2 * NVM Express device driver
3 * Copyright (c) 2011-2014, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 #include <linux/nvme.h>
16 #include <linux/bitops.h>
17 #include <linux/blkdev.h>
18 #include <linux/blk-mq.h>
19 #include <linux/cpu.h>
20 #include <linux/delay.h>
21 #include <linux/errno.h>
23 #include <linux/genhd.h>
24 #include <linux/hdreg.h>
25 #include <linux/idr.h>
26 #include <linux/init.h>
27 #include <linux/interrupt.h>
29 #include <linux/kdev_t.h>
30 #include <linux/kthread.h>
31 #include <linux/kernel.h>
32 #include <linux/list_sort.h>
34 #include <linux/module.h>
35 #include <linux/moduleparam.h>
36 #include <linux/pci.h>
37 #include <linux/poison.h>
38 #include <linux/ptrace.h>
39 #include <linux/sched.h>
40 #include <linux/slab.h>
41 #include <linux/t10-pi.h>
42 #include <linux/types.h>
44 #include <linux/io-64-nonatomic-lo-hi.h>
46 #define NVME_MINORS (1U << MINORBITS)
47 #define NVME_Q_DEPTH 1024
48 #define NVME_AQ_DEPTH 256
49 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
50 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
51 #define ADMIN_TIMEOUT (admin_timeout * HZ)
52 #define SHUTDOWN_TIMEOUT (shutdown_timeout * HZ)
54 static unsigned char admin_timeout = 60;
55 module_param(admin_timeout, byte, 0644);
56 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
58 unsigned char nvme_io_timeout = 30;
59 module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
60 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
62 static unsigned char shutdown_timeout = 5;
63 module_param(shutdown_timeout, byte, 0644);
64 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
66 static int nvme_major;
67 module_param(nvme_major, int, 0);
69 static int nvme_char_major;
70 module_param(nvme_char_major, int, 0);
72 static int use_threaded_interrupts;
73 module_param(use_threaded_interrupts, int, 0);
75 static bool use_cmb_sqes = true;
76 module_param(use_cmb_sqes, bool, 0644);
77 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
79 static DEFINE_SPINLOCK(dev_list_lock);
80 static LIST_HEAD(dev_list);
81 static struct task_struct *nvme_thread;
82 static struct workqueue_struct *nvme_workq;
83 static wait_queue_head_t nvme_kthread_wait;
85 static struct class *nvme_class;
87 static void nvme_reset_failed_dev(struct work_struct *ws);
88 static int nvme_reset(struct nvme_dev *dev);
89 static int nvme_process_cq(struct nvme_queue *nvmeq);
91 struct async_cmd_info {
92 struct kthread_work work;
93 struct kthread_worker *worker;
101 * An NVM Express queue. Each device has at least two (one for admin
102 * commands and one for I/O commands).
105 struct device *q_dmadev;
106 struct nvme_dev *dev;
107 char irqname[24]; /* nvme4294967295-65535\0 */
109 struct nvme_command *sq_cmds;
110 struct nvme_command __iomem *sq_cmds_io;
111 volatile struct nvme_completion *cqes;
112 struct blk_mq_tags **tags;
113 dma_addr_t sq_dma_addr;
114 dma_addr_t cq_dma_addr;
124 struct async_cmd_info cmdinfo;
128 * Check we didin't inadvertently grow the command struct
130 static inline void _nvme_check_size(void)
132 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
133 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
134 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
135 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
136 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
137 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
138 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
139 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
140 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
141 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
142 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
143 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
146 typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
147 struct nvme_completion *);
149 struct nvme_cmd_info {
150 nvme_completion_fn fn;
153 struct nvme_queue *nvmeq;
154 struct nvme_iod iod[0];
158 * Max size of iod being embedded in the request payload
160 #define NVME_INT_PAGES 2
161 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->page_size)
162 #define NVME_INT_MASK 0x01
165 * Will slightly overestimate the number of pages needed. This is OK
166 * as it only leads to a small amount of wasted memory for the lifetime of
169 static int nvme_npages(unsigned size, struct nvme_dev *dev)
171 unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
172 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
175 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
177 unsigned int ret = sizeof(struct nvme_cmd_info);
179 ret += sizeof(struct nvme_iod);
180 ret += sizeof(__le64 *) * nvme_npages(NVME_INT_BYTES(dev), dev);
181 ret += sizeof(struct scatterlist) * NVME_INT_PAGES;
186 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
187 unsigned int hctx_idx)
189 struct nvme_dev *dev = data;
190 struct nvme_queue *nvmeq = dev->queues[0];
192 WARN_ON(hctx_idx != 0);
193 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
194 WARN_ON(nvmeq->tags);
196 hctx->driver_data = nvmeq;
197 nvmeq->tags = &dev->admin_tagset.tags[0];
201 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
203 struct nvme_queue *nvmeq = hctx->driver_data;
208 static int nvme_admin_init_request(void *data, struct request *req,
209 unsigned int hctx_idx, unsigned int rq_idx,
210 unsigned int numa_node)
212 struct nvme_dev *dev = data;
213 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
214 struct nvme_queue *nvmeq = dev->queues[0];
221 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
222 unsigned int hctx_idx)
224 struct nvme_dev *dev = data;
225 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
228 nvmeq->tags = &dev->tagset.tags[hctx_idx];
230 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
231 hctx->driver_data = nvmeq;
235 static int nvme_init_request(void *data, struct request *req,
236 unsigned int hctx_idx, unsigned int rq_idx,
237 unsigned int numa_node)
239 struct nvme_dev *dev = data;
240 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
241 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
248 static void nvme_set_info(struct nvme_cmd_info *cmd, void *ctx,
249 nvme_completion_fn handler)
254 blk_mq_start_request(blk_mq_rq_from_pdu(cmd));
257 static void *iod_get_private(struct nvme_iod *iod)
259 return (void *) (iod->private & ~0x1UL);
263 * If bit 0 is set, the iod is embedded in the request payload.
265 static bool iod_should_kfree(struct nvme_iod *iod)
267 return (iod->private & NVME_INT_MASK) == 0;
270 /* Special values must be less than 0x1000 */
271 #define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
272 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
273 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
274 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
276 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
277 struct nvme_completion *cqe)
279 if (ctx == CMD_CTX_CANCELLED)
281 if (ctx == CMD_CTX_COMPLETED) {
282 dev_warn(nvmeq->q_dmadev,
283 "completed id %d twice on queue %d\n",
284 cqe->command_id, le16_to_cpup(&cqe->sq_id));
287 if (ctx == CMD_CTX_INVALID) {
288 dev_warn(nvmeq->q_dmadev,
289 "invalid id %d completed on queue %d\n",
290 cqe->command_id, le16_to_cpup(&cqe->sq_id));
293 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
296 static void *cancel_cmd_info(struct nvme_cmd_info *cmd, nvme_completion_fn *fn)
303 cmd->fn = special_completion;
304 cmd->ctx = CMD_CTX_CANCELLED;
308 static void async_req_completion(struct nvme_queue *nvmeq, void *ctx,
309 struct nvme_completion *cqe)
311 u32 result = le32_to_cpup(&cqe->result);
312 u16 status = le16_to_cpup(&cqe->status) >> 1;
314 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
315 ++nvmeq->dev->event_limit;
316 if (status != NVME_SC_SUCCESS)
319 switch (result & 0xff07) {
320 case NVME_AER_NOTICE_NS_CHANGED:
321 dev_info(nvmeq->q_dmadev, "rescanning\n");
322 schedule_work(&nvmeq->dev->scan_work);
324 dev_warn(nvmeq->q_dmadev, "async event result %08x\n", result);
328 static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
329 struct nvme_completion *cqe)
331 struct request *req = ctx;
333 u16 status = le16_to_cpup(&cqe->status) >> 1;
334 u32 result = le32_to_cpup(&cqe->result);
336 blk_mq_free_request(req);
338 dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
339 ++nvmeq->dev->abort_limit;
342 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
343 struct nvme_completion *cqe)
345 struct async_cmd_info *cmdinfo = ctx;
346 cmdinfo->result = le32_to_cpup(&cqe->result);
347 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
348 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
349 blk_mq_free_request(cmdinfo->req);
352 static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
355 struct request *req = blk_mq_tag_to_rq(*nvmeq->tags, tag);
357 return blk_mq_rq_to_pdu(req);
361 * Called with local interrupts disabled and the q_lock held. May not sleep.
363 static void *nvme_finish_cmd(struct nvme_queue *nvmeq, int tag,
364 nvme_completion_fn *fn)
366 struct nvme_cmd_info *cmd = get_cmd_from_tag(nvmeq, tag);
368 if (tag >= nvmeq->q_depth) {
369 *fn = special_completion;
370 return CMD_CTX_INVALID;
375 cmd->fn = special_completion;
376 cmd->ctx = CMD_CTX_COMPLETED;
381 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
382 * @nvmeq: The queue to use
383 * @cmd: The command to send
385 * Safe to use from interrupt context
387 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
388 struct nvme_command *cmd)
390 u16 tail = nvmeq->sq_tail;
392 if (nvmeq->sq_cmds_io)
393 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
395 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
397 if (++tail == nvmeq->q_depth)
399 writel(tail, nvmeq->q_db);
400 nvmeq->sq_tail = tail;
403 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
406 spin_lock_irqsave(&nvmeq->q_lock, flags);
407 __nvme_submit_cmd(nvmeq, cmd);
408 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
411 static __le64 **iod_list(struct nvme_iod *iod)
413 return ((void *)iod) + iod->offset;
416 static inline void iod_init(struct nvme_iod *iod, unsigned nbytes,
417 unsigned nseg, unsigned long private)
419 iod->private = private;
420 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
422 iod->length = nbytes;
426 static struct nvme_iod *
427 __nvme_alloc_iod(unsigned nseg, unsigned bytes, struct nvme_dev *dev,
428 unsigned long priv, gfp_t gfp)
430 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
431 sizeof(__le64 *) * nvme_npages(bytes, dev) +
432 sizeof(struct scatterlist) * nseg, gfp);
435 iod_init(iod, bytes, nseg, priv);
440 static struct nvme_iod *nvme_alloc_iod(struct request *rq, struct nvme_dev *dev,
443 unsigned size = !(rq->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(rq) :
444 sizeof(struct nvme_dsm_range);
445 struct nvme_iod *iod;
447 if (rq->nr_phys_segments <= NVME_INT_PAGES &&
448 size <= NVME_INT_BYTES(dev)) {
449 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(rq);
452 iod_init(iod, size, rq->nr_phys_segments,
453 (unsigned long) rq | NVME_INT_MASK);
457 return __nvme_alloc_iod(rq->nr_phys_segments, size, dev,
458 (unsigned long) rq, gfp);
461 static void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
463 const int last_prp = dev->page_size / 8 - 1;
465 __le64 **list = iod_list(iod);
466 dma_addr_t prp_dma = iod->first_dma;
468 if (iod->npages == 0)
469 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
470 for (i = 0; i < iod->npages; i++) {
471 __le64 *prp_list = list[i];
472 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
473 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
474 prp_dma = next_prp_dma;
477 if (iod_should_kfree(iod))
481 static int nvme_error_status(u16 status)
483 switch (status & 0x7ff) {
484 case NVME_SC_SUCCESS:
486 case NVME_SC_CAP_EXCEEDED:
493 #ifdef CONFIG_BLK_DEV_INTEGRITY
494 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
496 if (be32_to_cpu(pi->ref_tag) == v)
497 pi->ref_tag = cpu_to_be32(p);
500 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
502 if (be32_to_cpu(pi->ref_tag) == p)
503 pi->ref_tag = cpu_to_be32(v);
507 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
509 * The virtual start sector is the one that was originally submitted by the
510 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
511 * start sector may be different. Remap protection information to match the
512 * physical LBA on writes, and back to the original seed on reads.
