[PATCH v7 9/9] libnvdimm: Add DMA based blk-mq pmem driver

[Dave Jiang <dave.jiang@intel.com>]

Add a DMA supported blk-mq driver for pmem. This provides significant
CPU utilization reduction at the cost of some increased latency and
bandwidth reduction in some cases.  By default the current cpu-copy based
pmem driver will load, but this driver can be manually selected with a
modprobe configuration. The pmem driver will be using blk-mq with DMA
through the dmaengine API.

Numbers below are measured against pmem simulated via DRAM using
memmap=NN!SS.  DMA engine used is the ioatdma on Intel Skylake Xeon
platform.  Keep in mind the performance for persistent memory
will differ.
Fio 2.21 was used.

64k: 1 task queuedepth=1
CPU Read:  7631 MB/s  99.7% CPU    DMA Read: 2415 MB/s  54% CPU
CPU Write: 3552 MB/s  100% CPU     DMA Write 2173 MB/s  54% CPU

64k: 16 tasks queuedepth=16
CPU Read: 36800 MB/s  1593% CPU    DMA Read:  29100 MB/s  607% CPU
CPU Write 20900 MB/s  1589% CPU    DMA Write: 23400 MB/s  585% CPU

2M: 1 task queuedepth=1
CPU Read:  6013 MB/s  99.3% CPU    DMA Read:  7986 MB/s  59.3% CPU
CPU Write: 3579 MB/s  100% CPU     DMA Write: 5211 MB/s  58.3% CPU

2M: 16 tasks queuedepth=16
CPU Read:  18100 MB/s 1588% CPU    DMA Read:  21300 MB/s 180.9% CPU
CPU Write: 14100 MB/s 1594% CPU    DMA Write: 20400 MB/s 446.9% CPU

Also, due to a significant portion of the code being shared with the
pmem driver, the common code are broken out into a kernel module
called pmem_core to be shared between the two drivers.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
---
 drivers/nvdimm/Kconfig    |   18 ++
 drivers/nvdimm/Makefile   |    3 
 drivers/nvdimm/pmem.h     |    1 
 drivers/nvdimm/pmem_dma.c |  475 +++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 497 insertions(+)
 create mode 100644 drivers/nvdimm/pmem_dma.c

diff --git a/drivers/nvdimm/Kconfig b/drivers/nvdimm/Kconfig
index 01fe9e8..e0a4589 100644
--- a/drivers/nvdimm/Kconfig
+++ b/drivers/nvdimm/Kconfig
@@ -40,6 +40,24 @@ config BLK_DEV_PMEM
 
 	  Say Y if you want to use an NVDIMM
 
+config BLK_DEV_PMEM_DMA
+	tristate "PMEM: Persistent memory block device with DMA support"
+	depends on DMA_ENGINE
+	depends on BLK_DEV_PMEM=m || !BLK_DEV_PMEM
+	default LIBNVDIMM
+	select BLK_DEV_PMEM_CORE
+	help
+	  This driver utilizes DMA engines provided by the platform to
+	  help offload the data copying. The desire for this driver is to
+	  reduce CPU utilization with some sacrifice in latency and
+	  performance. Initial benchmarks on DRAM showed that as low
+	  as about 30% of the CPU was used for DMA vs doing CPU copy with
+	  some reduction in throughput. Be aware that when DAX is used,
+	  DMA is bypassed and only CPU is used. The path for using DMA is
+	  only through the normal block device path.
+
+	  Say Y if you want to use an NVDIMM
+
 config ND_BLK
 	tristate "BLK: Block data window (aperture) device support"
 	default LIBNVDIMM
diff --git a/drivers/nvdimm/Makefile b/drivers/nvdimm/Makefile
index 0ce99cf..cecc280 100644
--- a/drivers/nvdimm/Makefile
+++ b/drivers/nvdimm/Makefile
@@ -1,6 +1,7 @@
 obj-$(CONFIG_LIBNVDIMM) += libnvdimm.o
 obj-$(CONFIG_BLK_DEV_PMEM_CORE) += nd_pmem_core.o
 obj-$(CONFIG_BLK_DEV_PMEM) += nd_pmem.o
+obj-$(CONFIG_BLK_DEV_PMEM_DMA) += nd_pmem_dma.o
 obj-$(CONFIG_ND_BTT) += nd_btt.o
 obj-$(CONFIG_ND_BLK) += nd_blk.o
 obj-$(CONFIG_X86_PMEM_LEGACY) += nd_e820.o
@@ -9,6 +10,8 @@ nd_pmem_core-y := pmem_core.o
 
