From patchwork Wed Aug 30 20:56:23 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Dave Jiang X-Patchwork-Id: 9930777 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id AF82860383 for ; Wed, 30 Aug 2017 20:56:40 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 9A4D0287DB for ; Wed, 30 Aug 2017 20:56:40 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id 35D11287BF; Wed, 30 Aug 2017 20:56:40 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-1.9 required=2.0 tests=BAYES_00, RCVD_IN_DNSWL_NONE autolearn=ham version=3.3.1 Received: from ml01.01.org (ml01.01.org [198.145.21.10]) (using TLSv1.2 with cipher DHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mail.wl.linuxfoundation.org (Postfix) with ESMTPS id EE367287AD for ; Wed, 30 Aug 2017 20:56:35 +0000 (UTC) Received: from [127.0.0.1] (localhost [IPv6:::1]) by ml01.01.org (Postfix) with ESMTP id 2C81221E95DF5; Wed, 30 Aug 2017 13:56:22 -0700 (PDT) X-Original-To: linux-nvdimm@lists.01.org Delivered-To: linux-nvdimm@lists.01.org Received: from mga02.intel.com (mga02.intel.com [134.134.136.20]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ml01.01.org (Postfix) with ESMTPS id 966E921E95DE6 for ; Wed, 30 Aug 2017 13:56:20 -0700 (PDT) Received: from fmsmga005.fm.intel.com ([10.253.24.32]) by orsmga101.jf.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 30 Aug 2017 13:56:24 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.41,450,1498546800"; d="scan'208";a="145811876" Received: from djiang5-desk3.ch.intel.com ([143.182.137.38]) by fmsmga005.fm.intel.com with ESMTP; 30 Aug 2017 13:56:23 -0700 Subject: [PATCH v7 9/9] libnvdimm: Add DMA based blk-mq pmem driver From: Dave Jiang To: vinod.koul@intel.com, dan.j.williams@intel.com Date: Wed, 30 Aug 2017 13:56:23 -0700 Message-ID: <150412658373.69288.2856155675209269488.stgit@djiang5-desk3.ch.intel.com> In-Reply-To: <150412628764.69288.12074115435918322858.stgit@djiang5-desk3.ch.intel.com> References: <150412628764.69288.12074115435918322858.stgit@djiang5-desk3.ch.intel.com> User-Agent: StGit/0.17.1-dirty MIME-Version: 1.0 X-BeenThere: linux-nvdimm@lists.01.org X-Mailman-Version: 2.1.22 Precedence: list List-Id: "Linux-nvdimm developer list." List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: dmaengine@vger.kernel.org, hch@infradead.org, linux-nvdimm@lists.01.org Errors-To: linux-nvdimm-bounces@lists.01.org Sender: "Linux-nvdimm" X-Virus-Scanned: ClamAV using ClamSMTP 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 Reviewed-by: Ross Zwisler --- 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 +#include +#include +#include +#include +#include +#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");