[PATCH v7 0/5] Add shared workqueue support for idxd driver

[PATCH v7 0/5] Add shared workqueue support for idxd driver

[Dave Jiang]

v7:
- Add sign-off and review tag from Boris
Boris:
- Fixed up ENQCMDS patch
Vinod:
- Fix line formatting
- Add comment for completion address compare

v6:
Boris:
- Fixup MOBDIR64B inline asm input/output constraints

v5:
Boris:
- Fixup commit headers
- Fixup var names for movdir64b()
- Move enqcmds() to special_insns.h
- Fix up comments for enqcmds()
- Change enqcmds() to reflect instruction return. 0 as success, -EAGAIN for fail.

DavidL:
- Fixup enqcmds() gas constraints

v4:
- Rebased against latest dmaengine/next tree
- Split out enqcmd and pasid dependency.

V3:
- Rebased against latest dmaengine/next tree.
- Updated API doc with new kernel version and dates.
- Changed to allow driver to load without ENQCMD support.
- Break out some patches that can be sent ahead of this series for inclusion.

v2:
- Dropped device feature enabling (GregKH)
- Dropped PCI device feature enabling (Bjorn)
	- https://members.pcisig.com/wg/PCI-SIG/document/14237
- After some internal discussion, we have decided to hold off on the enabling of DMWR due to the
  following reasons. 1. Most first gen hw will not have the feature bits. 2. First gen hw that
  support the feature are all Root Complex integrated endpoints. 3. PCI devices that are not
  RCiEP’s with this capability won’t surface for a few years so we can wait until we can test the
  full code.
- Dropped special ioremap (hch)
- Added proper support for WQ flush (tony, dan)
- Changed descriptor submission to use sbitmap_queue for blocking. (dan)

Driver stage 1 postings for context: [1]

The patch series has compilation and functional dependency on Fenghua's "Tag application
address space for devices" patch series for the ENQCMD CPU command enumeration and the PASID MSR
support. [2] 

== Background ==
A typical DMA device requires the driver to translate application buffers to hardware addresses,
and a kernel-user transition to notify the hardware of new work. Shared Virtual Addressing (SVA)
allows the processor and device to use the same virtual addresses without requiring software to
translate between the address spaces. ENQCMD is a new instruction on Intel Platforms that allows
user applications to directly notify hardware of new work, much like how doorbells are used in
some hardware, but it carries a payload along with it. ENQCMDS is the supervisor version (ring0)
of ENQCMD.

== ENQCMDS ==
Introduce enqcmds(), a helper funciton that copies an input payload to a 64B aligned
destination and confirms whether the payload was accepted by the device or not.
enqcmds() wraps the new ENQCMDS CPU instruction. The ENQCMDS is a ring 0 CPU instruction that
performs similar to the ENQCMD instruction. Descriptor submission must use ENQCMD(S) for shared
workqueues (swq) on an Intel DSA device. 

== Shared WQ support ==
Introduce shared workqueue (swq) support for the idxd driver. The current idxd driver contains
dedicated workqueue (dwq) support only. A dwq accepts descriptors from a MOVDIR64B instruction.
MOVDIR64B is a posted instruction on the PCIe bus, it does not wait for any response from the
device. If the wq is full, submitted descriptors are dropped. A swq utilizes the ENQCMDS in
ring 0, which is a non-posted instruction. The zero flag would be set to 1 if the device rejects
the descriptor or if the wq is full. A swq can be shared between multiple users
(kernel or userspace) due to not having to keep track of the wq full condition for submission.
A swq requires PASID and can only run with SVA support. 

== IDXD SVA support ==
Add utilization of PASID to support Shared Virtual Addressing (SVA). With PASID support,
the descriptors can be programmed with host virtual address (HVA) rather than IOVA.
The hardware will work with the IOMMU in fulfilling page requests. With SVA support,
a user app using the char device interface can now submit descriptors without having to pin the
virtual memory range it wants to DMA in its own address space. 

The series does not add SVA support for the dmaengine subsystem. That support is coming at a
later time.

[1]: https://lore.kernel.org/lkml/157965011794.73301.15960052071729101309.stgit@djiang5-desk3.ch.intel.com/
[2]: https://lore.kernel.org/lkml/20200916080510.GA32552@8bytes.org/
[3]: https://software.intel.com/en-us/articles/intel-sdm
[4]: https://software.intel.com/en-us/download/intel-scalable-io-virtualization-technical-specification
[5]: https://software.intel.com/en-us/download/intel-data-streaming-accelerator-preliminary-architecture-specification
[6]: https://01.org/blogs/2019/introducing-intel-data-streaming-accelerator
[7]: https://intel.github.io/idxd/
[8]: https://github.com/intel/idxd-driver idxd-stage2

Dave Jiang (5):
  x86/asm: Carve out a generic movdir64b() helper for general usage
  x86/asm: Add an enqcmds() wrapper for the ENQCMDS instruction
  dmaengine: idxd: Add shared workqueue support
  dmaengine: idxd: Clean up descriptors with fault error
  dmaengine: idxd: Add ABI documentation for shared wq

 .../ABI/stable/sysfs-driver-dma-idxd          |  14 ++
 arch/x86/include/asm/io.h                     |  17 +-
 arch/x86/include/asm/special_insns.h          |  64 ++++++++
 drivers/dma/Kconfig                           |  10 ++
 drivers/dma/idxd/cdev.c                       |  49 +++++-
 drivers/dma/idxd/device.c                     |  91 ++++++++++-
 drivers/dma/idxd/dma.c                        |   9 --
 drivers/dma/idxd/idxd.h                       |  33 +++-
 drivers/dma/idxd/init.c                       |  92 ++++++++---
 drivers/dma/idxd/irq.c                        | 146 ++++++++++++++++--
 drivers/dma/idxd/registers.h                  |  14 ++
 drivers/dma/idxd/submit.c                     |  35 ++++-
 drivers/dma/idxd/sysfs.c                      | 127 +++++++++++++++
 13 files changed, 631 insertions(+), 70 deletions(-)

-- 
2.26.2

[PATCH v7 1/5] x86/asm: Carve out a generic movdir64b() helper for general usage

[Dave Jiang]

The MOVDIR64B instruction can be used by other wrapper instructions. Move
the asm code to special_insns.h and have iosubmit_cmds512() call the
asm function.

Reviewed-by: Tony Luck <tony.luck@intel.com>
Suggested-by: Michael Matz <matz@suse.de>
Reviewed-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
---
 arch/x86/include/asm/io.h            | 17 +++--------------
 arch/x86/include/asm/special_insns.h | 22 ++++++++++++++++++++++
 2 files changed, 25 insertions(+), 14 deletions(-)

diff --git a/arch/x86/include/asm/io.h b/arch/x86/include/asm/io.h
index e1aa17a468a8..d726459d08e5 100644
--- a/arch/x86/include/asm/io.h
+++ b/arch/x86/include/asm/io.h
@@ -401,7 +401,7 @@ extern bool phys_mem_access_encrypted(unsigned long phys_addr,
 
 /**
  * iosubmit_cmds512 - copy data to single MMIO location, in 512-bit units
- * @__dst: destination, in MMIO space (must be 512-bit aligned)
+ * @dst: destination, in MMIO space (must be 512-bit aligned)
  * @src: source
  * @count: number of 512 bits quantities to submit
  *
@@ -412,25 +412,14 @@ extern bool phys_mem_access_encrypted(unsigned long phys_addr,
  * Warning: Do not use this helper unless your driver has checked that the CPU
  * instruction is supported on the platform.
  */
-static inline void iosubmit_cmds512(void __iomem *__dst, const void *src,
+static inline void iosubmit_cmds512(void __iomem *dst, const void *src,
 				    size_t count)
 {
-	/*
-	 * Note that this isn't an "on-stack copy", just definition of "dst"
-	 * as a pointer to 64-bytes of stuff that is going to be overwritten.
-	 * In the MOVDIR64B case that may be needed as you can use the
-	 * MOVDIR64B instruction to copy arbitrary memory around. This trick
-	 * lets the compiler know how much gets clobbered.
-	 */
-	volatile struct { char _[64]; } *dst = __dst;
 	const u8 *from = src;
 	const u8 *end = from + count * 64;
 
 	while (from < end) {
-		/* MOVDIR64B [rdx], rax */
-		asm volatile(".byte 0x66, 0x0f, 0x38, 0xf8, 0x02"
-			     : "=m" (dst)
-			     : "d" (from), "a" (dst));
+		movdir64b(dst, from);
 		from += 64;
 	}
 }
diff --git a/arch/x86/include/asm/special_insns.h b/arch/x86/include/asm/special_insns.h
index 59a3e13204c3..2258c7d6e281 100644
--- a/arch/x86/include/asm/special_insns.h
+++ b/arch/x86/include/asm/special_insns.h
@@ -234,6 +234,28 @@ static inline void clwb(volatile void *__p)
 