 nd_pmem-y := pmem.o
 
+nd_pmem_dma-y := pmem_dma.o
+
 nd_btt-y := btt.o
 
 nd_blk-y := blk.o
diff --git a/drivers/nvdimm/pmem.h b/drivers/nvdimm/pmem.h
index 6df833e..ed83967 100644
--- a/drivers/nvdimm/pmem.h
+++ b/drivers/nvdimm/pmem.h
@@ -40,6 +40,7 @@ struct pmem_device {
 	struct gendisk		*disk;
 	struct blk_mq_tag_set	tag_set;
 	struct request_queue	*q;
+	unsigned int		sg_allocated;
 };
 
 static inline struct device *to_dev(struct pmem_device *pmem)
diff --git a/drivers/nvdimm/pmem_dma.c b/drivers/nvdimm/pmem_dma.c
new file mode 100644
index 0000000..a9c6c14
--- /dev/null
+++ b/drivers/nvdimm/pmem_dma.c
@@ -0,0 +1,475 @@
+/*
+ * Persistent Memory Block DMA Driver
+ * Copyright (c) 2017, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ */
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/blk-mq.h>
+#include <linux/dmaengine.h>
+#include <linux/dma-mapping.h>
+#include <linux/nodemask.h>
+#include "pmem.h"
+#include "pfn.h"
+#include "nd.h"
+
+/* After doing some measurements with various queue depth while running
+ * fio at 4k with 16 processes, it seems that a queue depth of 128
+ * provides the best performance. We can adjust this later when new
+ * data says otherwise.
+ */
+static int queue_depth = 128;
+
+struct pmem_cmd {
+	struct request *rq;
+	struct dma_chan *chan;
+	int sg_nents;
+	struct scatterlist sg[];
+};
+
+static void pmem_release_queue(void *data)
+{
+	struct pmem_device *pmem = data;
+
+	blk_cleanup_queue(pmem->q);
+	blk_mq_free_tag_set(&pmem->tag_set);
+}
+
+static void nd_pmem_dma_callback(void *data,
+		const struct dmaengine_result *res)
+{
+	struct pmem_cmd *cmd = data;
+	struct request *req = cmd->rq;
+	struct request_queue *q = req->q;
+	struct pmem_device *pmem = q->queuedata;
+	struct nd_region *nd_region = to_region(pmem);
+	struct device *dev = to_dev(pmem);
+	blk_status_t blk_status = BLK_STS_OK;
+
+	if (res) {
+		switch (res->result) {
+		case DMA_TRANS_READ_FAILED:
+		case DMA_TRANS_WRITE_FAILED:
+		case DMA_TRANS_ABORTED:
+			dev_dbg(dev, "bio failed\n");
+			blk_status = BLK_STS_IOERR;
+			break;
+		case DMA_TRANS_NOERROR:
+		default:
+			break;
+		}
+	}
+
+	if (req_op(req) == REQ_OP_WRITE && req->cmd_flags & REQ_FUA)
+		nvdimm_flush(nd_region);
+
+	blk_mq_end_request(cmd->rq, blk_status);
+}
+
+static int pmem_check_bad_pmem(struct pmem_cmd *cmd, bool is_write)
+{
+	struct request *req = cmd->rq;
+	struct request_queue *q = req->q;
+	struct pmem_device *pmem = q->queuedata;
+	struct bio_vec bvec;
+	struct req_iterator iter;
+
+	rq_for_each_segment(bvec, req, iter) {
+		sector_t sector = iter.iter.bi_sector;
+		unsigned int len = bvec.bv_len;
+		unsigned int off = bvec.bv_offset;
+
+		if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) {
+			if (is_write) {
+				struct page *page = bvec.bv_page;
+				phys_addr_t pmem_off = sector * 512 +
+					pmem->data_offset;
+				void *pmem_addr = pmem->virt_addr + pmem_off;
+
+		/*
+		 * Note that we write the data both before and after
+		 * clearing poison.  The write before clear poison
+		 * handles situations where the latest written data is
+		 * preserved and the clear poison operation simply marks
+		 * the address range as valid without changing the data.