514 * Type 0 and 3 do not have a ref tag, so no remapping required.
516 static void nvme_dif_remap(struct request *req,
517 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
519 struct nvme_ns *ns = req->rq_disk->private_data;
520 struct bio_integrity_payload *bip;
521 struct t10_pi_tuple *pi;
523 u32 i, nlb, ts, phys, virt;
525 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
528 bip = bio_integrity(req->bio);
532 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
535 virt = bip_get_seed(bip);
536 phys = nvme_block_nr(ns, blk_rq_pos(req));
537 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
538 ts = ns->disk->integrity->tuple_size;
540 for (i = 0; i < nlb; i++, virt++, phys++) {
541 pi = (struct t10_pi_tuple *)p;
542 dif_swap(phys, virt, pi);
548 static int nvme_noop_verify(struct blk_integrity_iter *iter)
553 static int nvme_noop_generate(struct blk_integrity_iter *iter)
558 struct blk_integrity nvme_meta_noop = {
559 .name = "NVME_META_NOOP",
560 .generate_fn = nvme_noop_generate,
561 .verify_fn = nvme_noop_verify,
564 static void nvme_init_integrity(struct nvme_ns *ns)
566 struct blk_integrity integrity;
568 switch (ns->pi_type) {
569 case NVME_NS_DPS_PI_TYPE3:
570 integrity = t10_pi_type3_crc;
572 case NVME_NS_DPS_PI_TYPE1:
573 case NVME_NS_DPS_PI_TYPE2:
574 integrity = t10_pi_type1_crc;
577 integrity = nvme_meta_noop;
580 integrity.tuple_size = ns->ms;
581 blk_integrity_register(ns->disk, &integrity);
582 blk_queue_max_integrity_segments(ns->queue, 1);
584 #else /* CONFIG_BLK_DEV_INTEGRITY */
585 static void nvme_dif_remap(struct request *req,
586 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
589 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
592 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
595 static void nvme_init_integrity(struct nvme_ns *ns)
600 static void req_completion(struct nvme_queue *nvmeq, void *ctx,
601 struct nvme_completion *cqe)
603 struct nvme_iod *iod = ctx;
604 struct request *req = iod_get_private(iod);
605 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
606 u16 status = le16_to_cpup(&cqe->status) >> 1;
607 bool requeue = false;
610 if (unlikely(status)) {
611 if (!(status & NVME_SC_DNR || blk_noretry_request(req))
612 && (jiffies - req->start_time) < req->timeout) {
616 blk_mq_requeue_request(req);
617 spin_lock_irqsave(req->q->queue_lock, flags);
618 if (!blk_queue_stopped(req->q))
619 blk_mq_kick_requeue_list(req->q);
620 spin_unlock_irqrestore(req->q->queue_lock, flags);
624 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
625 if (cmd_rq->ctx == CMD_CTX_CANCELLED)
630 error = nvme_error_status(status);
634 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
635 u32 result = le32_to_cpup(&cqe->result);
636 req->special = (void *)(uintptr_t)result;
640 dev_warn(nvmeq->dev->dev,
641 "completing aborted command with status:%04x\n",
646 dma_unmap_sg(nvmeq->dev->dev, iod->sg, iod->nents,
647 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
648 if (blk_integrity_rq(req)) {
649 if (!rq_data_dir(req))
650 nvme_dif_remap(req, nvme_dif_complete);
651 dma_unmap_sg(nvmeq->dev->dev, iod->meta_sg, 1,
652 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
655 nvme_free_iod(nvmeq->dev, iod);
657 if (likely(!requeue))
658 blk_mq_complete_request(req, error);
661 /* length is in bytes. gfp flags indicates whether we may sleep. */
662 static int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
663 int total_len, gfp_t gfp)
665 struct dma_pool *pool;
666 int length = total_len;
667 struct scatterlist *sg = iod->sg;
668 int dma_len = sg_dma_len(sg);
669 u64 dma_addr = sg_dma_address(sg);
670 u32 page_size = dev->page_size;
671 int offset = dma_addr & (page_size - 1);
673 __le64 **list = iod_list(iod);
677 length -= (page_size - offset);
681 dma_len -= (page_size - offset);
683 dma_addr += (page_size - offset);
686 dma_addr = sg_dma_address(sg);
687 dma_len = sg_dma_len(sg);
690 if (length <= page_size) {
691 iod->first_dma = dma_addr;
695 nprps = DIV_ROUND_UP(length, page_size);
696 if (nprps <= (256 / 8)) {
697 pool = dev->prp_small_pool;
700 pool = dev->prp_page_pool;
704 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
706 iod->first_dma = dma_addr;
708 return (total_len - length) + page_size;
711 iod->first_dma = prp_dma;
714 if (i == page_size >> 3) {
715 __le64 *old_prp_list = prp_list;
716 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
718 return total_len - length;
719 list[iod->npages++] = prp_list;
720 prp_list[0] = old_prp_list[i - 1];
721 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
724 prp_list[i++] = cpu_to_le64(dma_addr);
725 dma_len -= page_size;
726 dma_addr += page_size;
734 dma_addr = sg_dma_address(sg);
735 dma_len = sg_dma_len(sg);
741 static void nvme_submit_priv(struct nvme_queue *nvmeq, struct request *req,
742 struct nvme_iod *iod)
744 struct nvme_command cmnd;
746 memcpy(&cmnd, req->cmd, sizeof(cmnd));
747 cmnd.rw.command_id = req->tag;
748 if (req->nr_phys_segments) {
749 cmnd.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
750 cmnd.rw.prp2 = cpu_to_le64(iod->first_dma);
753 __nvme_submit_cmd(nvmeq, &cmnd);
757 * We reuse the small pool to allocate the 16-byte range here as it is not
758 * worth having a special pool for these or additional cases to handle freeing
761 static void nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
762 struct request *req, struct nvme_iod *iod)
764 struct nvme_dsm_range *range =
765 (struct nvme_dsm_range *)iod_list(iod)[0];
766 struct nvme_command cmnd;
768 range->cattr = cpu_to_le32(0);
769 range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift);
770 range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
772 memset(&cmnd, 0, sizeof(cmnd));
773 cmnd.dsm.opcode = nvme_cmd_dsm;
774 cmnd.dsm.command_id = req->tag;
775 cmnd.dsm.nsid = cpu_to_le32(ns->ns_id);
776 cmnd.dsm.prp1 = cpu_to_le64(iod->first_dma);
778 cmnd.dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
780 __nvme_submit_cmd(nvmeq, &cmnd);
783 static void nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
786 struct nvme_command cmnd;
788 memset(&cmnd, 0, sizeof(cmnd));
789 cmnd.common.opcode = nvme_cmd_flush;
790 cmnd.common.command_id = cmdid;
791 cmnd.common.nsid = cpu_to_le32(ns->ns_id);
793 __nvme_submit_cmd(nvmeq, &cmnd);
796 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod,
799 struct request *req = iod_get_private(iod);
800 struct nvme_command cmnd;
804 if (req->cmd_flags & REQ_FUA)
805 control |= NVME_RW_FUA;
806 if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
807 control |= NVME_RW_LR;
809 if (req->cmd_flags & REQ_RAHEAD)
810 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
812 memset(&cmnd, 0, sizeof(cmnd));
813 cmnd.rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
814 cmnd.rw.command_id = req->tag;
815 cmnd.rw.nsid = cpu_to_le32(ns->ns_id);
816 cmnd.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
817 cmnd.rw.prp2 = cpu_to_le64(iod->first_dma);
818 cmnd.rw.slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
819 cmnd.rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
822 switch (ns->pi_type) {
823 case NVME_NS_DPS_PI_TYPE3:
824 control |= NVME_RW_PRINFO_PRCHK_GUARD;
826 case NVME_NS_DPS_PI_TYPE1:
827 case NVME_NS_DPS_PI_TYPE2:
828 control |= NVME_RW_PRINFO_PRCHK_GUARD |
829 NVME_RW_PRINFO_PRCHK_REF;
830 cmnd.rw.reftag = cpu_to_le32(
831 nvme_block_nr(ns, blk_rq_pos(req)));
834 if (blk_integrity_rq(req))
836 cpu_to_le64(sg_dma_address(iod->meta_sg));
838 control |= NVME_RW_PRINFO_PRACT;
841 cmnd.rw.control = cpu_to_le16(control);
842 cmnd.rw.dsmgmt = cpu_to_le32(dsmgmt);
844 __nvme_submit_cmd(nvmeq, &cmnd);
850 * NOTE: ns is NULL when called on the admin queue.