 #define nop() asm volatile ("nop")
 
+/* The dst parameter must be 64-bytes aligned */
+static inline void movdir64b(void *dst, const void *src)
+{
+	const struct { char _[64]; } *__src = src;
+	struct { char _[64]; } *__dst = dst;
+
+	/*
+	 * MOVDIR64B %(rdx), rax.
+	 *
+	 * Both __src and __dst must be memory constraints in order to tell the
+	 * compiler that no other memory accesses should be reordered around
+	 * this one.
+	 *
+	 * Also, both must be supplied as lvalues because this tells
+	 * the compiler what the object is (its size) the instruction accesses.
+	 * I.e., not the pointers but what they point, thus the deref'ing '*'.
+	 */
+	asm volatile(".byte 0x66, 0x0f, 0x38, 0xf8, 0x02"
+		     : "+m" (*__dst)
+		     :  "m" (*__src), "a" (__dst), "d" (__src));
+}
+
 #endif /* __KERNEL__ */
 
 #endif /* _ASM_X86_SPECIAL_INSNS_H */
-- 
2.26.2

[PATCH v7 2/5] x86/asm: Add an enqcmds() wrapper for the ENQCMDS instruction

[Dave Jiang]

Currently, the MOVDIR64B instruction is used to atomically submit
64-byte work descriptors to devices. Although it can encounter errors
like device queue full, command not accepted, device not ready, etc when
writing to a device MMIO, MOVDIR64B can not report back on errors from
the device itself. This means that MOVDIR64B users need to separately
interact with a device to see if a descriptor was successfully queued,
which slows down device interactions.

ENQCMD and ENQCMDS also atomically submit 64-byte work descriptors
to devices. But, they *can* report back errors directly from the
device, such as if the device was busy, or device not enabled or does
not support the command. This immediate feedback from the submission
instruction itself reduces the number of interactions with the device
and can greatly increase efficiency.

ENQCMD can be used at any privilege level, but can effectively only
submit work on behalf of the current process. ENQCMDS is a ring0-only
instruction and can explicitly specify a process context instead of
being tied to the current process or needing to reprogram the IA32_PASID
MSR.

Use ENQCMDS for work submission within the kernel because a Process
Address ID (PASID) is setup to translate the kernel virtual address
space. This PASID is provided to ENQCMDS from the descriptor structure
submitted to the device and not retrieved from IA32_PASID MSR, which is
setup for the current user address space.

See Intel Software Developer’s Manual for more information on the
instructions.

 [ bp:
   - Make operand constraints like movdir64b() because both insns are
     basically doing the same thing, more or less.
   - Fixup comments and cleanup. ]

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Link: https://lkml.kernel.org/r/20200924180041.34056-3-dave.jiang@intel.com
---
 arch/x86/include/asm/special_insns.h | 42 ++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/arch/x86/include/asm/special_insns.h b/arch/x86/include/asm/special_insns.h
index 2258c7d6e281..83f7c1a391e0 100644
--- a/arch/x86/include/asm/special_insns.h
+++ b/arch/x86/include/asm/special_insns.h
@@ -256,6 +256,48 @@ static inline void movdir64b(void *dst, const void *src)
 		     :  "m" (*__src), "a" (__dst), "d" (__src));
 }
 
+/**
+ * enqcmds - Enqueue a command in supervisor (CPL0) mode
+ * @dst: destination, in MMIO space (must be 512-bit aligned)
+ * @src: 512 bits memory operand
+ *
+ * The ENQCMDS instruction allows software to write a 512-bit command to
+ * a 512-bit-aligned special MMIO region that supports the instruction.
+ * A return status is loaded into the ZF flag in the RFLAGS register.
+ * ZF = 0 equates to success, and ZF = 1 indicates retry or error.
+ *
+ * This function issues the ENQCMDS instruction to submit data from
+ * kernel space to MMIO space, in a unit of 512 bits. Order of data access
+ * is not guaranteed, nor is a memory barrier performed afterwards. It
+ * returns 0 on success and -EAGAIN on failure.
+ *
+ * Warning: Do not use this helper unless your driver has checked that the
+ * ENQCMDS instruction is supported on the platform and the device accepts
+ * ENQCMDS.
+ */
+static inline int enqcmds(void __iomem *dst, const void *src)
+{
+	const struct { char _[64]; } *__src = src;
+	struct { char _[64]; } *__dst = dst;
+	int zf;
+
+	/*
+	 * ENQCMDS %(rdx), rax
+	 *
+	 * See movdir64b()'s comment on operand specification.
+	 */
+	asm volatile(".byte 0xf3, 0x0f, 0x38, 0xf8, 0x02, 0x66, 0x90"
+		     CC_SET(z)
+		     : CC_OUT(z) (zf), "+m" (*__dst)
+		     : "m" (*__src), "a" (__dst), "d" (__src));
+
+	/* Submission failure is indicated via EFLAGS.ZF=1 */
+	if (zf)
+		return -EAGAIN;
+
+	return 0;
+}
+
 #endif /* __KERNEL__ */
 
 #endif /* _ASM_X86_SPECIAL_INSNS_H */
-- 
2.26.2

[PATCH v7 3/5] dmaengine: idxd: Add shared workqueue support

[Dave Jiang]

Add shared workqueue support that includes the support of Shared Virtual
memory (SVM) or in similar terms On Demand Paging (ODP). The shared
workqueue uses the enqcmds command in kernel and will respond with retry if
the workqueue is full. Shared workqueue only works when there is PASID
support from the IOMMU.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
---
 drivers/dma/Kconfig          |  10 +++
 drivers/dma/idxd/cdev.c      |  49 +++++++++++++-
 drivers/dma/idxd/device.c    |  91 ++++++++++++++++++++++---
 drivers/dma/idxd/dma.c       |   9 ---
 drivers/dma/idxd/idxd.h      |  28 +++++++-
 drivers/dma/idxd/init.c      |  91 +++++++++++++++++++------
 drivers/dma/idxd/registers.h |  14 ++++
 drivers/dma/idxd/submit.c    |  35 ++++++++--
 drivers/dma/idxd/sysfs.c     | 127 +++++++++++++++++++++++++++++++++++
 9 files changed, 410 insertions(+), 44 deletions(-)

diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
index 518a1437862a..6a908785a5f7 100644
--- a/drivers/dma/Kconfig
+++ b/drivers/dma/Kconfig
@@ -296,6 +296,16 @@ config INTEL_IDXD
 
 	  If unsure, say N.
 
+# Config symbol that collects all the dependencies that's necessary to
+# support shared virtual memory for the devices supported by idxd.
+config INTEL_IDXD_SVM
+	bool "Accelerator Shared Virtual Memory Support"
+	depends on INTEL_IDXD
+	depends on INTEL_IOMMU_SVM
+	depends on PCI_PRI
+	depends on PCI_PASID
+	depends on PCI_IOV
+
 config INTEL_IOATDMA
 	tristate "Intel I/OAT DMA support"
 	depends on PCI && X86_64
diff --git a/drivers/dma/idxd/cdev.c b/drivers/dma/idxd/cdev.c
index c3976156db2f..010b820d8f74 100644
--- a/drivers/dma/idxd/cdev.c
+++ b/drivers/dma/idxd/cdev.c
@@ -11,6 +11,7 @@
 #include <linux/cdev.h>
 #include <linux/fs.h>
 #include <linux/poll.h>
+#include <linux/iommu.h>
 #include <uapi/linux/idxd.h>
 #include "registers.h"
 #include "idxd.h"
@@ -32,7 +33,9 @@ static struct idxd_cdev_context ictx[IDXD_TYPE_MAX] = {
 struct idxd_user_context {
 	struct idxd_wq *wq;
 	struct task_struct *task;
+	unsigned int pasid;
 	unsigned int flags;
+	struct iommu_sva *sva;
 };
 
 enum idxd_cdev_cleanup {
@@ -75,6 +78,8 @@ static int idxd_cdev_open(struct inode *inode, struct file *filp)
 	struct idxd_wq *wq;
 	struct device *dev;
 	int rc = 0;
+	struct iommu_sva *sva;
+	unsigned int pasid;
 
 	wq = inode_wq(inode);
 	idxd = wq->idxd;
@@ -95,6 +100,34 @@ static int idxd_cdev_open(struct inode *inode, struct file *filp)
 
 	ctx->wq = wq;
 	filp->private_data = ctx;
+
+	if (device_pasid_enabled(idxd)) {
+		sva = iommu_sva_bind_device(dev, current->mm, NULL);
+		if (IS_ERR(sva)) {
+			rc = PTR_ERR(sva);
+			dev_err(dev, "pasid allocation failed: %d\n", rc);
+			goto failed;
+		}
+
+		pasid = iommu_sva_get_pasid(sva);
+		if (pasid == IOMMU_PASID_INVALID) {
+			iommu_sva_unbind_device(sva);
+			goto failed;
+		}
+
+		ctx->sva = sva;
+		ctx->pasid = pasid;
+
+		if (wq_dedicated(wq)) {
+			rc = idxd_wq_set_pasid(wq, pasid);
+			if (rc < 0) {
+				iommu_sva_unbind_device(sva);
+				dev_err(dev, "wq set pasid failed: %d\n", rc);
+				goto failed;
+			}
+		}
+	}
+
 	idxd_wq_get(wq);
 	mutex_unlock(&wq->wq_lock);
 	return 0;
@@ -111,13 +144,27 @@ static int idxd_cdev_release(struct inode *node, struct file *filep)
 	struct idxd_wq *wq = ctx->wq;
 	struct idxd_device *idxd = wq->idxd;
 	struct device *dev = &idxd->pdev->dev;
+	int rc;
 
 	dev_dbg(dev, "%s called\n", __func__);
 	filep->private_data = NULL;
 
 	/* Wait for in-flight operations to complete. */
-	idxd_wq_drain(wq);
+	if (wq_shared(wq)) {
+		idxd_device_drain_pasid(idxd, ctx->pasid);
+	} else {
+		if (device_pasid_enabled(idxd)) {
+			/* The wq disable in the disable pasid function will drain the wq */
+			rc = idxd_wq_disable_pasid(wq);
+			if (rc < 0)
+				dev_err(dev, "wq disable pasid failed.\n");
+		} else {
+			idxd_wq_drain(wq);
+		}
+	}
 
+	if (ctx->sva)
+		iommu_sva_unbind_device(ctx->sva);
 	kfree(ctx);
 	mutex_lock(&wq->wq_lock);
 	idxd_wq_put(wq);
diff --git a/drivers/dma/idxd/device.c b/drivers/dma/idxd/device.c
index 200b9109cacf..3c6c6df9d2e8 100644
--- a/drivers/dma/idxd/device.c
+++ b/drivers/dma/idxd/device.c
@@ -273,10 +273,9 @@ int idxd_wq_map_portal(struct idxd_wq *wq)
 	start = pci_resource_start(pdev, IDXD_WQ_BAR);
 	start = start + wq->id * IDXD_PORTAL_SIZE;
 
-	wq->dportal = devm_ioremap(dev, start, IDXD_PORTAL_SIZE);
-	if (!wq->dportal)
+	wq->portal = devm_ioremap(dev, start, IDXD_PORTAL_SIZE);
+	if (!wq->portal)
 		return -ENOMEM;
-	dev_dbg(dev, "wq %d portal mapped at %p\n", wq->id, wq->dportal);
 
 	return 0;
 }
@@ -285,7 +284,61 @@ void idxd_wq_unmap_portal(struct idxd_wq *wq)
 {
 	struct device *dev = &wq->idxd->pdev->dev;
 
-	devm_iounmap(dev, wq->dportal);
+	devm_iounmap(dev, wq->portal);
+}
+
+int idxd_wq_set_pasid(struct idxd_wq *wq, int pasid)
+{
+	struct idxd_device *idxd = wq->idxd;
+	int rc;
+	union wqcfg wqcfg;
+	unsigned int offset;
+	unsigned long flags;
+
+	rc = idxd_wq_disable(wq);
+	if (rc < 0)
+		return rc;
+
+	offset = WQCFG_OFFSET(idxd, wq->id, 2);
+	spin_lock_irqsave(&idxd->dev_lock, flags);
+	wqcfg.bits[2] = ioread32(idxd->reg_base + offset);
+	wqcfg.pasid_en = 1;
+	wqcfg.pasid = pasid;
+	iowrite32(wqcfg.bits[2], idxd->reg_base + offset);
+	spin_unlock_irqrestore(&idxd->dev_lock, flags);
+
+	rc = idxd_wq_enable(wq);
+	if (rc < 0)
+		return rc;
+
+	return 0;
+}
+
+int idxd_wq_disable_pasid(struct idxd_wq *wq)
+{
+	struct idxd_device *idxd = wq->idxd;
+	int rc;
+	union wqcfg wqcfg;
+	unsigned int offset;
+	unsigned long flags;
+
+	rc = idxd_wq_disable(wq);
+	if (rc < 0)
+		return rc;
+
+	offset = idxd->wqcfg_offset + wq->id * sizeof(wqcfg);
+	spin_lock_irqsave(&idxd->dev_lock, flags);
+	wqcfg.bits[2] = ioread32(idxd->reg_base + offset);
+	wqcfg.pasid_en = 0;
+	wqcfg.pasid = 0;
+	iowrite32(wqcfg.bits[2], idxd->reg_base + offset);
+	spin_unlock_irqrestore(&idxd->dev_lock, flags);
+
+	rc = idxd_wq_enable(wq);
+	if (rc < 0)
+		return rc;
+
+	return 0;
 }
 
 void idxd_wq_disable_cleanup(struct idxd_wq *wq)
@@ -468,6 +521,17 @@ void idxd_device_reset(struct idxd_device *idxd)
 	spin_unlock_irqrestore(&idxd->dev_lock, flags);
 }
 
+void idxd_device_drain_pasid(struct idxd_device *idxd, int pasid)
+{
+	struct device *dev = &idxd->pdev->dev;
+	u32 operand;
+
+	operand = pasid;
+	dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_DRAIN_PASID, operand);
+	idxd_cmd_exec(idxd, IDXD_CMD_DRAIN_PASID, operand, NULL);
+	dev_dbg(dev, "pasid %d drained\n", pasid);
+}
+
 /* Device configuration bits */
 static void idxd_group_config_write(struct idxd_group *group)
 {
@@ -553,11 +617,22 @@ static int idxd_wq_config_write(struct idxd_wq *wq)
 	wq->wqcfg.wq_thresh = wq->threshold;
 
 	/* byte 8-11 */
-	wq->wqcfg.priv = !!(wq->type == IDXD_WQT_KERNEL);
-	wq->wqcfg.mode = 1;
+	wq->wqcfg.priv = wq->type == IDXD_WQT_KERNEL ? 1 : 0;
+	if (wq_dedicated(wq))
+		wq->wqcfg.mode = 1;
+
+	if (device_pasid_enabled(idxd)) {
+		wq->wqcfg.pasid_en = 1;
+		if (wq->type == IDXD_WQT_KERNEL && wq_dedicated(wq))
+			wq->wqcfg.pasid = idxd->pasid;
+	}
 
 	wq->wqcfg.priority = wq->priority;
 
+	if (idxd->hw.gen_cap.block_on_fault &&
+	    test_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags))
+		wq->wqcfg.bof = 1;
+
 	/* bytes 12-15 */
 	wq->wqcfg.max_xfer_shift = ilog2(wq->max_xfer_bytes);
 	wq->wqcfg.max_batch_shift = ilog2(wq->max_batch_size);
@@ -665,8 +740,8 @@ static int idxd_wqs_setup(struct idxd_device *idxd)
 		if (!wq->size)
 			continue;
 
-		if (!wq_dedicated(wq)) {
-			dev_warn(dev, "No shared workqueue support.\n");
+		if (wq_shared(wq) && !device_swq_supported(idxd)) {
+			dev_warn(dev, "No shared wq support but configured.\n");
 			return -EINVAL;
 		}
 
diff --git a/drivers/dma/idxd/dma.c b/drivers/dma/idxd/dma.c
index 0c892cbd72e0..8ed2773d8285 100644
--- a/drivers/dma/idxd/dma.c
+++ b/drivers/dma/idxd/dma.c
@@ -61,8 +61,6 @@ static inline void idxd_prep_desc_common(struct idxd_wq *wq,
 					 u64 addr_f1, u64 addr_f2, u64 len,
 					 u64 compl, u32 flags)
 {
-	struct idxd_device *idxd = wq->idxd;
-
 	hw->flags = flags;
 	hw->opcode = opcode;
 	hw->src_addr = addr_f1;
@@ -70,13 +68,6 @@ static inline void idxd_prep_desc_common(struct idxd_wq *wq,
 	hw->xfer_size = len;
 	hw->priv = !!(wq->type == IDXD_WQT_KERNEL);
 	hw->completion_addr = compl;
-
-	/*
-	 * Descriptor completion vectors are 1-8 for MSIX. We will round
-	 * robin through the 8 vectors.
-	 */
-	wq->vec_ptr = (wq->vec_ptr % idxd->num_wq_irqs) + 1;
-	hw->int_handle =  wq->vec_ptr;
 }
 
 static struct dma_async_tx_descriptor *
diff --git a/drivers/dma/idxd/idxd.h b/drivers/dma/idxd/idxd.h
index c64df197e724..43a216c42d25 100644
--- a/drivers/dma/idxd/idxd.h
+++ b/drivers/dma/idxd/idxd.h
@@ -59,6 +59,7 @@ enum idxd_wq_state {
 
 enum idxd_wq_flag {
 	WQ_FLAG_DEDICATED = 0,
+	WQ_FLAG_BLOCK_ON_FAULT,
 };
 
 enum idxd_wq_type {
@@ -86,10 +87,11 @@ enum idxd_op_type {
 enum idxd_complete_type {
 	IDXD_COMPLETE_NORMAL = 0,
 	IDXD_COMPLETE_ABORT,
+	IDXD_COMPLETE_DEV_FAIL,
 };
 
 struct idxd_wq {
-	void __iomem *dportal;
+	void __iomem *portal;
 	struct device conf_dev;
 	struct idxd_cdev idxd_cdev;
 	struct idxd_device *idxd;
@@ -145,6 +147,7 @@ enum idxd_device_state {
 enum idxd_device_flag {
 	IDXD_FLAG_CONFIGURABLE = 0,
 	IDXD_FLAG_CMD_RUNNING,
+	IDXD_FLAG_PASID_ENABLED,
 };
 
 struct idxd_device {
@@ -167,6 +170,9 @@ struct idxd_device {
 	struct idxd_wq *wqs;
 	struct idxd_engine *engines;
 
+	struct iommu_sva *sva;
+	unsigned int pasid;
+
 	int num_groups;
 
 	u32 msix_perm_offset;
@@ -214,11 +220,28 @@ struct idxd_desc {
 
 extern struct bus_type dsa_bus_type;
 
+extern bool support_enqcmd;
+
 static inline bool wq_dedicated(struct idxd_wq *wq)
 {
 	return test_bit(WQ_FLAG_DEDICATED, &wq->flags);
 }
 
+static inline bool wq_shared(struct idxd_wq *wq)
+{
+	return !test_bit(WQ_FLAG_DEDICATED, &wq->flags);
+}
+
+static inline bool device_pasid_enabled(struct idxd_device *idxd)
+{
+	return test_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
+}
+
+static inline bool device_swq_supported(struct idxd_device *idxd)
+{
+	return (support_enqcmd && device_pasid_enabled(idxd));
+}
+
 enum idxd_portal_prot {
 	IDXD_PORTAL_UNLIMITED = 0,
 	IDXD_PORTAL_LIMITED,
@@ -287,6 +310,7 @@ void idxd_device_reset(struct idxd_device *idxd);
 void idxd_device_cleanup(struct idxd_device *idxd);
 int idxd_device_config(struct idxd_device *idxd);
 void idxd_device_wqs_clear_state(struct idxd_device *idxd);
+void idxd_device_drain_pasid(struct idxd_device *idxd, int pasid);
 
 /* work queue control */
 int idxd_wq_alloc_resources(struct idxd_wq *wq);
@@ -297,6 +321,8 @@ void idxd_wq_drain(struct idxd_wq *wq);
 int idxd_wq_map_portal(struct idxd_wq *wq);
 void idxd_wq_unmap_portal(struct idxd_wq *wq);
 void idxd_wq_disable_cleanup(struct idxd_wq *wq);
+int idxd_wq_set_pasid(struct idxd_wq *wq, int pasid);
+int idxd_wq_disable_pasid(struct idxd_wq *wq);
 
 /* submission */
 int idxd_submit_desc(struct idxd_wq *wq, struct idxd_desc *desc);
diff --git a/drivers/dma/idxd/init.c b/drivers/dma/idxd/init.c
index 11e5ce168177..626401a71fdd 100644
--- a/drivers/dma/idxd/init.c
+++ b/drivers/dma/idxd/init.c
@@ -14,6 +14,8 @@
 #include <linux/io-64-nonatomic-lo-hi.h>
 #include <linux/device.h>
 #include <linux/idr.h>
+#include <linux/intel-svm.h>
+#include <linux/iommu.h>
 #include <uapi/linux/idxd.h>
 #include <linux/dmaengine.h>
 #include "../dmaengine.h"
@@ -26,6 +28,8 @@ MODULE_AUTHOR("Intel Corporation");
 
 #define DRV_NAME "idxd"
 
+bool support_enqcmd;
+
 static struct idr idxd_idrs[IDXD_TYPE_MAX];
 static struct mutex idxd_idr_lock;
 
@@ -53,6 +57,7 @@ static int idxd_setup_interrupts(struct idxd_device *idxd)
 	struct idxd_irq_entry *irq_entry;
 	int i, msixcnt;
 	int rc = 0;
+	union msix_perm mperm;
 
 	msixcnt = pci_msix_vec_count(pdev);
 	if (msixcnt < 0) {
@@ -131,6 +136,13 @@ static int idxd_setup_interrupts(struct idxd_device *idxd)
 
 	idxd_unmask_error_interrupts(idxd);
 
+	/* Setup MSIX permission table */
+	mperm.bits = 0;
+	mperm.pasid = idxd->pasid;
+	mperm.pasid_en = device_pasid_enabled(idxd);
+	for (i = 1; i < msixcnt; i++)
+		iowrite32(mperm.bits, idxd->reg_base + idxd->msix_perm_offset + i * 8);
+
 	return 0;
 
  err_no_irq:
@@ -260,8 +272,7 @@ static void idxd_read_caps(struct idxd_device *idxd)
 	}
 }
 
-static struct idxd_device *idxd_alloc(struct pci_dev *pdev,
-				      void __iomem * const *iomap)
+static struct idxd_device *idxd_alloc(struct pci_dev *pdev)
 {
 	struct device *dev = &pdev->dev;
 	struct idxd_device *idxd;
@@ -271,12 +282,45 @@ static struct idxd_device *idxd_alloc(struct pci_dev *pdev,
 		return NULL;
 
 	idxd->pdev = pdev;
-	idxd->reg_base = iomap[IDXD_MMIO_BAR];
 	spin_lock_init(&idxd->dev_lock);
 
 	return idxd;
 }
 
+static int idxd_enable_system_pasid(struct idxd_device *idxd)
+{
+	int flags;
+	unsigned int pasid;
+	struct iommu_sva *sva;
+
+	flags = SVM_FLAG_SUPERVISOR_MODE;
+
+	sva = iommu_sva_bind_device(&idxd->pdev->dev, NULL, &flags);
+	if (IS_ERR(sva)) {
+		dev_warn(&idxd->pdev->dev,
+			 "iommu sva bind failed: %ld\n", PTR_ERR(sva));
+		return PTR_ERR(sva);
+	}
+
+	pasid = iommu_sva_get_pasid(sva);
+	if (pasid == IOMMU_PASID_INVALID) {
+		iommu_sva_unbind_device(sva);
+		return -ENODEV;
+	}
+
+	idxd->sva = sva;
+	idxd->pasid = pasid;
+	dev_dbg(&idxd->pdev->dev, "system pasid: %u\n", pasid);
+	return 0;
+}
+
+static void idxd_disable_system_pasid(struct idxd_device *idxd)
+{
+
+	iommu_sva_unbind_device(idxd->sva);
+	idxd->sva = NULL;
+}
+
 static int idxd_probe(struct idxd_device *idxd)
 {
 	struct pci_dev *pdev = idxd->pdev;
@@ -287,6 +331,14 @@ static int idxd_probe(struct idxd_device *idxd)
 	idxd_device_init_reset(idxd);
 	dev_dbg(dev, "IDXD reset complete\n");
 
+	if (IS_ENABLED(CONFIG_INTEL_IDXD_SVM)) {
+		rc = idxd_enable_system_pasid(idxd);
+		if (rc < 0)
+			dev_warn(dev, "Failed to enable PASID. No SVA support: %d\n", rc);
+		else
+			set_bit(IDXD_FLAG_PASID_ENABLED, &idxd->flags);
+	}
+
 	idxd_read_caps(idxd);
 	idxd_read_table_offsets(idxd);
 
@@ -317,29 +369,29 @@ static int idxd_probe(struct idxd_device *idxd)
 	idxd_mask_error_interrupts(idxd);
 	idxd_mask_msix_vectors(idxd);
  err_setup:
+	if (device_pasid_enabled(idxd))
+		idxd_disable_system_pasid(idxd);
 	return rc;
 }
 
 static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 {
-	void __iomem * const *iomap;
 	struct device *dev = &pdev->dev;
 	struct idxd_device *idxd;
 	int rc;
-	unsigned int mask;
 
 	rc = pcim_enable_device(pdev);
 	if (rc)
 		return rc;
 
-	dev_dbg(dev, "Mapping BARs\n");
-	mask = (1 << IDXD_MMIO_BAR);
-	rc = pcim_iomap_regions(pdev, mask, DRV_NAME);
-	if (rc)
-		return rc;
+	dev_dbg(dev, "Alloc IDXD context\n");
+	idxd = idxd_alloc(pdev);
+	if (!idxd)
+		return -ENOMEM;
 
-	iomap = pcim_iomap_table(pdev);
-	if (!iomap)
+	dev_dbg(dev, "Mapping BARs\n");
+	idxd->reg_base = pcim_iomap(pdev, IDXD_MMIO_BAR, 0);
+	if (!idxd->reg_base)
 		return -ENOMEM;
 
 	dev_dbg(dev, "Set DMA masks\n");
@@ -355,11 +407,6 @@ static int idxd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	if (rc)
 		return rc;
 
-	dev_dbg(dev, "Alloc IDXD context\n");
-	idxd = idxd_alloc(pdev, iomap);
-	if (!idxd)
-		return -ENOMEM;
-
 	idxd_set_type(idxd);
 
 	dev_dbg(dev, "Set PCI master\n");
@@ -447,6 +494,8 @@ static void idxd_remove(struct pci_dev *pdev)
 	dev_dbg(&pdev->dev, "%s called\n", __func__);
 	idxd_cleanup_sysfs(idxd);
 	idxd_shutdown(pdev);
+	if (device_pasid_enabled(idxd))
+		idxd_disable_system_pasid(idxd);
 	mutex_lock(&idxd_idr_lock);
 	idr_remove(&idxd_idrs[idxd->type], idxd->id);
 	mutex_unlock(&idxd_idr_lock);
@@ -465,7 +514,7 @@ static int __init idxd_init_module(void)
 	int err, i;
 
 	/*
-	 * If the CPU does not support write512, there's no point in
+	 * If the CPU does not support MOVDIR64B or ENQCMDS, there's no point in
 	 * enumerating the device. We can not utilize it.
 	 */
 	if (!boot_cpu_has(X86_FEATURE_MOVDIR64B)) {
@@ -473,8 +522,10 @@ static int __init idxd_init_module(void)
 		return -ENODEV;
 	}
 
-	pr_info("%s: Intel(R) Accelerator Devices Driver %s\n",
-		DRV_NAME, IDXD_DRIVER_VERSION);
+	if (!boot_cpu_has(X86_FEATURE_ENQCMD))
+		pr_warn("Platform does not have ENQCMD(S) support.\n");
+	else
+		support_enqcmd = true;
 
 	mutex_init(&idxd_idr_lock);
 	for (i = 0; i < IDXD_TYPE_MAX; i++)
diff --git a/drivers/dma/idxd/registers.h b/drivers/dma/idxd/registers.h
index a39e7ae6b3d9..a0df4f3fe1fb 100644
--- a/drivers/dma/idxd/registers.h
+++ b/drivers/dma/idxd/registers.h
@@ -333,4 +333,18 @@ union wqcfg {
 	};
 	u32 bits[8];
 } __packed;
+
+/*
+ * This macro calculates the offset into the WQCFG register
+ * idxd - struct idxd *
+ * n - wq id
+ * ofs - the index of the 32b dword for the config register
+ *
+ * The WQCFG register block is divided into groups per each wq. The n index
+ * allows us to move to the register group that's for that particular wq.
+ * Each register is 32bits. The ofs gives us the number of register to access.
+ */
+#define WQCFG_OFFSET(idxd_dev, n, ofs) ((idxd_dev)->wqcfg_offset +\
+					(n) * sizeof(union wqcfg) +\
+					sizeof(u32) * (ofs))
 #endif
diff --git a/drivers/dma/idxd/submit.c b/drivers/dma/idxd/submit.c
index 156a1ee233aa..efca5d8468a6 100644
--- a/drivers/dma/idxd/submit.c
+++ b/drivers/dma/idxd/submit.c
@@ -11,11 +11,22 @@
 static struct idxd_desc *__get_desc(struct idxd_wq *wq, int idx, int cpu)
 {
 	struct idxd_desc *desc;
+	struct idxd_device *idxd = wq->idxd;
 
 	desc = wq->descs[idx];
 	memset(desc->hw, 0, sizeof(struct dsa_hw_desc));
 	memset(desc->completion, 0, sizeof(struct dsa_completion_record));
 	desc->cpu = cpu;
+
+	if (device_pasid_enabled(idxd))
+		desc->hw->pasid = idxd->pasid;
+
+	/*
+	 * Descriptor completion vectors are 1-8 for MSIX. We will round
+	 * robin through the 8 vectors.
+	 */
+	wq->vec_ptr = (wq->vec_ptr % idxd->num_wq_irqs) + 1;
+	desc->hw->int_handle = wq->vec_ptr;
 	return desc;
 }
 
@@ -70,18 +81,32 @@ int idxd_submit_desc(struct idxd_wq *wq, struct idxd_desc *desc)
 	struct idxd_device *idxd = wq->idxd;
 	int vec = desc->hw->int_handle;
 	void __iomem *portal;
+	int rc;
 
 	if (idxd->state != IDXD_DEV_ENABLED)
 		return -EIO;
 
-	portal = wq->dportal + idxd_get_wq_portal_offset(IDXD_PORTAL_UNLIMITED);
+	portal = wq->portal + idxd_get_wq_portal_offset(IDXD_PORTAL_LIMITED);
+
 	/*
-	 * The wmb() flushes writes to coherent DMA data before possibly
-	 * triggering a DMA read. The wmb() is necessary even on UP because
-	 * the recipient is a device.
+	 * The wmb() flushes writes to coherent DMA data before
+	 * possibly triggering a DMA read. The wmb() is necessary
+	 * even on UP because the recipient is a device.
 	 */
 	wmb();
-	iosubmit_cmds512(portal, desc->hw, 1);
+	if (wq_dedicated(wq)) {
+		iosubmit_cmds512(portal, desc->hw, 1);
+	} else {
+		/*
+		 * It's not likely that we would receive queue full rejection
+		 * since the descriptor allocation gates at wq size. If we
+		 * receive a -EAGAIN, that means something went wrong such as the
+		 * device is not accepting descriptor at all.
+		 */
+		rc = enqcmds(portal, desc->hw);
+		if (rc < 0)
+			return rc;
+	}
 
 	/*
 	 * Pending the descriptor to the lockless list for the irq_entry
diff --git a/drivers/dma/idxd/sysfs.c b/drivers/dma/idxd/sysfs.c
index 07a5db06a29a..6d292eb79bf3 100644
--- a/drivers/dma/idxd/sysfs.c
+++ b/drivers/dma/idxd/sysfs.c
@@ -175,6 +175,30 @@ static int idxd_config_bus_probe(struct device *dev)
 			return -EINVAL;
 		}
 
+		/* Shared WQ checks */
+		if (wq_shared(wq)) {
+			if (!device_swq_supported(idxd)) {
+				dev_warn(dev,
+					 "PASID not enabled and shared WQ.\n");
+				mutex_unlock(&wq->wq_lock);
+				return -ENXIO;
+			}
+			/*
+			 * Shared wq with the threshold set to 0 means the user
+			 * did not set the threshold or transitioned from a
+			 * dedicated wq but did not set threshold. A value
+			 * of 0 would effectively disable the shared wq. The
+			 * driver does not allow a value of 0 to be set for
+			 * threshold via sysfs.
+			 */
+			if (wq->threshold == 0) {
+				dev_warn(dev,
+					 "Shared WQ and threshold 0.\n");
+				mutex_unlock(&wq->wq_lock);
+				return -EINVAL;
+			}
+		}
+
 		rc = idxd_wq_alloc_resources(wq);
 		if (rc < 0) {
 			mutex_unlock(&wq->wq_lock);
@@ -875,6 +899,8 @@ static ssize_t wq_mode_store(struct device *dev,
 	if (sysfs_streq(buf, "dedicated")) {
 		set_bit(WQ_FLAG_DEDICATED, &wq->flags);
 		wq->threshold = 0;
+	} else if (sysfs_streq(buf, "shared") && device_swq_supported(idxd)) {
+		clear_bit(WQ_FLAG_DEDICATED, &wq->flags);
 	} else {
 		return -EINVAL;
 	}
@@ -973,6 +999,87 @@ static ssize_t wq_priority_store(struct device *dev,
 static struct device_attribute dev_attr_wq_priority =
 		__ATTR(priority, 0644, wq_priority_show, wq_priority_store);
 
+static ssize_t wq_block_on_fault_show(struct device *dev,
+				      struct device_attribute *attr, char *buf)
+{
+	struct idxd_wq *wq = container_of(dev, struct idxd_wq, conf_dev);
+
+	return sprintf(buf, "%u\n",
+		       test_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags));
+}
+
+static ssize_t wq_block_on_fault_store(struct device *dev,
+				       struct device_attribute *attr,
+				       const char *buf, size_t count)
+{
+	struct idxd_wq *wq = container_of(dev, struct idxd_wq, conf_dev);
+	struct idxd_device *idxd = wq->idxd;
+	bool bof;
+	int rc;
+
+	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags))
+		return -EPERM;
+
+	if (wq->state != IDXD_WQ_DISABLED)
+		return -ENXIO;
+
+	rc = kstrtobool(buf, &bof);
+	if (rc < 0)
+		return rc;
+
+	if (bof)
+		set_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags);
+	else
+		clear_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags);
+
+	return count;
+}
+
+static struct device_attribute dev_attr_wq_block_on_fault =
+		__ATTR(block_on_fault, 0644, wq_block_on_fault_show,
+		       wq_block_on_fault_store);
+
+static ssize_t wq_threshold_show(struct device *dev,
+				 struct device_attribute *attr, char *buf)
+{
+	struct idxd_wq *wq = container_of(dev, struct idxd_wq, conf_dev);
+
+	return sprintf(buf, "%u\n", wq->threshold);
+}
+
+static ssize_t wq_threshold_store(struct device *dev,
+				  struct device_attribute *attr,
+				  const char *buf, size_t count)
+{
+	struct idxd_wq *wq = container_of(dev, struct idxd_wq, conf_dev);
+	struct idxd_device *idxd = wq->idxd;
+	unsigned int val;
+	int rc;
+
+	rc = kstrtouint(buf, 0, &val);
+	if (rc < 0)
+		return -EINVAL;
+
+	if (val > wq->size || val <= 0)
+		return -EINVAL;
+
+	if (!test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags))
+		return -EPERM;
+
+	if (wq->state != IDXD_WQ_DISABLED)
+		return -ENXIO;
+
+	if (test_bit(WQ_FLAG_DEDICATED, &wq->flags))
+		return -EINVAL;
+
+	wq->threshold = val;
+
+	return count;
+}
+
+static struct device_attribute dev_attr_wq_threshold =
+		__ATTR(threshold, 0644, wq_threshold_show, wq_threshold_store);
+
 static ssize_t wq_type_show(struct device *dev,
 			    struct device_attribute *attr, char *buf)
 {
@@ -1044,6 +1151,13 @@ static ssize_t wq_name_store(struct device *dev,
 	if (strlen(buf) > WQ_NAME_SIZE || strlen(buf) == 0)
 		return -EINVAL;
 
+	/*
+	 * This is temporarily placed here until we have SVM support for
+	 * dmaengine.
+	 */
+	if (wq->type == IDXD_WQT_KERNEL && device_pasid_enabled(wq->idxd))
+		return -EOPNOTSUPP;
+
 	memset(wq->name, 0, WQ_NAME_SIZE + 1);
 	strncpy(wq->name, buf, WQ_NAME_SIZE);
 	strreplace(wq->name, '\n', '\0');
@@ -1154,6 +1268,8 @@ static struct attribute *idxd_wq_attributes[] = {
 	&dev_attr_wq_mode.attr,
 	&dev_attr_wq_size.attr,
 	&dev_attr_wq_priority.attr,
+	&dev_attr_wq_block_on_fault.attr,
+	&dev_attr_wq_threshold.attr,
 	&dev_attr_wq_type.attr,
 	&dev_attr_wq_name.attr,
 	&dev_attr_wq_cdev_minor.attr,
@@ -1305,6 +1421,16 @@ static ssize_t clients_show(struct device *dev,
 }
 static DEVICE_ATTR_RO(clients);
 
+static ssize_t pasid_enabled_show(struct device *dev,
+				  struct device_attribute *attr, char *buf)
+{
+	struct idxd_device *idxd =
+		container_of(dev, struct idxd_device, conf_dev);
+
+	return sprintf(buf, "%u\n", device_pasid_enabled(idxd));
+}
+static DEVICE_ATTR_RO(pasid_enabled);
+
 static ssize_t state_show(struct device *dev,
 			  struct device_attribute *attr, char *buf)
 {
@@ -1424,6 +1550,7 @@ static struct attribute *idxd_device_attributes[] = {
 	&dev_attr_gen_cap.attr,
 	&dev_attr_configurable.attr,
 	&dev_attr_clients.attr,
+	&dev_attr_pasid_enabled.attr,
 	&dev_attr_state.attr,
 	&dev_attr_errors.attr,
 	&dev_attr_max_tokens.attr,
-- 
2.26.2

[PATCH v7 4/5] dmaengine: idxd: Clean up descriptors with fault error

[Dave Jiang]

Add code to "complete" a descriptor when the descriptor or its completion
address hit a fault error when SVA mode is being used. This error can be
triggered due to bad programming by the user. A lock is introduced in order
to protect the descriptor completion lists since the fault handler will run
from the system work queue after being scheduled in the interrupt handler.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
---
 drivers/dma/idxd/idxd.h |   5 ++
 drivers/dma/idxd/init.c |   1 +
 drivers/dma/idxd/irq.c  | 146 ++++++++++++++++++++++++++++++++++++----
 3 files changed, 140 insertions(+), 12 deletions(-)

diff --git a/drivers/dma/idxd/idxd.h b/drivers/dma/idxd/idxd.h
index 43a216c42d25..b64b6266ca97 100644
--- a/drivers/dma/idxd/idxd.h
+++ b/drivers/dma/idxd/idxd.h
@@ -34,6 +34,11 @@ struct idxd_irq_entry {
 	int id;
 	struct llist_head pending_llist;
 	struct list_head work_list;
+	/*
+	 * Lock to protect access between irq thread process descriptor
+	 * and irq thread processing error descriptor.
+	 */
+	spinlock_t list_lock;
 };
 
 struct idxd_group {
diff --git a/drivers/dma/idxd/init.c b/drivers/dma/idxd/init.c
index 626401a71fdd..1bb7637b02eb 100644
--- a/drivers/dma/idxd/init.c
+++ b/drivers/dma/idxd/init.c
@@ -97,6 +97,7 @@ static int idxd_setup_interrupts(struct idxd_device *idxd)
 	for (i = 0; i < msixcnt; i++) {
 		idxd->irq_entries[i].id = i;
 		idxd->irq_entries[i].idxd = idxd;
+		spin_lock_init(&idxd->irq_entries[i].list_lock);
 	}
 
 	msix = &idxd->msix_entries[0];
diff --git a/drivers/dma/idxd/irq.c b/drivers/dma/idxd/irq.c
index 17a65a13fb64..593a2f6ed16c 100644
--- a/drivers/dma/idxd/irq.c
+++ b/drivers/dma/idxd/irq.c
@@ -11,6 +11,24 @@
 #include "idxd.h"
 #include "registers.h"
 
+enum irq_work_type {
+	IRQ_WORK_NORMAL = 0,
+	IRQ_WORK_PROCESS_FAULT,
+};
+
+struct idxd_fault {
+	struct work_struct work;
+	u64 addr;
+	struct idxd_device *idxd;
+};
+
+static int irq_process_work_list(struct idxd_irq_entry *irq_entry,
+				 enum irq_work_type wtype,
+				 int *processed, u64 data);
+static int irq_process_pending_llist(struct idxd_irq_entry *irq_entry,
+				     enum irq_work_type wtype,
+				     int *processed, u64 data);
+
 static void idxd_device_reinit(struct work_struct *work)
 {
 	struct idxd_device *idxd = container_of(work, struct idxd_device, work);
@@ -44,6 +62,46 @@ static void idxd_device_reinit(struct work_struct *work)
 	idxd_device_wqs_clear_state(idxd);
 }
 
+static void idxd_device_fault_work(struct work_struct *work)
+{
+	struct idxd_fault *fault = container_of(work, struct idxd_fault, work);
+	struct idxd_irq_entry *ie;
+	int i;
+	int processed;
+	int irqcnt = fault->idxd->num_wq_irqs + 1;
+
+	for (i = 1; i < irqcnt; i++) {
+		ie = &fault->idxd->irq_entries[i];
+		irq_process_work_list(ie, IRQ_WORK_PROCESS_FAULT,
+				      &processed, fault->addr);
+		if (processed)
+			break;
+
+		irq_process_pending_llist(ie, IRQ_WORK_PROCESS_FAULT,
+					  &processed, fault->addr);
+		if (processed)
+			break;
+	}
+
+	kfree(fault);
+}
+
+static int idxd_device_schedule_fault_process(struct idxd_device *idxd,
+					      u64 fault_addr)
+{
+	struct idxd_fault *fault;
+
+	fault = kmalloc(sizeof(*fault), GFP_ATOMIC);
+	if (!fault)
+		return -ENOMEM;
+
+	fault->addr = fault_addr;
+	fault->idxd = idxd;
+	INIT_WORK(&fault->work, idxd_device_fault_work);
+	queue_work(idxd->wq, &fault->work);
+	return 0;
+}
+
 irqreturn_t idxd_irq_handler(int vec, void *data)
 {
 	struct idxd_irq_entry *irq_entry = data;
@@ -125,6 +183,15 @@ irqreturn_t idxd_misc_thread(int vec, void *data)
 	if (!err)
 		goto out;
 
+	/*
+	 * This case should rarely happen and typically is due to software
+	 * programming error by the driver.
+	 */
+	if (idxd->sw_err.valid &&
+	    idxd->sw_err.desc_valid &&
+	    idxd->sw_err.fault_addr)
+		idxd_device_schedule_fault_process(idxd, idxd->sw_err.fault_addr);
+
 	gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
 	if (gensts.state == IDXD_DEVICE_STATE_HALT) {
 		idxd->state = IDXD_DEV_HALTED;
@@ -152,57 +219,110 @@ irqreturn_t idxd_misc_thread(int vec, void *data)
 	return IRQ_HANDLED;
 }
 
+static bool process_fault(struct idxd_desc *desc, u64 fault_addr)
+{
+	/*
+	 * Completion address can be bad as well. Check fault address match for descriptor
+	 * and completion address.
+	 */
+	if ((u64)desc->hw == fault_addr ||
+	    (u64)desc->completion == fault_addr) {
+		idxd_dma_complete_txd(desc, IDXD_COMPLETE_DEV_FAIL);
+		return true;
+	}
+
+	return false;
+}
+
+static bool complete_desc(struct idxd_desc *desc)
+{
+	if (desc->completion->status) {
+		idxd_dma_complete_txd(desc, IDXD_COMPLETE_NORMAL);
+		return true;
+	}
+
+	return false;
+}
+
 static int irq_process_pending_llist(struct idxd_irq_entry *irq_entry,
-				     int *processed)
+				     enum irq_work_type wtype,
+				     int *processed, u64 data)
 {
 	struct idxd_desc *desc, *t;
 	struct llist_node *head;
 	int queued = 0;
+	bool completed = false;
+	unsigned long flags;
 
 	*processed = 0;
 	head = llist_del_all(&irq_entry->pending_llist);
 	if (!head)
-		return 0;
+		goto out;
 
 	llist_for_each_entry_safe(desc, t, head, llnode) {
-		if (desc->completion->status) {
-			idxd_dma_complete_txd(desc, IDXD_COMPLETE_NORMAL);
+		if (wtype == IRQ_WORK_NORMAL)
+			completed = complete_desc(desc);
+		else if (wtype == IRQ_WORK_PROCESS_FAULT)
+			completed = process_fault(desc, data);
+
+		if (completed) {
 			idxd_free_desc(desc->wq, desc);
 			(*processed)++;
+			if (wtype == IRQ_WORK_PROCESS_FAULT)
+				break;
 		} else {
-			list_add_tail(&desc->list, &irq_entry->work_list);
+			spin_lock_irqsave(&irq_entry->list_lock, flags);
+			list_add_tail(&desc->list,
+				      &irq_entry->work_list);
+			spin_unlock_irqrestore(&irq_entry->list_lock, flags);
 			queued++;
 		}
 	}
 
+ out:
 	return queued;
 }
 
 static int irq_process_work_list(struct idxd_irq_entry *irq_entry,
-				 int *processed)
+				 enum irq_work_type wtype,
+				 int *processed, u64 data)
 {
 	struct list_head *node, *next;
 	int queued = 0;
+	bool completed = false;
+	unsigned long flags;
 
 	*processed = 0;
+	spin_lock_irqsave(&irq_entry->list_lock, flags);
 	if (list_empty(&irq_entry->work_list))
-		return 0;
+		goto out;
 
 	list_for_each_safe(node, next, &irq_entry->work_list) {
 		struct idxd_desc *desc =
 			container_of(node, struct idxd_desc, list);
 
-		if (desc->completion->status) {
+		spin_unlock_irqrestore(&irq_entry->list_lock, flags);
+		if (wtype == IRQ_WORK_NORMAL)
+			completed = complete_desc(desc);
+		else if (wtype == IRQ_WORK_PROCESS_FAULT)
+			completed = process_fault(desc, data);
+
+		if (completed) {
+			spin_lock_irqsave(&irq_entry->list_lock, flags);
 			list_del(&desc->list);
-			/* process and callback */
-			idxd_dma_complete_txd(desc, IDXD_COMPLETE_NORMAL);
+			spin_unlock_irqrestore(&irq_entry->list_lock, flags);
 			idxd_free_desc(desc->wq, desc);
 			(*processed)++;
+			if (wtype == IRQ_WORK_PROCESS_FAULT)
+				return queued;
 		} else {
 			queued++;
 		}
+		spin_lock_irqsave(&irq_entry->list_lock, flags);
 	}
 
+ out:
+	spin_unlock_irqrestore(&irq_entry->list_lock, flags);
 	return queued;
 }
 
@@ -230,12 +350,14 @@ static int idxd_desc_process(struct idxd_irq_entry *irq_entry)
 	 * 5. Repeat until no more descriptors.
 	 */
 	do {
-		rc = irq_process_work_list(irq_entry, &processed);
+		rc = irq_process_work_list(irq_entry, IRQ_WORK_NORMAL,
+					   &processed, 0);
 		total += processed;
 		if (rc != 0)
 			continue;
 
-		rc = irq_process_pending_llist(irq_entry, &processed);
+		rc = irq_process_pending_llist(irq_entry, IRQ_WORK_NORMAL,
+					       &processed, 0);
 		total += processed;
 	} while (rc != 0);
 
-- 
2.26.2

[PATCH v7 5/5] dmaengine: idxd: Add ABI documentation for shared wq

[Dave Jiang]

Add the sysfs attribute bits in ABI/stable for shared wq support.

Signed-off-by: Jing Lin <jing.lin@intel.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
---
 Documentation/ABI/stable/sysfs-driver-dma-idxd | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/Documentation/ABI/stable/sysfs-driver-dma-idxd b/Documentation/ABI/stable/sysfs-driver-dma-idxd
index b44183880935..42d3dc03ffea 100644
--- a/Documentation/ABI/stable/sysfs-driver-dma-idxd
+++ b/Documentation/ABI/stable/sysfs-driver-dma-idxd
@@ -77,6 +77,13 @@ Contact:        dmaengine@vger.kernel.org
 Description:    The operation capability bit mask specify the operation types
 		supported by the this device.
 
+What:		/sys/bus/dsa/devices/dsa<m>/pasid_enabled
+Date:		Sep 17, 2020
+KernelVersion:	5.10.0
+Contact:	dmaengine@vger.kernel.org
+Description:	To indicate if PASID (process address space identifier) is
+		enabled or not for this device.
+
 What:           /sys/bus/dsa/devices/dsa<m>/state
 Date:           Oct 25, 2019
 KernelVersion:  5.6.0
@@ -122,6 +129,13 @@ KernelVersion:	5.10.0
 Contact:	dmaengine@vger.kernel.org
 Description:	The last executed device administrative command's status/error.
 
+What:		/sys/bus/dsa/devices/wq<m>.<n>/block_on_fault
+Date:		Sept 17, 2020
+KernelVersion:	5.10.0
+Contact:	dmaengine@vger.kernel.org
+Description:	To indicate block on fault is allowed or not for the work queue
+		to support on demand paging.
+
 What:           /sys/bus/dsa/devices/wq<m>.<n>/group_id
 Date:           Oct 25, 2019
 KernelVersion:  5.6.0
-- 
2.26.2

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Vinod Koul]

On 05-10-20, 08:11, Dave Jiang wrote:

> == Background ==
> A typical DMA device requires the driver to translate application buffers to hardware addresses,
> and a kernel-user transition to notify the hardware of new work. Shared Virtual Addressing (SVA)
> allows the processor and device to use the same virtual addresses without requiring software to
> translate between the address spaces. ENQCMD is a new instruction on Intel Platforms that allows
> user applications to directly notify hardware of new work, much like how doorbells are used in
> some hardware, but it carries a payload along with it. ENQCMDS is the supervisor version (ring0)
> of ENQCMD.
> 
> == ENQCMDS ==
> Introduce enqcmds(), a helper funciton that copies an input payload to a 64B aligned
> destination and confirms whether the payload was accepted by the device or not.
> enqcmds() wraps the new ENQCMDS CPU instruction. The ENQCMDS is a ring 0 CPU instruction that
> performs similar to the ENQCMD instruction. Descriptor submission must use ENQCMD(S) for shared
> workqueues (swq) on an Intel DSA device. 
> 
> == Shared WQ support ==
> Introduce shared workqueue (swq) support for the idxd driver. The current idxd driver contains
> dedicated workqueue (dwq) support only. A dwq accepts descriptors from a MOVDIR64B instruction.
> MOVDIR64B is a posted instruction on the PCIe bus, it does not wait for any response from the
> device. If the wq is full, submitted descriptors are dropped. A swq utilizes the ENQCMDS in
> ring 0, which is a non-posted instruction. The zero flag would be set to 1 if the device rejects
> the descriptor or if the wq is full. A swq can be shared between multiple users
> (kernel or userspace) due to not having to keep track of the wq full condition for submission.
> A swq requires PASID and can only run with SVA support. 
> 
> == IDXD SVA support ==
> Add utilization of PASID to support Shared Virtual Addressing (SVA). With PASID support,
> the descriptors can be programmed with host virtual address (HVA) rather than IOVA.
> The hardware will work with the IOMMU in fulfilling page requests. With SVA support,
> a user app using the char device interface can now submit descriptors without having to pin the
> virtual memory range it wants to DMA in its own address space. 
> 
> The series does not add SVA support for the dmaengine subsystem. That support is coming at a
> later time.

Applied, thanks

-- 
~Vinod

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Borislav Petkov]

On Wed, Oct 07, 2020 at 12:31:32PM +0530, Vinod Koul wrote:
> Applied, thanks

I'm tired of repeating what you should've done - your branch doesn't
even build. How did you test it?

Also, what happens if Linus merges your branch first, before tip?

Oh boy.

In file included from ./arch/x86/include/asm/tsc.h:9,
                 from ./arch/x86/include/asm/timex.h:6,
                 from ./include/linux/timex.h:65,
                 from ./include/linux/time32.h:13,
                 from ./include/linux/time.h:73,
                 from ./include/linux/stat.h:19,
                 from ./include/linux/module.h:13,
                 from drivers/dma/idxd/init.c:5:
drivers/dma/idxd/init.c: In function ‘idxd_init_module’:
drivers/dma/idxd/init.c:526:20: error: ‘X86_FEATURE_ENQCMD’ undeclared (first use in this function); did you mean ‘X86_FEATURE_PCID’?
  526 |  if (!boot_cpu_has(X86_FEATURE_ENQCMD))
      |                    ^~~~~~~~~~~~~~~~~~
./arch/x86/include/asm/cpufeature.h:118:24: note: in definition of macro ‘cpu_has’
  118 |  (__builtin_constant_p(bit) && REQUIRED_MASK_BIT_SET(bit) ? 1 : \
      |                        ^~~
drivers/dma/idxd/init.c:526:7: note: in expansion of macro ‘boot_cpu_has’
  526 |  if (!boot_cpu_has(X86_FEATURE_ENQCMD))
      |       ^~~~~~~~~~~~
drivers/dma/idxd/init.c:526:20: note: each undeclared identifier is reported only once for each function it appears in
  526 |  if (!boot_cpu_has(X86_FEATURE_ENQCMD))
      |                    ^~~~~~~~~~~~~~~~~~
./arch/x86/include/asm/cpufeature.h:118:24: note: in definition of macro ‘cpu_has’
  118 |  (__builtin_constant_p(bit) && REQUIRED_MASK_BIT_SET(bit) ? 1 : \
      |                        ^~~
drivers/dma/idxd/init.c:526:7: note: in expansion of macro ‘boot_cpu_has’
  526 |  if (!boot_cpu_has(X86_FEATURE_ENQCMD))
      |       ^~~~~~~~~~~~
make[3]: *** [scripts/Makefile.build:283: drivers/dma/idxd/init.o] Error 1
make[3]: *** Waiting for unfinished jobs....
make[2]: *** [scripts/Makefile.build:500: drivers/dma/idxd] Error 2
make[2]: *** Waiting for unfinished jobs....
make[1]: *** [scripts/Makefile.build:500: drivers/dma] Error 2
make[1]: *** Waiting for unfinished jobs....
make: *** [Makefile:1788: drivers] Error 2
make: *** Waiting for unfinished jobs....

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Vinod Koul]

On 07-10-20, 10:48, Borislav Petkov wrote:
> On Wed, Oct 07, 2020 at 12:31:32PM +0530, Vinod Koul wrote:
> > Applied, thanks
> 
> I'm tired of repeating what you should've done - your branch doesn't
> even build. How did you test it?

Right my build failed for x86 and I have dropped these now. I would have
expected the dependency to be a signed tag to be cross merged when I was
asked to merge this.

-- 
~Vinod

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Borislav Petkov]

On Wed, Oct 07, 2020 at 03:23:13PM +0530, Vinod Koul wrote:
> Right my build failed for x86 and I have dropped these now. I would have
> expected the dependency to be a signed tag to be cross merged when I was
> asked to merge this.

I can give you a signed tag is you prefer but that's usually not
necessary. You can simply merge tip's x86/pasid branch, then apply those
ontop and test.

HTH.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Vinod Koul]

On 07-10-20, 12:04, Borislav Petkov wrote:
> On Wed, Oct 07, 2020 at 03:23:13PM +0530, Vinod Koul wrote:
> > Right my build failed for x86 and I have dropped these now. I would have
> > expected the dependency to be a signed tag to be cross merged when I was
> > asked to merge this.
> 
> I can give you a signed tag is you prefer but that's usually not

That would be better, signed tags are preferred

> necessary. You can simply merge tip's x86/pasid branch, then apply those
> ontop and test.

While at it, it would be good if x86 patches of this series come from
your tree, that makes more sense if we are doing a cross merge

Thanks
-- 
~Vinod

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Borislav Petkov]

On Wed, Oct 07, 2020 at 08:27:33PM +0530, Vinod Koul wrote:
> That would be better, signed tags are preferred

...

> While at it, it would be good if x86 patches of this series come from
> your tree, that makes more sense if we are doing a cross merge

All done, here it is:

https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/tag/?h=x86_pasid_for_5.10

That's going to be the tag I send to Linus next week too. I'll send it
on Monday when the merge window opens so that he can merge it before
your branch.

HTH.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH v7 0/5] Add shared workqueue support for idxd driver

[Dave Jiang]

On 10/7/2020 12:01 AM, Vinod Koul wrote:
> On 05-10-20, 08:11, Dave Jiang wrote:
> 
>> == Background ==
>> A typical DMA device requires the driver to translate application buffers to hardware addresses,
>> and a kernel-user transition to notify the hardware of new work. Shared Virtual Addressing (SVA)
>> allows the processor and device to use the same virtual addresses without requiring software to
>> translate between the address spaces. ENQCMD is a new instruction on Intel Platforms that allows
>> user applications to directly notify hardware of new work, much like how doorbells are used in
>> some hardware, but it carries a payload along with it. ENQCMDS is the supervisor version (ring0)
>> of ENQCMD.
>>
>> == ENQCMDS ==
>> Introduce enqcmds(), a helper funciton that copies an input payload to a 64B aligned
>> destination and confirms whether the payload was accepted by the device or not.
>> enqcmds() wraps the new ENQCMDS CPU instruction. The ENQCMDS is a ring 0 CPU instruction that
>> performs similar to the ENQCMD instruction. Descriptor submission must use ENQCMD(S) for shared
>> workqueues (swq) on an Intel DSA device.
>>
>> == Shared WQ support ==
>> Introduce shared workqueue (swq) support for the idxd driver. The current idxd driver contains
>> dedicated workqueue (dwq) support only. A dwq accepts descriptors from a MOVDIR64B instruction.
>> MOVDIR64B is a posted instruction on the PCIe bus, it does not wait for any response from the
>> device. If the wq is full, submitted descriptors are dropped. A swq utilizes the ENQCMDS in
>> ring 0, which is a non-posted instruction. The zero flag would be set to 1 if the device rejects
>> the descriptor or if the wq is full. A swq can be shared between multiple users
>> (kernel or userspace) due to not having to keep track of the wq full condition for submission.
>> A swq requires PASID and can only run with SVA support.
>>
>> == IDXD SVA support ==
>> Add utilization of PASID to support Shared Virtual Addressing (SVA). With PASID support,
>> the descriptors can be programmed with host virtual address (HVA) rather than IOVA.
>> The hardware will work with the IOMMU in fulfilling page requests. With SVA support,
>> a user app using the char device interface can now submit descriptors without having to pin the
>> virtual memory range it wants to DMA in its own address space.
>>
>> The series does not add SVA support for the dmaengine subsystem. That support is coming at a
>> later time.
> 
> Applied, thanks
> 

Thanks Vinod!