+		 * In this case application software can assume that an
+		 * interrupted write will either return the new good
+		 * data or an error.
+		 *
+		 * However, if pmem_clear_poison() leaves the data in an
+		 * indeterminate state we need to perform the write
+		 * after clear poison.
+		 */
+				flush_dcache_page(page);
+				write_pmem(pmem_addr, page, off, len);
+				pmem_clear_poison(pmem, pmem_off, len);
+				write_pmem(pmem_addr, page, off, len);
+			} else
+				return -EIO;
+		}
+	}
+
+	return 0;
+}
+
+static blk_status_t pmem_handle_cmd_dma(struct pmem_cmd *cmd, bool is_write)
+{
+	struct request *req = cmd->rq;
+	struct request_queue *q = req->q;
+	struct pmem_device *pmem = q->queuedata;
+	struct device *dev = to_dev(pmem);
+	phys_addr_t pmem_off = blk_rq_pos(req) * 512 + pmem->data_offset;
+	void *pmem_addr = pmem->virt_addr + pmem_off;
+	size_t len;
+	struct dma_device *dma = cmd->chan->device;
+	struct dmaengine_unmap_data *unmap;
+	dma_cookie_t cookie;
+	struct dma_async_tx_descriptor *txd;
+	struct page *page;
+	unsigned int off;
+	int rc;
+	blk_status_t blk_status = BLK_STS_OK;
+	enum dma_data_direction dir;
+	dma_addr_t dma_addr;
+
+	rc = pmem_check_bad_pmem(cmd, is_write);
+	if (rc < 0) {
+		blk_status = BLK_STS_IOERR;
+		goto err;
+	}
+
+	unmap = dmaengine_get_unmap_data(dma->dev, 2, GFP_NOWAIT);
+	if (!unmap) {
+		dev_dbg(dev, "failed to get dma unmap data\n");
+		blk_status = BLK_STS_IOERR;
+		goto err;
+	}
+
+	/*
+	 * If reading from pmem, writing to scatterlist,
+	 * and if writing to pmem, reading from scatterlist.
+	 */
+	dir = is_write ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
+	cmd->sg_nents = blk_rq_map_sg(req->q, req, cmd->sg);
+	if (cmd->sg_nents < 1) {
+		blk_status = BLK_STS_IOERR;
+		goto err;
+	}
+
+	WARN_ON_ONCE(cmd->sg_nents > pmem->sg_allocated);
+
+	rc = dma_map_sg(dma->dev, cmd->sg, cmd->sg_nents, dir);
+	if (rc < 1) {
+		dev_dbg(dma->dev, "DMA scatterlist mapping error\n");
+		blk_status = BLK_STS_IOERR;
+		goto err;
+	}
+
+	unmap->unmap_sg.sg = cmd->sg;
+	unmap->sg_nents = cmd->sg_nents;
+	if (is_write)
+		unmap->from_sg = 1;
+	else
+		unmap->to_sg = 1;
+
+	len = blk_rq_payload_bytes(req);
+	page = virt_to_page(pmem_addr);
+	off = offset_in_page(pmem_addr);
+	dir = is_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
+	dma_addr = dma_map_page(dma->dev, page, off, len, dir);
+	if (dma_mapping_error(dma->dev, unmap->addr[0])) {
+		dev_dbg(dma->dev, "DMA buffer mapping error\n");
+		blk_status = BLK_STS_IOERR;
+		goto err_unmap;
+	}
+
+	unmap->unmap_sg.buf_phys = dma_addr;
+	unmap->len = len;
+	if (is_write)
+		unmap->to_cnt = 1;
+	else
+		unmap->from_cnt = 1;
+
+	txd = dmaengine_prep_dma_memcpy_sg(cmd->chan,
+				cmd->sg, cmd->sg_nents, dma_addr,
+				!is_write, DMA_PREP_INTERRUPT);
+	if (!txd) {
+		dev_dbg(dma->dev, "dma prep failed\n");
+		blk_status = BLK_STS_IOERR;
+		goto err_unmap;
+	}
+
+	txd->callback_result = nd_pmem_dma_callback;
+	txd->callback_param = cmd;
+	dma_set_unmap(txd, unmap);
+	dmaengine_unmap_put(unmap);
+	cookie = dmaengine_submit(txd);
+	if (dma_submit_error(cookie)) {
+		dev_dbg(dma->dev, "dma submit error\n");
+		blk_status = BLK_STS_IOERR;
+		goto err_set_unmap;