852 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
853 const struct blk_mq_queue_data *bd)
855 struct nvme_ns *ns = hctx->queue->queuedata;
856 struct nvme_queue *nvmeq = hctx->driver_data;
857 struct nvme_dev *dev = nvmeq->dev;
858 struct request *req = bd->rq;
859 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
860 struct nvme_iod *iod;
861 enum dma_data_direction dma_dir;
864 * If formated with metadata, require the block layer provide a buffer
865 * unless this namespace is formated such that the metadata can be
866 * stripped/generated by the controller with PRACT=1.
868 if (ns && ns->ms && !blk_integrity_rq(req)) {
869 if (!(ns->pi_type && ns->ms == 8) &&
870 req->cmd_type != REQ_TYPE_DRV_PRIV) {
871 blk_mq_complete_request(req, -EFAULT);
872 return BLK_MQ_RQ_QUEUE_OK;
876 iod = nvme_alloc_iod(req, dev, GFP_ATOMIC);
878 return BLK_MQ_RQ_QUEUE_BUSY;
880 if (req->cmd_flags & REQ_DISCARD) {
883 * We reuse the small pool to allocate the 16-byte range here
884 * as it is not worth having a special pool for these or
885 * additional cases to handle freeing the iod.
887 range = dma_pool_alloc(dev->prp_small_pool, GFP_ATOMIC,
891 iod_list(iod)[0] = (__le64 *)range;
893 } else if (req->nr_phys_segments) {
894 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
896 sg_init_table(iod->sg, req->nr_phys_segments);
897 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
901 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
904 if (blk_rq_bytes(req) !=
905 nvme_setup_prps(dev, iod, blk_rq_bytes(req), GFP_ATOMIC)) {
906 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
909 if (blk_integrity_rq(req)) {
910 if (blk_rq_count_integrity_sg(req->q, req->bio) != 1)
913 sg_init_table(iod->meta_sg, 1);
914 if (blk_rq_map_integrity_sg(
915 req->q, req->bio, iod->meta_sg) != 1)
918 if (rq_data_dir(req))
919 nvme_dif_remap(req, nvme_dif_prep);
921 if (!dma_map_sg(nvmeq->q_dmadev, iod->meta_sg, 1, dma_dir))
926 nvme_set_info(cmd, iod, req_completion);
927 spin_lock_irq(&nvmeq->q_lock);
928 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
929 nvme_submit_priv(nvmeq, req, iod);
930 else if (req->cmd_flags & REQ_DISCARD)
931 nvme_submit_discard(nvmeq, ns, req, iod);
932 else if (req->cmd_flags & REQ_FLUSH)
933 nvme_submit_flush(nvmeq, ns, req->tag);
935 nvme_submit_iod(nvmeq, iod, ns);
937 nvme_process_cq(nvmeq);
938 spin_unlock_irq(&nvmeq->q_lock);
939 return BLK_MQ_RQ_QUEUE_OK;
942 nvme_free_iod(dev, iod);
943 return BLK_MQ_RQ_QUEUE_ERROR;
945 nvme_free_iod(dev, iod);
946 return BLK_MQ_RQ_QUEUE_BUSY;
949 static int nvme_process_cq(struct nvme_queue *nvmeq)
953 head = nvmeq->cq_head;
954 phase = nvmeq->cq_phase;
958 nvme_completion_fn fn;
959 struct nvme_completion cqe = nvmeq->cqes[head];
960 if ((le16_to_cpu(cqe.status) & 1) != phase)
962 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
963 if (++head == nvmeq->q_depth) {
967 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
968 fn(nvmeq, ctx, &cqe);
971 /* If the controller ignores the cq head doorbell and continuously
972 * writes to the queue, it is theoretically possible to wrap around
973 * the queue twice and mistakenly return IRQ_NONE. Linux only
974 * requires that 0.1% of your interrupts are handled, so this isn't
977 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
980 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
981 nvmeq->cq_head = head;
982 nvmeq->cq_phase = phase;
988 static irqreturn_t nvme_irq(int irq, void *data)
991 struct nvme_queue *nvmeq = data;
992 spin_lock(&nvmeq->q_lock);
993 nvme_process_cq(nvmeq);
994 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
996 spin_unlock(&nvmeq->q_lock);
1000 static irqreturn_t nvme_irq_check(int irq, void *data)
1002 struct nvme_queue *nvmeq = data;
1003 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
1004 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
1006 return IRQ_WAKE_THREAD;
1010 * Returns 0 on success. If the result is negative, it's a Linux error code;
1011 * if the result is positive, it's an NVM Express status code
1013 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1014 void *buffer, void __user *ubuffer, unsigned bufflen,
1015 u32 *result, unsigned timeout)
1017 bool write = cmd->common.opcode & 1;
1018 struct bio *bio = NULL;
1019 struct request *req;
1022 req = blk_mq_alloc_request(q, write, GFP_KERNEL, false);
1024 return PTR_ERR(req);
1026 req->cmd_type = REQ_TYPE_DRV_PRIV;
1027 req->cmd_flags |= REQ_FAILFAST_DRIVER;
1028 req->__data_len = 0;
1029 req->__sector = (sector_t) -1;
1030 req->bio = req->biotail = NULL;
1032 req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
1034 req->cmd = (unsigned char *)cmd;
1035 req->cmd_len = sizeof(struct nvme_command);
1036 req->special = (void *)0;
1038 if (buffer && bufflen) {
1039 ret = blk_rq_map_kern(q, req, buffer, bufflen, __GFP_WAIT);
1042 } else if (ubuffer && bufflen) {
1043 ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen, __GFP_WAIT);
1049 blk_execute_rq(req->q, NULL, req, 0);
1051 blk_rq_unmap_user(bio);
1053 *result = (u32)(uintptr_t)req->special;
1056 blk_mq_free_request(req);
1060 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1061 void *buffer, unsigned bufflen)
1063 return __nvme_submit_sync_cmd(q, cmd, buffer, NULL, bufflen, NULL, 0);
1066 static int nvme_submit_async_admin_req(struct nvme_dev *dev)
1068 struct nvme_queue *nvmeq = dev->queues[0];
1069 struct nvme_command c;
1070 struct nvme_cmd_info *cmd_info;
1071 struct request *req;
1073 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, true);
1075 return PTR_ERR(req);
1077 req->cmd_flags |= REQ_NO_TIMEOUT;
1078 cmd_info = blk_mq_rq_to_pdu(req);
1079 nvme_set_info(cmd_info, NULL, async_req_completion);
1081 memset(&c, 0, sizeof(c));
1082 c.common.opcode = nvme_admin_async_event;
1083 c.common.command_id = req->tag;
1085 blk_mq_free_request(req);
1086 __nvme_submit_cmd(nvmeq, &c);
1090 static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
1091 struct nvme_command *cmd,
1092 struct async_cmd_info *cmdinfo, unsigned timeout)
1094 struct nvme_queue *nvmeq = dev->queues[0];
1095 struct request *req;
1096 struct nvme_cmd_info *cmd_rq;
1098 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
1100 return PTR_ERR(req);
1102 req->timeout = timeout;
1103 cmd_rq = blk_mq_rq_to_pdu(req);
1105 nvme_set_info(cmd_rq, cmdinfo, async_completion);
1106 cmdinfo->status = -EINTR;
1108 cmd->common.command_id = req->tag;
1110 nvme_submit_cmd(nvmeq, cmd);
1114 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1116 struct nvme_command c;
1118 memset(&c, 0, sizeof(c));
1119 c.delete_queue.opcode = opcode;
1120 c.delete_queue.qid = cpu_to_le16(id);
1122 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1125 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1126 struct nvme_queue *nvmeq)
1128 struct nvme_command c;
1129 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1132 * Note: we (ab)use the fact the the prp fields survive if no data
1133 * is attached to the request.
1135 memset(&c, 0, sizeof(c));
1136 c.create_cq.opcode = nvme_admin_create_cq;
1137 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1138 c.create_cq.cqid = cpu_to_le16(qid);
1139 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1140 c.create_cq.cq_flags = cpu_to_le16(flags);
1141 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1143 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1146 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1147 struct nvme_queue *nvmeq)
1149 struct nvme_command c;
1150 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1153 * Note: we (ab)use the fact the the prp fields survive if no data
1154 * is attached to the request.
1156 memset(&c, 0, sizeof(c));
1157 c.create_sq.opcode = nvme_admin_create_sq;
1158 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1159 c.create_sq.sqid = cpu_to_le16(qid);
1160 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1161 c.create_sq.sq_flags = cpu_to_le16(flags);
1162 c.create_sq.cqid = cpu_to_le16(qid);
1164 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1167 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1169 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1172 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1174 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1177 int nvme_identify_ctrl(struct nvme_dev *dev, struct nvme_id_ctrl **id)
1179 struct nvme_command c = { };
1182 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1183 c.identify.opcode = nvme_admin_identify;
1184 c.identify.cns = cpu_to_le32(1);
1186 *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1190 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1191 sizeof(struct nvme_id_ctrl));
1197 int nvme_identify_ns(struct nvme_dev *dev, unsigned nsid,
1198 struct nvme_id_ns **id)
1200 struct nvme_command c = { };
1203 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1204 c.identify.opcode = nvme_admin_identify,
1205 c.identify.nsid = cpu_to_le32(nsid),
1207 *id = kmalloc(sizeof(struct nvme_id_ns), GFP_KERNEL);
1211 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1212 sizeof(struct nvme_id_ns));
1218 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1219 dma_addr_t dma_addr, u32 *result)
1221 struct nvme_command c;
1223 memset(&c, 0, sizeof(c));
1224 c.features.opcode = nvme_admin_get_features;
1225 c.features.nsid = cpu_to_le32(nsid);
1226 c.features.prp1 = cpu_to_le64(dma_addr);
1227 c.features.fid = cpu_to_le32(fid);
1229 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1233 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1234 dma_addr_t dma_addr, u32 *result)
1236 struct nvme_command c;
1238 memset(&c, 0, sizeof(c));
1239 c.features.opcode = nvme_admin_set_features;
1240 c.features.prp1 = cpu_to_le64(dma_addr);
1241 c.features.fid = cpu_to_le32(fid);
1242 c.features.dword11 = cpu_to_le32(dword11);
1244 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1248 int nvme_get_log_page(struct nvme_dev *dev, struct nvme_smart_log **log)
1250 struct nvme_command c = { };
1253 c.common.opcode = nvme_admin_get_log_page,
1254 c.common.nsid = cpu_to_le32(0xFFFFFFFF),
1255 c.common.cdw10[0] = cpu_to_le32(
1256 (((sizeof(struct nvme_smart_log) / 4) - 1) << 16) |
1259 *log = kmalloc(sizeof(struct nvme_smart_log), GFP_KERNEL);
1263 error = nvme_submit_sync_cmd(dev->admin_q, &c, *log,
1264 sizeof(struct nvme_smart_log));
1271 * nvme_abort_req - Attempt aborting a request
1273 * Schedule controller reset if the command was already aborted once before and
1274 * still hasn't been returned to the driver, or if this is the admin queue.
