnerable as host page cache.
> 
> 
> Countermeasures: which host-side countermeasures can be designed would
> depend on which countermeasures are used guest-side - we would need to
> make sure they are not broken by pmem.  For "Preventing Efficient
> Eviction while Increasing the System Performance" modifying the host
> implementation to ensure that pmem device bypasses the host page cache
> would seem to address the security problem.Similarly, ensuring that a
> real memory device (e.g. DAX, RAM such as hugetlbfs, pmem for nested
> virt) is used for pmem would make the memory locking countermeasure
> work.  Whether with such limitations the device is still useful
> performance wise is an open question.  These questions probably should
> be addressed in the documentation, spec and possible qemu code.
> 
> 
> 
> Severity of the security implications: some people argue that the
> security implications of the page cache leaks are minor.  I do not have
> an opinion on this: the severity would seem to depend on the specific
> configuration.
> 
> 
> Other security implications: recent discussion seems to suggest there
> are other concerns around e.g. resource management and thus DOS
> potential. If that's so, it's a matter for a separate discussion
> as I didn't look into that in depth.
> 
> Some or all of the above might be based on a misunderstanding of the
> current pmem code, the whitepaper and linux page cache in general.
> If so I apologise, do not hesitate to call out any mistakes.

I agree similar to any guest VM(without virtio-pmem) or any host 
userspace process, virtio-pmem may also have some security implications.
We need to document these in virtio-pmem host device specification
and in qemu code. This is to make sure while creating virtio-pmem
device, we are aware of these implications and create/use 
host side device backing file accordingly as per the use-case
described by David here [1].

I will document these in next version of my patch series.

Hello Dave,
Are we okay with this?

Thank you everyone for the discussion.

[1] https://marc.info/?l=linux-kernel&m=154946989419403&w=2

Best regards,
Pankaj    

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-11  7:29   ` [Qemu-devel] " Pankaj Gupta
@ 2019-02-11 22:29     ` Dave Chinner
  2019-02-11 22:58       ` David Hildenbrand
  0 siblings, 1 reply; 61+ messages in thread
From: Dave Chinner @ 2019-02-11 22:29 UTC (permalink / raw)
  To: Pankaj Gupta
  Cc: Michael S. Tsirkin, dchinner, jack, kvm, david, linux-nvdimm,
	jasowang, qemu-devel, virtualization, adilger kernel, zwisler,
	Andrea Arcangeli, dave jiang, darrick wong, vishal l verma,
	willy, hch, linux-acpi, jmoyer, nilal, riel, stefanha, imammedo,
	dan j williams, lcapitulino, kwolf, linux-ext4, tytso,
	xiaoguangrong eric, rjw, linux-kernel, linux-xfs, linux-fsdevel,
	pbonzini

On Mon, Feb 11, 2019 at 02:29:46AM -0500, Pankaj Gupta wrote:
> Hello Dave,
> Are we okay with this?

Sure.

I'm not sure I agree with all the analysis presented, but, well, I
haven't looked any deeper because I'm tired of being shouted at and
being called argumentative for daring to ask hard questions about
this topic....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-11 22:29     ` Dave Chinner
@ 2019-02-11 22:58       ` David Hildenbrand
  2019-02-11 23:07         ` Michael S. Tsirkin
  0 siblings, 1 reply; 61+ messages in thread
From: David Hildenbrand @ 2019-02-11 22:58 UTC (permalink / raw)
  To: Dave Chinner, Pankaj Gupta
  Cc: Michael S. Tsirkin, dchinner, jack, kvm, linux-nvdimm, jasowang,
	qemu-devel, virtualization, adilger kernel, zwisler,
	Andrea Arcangeli, dave jiang, darrick wong, vishal l verma,
	willy, hch, linux-acpi, jmoyer, nilal, riel, stefanha, imammedo,
	dan j williams, lcapitulino, kwolf, linux-ext4, tytso,
	xiaoguangrong eric, rjw, linux-kernel, linux-xfs, linux-fsdevel,
	pbonzini

On 11.02.19 23:29, Dave Chinner wrote:
> On Mon, Feb 11, 2019 at 02:29:46AM -0500, Pankaj Gupta wrote:
>> Hello Dave,
>> Are we okay with this?
> 
> Sure.
> 
> I'm not sure I agree with all the analysis presented, but, well, I
> haven't looked any deeper because I'm tired of being shouted at and
> being called argumentative for daring to ask hard questions about
> this topic....

I think if you have concerns, they should definitely be discussed.
Making people frustrated that review code is not what we want. Not at all.

I suggest that Pankaj properly documents what we found out so far about
security concerns and properly describes intended use cases and answers
other questions you had in the cover letter / documentation of the
follow up series.

Thanks Dave!

> 
> Cheers,
> 
> Dave.
> 


-- 

Thanks,

David / dhildenb

^ permalink raw reply	[flat|nested] 61+ messages in thread

* Re: [Qemu-devel] security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device)
  2019-02-11 22:58       ` David Hildenbrand
@ 2019-02-11 23:07         ` Michael S. Tsirkin
  0 siblings, 0 replies; 61+ messages in thread
From: Michael S. Tsirkin @ 2019-02-11 23:07 UTC (permalink / raw)
  To: David Hildenbrand
  Cc: Dave Chinner, Pankaj Gupta, dchinner, jack, kvm, linux-nvdimm,
	jasowang, qemu-devel, virtualization, adilger kernel, zwisler,
	Andrea Arcangeli, dave jiang, darrick wong, vishal l verma,
	willy, hch, linux-acpi, jmoyer, nilal, riel, stefanha, imammedo,
	dan j williams, lcapitulino, kwolf, linux-ext4, tytso,
	xiaoguangrong eric, rjw, linux-kernel, linux-xfs, linux-fsdevel,
	pbonzini

On Mon, Feb 11, 2019 at 11:58:15PM +0100, David Hildenbrand wrote:
> On 11.02.19 23:29, Dave Chinner wrote:
> > On Mon, Feb 11, 2019 at 02:29:46AM -0500, Pankaj Gupta wrote:
> >> Hello Dave,
> >> Are we okay with this?
> > 
> > Sure.
> > 
> > I'm not sure I agree with all the analysis presented, but, well, I
> > haven't looked any deeper because I'm tired of being shouted at and
> > being called argumentative for daring to ask hard questions about
> > this topic....
> 
> I think if you have concerns, they should definitely be discussed.
> Making people frustrated that review code is not what we want. Not at all.
> 
> I suggest that Pankaj properly documents what we found out so far about
> security concerns and properly describes intended use cases and answers
> other questions you had in the cover letter / documentation of the
> follow up series.
> 
> Thanks Dave!

Right. Also, there's an open question that you posed:
	Also other storage devices have caches like that (well, the caches size
	depends on the device) - thinking especially about storage systems -
	which would in my opinion, also allow similar leaks. How are such
	security concerns handled there? Are they different (besides eventually
	access speed)?
and that needs some looking into, and reporting on.


> > 
> > Cheers,
> > 
> > Dave.
> > 
> 
> 
> -- 
> 
> Thanks,
> 
> David / dhildenb

^ permalink raw reply	[flat|nested] 61+ messages in thread

end of thread, back to index

Thread overview: 61+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-01-09 14:47 [PATCH v3 0/5] kvm "virtio pmem" device Pankaj Gupta
2019-01-09 14:47 ` [PATCH v3 1/5] libnvdimm: nd_region flush callback support Pankaj Gupta
2019-01-09 14:47 ` [PATCH v3 2/5] virtio-pmem: Add virtio pmem driver Pankaj Gupta
2019-01-14 15:54   ` Michael S. Tsirkin
2019-01-14 15:54     ` Michael S. Tsirkin
     [not found]     ` <20190114105314-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
2019-01-15  6:33       ` Pankaj Gupta
2019-01-15  6:33         ` Pankaj Gupta
2019-01-09 14:47 ` [PATCH v3 3/5] libnvdimm: add nd_region buffered dax_dev flag Pankaj Gupta
2019-01-09 14:47 ` [PATCH v3 4/5] ext4: disable map_sync for virtio pmem Pankaj Gupta
2019-01-09 14:47 ` [PATCH v3 5/5] xfs: " Pankaj Gupta
2019-01-09 14:47   ` Pankaj Gupta
2019-01-09 16:26   ` Darrick J. Wong
2019-01-09 18:08     ` Pankaj Gupta
     [not found] ` <20190109144736.17452-1-pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-10  1:26   ` [PATCH v3 0/5] kvm "virtio pmem" device Dave Chinner
2019-01-10  1:26     ` Dave Chinner
2019-01-10  2:40     ` Rik van Riel
2019-01-10  2:40       ` Rik van Riel
2019-01-10 10:17     ` Jan Kara
2019-01-10 10:17       ` Jan Kara
     [not found]       ` <20190110101757.GC15790-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
2019-01-13  1:38         ` Pankaj Gupta
2019-01-13  1:38           ` Pankaj Gupta
     [not found]           ` <1354249849.63357171.1547343519970.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-13  1:43             ` Dan Williams
2019-01-13  1:43               ` Dan Williams
     [not found]               ` <CAPcyv4hwcgTUpgNCefCGu4DvgkYBp5b=f+hJ+FC=s5APYKoycg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
2019-01-13  2:17                 ` [Qemu-devel] " Pankaj Gupta
2019-01-13  2:17                   ` Pankaj Gupta
     [not found]                   ` <540171952.63371441.1547345866585.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-14  9:55                     ` Jan Kara
2019-01-14  9:55                       ` Jan Kara
     [not found]                       ` <20190114095520.GC13316-4I4JzKEfoa/jFM9bn6wA6Q@public.gmane.org>
2019-01-14 10:16                         ` Pankaj Gupta
2019-01-14 10:16                           ` Pankaj Gupta
2019-01-11  7:45     ` Pankaj Gupta
2019-01-11  7:45       ` Pankaj Gupta
     [not found]       ` <1326478078.61913951.1547192704870.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-13 23:29         ` Dave Chinner
2019-01-13 23:29           ` Dave Chinner
2019-01-13 23:38           ` Matthew Wilcox
2019-01-13 23:38             ` Matthew Wilcox
     [not found]             ` <20190113233820.GX6310-PfSpb0PWhxZc2C7mugBRk2EX/6BAtgUQ@public.gmane.org>
2019-01-14  2:50               ` Dave Chinner
2019-01-14  2:50                 ` Dave Chinner
2019-01-14  7:15               ` Pankaj Gupta
2019-01-14  7:15                 ` Pankaj Gupta
     [not found]                 ` <942065073.64011540.1547450140670.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-14 21:25                   ` Dave Chinner
2019-01-14 21:25                     ` Dave Chinner
2019-01-14 21:35                     ` Dan Williams
2019-01-14 21:35                       ` Dan Williams
     [not found]                       ` <CAPcyv4jtPcLV-s0sKNHwwk0ug7GLBV6699dpm1h3r2xSo879dg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
2019-01-14 22:21                         ` Dave Chinner
2019-01-14 22:21                           ` Dave Chinner
2019-01-15  2:19                           ` Michael S. Tsirkin
2019-01-15  2:19                             ` Michael S. Tsirkin
     [not found]                             ` <20190114205031-mutt-send-email-mst-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
2019-01-15  5:37                               ` [Qemu-devel] " Pankaj Gupta
2019-01-15  5:37                                 ` Pankaj Gupta
2019-01-15  5:35                           ` Pankaj Gupta
2019-01-15  5:35                             ` Pankaj Gupta
     [not found]                             ` <1684638419.64320214.1547530506805.JavaMail.zimbra-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2019-01-15 20:42                               ` Dave Chinner
2019-01-15 20:42                                 ` Dave Chinner
2019-02-04 22:56 ` security implications of caching with virtio pmem (was Re: [PATCH v3 0/5] kvm "virtio pmem" device) Michael S. Tsirkin
2019-02-05  7:29   ` [Qemu-devel] " Pankaj Gupta
2019-02-06 14:00   ` David Hildenbrand
2019-02-06 18:01     ` Michael S. Tsirkin
2019-02-11  7:29   ` [Qemu-devel] " Pankaj Gupta
2019-02-11 22:29     ` Dave Chinner
2019-02-11 22:58       ` David Hildenbrand
2019-02-11 23:07         ` Michael S. Tsirkin

Linux-Fsdevel Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-fsdevel/0 linux-fsdevel/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-fsdevel linux-fsdevel/ https://lore.kernel.org/linux-fsdevel \
		linux-fsdevel@vger.kernel.org linux-fsdevel@archiver.kernel.org
	public-inbox-index linux-fsdevel

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-fsdevel


AGPL code for this site: git clone https://public-inbox.org/ public-inbox