+	}
+
+	dma_async_issue_pending(cmd->chan);
+	return BLK_STS_OK;
+
+err_set_unmap:
+	dmaengine_unmap_put(unmap);
+err_unmap:
+	dmaengine_unmap_put(unmap);
+err:
+	blk_mq_end_request(cmd->rq, blk_status);
+	return blk_status;
+}
+
+static blk_status_t pmem_handle_cmd(struct pmem_cmd *cmd, bool is_write)
+{
+	struct request *req = cmd->rq;
+	struct request_queue *q = req->q;
+	struct pmem_device *pmem = q->queuedata;
+	struct nd_region *nd_region = to_region(pmem);
+	struct bio_vec bvec;
+	struct req_iterator iter;
+	blk_status_t blk_status = BLK_STS_OK;
+
+	rq_for_each_segment(bvec, req, iter) {
+		blk_status = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
+				bvec.bv_offset, is_write,
+				iter.iter.bi_sector);
+		if (blk_status != BLK_STS_OK)
+			break;
+	}
+
+	if (is_write && req->cmd_flags & REQ_FUA)
+		nvdimm_flush(nd_region);
+
+	blk_mq_end_request(cmd->rq, blk_status);
+
+	return blk_status;
+}
+
+typedef blk_status_t (*pmem_do_io)(struct pmem_cmd *cmd, bool is_write);
+
+static blk_status_t pmem_queue_rq(struct blk_mq_hw_ctx *hctx,
+		const struct blk_mq_queue_data *bd)
+{
+	struct pmem_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
+	struct request *req = cmd->rq = bd->rq;
+	struct request_queue *q = req->q;
+	struct pmem_device *pmem = q->queuedata;
+	struct nd_region *nd_region = to_region(pmem);
+	blk_status_t blk_status = BLK_STS_OK;
+	pmem_do_io do_io;
+
+	blk_mq_start_request(req);
+	cmd->chan = dma_find_channel(DMA_MEMCPY_SG);
+	if (cmd->chan)
+		do_io = pmem_handle_cmd_dma;
+	else
+		do_io = pmem_handle_cmd;
+
+	switch (req_op(req)) {
+	case REQ_PREFLUSH:
+		nvdimm_flush(nd_region);
+		blk_mq_end_request(cmd->rq, BLK_STS_OK);
+		break;
+	case REQ_OP_READ:
+		blk_status = do_io(cmd, false);
+		break;
+	case REQ_OP_WRITE:
+		blk_status = do_io(cmd, true);
+		break;
+	default:
+		blk_status = BLK_STS_NOTSUPP;
+		break;
+	}
+
+	if (blk_status != BLK_STS_OK)
+		blk_mq_end_request(cmd->rq, blk_status);
+
+	return blk_status;
+}
+
+static const struct blk_mq_ops pmem_mq_ops = {
+	.queue_rq	= pmem_queue_rq,
+};
+
+static const struct attribute_group *pmem_attribute_groups[] = {
+	&dax_attribute_group,
+	NULL,
+};
+
+static const struct block_device_operations pmem_fops = {
+	.owner =		THIS_MODULE,
+	.rw_page =		pmem_rw_page,
+	.revalidate_disk =	nvdimm_revalidate_disk,
+};
+
+static const struct dax_operations pmem_dax_ops = {
+	.direct_access = pmem_dax_direct_access,
+	.copy_from_iter = pmem_copy_from_iter,
+	.flush = pmem_dax_flush,
+};
+
+static bool pmem_dma_filter_fn(struct dma_chan *chan, void *node)
+{
+	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
+}
+
+static int pmem_attach_disk(struct device *dev,
+		struct nd_namespace_common *ndns)
+{
+	struct pmem_device *pmem;
+	int rc;
+	struct dma_chan *chan = NULL;
+	int has_dma;
+
+	pmem = pmem_core_setup_pmem(dev, ndns);
+	if (!pmem)
+		return -ENXIO;
+
+	chan = dma_find_channel(DMA_MEMCPY_SG);
+	if (!chan)
+		dev_warn(dev, "Forced back to CPU, no DMA\n");
+
+	has_dma = 1;
+	pmem->tag_set.ops = &pmem_mq_ops;
+	if (has_dma) {
+		dma_cap_mask_t dma_mask;
+		int node = 0, count;
+
+		dma_cap_zero(dma_mask);
+		dma_cap_set(DMA_MEMCPY_SG, dma_mask);