1276 static void nvme_abort_req(struct request *req)
1278 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1279 struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1280 struct nvme_dev *dev = nvmeq->dev;
1281 struct request *abort_req;
1282 struct nvme_cmd_info *abort_cmd;
1283 struct nvme_command cmd;
1285 if (!nvmeq->qid || cmd_rq->aborted) {
1286 unsigned long flags;
1288 spin_lock_irqsave(&dev_list_lock, flags);
1289 if (work_busy(&dev->reset_work))
1291 list_del_init(&dev->node);
1292 dev_warn(dev->dev, "I/O %d QID %d timeout, reset controller\n",
1293 req->tag, nvmeq->qid);
1294 dev->reset_workfn = nvme_reset_failed_dev;
1295 queue_work(nvme_workq, &dev->reset_work);
1297 spin_unlock_irqrestore(&dev_list_lock, flags);
1301 if (!dev->abort_limit)
1304 abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1306 if (IS_ERR(abort_req))
1309 abort_cmd = blk_mq_rq_to_pdu(abort_req);
1310 nvme_set_info(abort_cmd, abort_req, abort_completion);
1312 memset(&cmd, 0, sizeof(cmd));
1313 cmd.abort.opcode = nvme_admin_abort_cmd;
1314 cmd.abort.cid = req->tag;
1315 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1316 cmd.abort.command_id = abort_req->tag;
1319 cmd_rq->aborted = 1;
1321 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1323 nvme_submit_cmd(dev->queues[0], &cmd);
1326 static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved)
1328 struct nvme_queue *nvmeq = data;
1330 nvme_completion_fn fn;
1331 struct nvme_cmd_info *cmd;
1332 struct nvme_completion cqe;
1334 if (!blk_mq_request_started(req))
1337 cmd = blk_mq_rq_to_pdu(req);
1339 if (cmd->ctx == CMD_CTX_CANCELLED)
1342 if (blk_queue_dying(req->q))
1343 cqe.status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1345 cqe.status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1348 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1349 req->tag, nvmeq->qid);
1350 ctx = cancel_cmd_info(cmd, &fn);
1351 fn(nvmeq, ctx, &cqe);
1354 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1356 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1357 struct nvme_queue *nvmeq = cmd->nvmeq;
1359 dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1361 spin_lock_irq(&nvmeq->q_lock);
1362 nvme_abort_req(req);
1363 spin_unlock_irq(&nvmeq->q_lock);
1366 * The aborted req will be completed on receiving the abort req.
1367 * We enable the timer again. If hit twice, it'll cause a device reset,
1368 * as the device then is in a faulty state.
1370 return BLK_EH_RESET_TIMER;
1373 static void nvme_free_queue(struct nvme_queue *nvmeq)
1375 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1376 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1378 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1379 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1383 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1387 for (i = dev->queue_count - 1; i >= lowest; i--) {
1388 struct nvme_queue *nvmeq = dev->queues[i];
1390 dev->queues[i] = NULL;
1391 nvme_free_queue(nvmeq);
1396 * nvme_suspend_queue - put queue into suspended state
1397 * @nvmeq - queue to suspend
1399 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1403 spin_lock_irq(&nvmeq->q_lock);
1404 if (nvmeq->cq_vector == -1) {
1405 spin_unlock_irq(&nvmeq->q_lock);
1408 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1409 nvmeq->dev->online_queues--;
1410 nvmeq->cq_vector = -1;
1411 spin_unlock_irq(&nvmeq->q_lock);
1413 if (!nvmeq->qid && nvmeq->dev->admin_q)
1414 blk_mq_freeze_queue_start(nvmeq->dev->admin_q);
1416 irq_set_affinity_hint(vector, NULL);
1417 free_irq(vector, nvmeq);
1422 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1424 spin_lock_irq(&nvmeq->q_lock);
1425 if (nvmeq->tags && *nvmeq->tags)
1426 blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq);
1427 spin_unlock_irq(&nvmeq->q_lock);
1430 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1432 struct nvme_queue *nvmeq = dev->queues[qid];
1436 if (nvme_suspend_queue(nvmeq))
1439 /* Don't tell the adapter to delete the admin queue.
1440 * Don't tell a removed adapter to delete IO queues. */
1441 if (qid && readl(&dev->bar->csts) != -1) {
1442 adapter_delete_sq(dev, qid);
1443 adapter_delete_cq(dev, qid);
1446 spin_lock_irq(&nvmeq->q_lock);
1447 nvme_process_cq(nvmeq);
1448 spin_unlock_irq(&nvmeq->q_lock);
1451 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1454 int q_depth = dev->q_depth;
1455 unsigned q_size_aligned = roundup(q_depth * entry_size, dev->page_size);
1457 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1458 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1459 mem_per_q = round_down(mem_per_q, dev->page_size);
1460 q_depth = div_u64(mem_per_q, entry_size);
1463 * Ensure the reduced q_depth is above some threshold where it
1464 * would be better to map queues in system memory with the
1474 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1477 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1478 unsigned offset = (qid - 1) *
1479 roundup(SQ_SIZE(depth), dev->page_size);
1480 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1481 nvmeq->sq_cmds_io = dev->cmb + offset;
1483 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1484 &nvmeq->sq_dma_addr, GFP_KERNEL);
1485 if (!nvmeq->sq_cmds)
1492 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1495 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1499 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1500 &nvmeq->cq_dma_addr, GFP_KERNEL);
1504 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1507 nvmeq->q_dmadev = dev->dev;
1509 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1510 dev->instance, qid);
1511 spin_lock_init(&nvmeq->q_lock);
1513 nvmeq->cq_phase = 1;
1514 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1515 nvmeq->q_depth = depth;
1517 nvmeq->cq_vector = -1;
1518 dev->queues[qid] = nvmeq;
1520 /* make sure queue descriptor is set before queue count, for kthread */
1527 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1528 nvmeq->cq_dma_addr);
1534 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1537 if (use_threaded_interrupts)
1538 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1539 nvme_irq_check, nvme_irq, IRQF_SHARED,
1541 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1542 IRQF_SHARED, name, nvmeq);
1545 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1547 struct nvme_dev *dev = nvmeq->dev;
1549 spin_lock_irq(&nvmeq->q_lock);
1552 nvmeq->cq_phase = 1;
1553 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1554 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1555 dev->online_queues++;
1556 spin_unlock_irq(&nvmeq->q_lock);
1559 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1561 struct nvme_dev *dev = nvmeq->dev;
1564 nvmeq->cq_vector = qid - 1;
1565 result = adapter_alloc_cq(dev, qid, nvmeq);
1569 result = adapter_alloc_sq(dev, qid, nvmeq);
1573 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1577 nvme_init_queue(nvmeq, qid);
1581 adapter_delete_sq(dev, qid);
1583 adapter_delete_cq(dev, qid);
1587 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1589 unsigned long timeout;
1590 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1592 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1594 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1596 if (fatal_signal_pending(current))
1598 if (time_after(jiffies, timeout)) {
1600 "Device not ready; aborting %s\n", enabled ?
1601 "initialisation" : "reset");
1610 * If the device has been passed off to us in an enabled state, just clear
1611 * the enabled bit. The spec says we should set the 'shutdown notification
1612 * bits', but doing so may cause the device to complete commands to the
1613 * admin queue ... and we don't know what memory that might be pointing at!
1615 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1617 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1618 dev->ctrl_config &= ~NVME_CC_ENABLE;
1619 writel(dev->ctrl_config, &dev->bar->cc);
1621 return nvme_wait_ready(dev, cap, false);
1624 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1626 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1627 dev->ctrl_config |= NVME_CC_ENABLE;
1628 writel(dev->ctrl_config, &dev->bar->cc);
1630 return nvme_wait_ready(dev, cap, true);
1633 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1635 unsigned long timeout;
1637 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1638 dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1640 writel(dev->ctrl_config, &dev->bar->cc);
1642 timeout = SHUTDOWN_TIMEOUT + jiffies;
1643 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1644 NVME_CSTS_SHST_CMPLT) {
1646 if (fatal_signal_pending(current))
1648 if (time_after(jiffies, timeout)) {
1650 "Device shutdown incomplete; abort shutdown\n");
1658 static struct blk_mq_ops nvme_mq_admin_ops = {
1659 .queue_rq = nvme_queue_rq,
1660 .map_queue = blk_mq_map_queue,
1661 .init_hctx = nvme_admin_init_hctx,
1662 .exit_hctx = nvme_admin_exit_hctx,
1663 .init_request = nvme_admin_init_request,
1664 .timeout = nvme_timeout,
1667 static struct blk_mq_ops nvme_mq_ops = {
1668 .queue_rq = nvme_queue_rq,
1669 .map_queue = blk_mq_map_queue,
1670 .init_hctx = nvme_init_hctx,
1671 .init_request = nvme_init_request,
1672 .timeout = nvme_timeout,
1675 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1677 if (dev->admin_q && !blk_queue_dying(dev->admin_q)) {
1678 blk_cleanup_queue(dev->admin_q);
1679 blk_mq_free_tag_set(&dev->admin_tagset);
1683 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1685 if (!dev->admin_q) {
1686 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1687 dev->admin_tagset.nr_hw_queues = 1;
1688 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1689 dev->admin_tagset.reserved_tags = 1;
1690 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1691 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1692 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1693 dev->admin_tagset.driver_data = dev;
1695 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1698 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1699 if (IS_ERR(dev->admin_q)) {
1700 blk_mq_free_tag_set(&dev->admin_tagset);
1703 if (!blk_get_queue(dev->admin_q)) {
1704 nvme_dev_remove_admin(dev);
1705 dev->admin_q = NULL;
1709 blk_mq_unfreeze_queue(dev->admin_q);
1714 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1718 u64 cap = readq(&dev->bar->cap);
1719 struct nvme_queue *nvmeq;
1720 unsigned page_shift = PAGE_SHIFT;
1721 unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1722 unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1724 if (page_shift < dev_page_min) {
1726 "Minimum device page size (%u) too large for "
1727 "host (%u)\n", 1 << dev_page_min,
1731 if (page_shift > dev_page_max) {
1733 "Device maximum page size (%u) smaller than "
1734 "host (%u); enabling work-around\n",
1735 1 << dev_page_max, 1 << page_shift);
1736 page_shift = dev_page_max;
1739 dev->subsystem = readl(&dev->bar->vs) >= NVME_VS(1, 1) ?
1740 NVME_CAP_NSSRC(cap) : 0;
1742 if (dev->subsystem && (readl(&dev->bar->csts) & NVME_CSTS_NSSRO))
1743 writel(NVME_CSTS_NSSRO, &dev->bar->csts);
1745 result = nvme_disable_ctrl(dev, cap);
1749 nvmeq = dev->queues[0];
1751 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1756 aqa = nvmeq->q_depth - 1;
1759 dev->page_size = 1 << page_shift;
1761 dev->ctrl_config = NVME_CC_CSS_NVM;
1762 dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1763 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1764 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1766 writel(aqa, &dev->bar->aqa);
1767 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1768 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1770 result = nvme_enable_ctrl(dev, cap);
1774 nvmeq->cq_vector = 0;
1775 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1777 nvmeq->cq_vector = -1;
1784 nvme_free_queues(dev, 0);
1788 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1790 struct nvme_dev *dev = ns->dev;
1791 struct nvme_user_io io;
1792 struct nvme_command c;
1793 unsigned length, meta_len;
1795 dma_addr_t meta_dma = 0;
1797 void __user *metadata;
1799 if (copy_from_user(&io, uio, sizeof(io)))
1802 switch (io.opcode) {
1803 case nvme_cmd_write:
1805 case nvme_cmd_compare:
1811 length = (io.nblocks + 1) << ns->lba_shift;
1812 meta_len = (io.nblocks + 1) * ns->ms;
1813 metadata = (void __user *)(uintptr_t)io.metadata;
1814 write = io.opcode & 1;
1821 if (((io.metadata & 3) || !io.metadata) && !ns->ext)
1824 meta = dma_alloc_coherent(dev->dev, meta_len,
1825 &meta_dma, GFP_KERNEL);
1832 if (copy_from_user(meta, metadata, meta_len)) {
1839 memset(&c, 0, sizeof(c));
1840 c.rw.opcode = io.opcode;
1841 c.rw.flags = io.flags;
1842 c.rw.nsid = cpu_to_le32(ns->ns_id);
1843 c.rw.slba = cpu_to_le64(io.slba);
1844 c.rw.length = cpu_to_le16(io.nblocks);
1845 c.rw.control = cpu_to_le16(io.control);
1846 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1847 c.rw.reftag = cpu_to_le32(io.reftag);
1848 c.rw.apptag = cpu_to_le16(io.apptag);
1849 c.rw.appmask = cpu_to_le16(io.appmask);
1850 c.rw.metadata = cpu_to_le64(meta_dma);
1852 status = __nvme_submit_sync_cmd(ns->queue, &c, NULL,
1853 (void __user *)(uintptr_t)io.addr, length, NULL, 0);
1856 if (status == NVME_SC_SUCCESS && !write) {
1857 if (copy_to_user(metadata, meta, meta_len))
1860 dma_free_coherent(dev->dev, meta_len, meta, meta_dma);
1865 static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1866 struct nvme_passthru_cmd __user *ucmd)
1868 struct nvme_passthru_cmd cmd;
1869 struct nvme_command c;
1870 unsigned timeout = 0;
1873 if (!capable(CAP_SYS_ADMIN))
1875 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1878 memset(&c, 0, sizeof(c));
1879 c.common.opcode = cmd.opcode;
1880 c.common.flags = cmd.flags;
1881 c.common.nsid = cpu_to_le32(cmd.nsid);
1882 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1883 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1884 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1885 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1886 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1887 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1888 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1889 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1892 timeout = msecs_to_jiffies(cmd.timeout_ms);
1894 status = __nvme_submit_sync_cmd(ns ? ns->queue : dev->admin_q, &c,
1895 NULL, (void __user *)(uintptr_t)cmd.addr, cmd.data_len,
1896 &cmd.result, timeout);
1898 if (put_user(cmd.result, &ucmd->result))
1905 static int nvme_subsys_reset(struct nvme_dev *dev)
1907 if (!dev->subsystem)
1910 writel(0x4E564D65, &dev->bar->nssr); /* "NVMe" */
1914 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1917 struct nvme_ns *ns = bdev->bd_disk->private_data;
1921 force_successful_syscall_return();
1923 case NVME_IOCTL_ADMIN_CMD:
1924 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1925 case NVME_IOCTL_IO_CMD:
1926 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1927 case NVME_IOCTL_SUBMIT_IO:
1928 return nvme_submit_io(ns, (void __user *)arg);
1929 case SG_GET_VERSION_NUM:
1930 return nvme_sg_get_version_num((void __user *)arg);
1932 return nvme_sg_io(ns, (void __user *)arg);
1938 #ifdef CONFIG_COMPAT
1939 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1940 unsigned int cmd, unsigned long arg)
1944 return -ENOIOCTLCMD;
1946 return nvme_ioctl(bdev, mode, cmd, arg);
1949 #define nvme_compat_ioctl NULL
1952 static int nvme_open(struct block_device *bdev, fmode_t mode)
1957 spin_lock(&dev_list_lock);
1958 ns = bdev->bd_disk->private_data;
1961 else if (!kref_get_unless_zero(&ns->dev->kref))
1963 spin_unlock(&dev_list_lock);
1968 static void nvme_free_dev(struct kref *kref);
1970 static void nvme_release(struct gendisk *disk, fmode_t mode)
1972 struct nvme_ns *ns = disk->private_data;
1973 struct nvme_dev *dev = ns->dev;
1975 kref_put(&dev->kref, nvme_free_dev);
1978 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1980 /* some standard values */
1981 geo->heads = 1 << 6;
1982 geo->sectors = 1 << 5;
1983 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1987 static void nvme_config_discard(struct nvme_ns *ns)
1989 u32 logical_block_size = queue_logical_block_size(ns->queue);
1990 ns->queue->limits.discard_zeroes_data = 0;
1991 ns->queue->limits.discard_alignment = logical_block_size;
1992 ns->queue->limits.discard_granularity = logical_block_size;
1993 blk_queue_max_discard_sectors(ns->queue, 0xffffffff);
1994 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1997 static int nvme_revalidate_disk(struct gendisk *disk)
1999 struct nvme_ns *ns = disk->private_data;
2000 struct nvme_dev *dev = ns->dev;
2001 struct nvme_id_ns *id;
2006 if (nvme_identify_ns(dev, ns->ns_id, &id)) {
2007 dev_warn(dev->dev, "%s: Identify failure nvme%dn%d\n", __func__,
2008 dev->instance, ns->ns_id);
2011 if (id->ncap == 0) {
2017 lbaf = id->flbas & NVME_NS_FLBAS_LBA_MASK;
2018 ns->lba_shift = id->lbaf[lbaf].ds;
2019 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
2020 ns->ext = ns->ms && (id->flbas & NVME_NS_FLBAS_META_EXT);
2023 * If identify namespace failed, use default 512 byte block size so
2024 * block layer can use before failing read/write for 0 capacity.
2026 if (ns->lba_shift == 0)
2028 bs = 1 << ns->lba_shift;
2030 /* XXX: PI implementation requires metadata equal t10 pi tuple size */
2031 pi_type = ns->ms == sizeof(struct t10_pi_tuple) ?
2032 id->dps & NVME_NS_DPS_PI_MASK : 0;
2034 if (blk_get_integrity(disk) && (ns->pi_type != pi_type ||
2036 bs != queue_logical_block_size(disk->queue) ||
2037 (ns->ms && ns->ext)))
2038 blk_integrity_unregister(disk);
2040 ns->pi_type = pi_type;
2041 blk_queue_logical_block_size(ns->queue, bs);
2043 if (ns->ms && !blk_get_integrity(disk) && (disk->flags & GENHD_FL_UP) &&
2045 nvme_init_integrity(ns);
2047 if (ns->ms && !(ns->ms == 8 && ns->pi_type) && !blk_get_integrity(disk))
2048 set_capacity(disk, 0);
2050 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
2052 if (dev->oncs & NVME_CTRL_ONCS_DSM)
2053 nvme_config_discard(ns);
2059 static const struct block_device_operations nvme_fops = {
2060 .owner = THIS_MODULE,
2061 .ioctl = nvme_ioctl,
2062 .compat_ioctl = nvme_compat_ioctl,
2064 .release = nvme_release,
2065 .getgeo = nvme_getgeo,
2066 .revalidate_disk= nvme_revalidate_disk,
2069 static int nvme_kthread(void *data)
2071 struct nvme_dev *dev, *next;
2073 while (!kthread_should_stop()) {
2074 set_current_state(TASK_INTERRUPTIBLE);
2075 spin_lock(&dev_list_lock);
2076 list_for_each_entry_safe(dev, next, &dev_list, node) {
2078 u32 csts = readl(&dev->bar->csts);
2080 if ((dev->subsystem && (csts & NVME_CSTS_NSSRO)) ||
2081 csts & NVME_CSTS_CFS) {
2082 if (work_busy(&dev->reset_work))
2084 list_del_init(&dev->node);
2086 "Failed status: %x, reset controller\n",
2087 readl(&dev->bar->csts));
2088 dev->reset_workfn = nvme_reset_failed_dev;
2089 queue_work(nvme_workq, &dev->reset_work);
2092 for (i = 0; i < dev->queue_count; i++) {
2093 struct nvme_queue *nvmeq = dev->queues[i];
2096 spin_lock_irq(&nvmeq->q_lock);
2097 nvme_process_cq(nvmeq);
2099 while ((i == 0) && (dev->event_limit > 0)) {
2100 if (nvme_submit_async_admin_req(dev))
2104 spin_unlock_irq(&nvmeq->q_lock);
2107 spin_unlock(&dev_list_lock);
2108 schedule_timeout(round_jiffies_relative(HZ));
2113 static void nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid)
2116 struct gendisk *disk;
2117 int node = dev_to_node(dev->dev);
2119 ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
2123 ns->queue = blk_mq_init_queue(&dev->tagset);
2124 if (IS_ERR(ns->queue))
2126 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
2127 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
2129 ns->queue->queuedata = ns;
2131 disk = alloc_disk_node(0, node);
2133 goto out_free_queue;
2137 ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
2138 list_add_tail(&ns->list, &dev->namespaces);
2140 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2141 if (dev->max_hw_sectors) {
2142 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2143 blk_queue_max_segments(ns->queue,
2144 ((dev->max_hw_sectors << 9) / dev->page_size) + 1);
2146 if (dev->stripe_size)
2147 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
2148 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2149 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2150 blk_queue_virt_boundary(ns->queue, dev->page_size - 1);
2152 disk->major = nvme_major;
2153 disk->first_minor = 0;
2154 disk->fops = &nvme_fops;
2155 disk->private_data = ns;
2156 disk->queue = ns->queue;
2157 disk->driverfs_dev = dev->device;
2158 disk->flags = GENHD_FL_EXT_DEVT;
2159 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2162 * Initialize capacity to 0 until we establish the namespace format and
2163 * setup integrity extentions if necessary. The revalidate_disk after
2164 * add_disk allows the driver to register with integrity if the format
2167 set_capacity(disk, 0);
2168 if (nvme_revalidate_disk(ns->disk))
2173 struct block_device *bd = bdget_disk(ns->disk, 0);
2176 if (blkdev_get(bd, FMODE_READ, NULL)) {
2180 blkdev_reread_part(bd);
2181 blkdev_put(bd, FMODE_READ);
2186 list_del(&ns->list);
2188 blk_cleanup_queue(ns->queue);
2193 static void nvme_create_io_queues(struct nvme_dev *dev)
2197 for (i = dev->queue_count; i <= dev->max_qid; i++)
2198 if (!nvme_alloc_queue(dev, i, dev->q_depth))
2201 for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
2202 if (nvme_create_queue(dev->queues[i], i))
2206 static int set_queue_count(struct nvme_dev *dev, int count)
2210 u32 q_count = (count - 1) | ((count - 1) << 16);
2212 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2217 dev_err(dev->dev, "Could not set queue count (%d)\n", status);
2220 return min(result & 0xffff, result >> 16) + 1;
2223 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
2225 u64 szu, size, offset;
2227 resource_size_t bar_size;
2228 struct pci_dev *pdev = to_pci_dev(dev->dev);
2230 dma_addr_t dma_addr;
2235 dev->cmbsz = readl(&dev->bar->cmbsz);
2236 if (!(NVME_CMB_SZ(dev->cmbsz)))
2239 cmbloc = readl(&dev->bar->cmbloc);
2241 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
2242 size = szu * NVME_CMB_SZ(dev->cmbsz);
2243 offset = szu * NVME_CMB_OFST(cmbloc);
2244 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(cmbloc));
2246 if (offset > bar_size)
2250 * Controllers may support a CMB size larger than their BAR,
2251 * for example, due to being behind a bridge. Reduce the CMB to
2252 * the reported size of the BAR
2254 if (size > bar_size - offset)
2255 size = bar_size - offset;
2257 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(cmbloc)) + offset;
2258 cmb = ioremap_wc(dma_addr, size);
2262 dev->cmb_dma_addr = dma_addr;
2263 dev->cmb_size = size;
2267 static inline void nvme_release_cmb(struct nvme_dev *dev)
2275 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2277 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2280 static int nvme_setup_io_queues(struct nvme_dev *dev)
2282 struct nvme_queue *adminq = dev->queues[0];
2283 struct pci_dev *pdev = to_pci_dev(dev->dev);
2284 int result, i, vecs, nr_io_queues, size;
2286 nr_io_queues = num_possible_cpus();
2287 result = set_queue_count(dev, nr_io_queues);
2290 if (result < nr_io_queues)
2291 nr_io_queues = result;
2293 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
2294 result = nvme_cmb_qdepth(dev, nr_io_queues,
2295 sizeof(struct nvme_command));
2297 dev->q_depth = result;
2299 nvme_release_cmb(dev);
2302 size = db_bar_size(dev, nr_io_queues);
2306 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2309 if (!--nr_io_queues)
2311 size = db_bar_size(dev, nr_io_queues);
2313 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2314 adminq->q_db = dev->dbs;
2317 /* Deregister the admin queue's interrupt */
2318 free_irq(dev->entry[0].vector, adminq);
2321 * If we enable msix early due to not intx, disable it again before
2322 * setting up the full range we need.
2325 pci_disable_msix(pdev);
2327 for (i = 0; i < nr_io_queues; i++)
2328 dev->entry[i].entry = i;
2329 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2331 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2335 for (i = 0; i < vecs; i++)
2336 dev->entry[i].vector = i + pdev->irq;
2341 * Should investigate if there's a performance win from allocating
2342 * more queues than interrupt vectors; it might allow the submission
2343 * path to scale better, even if the receive path is limited by the
2344 * number of interrupts.
2346 nr_io_queues = vecs;
2347 dev->max_qid = nr_io_queues;
2349 result = queue_request_irq(dev, adminq, adminq->irqname);
2351 adminq->cq_vector = -1;
2355 /* Free previously allocated queues that are no longer usable */
2356 nvme_free_queues(dev, nr_io_queues + 1);
2357 nvme_create_io_queues(dev);
2362 nvme_free_queues(dev, 1);
2366 static void nvme_free_namespace(struct nvme_ns *ns)
2368 list_del(&ns->list);
2370 spin_lock(&dev_list_lock);
2371 ns->disk->private_data = NULL;
2372 spin_unlock(&dev_list_lock);
2378 static int ns_cmp(void *priv, struct list_head *a, struct list_head *b)
2380 struct nvme_ns *nsa = container_of(a, struct nvme_ns, list);
2381 struct nvme_ns *nsb = container_of(b, struct nvme_ns, list);
2383 return nsa->ns_id - nsb->ns_id;
2386 static struct nvme_ns *nvme_find_ns(struct nvme_dev *dev, unsigned nsid)
2390 list_for_each_entry(ns, &dev->namespaces, list) {
2391 if (ns->ns_id == nsid)
2393 if (ns->ns_id > nsid)
2399 static inline bool nvme_io_incapable(struct nvme_dev *dev)
2401 return (!dev->bar || readl(&dev->bar->csts) & NVME_CSTS_CFS ||
2402 dev->online_queues < 2);
2405 static void nvme_ns_remove(struct nvme_ns *ns)
2407 bool kill = nvme_io_incapable(ns->dev) && !blk_queue_dying(ns->queue);
2410 blk_set_queue_dying(ns->queue);
2411 if (ns->disk->flags & GENHD_FL_UP) {
2412 if (blk_get_integrity(ns->disk))
2413 blk_integrity_unregister(ns->disk);
2414 del_gendisk(ns->disk);
2416 if (kill || !blk_queue_dying(ns->queue)) {
2417 blk_mq_abort_requeue_list(ns->queue);
2418 blk_cleanup_queue(ns->queue);
2422 static void nvme_scan_namespaces(struct nvme_dev *dev, unsigned nn)
2424 struct nvme_ns *ns, *next;
2427 for (i = 1; i <= nn; i++) {
2428 ns = nvme_find_ns(dev, i);
2430 if (revalidate_disk(ns->disk)) {
2432 nvme_free_namespace(ns);
2435 nvme_alloc_ns(dev, i);
2437 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2438 if (ns->ns_id > nn) {
2440 nvme_free_namespace(ns);
2443 list_sort(NULL, &dev->namespaces, ns_cmp);
2446 static void nvme_set_irq_hints(struct nvme_dev *dev)
2448 struct nvme_queue *nvmeq;
2451 for (i = 0; i < dev->online_queues; i++) {
2452 nvmeq = dev->queues[i];
2454 if (!nvmeq->tags || !(*nvmeq->tags))
2457 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2458 blk_mq_tags_cpumask(*nvmeq->tags));
2462 static void nvme_dev_scan(struct work_struct *work)
2464 struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work);
2465 struct nvme_id_ctrl *ctrl;
2467 if (!dev->tagset.tags)
2469 if (nvme_identify_ctrl(dev, &ctrl))
2471 nvme_scan_namespaces(dev, le32_to_cpup(&ctrl->nn));
2473 nvme_set_irq_hints(dev);
2477 * Return: error value if an error occurred setting up the queues or calling
2478 * Identify Device. 0 if these succeeded, even if adding some of the
2479 * namespaces failed. At the moment, these failures are silent. TBD which
2480 * failures should be reported.
2482 static int nvme_dev_add(struct nvme_dev *dev)
2484 struct pci_dev *pdev = to_pci_dev(dev->dev);
2486 struct nvme_id_ctrl *ctrl;
2487 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2489 res = nvme_identify_ctrl(dev, &ctrl);
2491 dev_err(dev->dev, "Identify Controller failed (%d)\n", res);
2495 dev->oncs = le16_to_cpup(&ctrl->oncs);
2496 dev->abort_limit = ctrl->acl + 1;
2497 dev->vwc = ctrl->vwc;
2498 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2499 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2500 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2502 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2503 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2504 (pdev->device == 0x0953) && ctrl->vs[3]) {
2505 unsigned int max_hw_sectors;
2507 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2508 max_hw_sectors = dev->stripe_size >> (shift - 9);
2509 if (dev->max_hw_sectors) {
2510 dev->max_hw_sectors = min(max_hw_sectors,
2511 dev->max_hw_sectors);
2513 dev->max_hw_sectors = max_hw_sectors;
2517 if (!dev->tagset.tags) {
2518 dev->tagset.ops = &nvme_mq_ops;
2519 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2520 dev->tagset.timeout = NVME_IO_TIMEOUT;
2521 dev->tagset.numa_node = dev_to_node(dev->dev);
2522 dev->tagset.queue_depth =
2523 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2524 dev->tagset.cmd_size = nvme_cmd_size(dev);
2525 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2526 dev->tagset.driver_data = dev;
2528 if (blk_mq_alloc_tag_set(&dev->tagset))
2531 schedule_work(&dev->scan_work);
2535 static int nvme_dev_map(struct nvme_dev *dev)
2538 int bars, result = -ENOMEM;
2539 struct pci_dev *pdev = to_pci_dev(dev->dev);
2541 if (pci_enable_device_mem(pdev))
2544 dev->entry[0].vector = pdev->irq;
2545 pci_set_master(pdev);
2546 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2550 if (pci_request_selected_regions(pdev, bars, "nvme"))
2553 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
2554 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
2557 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2561 if (readl(&dev->bar->csts) == -1) {
2567 * Some devices don't advertse INTx interrupts, pre-enable a single
2568 * MSIX vec for setup. We'll adjust this later.
2571 result = pci_enable_msix(pdev, dev->entry, 1);
2576 cap = readq(&dev->bar->cap);
2577 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2578 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2579 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2580 if (readl(&dev->bar->vs) >= NVME_VS(1, 2))
2581 dev->cmb = nvme_map_cmb(dev);
2589 pci_release_regions(pdev);
2591 pci_disable_device(pdev);
2595 static void nvme_dev_unmap(struct nvme_dev *dev)
2597 struct pci_dev *pdev = to_pci_dev(dev->dev);
2599 if (pdev->msi_enabled)
2600 pci_disable_msi(pdev);
2601 else if (pdev->msix_enabled)
2602 pci_disable_msix(pdev);
2607 pci_release_regions(pdev);
2610 if (pci_is_enabled(pdev))
2611 pci_disable_device(pdev);
2614 struct nvme_delq_ctx {
2615 struct task_struct *waiter;
2616 struct kthread_worker *worker;
2620 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2622 dq->waiter = current;
2626 set_current_state(TASK_KILLABLE);
2627 if (!atomic_read(&dq->refcount))
2629 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2630 fatal_signal_pending(current)) {
2632 * Disable the controller first since we can't trust it
2633 * at this point, but leave the admin queue enabled
2634 * until all queue deletion requests are flushed.
2635 * FIXME: This may take a while if there are more h/w
2636 * queues than admin tags.
2638 set_current_state(TASK_RUNNING);
2639 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2640 nvme_clear_queue(dev->queues[0]);
2641 flush_kthread_worker(dq->worker);
2642 nvme_disable_queue(dev, 0);
2646 set_current_state(TASK_RUNNING);
2649 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2651 atomic_dec(&dq->refcount);
2653 wake_up_process(dq->waiter);
2656 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2658 atomic_inc(&dq->refcount);
2662 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2664 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2668 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2669 kthread_work_func_t fn)
2671 struct nvme_command c;
2673 memset(&c, 0, sizeof(c));
2674 c.delete_queue.opcode = opcode;
2675 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2677 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2678 return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2682 static void nvme_del_cq_work_handler(struct kthread_work *work)
2684 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2686 nvme_del_queue_end(nvmeq);
2689 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2691 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2692 nvme_del_cq_work_handler);
2695 static void nvme_del_sq_work_handler(struct kthread_work *work)
2697 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2699 int status = nvmeq->cmdinfo.status;
2702 status = nvme_delete_cq(nvmeq);
2704 nvme_del_queue_end(nvmeq);
2707 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2709 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2710 nvme_del_sq_work_handler);
2713 static void nvme_del_queue_start(struct kthread_work *work)
2715 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2717 if (nvme_delete_sq(nvmeq))
2718 nvme_del_queue_end(nvmeq);
2721 static void nvme_disable_io_queues(struct nvme_dev *dev)
2724 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2725 struct nvme_delq_ctx dq;
2726 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2727 &worker, "nvme%d", dev->instance);
2729 if (IS_ERR(kworker_task)) {
2731 "Failed to create queue del task\n");
2732 for (i = dev->queue_count - 1; i > 0; i--)
2733 nvme_disable_queue(dev, i);
2738 atomic_set(&dq.refcount, 0);
2739 dq.worker = &worker;
2740 for (i = dev->queue_count - 1; i > 0; i--) {
2741 struct nvme_queue *nvmeq = dev->queues[i];
2743 if (nvme_suspend_queue(nvmeq))
2745 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2746 nvmeq->cmdinfo.worker = dq.worker;
2747 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2748 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2750 nvme_wait_dq(&dq, dev);
2751 kthread_stop(kworker_task);
2755 * Remove the node from the device list and check
2756 * for whether or not we need to stop the nvme_thread.
2758 static void nvme_dev_list_remove(struct nvme_dev *dev)
2760 struct task_struct *tmp = NULL;
2762 spin_lock(&dev_list_lock);
2763 list_del_init(&dev->node);
2764 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2768 spin_unlock(&dev_list_lock);
2774 static void nvme_freeze_queues(struct nvme_dev *dev)
2778 list_for_each_entry(ns, &dev->namespaces, list) {
2779 blk_mq_freeze_queue_start(ns->queue);
2781 spin_lock_irq(ns->queue->queue_lock);
2782 queue_flag_set(QUEUE_FLAG_STOPPED, ns->queue);
2783 spin_unlock_irq(ns->queue->queue_lock);
2785 blk_mq_cancel_requeue_work(ns->queue);
2786 blk_mq_stop_hw_queues(ns->queue);
2790 static void nvme_unfreeze_queues(struct nvme_dev *dev)
2794 list_for_each_entry(ns, &dev->namespaces, list) {
2795 queue_flag_clear_unlocked(QUEUE_FLAG_STOPPED, ns->queue);
2796 blk_mq_unfreeze_queue(ns->queue);
2797 blk_mq_start_stopped_hw_queues(ns->queue, true);
2798 blk_mq_kick_requeue_list(ns->queue);
2802 static void nvme_dev_shutdown(struct nvme_dev *dev)
2807 nvme_dev_list_remove(dev);
2810 nvme_freeze_queues(dev);
2811 csts = readl(&dev->bar->csts);
2813 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2814 for (i = dev->queue_count - 1; i >= 0; i--) {
2815 struct nvme_queue *nvmeq = dev->queues[i];
2816 nvme_suspend_queue(nvmeq);
2819 nvme_disable_io_queues(dev);
2820 nvme_shutdown_ctrl(dev);
2821 nvme_disable_queue(dev, 0);
2823 nvme_dev_unmap(dev);
2825 for (i = dev->queue_count - 1; i >= 0; i--)
2826 nvme_clear_queue(dev->queues[i]);
2829 static void nvme_dev_remove(struct nvme_dev *dev)
2833 list_for_each_entry(ns, &dev->namespaces, list)
2837 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2839 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2840 PAGE_SIZE, PAGE_SIZE, 0);
2841 if (!dev->prp_page_pool)
2844 /* Optimisation for I/Os between 4k and 128k */
2845 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2847 if (!dev->prp_small_pool) {
2848 dma_pool_destroy(dev->prp_page_pool);
2854 static void nvme_release_prp_pools(struct nvme_dev *dev)
2856 dma_pool_destroy(dev->prp_page_pool);
2857 dma_pool_destroy(dev->prp_small_pool);
2860 static DEFINE_IDA(nvme_instance_ida);
2862 static int nvme_set_instance(struct nvme_dev *dev)
2864 int instance, error;
2867 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2870 spin_lock(&dev_list_lock);
2871 error = ida_get_new(&nvme_instance_ida, &instance);
2872 spin_unlock(&dev_list_lock);
2873 } while (error == -EAGAIN);
2878 dev->instance = instance;
2882 static void nvme_release_instance(struct nvme_dev *dev)
2884 spin_lock(&dev_list_lock);
2885 ida_remove(&nvme_instance_ida, dev->instance);
2886 spin_unlock(&dev_list_lock);
2889 static void nvme_free_namespaces(struct nvme_dev *dev)
2891 struct nvme_ns *ns, *next;
2893 list_for_each_entry_safe(ns, next, &dev->namespaces, list)
2894 nvme_free_namespace(ns);
2897 static void nvme_free_dev(struct kref *kref)
2899 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2901 put_device(dev->dev);
2902 put_device(dev->device);
2903 nvme_free_namespaces(dev);
2904 nvme_release_instance(dev);
2905 if (dev->tagset.tags)
2906 blk_mq_free_tag_set(&dev->tagset);
2908 blk_put_queue(dev->admin_q);
2914 static int nvme_dev_open(struct inode *inode, struct file *f)
2916 struct nvme_dev *dev;
2917 int instance = iminor(inode);
2920 spin_lock(&dev_list_lock);
2921 list_for_each_entry(dev, &dev_list, node) {
2922 if (dev->instance == instance) {
2923 if (!dev->admin_q) {
2927 if (!kref_get_unless_zero(&dev->kref))
2929 f->private_data = dev;
2934 spin_unlock(&dev_list_lock);
2939 static int nvme_dev_release(struct inode *inode, struct file *f)
2941 struct nvme_dev *dev = f->private_data;
2942 kref_put(&dev->kref, nvme_free_dev);
2946 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2948 struct nvme_dev *dev = f->private_data;
2952 case NVME_IOCTL_ADMIN_CMD:
2953 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2954 case NVME_IOCTL_IO_CMD:
2955 if (list_empty(&dev->namespaces))
2957 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2958 return nvme_user_cmd(dev, ns, (void __user *)arg);
2959 case NVME_IOCTL_RESET:
2960 dev_warn(dev->dev, "resetting controller\n");
2961 return nvme_reset(dev);
2962 case NVME_IOCTL_SUBSYS_RESET:
2963 return nvme_subsys_reset(dev);
2969 static const struct file_operations nvme_dev_fops = {
2970 .owner = THIS_MODULE,
2971 .open = nvme_dev_open,
2972 .release = nvme_dev_release,
2973 .unlocked_ioctl = nvme_dev_ioctl,
2974 .compat_ioctl = nvme_dev_ioctl,
2977 static int nvme_dev_start(struct nvme_dev *dev)
2980 bool start_thread = false;
2982 result = nvme_dev_map(dev);
2986 result = nvme_configure_admin_queue(dev);
2990 spin_lock(&dev_list_lock);
2991 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2992 start_thread = true;
2995 list_add(&dev->node, &dev_list);
2996 spin_unlock(&dev_list_lock);
2999 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
3000 wake_up_all(&nvme_kthread_wait);
3002 wait_event_killable(nvme_kthread_wait, nvme_thread);
3004 if (IS_ERR_OR_NULL(nvme_thread)) {
3005 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
3009 nvme_init_queue(dev->queues[0], 0);
3010 result = nvme_alloc_admin_tags(dev);
3014 result = nvme_setup_io_queues(dev);
3018 dev->event_limit = 1;
3022 nvme_dev_remove_admin(dev);
3023 blk_put_queue(dev->admin_q);
3024 dev->admin_q = NULL;
3025 dev->queues[0]->tags = NULL;
3027 nvme_disable_queue(dev, 0);
3028 nvme_dev_list_remove(dev);
3030 nvme_dev_unmap(dev);
3034 static int nvme_remove_dead_ctrl(void *arg)
3036 struct nvme_dev *dev = (struct nvme_dev *)arg;
3037 struct pci_dev *pdev = to_pci_dev(dev->dev);
3039 if (pci_get_drvdata(pdev))
3040 pci_stop_and_remove_bus_device_locked(pdev);
3041 kref_put(&dev->kref, nvme_free_dev);
3045 static void nvme_remove_disks(struct work_struct *ws)
3047 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
3049 nvme_free_queues(dev, 1);
3050 nvme_dev_remove(dev);
3053 static int nvme_dev_resume(struct nvme_dev *dev)
3057 ret = nvme_dev_start(dev);
3060 if (dev->online_queues < 2) {
3061 spin_lock(&dev_list_lock);
3062 dev->reset_workfn = nvme_remove_disks;
3063 queue_work(nvme_workq, &dev->reset_work);
3064 spin_unlock(&dev_list_lock);
3066 nvme_unfreeze_queues(dev);
3072 static void nvme_dead_ctrl(struct nvme_dev *dev)
3074 dev_warn(dev->dev, "Device failed to resume\n");
3075 kref_get(&dev->kref);
3076 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
3079 "Failed to start controller remove task\n");
3080 kref_put(&dev->kref, nvme_free_dev);
3084 static void nvme_dev_reset(struct nvme_dev *dev)
3086 bool in_probe = work_busy(&dev->probe_work);
3088 nvme_dev_shutdown(dev);
3090 /* Synchronize with device probe so that work will see failure status
3091 * and exit gracefully without trying to schedule another reset */
3092 flush_work(&dev->probe_work);
3094 /* Fail this device if reset occured during probe to avoid
3095 * infinite initialization loops. */
3097 nvme_dead_ctrl(dev);
3100 /* Schedule device resume asynchronously so the reset work is available
3101 * to cleanup errors that may occur during reinitialization */
3102 schedule_work(&dev->probe_work);
3105 static void nvme_reset_failed_dev(struct work_struct *ws)
3107 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
3108 nvme_dev_reset(dev);
3111 static void nvme_reset_workfn(struct work_struct *work)
3113 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
3114 dev->reset_workfn(work);
3117 static int nvme_reset(struct nvme_dev *dev)
3121 if (!dev->admin_q || blk_queue_dying(dev->admin_q))
3124 spin_lock(&dev_list_lock);
3125 if (!work_pending(&dev->reset_work)) {
3126 dev->reset_workfn = nvme_reset_failed_dev;
3127 queue_work(nvme_workq, &dev->reset_work);
3130 spin_unlock(&dev_list_lock);
3133 flush_work(&dev->reset_work);
3134 flush_work(&dev->probe_work);
3141 static ssize_t nvme_sysfs_reset(struct device *dev,
3142 struct device_attribute *attr, const char *buf,
3145 struct nvme_dev *ndev = dev_get_drvdata(dev);
3148 ret = nvme_reset(ndev);
3154 static DEVICE_ATTR(reset_controller, S_IWUSR, NULL, nvme_sysfs_reset);
3156 static void nvme_async_probe(struct work_struct *work);
3157 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3159 int node, result = -ENOMEM;
3160 struct nvme_dev *dev;
3162 node = dev_to_node(&pdev->dev);
3163 if (node == NUMA_NO_NODE)
3164 set_dev_node(&pdev->dev, 0);
3166 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
3169 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
3173 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
3178 INIT_LIST_HEAD(&dev->namespaces);
3179 dev->reset_workfn = nvme_reset_failed_dev;
3180 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
3181 dev->dev = get_device(&pdev->dev);
3182 pci_set_drvdata(pdev, dev);
3183 result = nvme_set_instance(dev);
3187 result = nvme_setup_prp_pools(dev);
3191 kref_init(&dev->kref);
3192 dev->device = device_create(nvme_class, &pdev->dev,
3193 MKDEV(nvme_char_major, dev->instance),
3194 dev, "nvme%d", dev->instance);
3195 if (IS_ERR(dev->device)) {
3196 result = PTR_ERR(dev->device);
3199 get_device(dev->device);
3200 dev_set_drvdata(dev->device, dev);
3202 result = device_create_file(dev->device, &dev_attr_reset_controller);
3206 INIT_LIST_HEAD(&dev->node);
3207 INIT_WORK(&dev->scan_work, nvme_dev_scan);
3208 INIT_WORK(&dev->probe_work, nvme_async_probe);
3209 schedule_work(&dev->probe_work);
3213 device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3214 put_device(dev->device);
3216 nvme_release_prp_pools(dev);
3218 nvme_release_instance(dev);
3220 put_device(dev->dev);
3228 static void nvme_async_probe(struct work_struct *work)
3230 struct nvme_dev *dev = container_of(work, struct nvme_dev, probe_work);
3232 if (nvme_dev_resume(dev) && !work_busy(&dev->reset_work))
3233 nvme_dead_ctrl(dev);
3236 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
3238 struct nvme_dev *dev = pci_get_drvdata(pdev);
3241 nvme_dev_shutdown(dev);
3243 nvme_dev_resume(dev);
3246 static void nvme_shutdown(struct pci_dev *pdev)
3248 struct nvme_dev *dev = pci_get_drvdata(pdev);
3249 nvme_dev_shutdown(dev);
3252 static void nvme_remove(struct pci_dev *pdev)
3254 struct nvme_dev *dev = pci_get_drvdata(pdev);
3256 spin_lock(&dev_list_lock);
3257 list_del_init(&dev->node);
3258 spin_unlock(&dev_list_lock);
3260 pci_set_drvdata(pdev, NULL);
3261 flush_work(&dev->probe_work);
3262 flush_work(&dev->reset_work);
3263 flush_work(&dev->scan_work);
3264 device_remove_file(dev->device, &dev_attr_reset_controller);
3265 nvme_dev_remove(dev);
3266 nvme_dev_shutdown(dev);
3267 nvme_dev_remove_admin(dev);
3268 device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3269 nvme_free_queues(dev, 0);
3270 nvme_release_cmb(dev);
3271 nvme_release_prp_pools(dev);
3272 kref_put(&dev->kref, nvme_free_dev);
3275 /* These functions are yet to be implemented */
3276 #define nvme_error_detected NULL
3277 #define nvme_dump_registers NULL
3278 #define nvme_link_reset NULL
3279 #define nvme_slot_reset NULL
3280 #define nvme_error_resume NULL
3282 #ifdef CONFIG_PM_SLEEP
3283 static int nvme_suspend(struct device *dev)
3285 struct pci_dev *pdev = to_pci_dev(dev);
3286 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3288 nvme_dev_shutdown(ndev);
3292 static int nvme_resume(struct device *dev)
3294 struct pci_dev *pdev = to_pci_dev(dev);
3295 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3297 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
3298 ndev->reset_workfn = nvme_reset_failed_dev;
3299 queue_work(nvme_workq, &ndev->reset_work);
3305 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
3307 static const struct pci_error_handlers nvme_err_handler = {
3308 .error_detected = nvme_error_detected,
3309 .mmio_enabled = nvme_dump_registers,
3310 .link_reset = nvme_link_reset,
3311 .slot_reset = nvme_slot_reset,
3312 .resume = nvme_error_resume,
3313 .reset_notify = nvme_reset_notify,
3316 /* Move to pci_ids.h later */
3317 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
3319 static const struct pci_device_id nvme_id_table[] = {
3320 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3323 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3325 static struct pci_driver nvme_driver = {
3327 .id_table = nvme_id_table,
3328 .probe = nvme_probe,
3329 .remove = nvme_remove,
3330 .shutdown = nvme_shutdown,
3332 .pm = &nvme_dev_pm_ops,
3334 .err_handler = &nvme_err_handler,
3337 static int __init nvme_init(void)
3341 init_waitqueue_head(&nvme_kthread_wait);
3343 nvme_workq = create_singlethread_workqueue("nvme");
3347 result = register_blkdev(nvme_major, "nvme");
3350 else if (result > 0)
3351 nvme_major = result;
3353 result = __register_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme",
3356 goto unregister_blkdev;
3357 else if (result > 0)
3358 nvme_char_major = result;
3360 nvme_class = class_create(THIS_MODULE, "nvme");
3361 if (IS_ERR(nvme_class)) {
3362 result = PTR_ERR(nvme_class);
3363 goto unregister_chrdev;
3366 result = pci_register_driver(&nvme_driver);
3372 class_destroy(nvme_class);
3374 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3376 unregister_blkdev(nvme_major, "nvme");
3378 destroy_workqueue(nvme_workq);
3382 static void __exit nvme_exit(void)
3384 pci_unregister_driver(&nvme_driver);
3385 unregister_blkdev(nvme_major, "nvme");
3386 destroy_workqueue(nvme_workq);
3387 class_destroy(nvme_class);
3388 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3389 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3393 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3394 MODULE_LICENSE("GPL");
3395 MODULE_VERSION("1.0");
3396 module_init(nvme_init);
3397 module_exit(nvme_exit);