+		count = dma_get_channel_count(&dma_mask, pmem_dma_filter_fn,
+				(void *)(unsigned long)node);
+		if (count)
+			pmem->tag_set.nr_hw_queues = count;
+		else {
+			has_dma = 0;
+			pmem->tag_set.nr_hw_queues = num_online_cpus();
+		}
+	} else
+		pmem->tag_set.nr_hw_queues = num_online_cpus();
+
+	dev_dbg(dev, "%d HW queues allocated\n", pmem->tag_set.nr_hw_queues);
+
+	pmem->tag_set.queue_depth = queue_depth;
+	pmem->tag_set.numa_node = dev_to_node(dev);
+
+	if (has_dma) {
+		pmem->sg_allocated = (SZ_4K - sizeof(struct pmem_cmd)) /
+			sizeof(struct scatterlist);
+		pmem->tag_set.cmd_size = sizeof(struct pmem_cmd) +
+			sizeof(struct scatterlist) * pmem->sg_allocated;
+	} else
+		pmem->tag_set.cmd_size = sizeof(struct pmem_cmd);
+
+	pmem->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
+	pmem->tag_set.driver_data = pmem;
+
+	rc = blk_mq_alloc_tag_set(&pmem->tag_set);
+	if (rc < 0)
+		return rc;
+
+	pmem->q = blk_mq_init_queue(&pmem->tag_set);
+	if (IS_ERR(pmem->q)) {
+		blk_mq_free_tag_set(&pmem->tag_set);
+		return -ENOMEM;
+	}
+
+	pmem_core_setup_queue(dev, pmem, ndns);
+
+	if (has_dma) {
+		u64 xfercap = dma_get_desc_xfercap(chan);
+
+		/* set it to some sane size if DMA driver didn't export */
+		if (xfercap == 0)
+			xfercap = SZ_1M;
+
+		dev_dbg(dev, "xfercap: %#llx\n", xfercap);
+		/* max xfer size is per_descriptor_cap * num_of_sg */
+		blk_queue_max_hw_sectors(pmem->q,
+				pmem->sg_allocated * xfercap / 512);
+		blk_queue_max_segments(pmem->q, pmem->sg_allocated);
+	}
+		blk_queue_max_hw_sectors(pmem->q, UINT_MAX);
+
+	if (devm_add_action_or_reset(dev, pmem_release_queue, pmem)) {
+		pmem_release_queue(pmem);
+		return -ENOMEM;
+	}
+
+	rc = pmem_core_remap_pages(dev, pmem, ndns);
+	if (rc < 0)
+		return rc;
+
+	rc = pmem_core_setup_disk(dev, pmem, ndns, &pmem_fops,
+			&pmem_dax_ops, pmem_attribute_groups);
+	if (rc < 0)
+		return rc;
+
+	return 0;
+}
+
+static int nd_pmem_probe(struct device *dev)
+{
+	struct nd_namespace_common *ndns;
+
+	ndns = nvdimm_namespace_common_probe(dev);
+	if (IS_ERR(ndns))
+		return PTR_ERR(ndns);
+
+	if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
+		return -ENXIO;
+
+	if (is_nd_btt(dev))
+		return nvdimm_namespace_attach_btt(ndns);
+
+	if (is_nd_pfn(dev))
+		return pmem_attach_disk(dev, ndns);
+
+	/* if we find a valid info-block we'll come back as that personality */
+	if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
+			|| nd_dax_probe(dev, ndns) == 0)
+		return -ENXIO;
+
+	/* ...otherwise we're just a raw pmem device */
+	return pmem_attach_disk(dev, ndns);
+}
+
+static struct nd_device_driver nd_pmem_driver = {
+	.probe = nd_pmem_probe,
+	.remove = nd_pmem_remove,
+	.notify = nd_pmem_notify,
+	.shutdown = nd_pmem_shutdown,
+	.drv = {
+		.name = "nd_pmem",
+	},
+	.type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
+};
+
+static int __init pmem_init(void)
+{
+	dmaengine_get();
+	return nd_driver_register(&nd_pmem_driver);
+}
+module_init(pmem_init);
+
+static void pmem_exit(void)
+{
+	dmaengine_put();
+	driver_unregister(&nd_pmem_driver.drv);
+}
+module_exit(pmem_exit);
+
+MODULE_SOFTDEP("pre: dmaengine");
+MODULE_LICENSE("GPL v2");

_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm