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>
33 #include <linux/module.h>
34 #include <linux/moduleparam.h>
35 #include <linux/pci.h>
36 #include <linux/poison.h>
37 #include <linux/ptrace.h>
38 #include <linux/sched.h>
39 #include <linux/slab.h>
40 #include <linux/t10-pi.h>
41 #include <linux/types.h>
43 #include <asm-generic/io-64-nonatomic-lo-hi.h>
45 #define NVME_MINORS (1U << MINORBITS)
46 #define NVME_Q_DEPTH 1024
47 #define NVME_AQ_DEPTH 256
48 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
49 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
50 #define ADMIN_TIMEOUT (admin_timeout * HZ)
51 #define SHUTDOWN_TIMEOUT (shutdown_timeout * HZ)
53 static unsigned char admin_timeout = 60;
54 module_param(admin_timeout, byte, 0644);
55 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
57 unsigned char nvme_io_timeout = 30;
58 module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
59 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
61 static unsigned char shutdown_timeout = 5;
62 module_param(shutdown_timeout, byte, 0644);
63 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
65 static int nvme_major;
66 module_param(nvme_major, int, 0);
68 static int nvme_char_major;
69 module_param(nvme_char_major, int, 0);
71 static int use_threaded_interrupts;
72 module_param(use_threaded_interrupts, int, 0);
74 static DEFINE_SPINLOCK(dev_list_lock);
75 static LIST_HEAD(dev_list);
76 static struct task_struct *nvme_thread;
77 static struct workqueue_struct *nvme_workq;
78 static wait_queue_head_t nvme_kthread_wait;
80 static struct class *nvme_class;
82 static void nvme_reset_failed_dev(struct work_struct *ws);
83 static int nvme_process_cq(struct nvme_queue *nvmeq);
85 struct async_cmd_info {
86 struct kthread_work work;
87 struct kthread_worker *worker;
95 * An NVM Express queue. Each device has at least two (one for admin
96 * commands and one for I/O commands).
99 struct device *q_dmadev;
100 struct nvme_dev *dev;
101 char irqname[24]; /* nvme4294967295-65535\0 */
103 struct nvme_command *sq_cmds;
104 volatile struct nvme_completion *cqes;
105 dma_addr_t sq_dma_addr;
106 dma_addr_t cq_dma_addr;
116 struct async_cmd_info cmdinfo;
117 struct blk_mq_hw_ctx *hctx;
121 * Check we didin't inadvertently grow the command struct
123 static inline void _nvme_check_size(void)
125 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
126 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
127 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
128 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
129 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
130 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
131 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
132 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
133 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
134 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
135 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
136 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
139 typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
140 struct nvme_completion *);
142 struct nvme_cmd_info {
143 nvme_completion_fn fn;
146 struct nvme_queue *nvmeq;
147 struct nvme_iod iod[0];
151 * Max size of iod being embedded in the request payload
153 #define NVME_INT_PAGES 2
154 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->page_size)
155 #define NVME_INT_MASK 0x01
158 * Will slightly overestimate the number of pages needed. This is OK
159 * as it only leads to a small amount of wasted memory for the lifetime of
162 static int nvme_npages(unsigned size, struct nvme_dev *dev)
164 unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
165 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
168 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
170 unsigned int ret = sizeof(struct nvme_cmd_info);
172 ret += sizeof(struct nvme_iod);
173 ret += sizeof(__le64 *) * nvme_npages(NVME_INT_BYTES(dev), dev);
174 ret += sizeof(struct scatterlist) * NVME_INT_PAGES;
179 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
180 unsigned int hctx_idx)
182 struct nvme_dev *dev = data;
183 struct nvme_queue *nvmeq = dev->queues[0];
185 WARN_ON(nvmeq->hctx);
187 hctx->driver_data = nvmeq;
191 static int nvme_admin_init_request(void *data, struct request *req,
192 unsigned int hctx_idx, unsigned int rq_idx,
193 unsigned int numa_node)
195 struct nvme_dev *dev = data;
196 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
197 struct nvme_queue *nvmeq = dev->queues[0];
204 static void nvme_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
206 struct nvme_queue *nvmeq = hctx->driver_data;
211 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
212 unsigned int hctx_idx)
214 struct nvme_dev *dev = data;
215 struct nvme_queue *nvmeq = dev->queues[
216 (hctx_idx % dev->queue_count) + 1];
221 /* nvmeq queues are shared between namespaces. We assume here that
222 * blk-mq map the tags so they match up with the nvme queue tags. */
223 WARN_ON(nvmeq->hctx->tags != hctx->tags);
225 hctx->driver_data = nvmeq;
229 static int nvme_init_request(void *data, struct request *req,
230 unsigned int hctx_idx, unsigned int rq_idx,
231 unsigned int numa_node)
233 struct nvme_dev *dev = data;
234 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
235 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
242 static void nvme_set_info(struct nvme_cmd_info *cmd, void *ctx,
243 nvme_completion_fn handler)
248 blk_mq_start_request(blk_mq_rq_from_pdu(cmd));
251 static void *iod_get_private(struct nvme_iod *iod)
253 return (void *) (iod->private & ~0x1UL);
257 * If bit 0 is set, the iod is embedded in the request payload.
259 static bool iod_should_kfree(struct nvme_iod *iod)
261 return (iod->private & NVME_INT_MASK) == 0;
264 /* Special values must be less than 0x1000 */
265 #define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
266 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
267 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
268 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
270 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
271 struct nvme_completion *cqe)
273 if (ctx == CMD_CTX_CANCELLED)
275 if (ctx == CMD_CTX_COMPLETED) {
276 dev_warn(nvmeq->q_dmadev,
277 "completed id %d twice on queue %d\n",
278 cqe->command_id, le16_to_cpup(&cqe->sq_id));
281 if (ctx == CMD_CTX_INVALID) {
282 dev_warn(nvmeq->q_dmadev,
283 "invalid id %d completed on queue %d\n",
284 cqe->command_id, le16_to_cpup(&cqe->sq_id));
287 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
290 static void *cancel_cmd_info(struct nvme_cmd_info *cmd, nvme_completion_fn *fn)
297 cmd->fn = special_completion;
298 cmd->ctx = CMD_CTX_CANCELLED;
302 static void async_req_completion(struct nvme_queue *nvmeq, void *ctx,
303 struct nvme_completion *cqe)
305 u32 result = le32_to_cpup(&cqe->result);
306 u16 status = le16_to_cpup(&cqe->status) >> 1;
308 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
309 ++nvmeq->dev->event_limit;
310 if (status == NVME_SC_SUCCESS)
311 dev_warn(nvmeq->q_dmadev,
312 "async event result %08x\n", result);
315 static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
316 struct nvme_completion *cqe)
318 struct request *req = ctx;
320 u16 status = le16_to_cpup(&cqe->status) >> 1;
321 u32 result = le32_to_cpup(&cqe->result);
323 blk_mq_free_hctx_request(nvmeq->hctx, req);
325 dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
326 ++nvmeq->dev->abort_limit;
329 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
330 struct nvme_completion *cqe)
332 struct async_cmd_info *cmdinfo = ctx;
333 cmdinfo->result = le32_to_cpup(&cqe->result);
334 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
335 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
336 blk_mq_free_hctx_request(nvmeq->hctx, cmdinfo->req);
339 static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
342 struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
343 struct request *req = blk_mq_tag_to_rq(hctx->tags, tag);
345 return blk_mq_rq_to_pdu(req);
349 * Called with local interrupts disabled and the q_lock held. May not sleep.
351 static void *nvme_finish_cmd(struct nvme_queue *nvmeq, int tag,
352 nvme_completion_fn *fn)
354 struct nvme_cmd_info *cmd = get_cmd_from_tag(nvmeq, tag);
356 if (tag >= nvmeq->q_depth) {
357 *fn = special_completion;
358 return CMD_CTX_INVALID;
363 cmd->fn = special_completion;
364 cmd->ctx = CMD_CTX_COMPLETED;
369 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
370 * @nvmeq: The queue to use
371 * @cmd: The command to send
373 * Safe to use from interrupt context
375 static int __nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
377 u16 tail = nvmeq->sq_tail;
379 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
380 if (++tail == nvmeq->q_depth)
382 writel(tail, nvmeq->q_db);
383 nvmeq->sq_tail = tail;
388 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
392 spin_lock_irqsave(&nvmeq->q_lock, flags);
393 ret = __nvme_submit_cmd(nvmeq, cmd);
394 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
398 static __le64 **iod_list(struct nvme_iod *iod)
400 return ((void *)iod) + iod->offset;
403 static inline void iod_init(struct nvme_iod *iod, unsigned nbytes,
404 unsigned nseg, unsigned long private)
406 iod->private = private;
407 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
409 iod->length = nbytes;
413 static struct nvme_iod *
414 __nvme_alloc_iod(unsigned nseg, unsigned bytes, struct nvme_dev *dev,
415 unsigned long priv, gfp_t gfp)
417 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
418 sizeof(__le64 *) * nvme_npages(bytes, dev) +
419 sizeof(struct scatterlist) * nseg, gfp);
422 iod_init(iod, bytes, nseg, priv);
427 static struct nvme_iod *nvme_alloc_iod(struct request *rq, struct nvme_dev *dev,
430 unsigned size = !(rq->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(rq) :
431 sizeof(struct nvme_dsm_range);
432 struct nvme_iod *iod;
434 if (rq->nr_phys_segments <= NVME_INT_PAGES &&
435 size <= NVME_INT_BYTES(dev)) {
436 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(rq);
439 iod_init(iod, size, rq->nr_phys_segments,
440 (unsigned long) rq | NVME_INT_MASK);
444 return __nvme_alloc_iod(rq->nr_phys_segments, size, dev,
445 (unsigned long) rq, gfp);
448 static void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
450 const int last_prp = dev->page_size / 8 - 1;
452 __le64 **list = iod_list(iod);
453 dma_addr_t prp_dma = iod->first_dma;
455 if (iod->npages == 0)
456 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
457 for (i = 0; i < iod->npages; i++) {
458 __le64 *prp_list = list[i];
459 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
460 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
461 prp_dma = next_prp_dma;
464 if (iod_should_kfree(iod))
468 static int nvme_error_status(u16 status)
470 switch (status & 0x7ff) {
471 case NVME_SC_SUCCESS:
473 case NVME_SC_CAP_EXCEEDED:
480 #ifdef CONFIG_BLK_DEV_INTEGRITY
481 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
483 if (be32_to_cpu(pi->ref_tag) == v)
484 pi->ref_tag = cpu_to_be32(p);
487 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
489 if (be32_to_cpu(pi->ref_tag) == p)
490 pi->ref_tag = cpu_to_be32(v);
494 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
496 * The virtual start sector is the one that was originally submitted by the
497 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
498 * start sector may be different. Remap protection information to match the
499 * physical LBA on writes, and back to the original seed on reads.
501 * Type 0 and 3 do not have a ref tag, so no remapping required.
503 static void nvme_dif_remap(struct request *req,
504 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
506 struct nvme_ns *ns = req->rq_disk->private_data;
507 struct bio_integrity_payload *bip;
508 struct t10_pi_tuple *pi;
510 u32 i, nlb, ts, phys, virt;
512 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
515 bip = bio_integrity(req->bio);
519 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
522 virt = bip_get_seed(bip);
523 phys = nvme_block_nr(ns, blk_rq_pos(req));
524 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
525 ts = ns->disk->integrity->tuple_size;
527 for (i = 0; i < nlb; i++, virt++, phys++) {
528 pi = (struct t10_pi_tuple *)p;
529 dif_swap(phys, virt, pi);
535 static int nvme_noop_verify(struct blk_integrity_iter *iter)
540 static int nvme_noop_generate(struct blk_integrity_iter *iter)
545 struct blk_integrity nvme_meta_noop = {
546 .name = "NVME_META_NOOP",
547 .generate_fn = nvme_noop_generate,
548 .verify_fn = nvme_noop_verify,
551 static void nvme_init_integrity(struct nvme_ns *ns)
553 struct blk_integrity integrity;
555 switch (ns->pi_type) {
556 case NVME_NS_DPS_PI_TYPE3:
557 integrity = t10_pi_type3_crc;
559 case NVME_NS_DPS_PI_TYPE1:
560 case NVME_NS_DPS_PI_TYPE2:
561 integrity = t10_pi_type1_crc;
564 integrity = nvme_meta_noop;
567 integrity.tuple_size = ns->ms;
568 blk_integrity_register(ns->disk, &integrity);
569 blk_queue_max_integrity_segments(ns->queue, 1);
571 #else /* CONFIG_BLK_DEV_INTEGRITY */
572 static void nvme_dif_remap(struct request *req,
573 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
576 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
579 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
582 static void nvme_init_integrity(struct nvme_ns *ns)
587 static void req_completion(struct nvme_queue *nvmeq, void *ctx,
588 struct nvme_completion *cqe)
590 struct nvme_iod *iod = ctx;
591 struct request *req = iod_get_private(iod);
592 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
594 u16 status = le16_to_cpup(&cqe->status) >> 1;
596 if (unlikely(status)) {
597 if (!(status & NVME_SC_DNR || blk_noretry_request(req))
598 && (jiffies - req->start_time) < req->timeout) {
601 blk_mq_requeue_request(req);
602 spin_lock_irqsave(req->q->queue_lock, flags);
603 if (!blk_queue_stopped(req->q))
604 blk_mq_kick_requeue_list(req->q);
605 spin_unlock_irqrestore(req->q->queue_lock, flags);
608 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
609 req->errors = status;
611 req->errors = nvme_error_status(status);
615 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
616 u32 result = le32_to_cpup(&cqe->result);
617 req->special = (void *)(uintptr_t)result;
621 dev_warn(nvmeq->dev->dev,
622 "completing aborted command with status:%04x\n",
626 dma_unmap_sg(nvmeq->dev->dev, iod->sg, iod->nents,
627 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
628 if (blk_integrity_rq(req)) {
629 if (!rq_data_dir(req))
630 nvme_dif_remap(req, nvme_dif_complete);
631 dma_unmap_sg(nvmeq->dev->dev, iod->meta_sg, 1,
632 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
635 nvme_free_iod(nvmeq->dev, iod);
637 blk_mq_complete_request(req);
640 /* length is in bytes. gfp flags indicates whether we may sleep. */
641 static int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
642 int total_len, gfp_t gfp)
644 struct dma_pool *pool;
645 int length = total_len;
646 struct scatterlist *sg = iod->sg;
647 int dma_len = sg_dma_len(sg);
648 u64 dma_addr = sg_dma_address(sg);
649 u32 page_size = dev->page_size;
650 int offset = dma_addr & (page_size - 1);
652 __le64 **list = iod_list(iod);
656 length -= (page_size - offset);
660 dma_len -= (page_size - offset);
662 dma_addr += (page_size - offset);
665 dma_addr = sg_dma_address(sg);
666 dma_len = sg_dma_len(sg);
669 if (length <= page_size) {
670 iod->first_dma = dma_addr;
674 nprps = DIV_ROUND_UP(length, page_size);
675 if (nprps <= (256 / 8)) {
676 pool = dev->prp_small_pool;
679 pool = dev->prp_page_pool;
683 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
685 iod->first_dma = dma_addr;
687 return (total_len - length) + page_size;
690 iod->first_dma = prp_dma;
693 if (i == page_size >> 3) {
694 __le64 *old_prp_list = prp_list;
695 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
697 return total_len - length;
698 list[iod->npages++] = prp_list;
699 prp_list[0] = old_prp_list[i - 1];
700 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
703 prp_list[i++] = cpu_to_le64(dma_addr);
704 dma_len -= page_size;
705 dma_addr += page_size;
713 dma_addr = sg_dma_address(sg);
714 dma_len = sg_dma_len(sg);
720 static void nvme_submit_priv(struct nvme_queue *nvmeq, struct request *req,
721 struct nvme_iod *iod)
723 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
725 memcpy(cmnd, req->cmd, sizeof(struct nvme_command));
726 cmnd->rw.command_id = req->tag;
727 if (req->nr_phys_segments) {
728 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
729 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
732 if (++nvmeq->sq_tail == nvmeq->q_depth)
734 writel(nvmeq->sq_tail, nvmeq->q_db);
738 * We reuse the small pool to allocate the 16-byte range here as it is not
739 * worth having a special pool for these or additional cases to handle freeing
742 static void nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
743 struct request *req, struct nvme_iod *iod)
745 struct nvme_dsm_range *range =
746 (struct nvme_dsm_range *)iod_list(iod)[0];
747 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
749 range->cattr = cpu_to_le32(0);
750 range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift);
751 range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
753 memset(cmnd, 0, sizeof(*cmnd));
754 cmnd->dsm.opcode = nvme_cmd_dsm;
755 cmnd->dsm.command_id = req->tag;
756 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
757 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
759 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
761 if (++nvmeq->sq_tail == nvmeq->q_depth)
763 writel(nvmeq->sq_tail, nvmeq->q_db);
766 static void nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
769 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
771 memset(cmnd, 0, sizeof(*cmnd));
772 cmnd->common.opcode = nvme_cmd_flush;
773 cmnd->common.command_id = cmdid;
774 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
776 if (++nvmeq->sq_tail == nvmeq->q_depth)
778 writel(nvmeq->sq_tail, nvmeq->q_db);
781 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod,
784 struct request *req = iod_get_private(iod);
785 struct nvme_command *cmnd;
789 if (req->cmd_flags & REQ_FUA)
790 control |= NVME_RW_FUA;
791 if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
792 control |= NVME_RW_LR;
794 if (req->cmd_flags & REQ_RAHEAD)
795 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
797 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
798 memset(cmnd, 0, sizeof(*cmnd));
800 cmnd->rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
801 cmnd->rw.command_id = req->tag;
802 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
803 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
804 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
805 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
806 cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
808 if (blk_integrity_rq(req)) {
809 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(iod->meta_sg));
810 switch (ns->pi_type) {
811 case NVME_NS_DPS_PI_TYPE3:
812 control |= NVME_RW_PRINFO_PRCHK_GUARD;
814 case NVME_NS_DPS_PI_TYPE1:
815 case NVME_NS_DPS_PI_TYPE2:
816 control |= NVME_RW_PRINFO_PRCHK_GUARD |
817 NVME_RW_PRINFO_PRCHK_REF;
818 cmnd->rw.reftag = cpu_to_le32(
819 nvme_block_nr(ns, blk_rq_pos(req)));
823 control |= NVME_RW_PRINFO_PRACT;
825 cmnd->rw.control = cpu_to_le16(control);
826 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
828 if (++nvmeq->sq_tail == nvmeq->q_depth)
830 writel(nvmeq->sq_tail, nvmeq->q_db);
836 * NOTE: ns is NULL when called on the admin queue.
838 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
839 const struct blk_mq_queue_data *bd)
841 struct nvme_ns *ns = hctx->queue->queuedata;
842 struct nvme_queue *nvmeq = hctx->driver_data;
843 struct nvme_dev *dev = nvmeq->dev;
844 struct request *req = bd->rq;
845 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
846 struct nvme_iod *iod;
847 enum dma_data_direction dma_dir;
850 * If formated with metadata, require the block layer provide a buffer
851 * unless this namespace is formated such that the metadata can be
852 * stripped/generated by the controller with PRACT=1.
854 if (ns && ns->ms && !blk_integrity_rq(req)) {
855 if (!(ns->pi_type && ns->ms == 8)) {
856 req->errors = -EFAULT;
857 blk_mq_complete_request(req);
858 return BLK_MQ_RQ_QUEUE_OK;
862 iod = nvme_alloc_iod(req, dev, GFP_ATOMIC);
864 return BLK_MQ_RQ_QUEUE_BUSY;
866 if (req->cmd_flags & REQ_DISCARD) {
869 * We reuse the small pool to allocate the 16-byte range here
870 * as it is not worth having a special pool for these or
871 * additional cases to handle freeing the iod.
873 range = dma_pool_alloc(dev->prp_small_pool, GFP_ATOMIC,
877 iod_list(iod)[0] = (__le64 *)range;
879 } else if (req->nr_phys_segments) {
880 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
882 sg_init_table(iod->sg, req->nr_phys_segments);
883 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
887 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
890 if (blk_rq_bytes(req) !=
891 nvme_setup_prps(dev, iod, blk_rq_bytes(req), GFP_ATOMIC)) {
892 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
895 if (blk_integrity_rq(req)) {
896 if (blk_rq_count_integrity_sg(req->q, req->bio) != 1)
899 sg_init_table(iod->meta_sg, 1);
900 if (blk_rq_map_integrity_sg(
901 req->q, req->bio, iod->meta_sg) != 1)
904 if (rq_data_dir(req))
905 nvme_dif_remap(req, nvme_dif_prep);
907 if (!dma_map_sg(nvmeq->q_dmadev, iod->meta_sg, 1, dma_dir))
912 nvme_set_info(cmd, iod, req_completion);
913 spin_lock_irq(&nvmeq->q_lock);
914 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
915 nvme_submit_priv(nvmeq, req, iod);
916 else if (req->cmd_flags & REQ_DISCARD)
917 nvme_submit_discard(nvmeq, ns, req, iod);
918 else if (req->cmd_flags & REQ_FLUSH)
919 nvme_submit_flush(nvmeq, ns, req->tag);
921 nvme_submit_iod(nvmeq, iod, ns);
923 nvme_process_cq(nvmeq);
924 spin_unlock_irq(&nvmeq->q_lock);
925 return BLK_MQ_RQ_QUEUE_OK;
928 nvme_free_iod(dev, iod);
929 return BLK_MQ_RQ_QUEUE_ERROR;
931 nvme_free_iod(dev, iod);
932 return BLK_MQ_RQ_QUEUE_BUSY;
935 static int nvme_process_cq(struct nvme_queue *nvmeq)
939 head = nvmeq->cq_head;
940 phase = nvmeq->cq_phase;
944 nvme_completion_fn fn;
945 struct nvme_completion cqe = nvmeq->cqes[head];
946 if ((le16_to_cpu(cqe.status) & 1) != phase)
948 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
949 if (++head == nvmeq->q_depth) {
953 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
954 fn(nvmeq, ctx, &cqe);
957 /* If the controller ignores the cq head doorbell and continuously
958 * writes to the queue, it is theoretically possible to wrap around
959 * the queue twice and mistakenly return IRQ_NONE. Linux only
960 * requires that 0.1% of your interrupts are handled, so this isn't
963 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
966 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
967 nvmeq->cq_head = head;
968 nvmeq->cq_phase = phase;
974 static irqreturn_t nvme_irq(int irq, void *data)
977 struct nvme_queue *nvmeq = data;
978 spin_lock(&nvmeq->q_lock);
979 nvme_process_cq(nvmeq);
980 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
982 spin_unlock(&nvmeq->q_lock);
986 static irqreturn_t nvme_irq_check(int irq, void *data)
988 struct nvme_queue *nvmeq = data;
989 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
990 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
992 return IRQ_WAKE_THREAD;
996 * Returns 0 on success. If the result is negative, it's a Linux error code;
997 * if the result is positive, it's an NVM Express status code
999 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1000 void *buffer, void __user *ubuffer, unsigned bufflen,
1001 u32 *result, unsigned timeout)
1003 bool write = cmd->common.opcode & 1;
1004 struct bio *bio = NULL;
1005 struct request *req;
1008 req = blk_mq_alloc_request(q, write, GFP_KERNEL, false);
1010 return PTR_ERR(req);
1012 req->cmd_type = REQ_TYPE_DRV_PRIV;
1013 req->cmd_flags = REQ_FAILFAST_DRIVER;
1014 req->__data_len = 0;
1015 req->__sector = (sector_t) -1;
1016 req->bio = req->biotail = NULL;
1018 req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
1020 req->cmd = (unsigned char *)cmd;
1021 req->cmd_len = sizeof(struct nvme_command);
1022 req->special = (void *)0;
1024 if (buffer && bufflen) {
1025 ret = blk_rq_map_kern(q, req, buffer, bufflen, __GFP_WAIT);
1028 } else if (ubuffer && bufflen) {
1029 ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen, __GFP_WAIT);
1035 blk_execute_rq(req->q, NULL, req, 0);
1037 blk_rq_unmap_user(bio);
1039 *result = (u32)(uintptr_t)req->special;
1042 blk_mq_free_request(req);
1046 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1047 void *buffer, unsigned bufflen)
1049 return __nvme_submit_sync_cmd(q, cmd, buffer, NULL, bufflen, NULL, 0);
1052 static int nvme_submit_async_admin_req(struct nvme_dev *dev)
1054 struct nvme_queue *nvmeq = dev->queues[0];
1055 struct nvme_command c;
1056 struct nvme_cmd_info *cmd_info;
1057 struct request *req;
1059 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, true);
1061 return PTR_ERR(req);
1063 req->cmd_flags |= REQ_NO_TIMEOUT;
1064 cmd_info = blk_mq_rq_to_pdu(req);
1065 nvme_set_info(cmd_info, NULL, async_req_completion);
1067 memset(&c, 0, sizeof(c));
1068 c.common.opcode = nvme_admin_async_event;
1069 c.common.command_id = req->tag;
1071 blk_mq_free_hctx_request(nvmeq->hctx, req);
1072 return __nvme_submit_cmd(nvmeq, &c);
1075 static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
1076 struct nvme_command *cmd,
1077 struct async_cmd_info *cmdinfo, unsigned timeout)
1079 struct nvme_queue *nvmeq = dev->queues[0];
1080 struct request *req;
1081 struct nvme_cmd_info *cmd_rq;
1083 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
1085 return PTR_ERR(req);
1087 req->timeout = timeout;
1088 cmd_rq = blk_mq_rq_to_pdu(req);
1090 nvme_set_info(cmd_rq, cmdinfo, async_completion);
1091 cmdinfo->status = -EINTR;
1093 cmd->common.command_id = req->tag;
1095 return nvme_submit_cmd(nvmeq, cmd);
1098 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1100 struct nvme_command c;
1102 memset(&c, 0, sizeof(c));
1103 c.delete_queue.opcode = opcode;
1104 c.delete_queue.qid = cpu_to_le16(id);
1106 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1109 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1110 struct nvme_queue *nvmeq)
1112 struct nvme_command c;
1113 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1116 * Note: we (ab)use the fact the the prp fields survive if no data
1117 * is attached to the request.
1119 memset(&c, 0, sizeof(c));
1120 c.create_cq.opcode = nvme_admin_create_cq;
1121 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1122 c.create_cq.cqid = cpu_to_le16(qid);
1123 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1124 c.create_cq.cq_flags = cpu_to_le16(flags);
1125 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1127 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1130 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1131 struct nvme_queue *nvmeq)
1133 struct nvme_command c;
1134 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1137 * Note: we (ab)use the fact the the prp fields survive if no data
1138 * is attached to the request.
1140 memset(&c, 0, sizeof(c));
1141 c.create_sq.opcode = nvme_admin_create_sq;
1142 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1143 c.create_sq.sqid = cpu_to_le16(qid);
1144 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1145 c.create_sq.sq_flags = cpu_to_le16(flags);
1146 c.create_sq.cqid = cpu_to_le16(qid);
1148 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1151 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1153 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1156 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1158 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1161 int nvme_identify_ctrl(struct nvme_dev *dev, struct nvme_id_ctrl **id)
1163 struct nvme_command c = {
1164 .identify.opcode = nvme_admin_identify,
1165 .identify.cns = cpu_to_le32(1),
1169 *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1173 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1174 sizeof(struct nvme_id_ctrl));
1180 int nvme_identify_ns(struct nvme_dev *dev, unsigned nsid,
1181 struct nvme_id_ns **id)
1183 struct nvme_command c = {
1184 .identify.opcode = nvme_admin_identify,
1185 .identify.nsid = cpu_to_le32(nsid),
1189 *id = kmalloc(sizeof(struct nvme_id_ns), GFP_KERNEL);
1193 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1194 sizeof(struct nvme_id_ns));
1200 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1201 dma_addr_t dma_addr, u32 *result)
1203 struct nvme_command c;
1205 memset(&c, 0, sizeof(c));
1206 c.features.opcode = nvme_admin_get_features;
1207 c.features.nsid = cpu_to_le32(nsid);
1208 c.features.prp1 = cpu_to_le64(dma_addr);
1209 c.features.fid = cpu_to_le32(fid);
1211 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1215 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1216 dma_addr_t dma_addr, u32 *result)
1218 struct nvme_command c;
1220 memset(&c, 0, sizeof(c));
1221 c.features.opcode = nvme_admin_set_features;
1222 c.features.prp1 = cpu_to_le64(dma_addr);
1223 c.features.fid = cpu_to_le32(fid);
1224 c.features.dword11 = cpu_to_le32(dword11);
1226 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1230 int nvme_get_log_page(struct nvme_dev *dev, struct nvme_smart_log **log)
1232 struct nvme_command c = {
1233 .common.opcode = nvme_admin_get_log_page,
1234 .common.nsid = cpu_to_le32(0xFFFFFFFF),
1235 .common.cdw10[0] = cpu_to_le32(
1236 (((sizeof(struct nvme_smart_log) / 4) - 1) << 16) |
1241 *log = kmalloc(sizeof(struct nvme_smart_log), GFP_KERNEL);
1245 error = nvme_submit_sync_cmd(dev->admin_q, &c, *log,
1246 sizeof(struct nvme_smart_log));
1253 * nvme_abort_req - Attempt aborting a request
1255 * Schedule controller reset if the command was already aborted once before and
1256 * still hasn't been returned to the driver, or if this is the admin queue.
1258 static void nvme_abort_req(struct request *req)
1260 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1261 struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1262 struct nvme_dev *dev = nvmeq->dev;
1263 struct request *abort_req;
1264 struct nvme_cmd_info *abort_cmd;
1265 struct nvme_command cmd;
1267 if (!nvmeq->qid || cmd_rq->aborted) {
1268 unsigned long flags;
1270 spin_lock_irqsave(&dev_list_lock, flags);
1271 if (work_busy(&dev->reset_work))
1273 list_del_init(&dev->node);
1274 dev_warn(dev->dev, "I/O %d QID %d timeout, reset controller\n",
1275 req->tag, nvmeq->qid);
1276 dev->reset_workfn = nvme_reset_failed_dev;
1277 queue_work(nvme_workq, &dev->reset_work);
1279 spin_unlock_irqrestore(&dev_list_lock, flags);
1283 if (!dev->abort_limit)
1286 abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1288 if (IS_ERR(abort_req))
1291 abort_cmd = blk_mq_rq_to_pdu(abort_req);
1292 nvme_set_info(abort_cmd, abort_req, abort_completion);
1294 memset(&cmd, 0, sizeof(cmd));
1295 cmd.abort.opcode = nvme_admin_abort_cmd;
1296 cmd.abort.cid = req->tag;
1297 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1298 cmd.abort.command_id = abort_req->tag;
1301 cmd_rq->aborted = 1;
1303 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1305 if (nvme_submit_cmd(dev->queues[0], &cmd) < 0) {
1306 dev_warn(nvmeq->q_dmadev,
1307 "Could not abort I/O %d QID %d",
1308 req->tag, nvmeq->qid);
1309 blk_mq_free_request(abort_req);
1313 static void nvme_cancel_queue_ios(struct blk_mq_hw_ctx *hctx,
1314 struct request *req, void *data, bool reserved)
1316 struct nvme_queue *nvmeq = data;
1318 nvme_completion_fn fn;
1319 struct nvme_cmd_info *cmd;
1320 struct nvme_completion cqe;
1322 if (!blk_mq_request_started(req))
1325 cmd = blk_mq_rq_to_pdu(req);
1327 if (cmd->ctx == CMD_CTX_CANCELLED)
1330 if (blk_queue_dying(req->q))
1331 cqe.status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1333 cqe.status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1336 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1337 req->tag, nvmeq->qid);
1338 ctx = cancel_cmd_info(cmd, &fn);
1339 fn(nvmeq, ctx, &cqe);
1342 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1344 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1345 struct nvme_queue *nvmeq = cmd->nvmeq;
1347 dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1349 spin_lock_irq(&nvmeq->q_lock);
1350 nvme_abort_req(req);
1351 spin_unlock_irq(&nvmeq->q_lock);
1354 * The aborted req will be completed on receiving the abort req.
1355 * We enable the timer again. If hit twice, it'll cause a device reset,
1356 * as the device then is in a faulty state.
1358 return BLK_EH_RESET_TIMER;
1361 static void nvme_free_queue(struct nvme_queue *nvmeq)
1363 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1364 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1365 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1366 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1370 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1374 for (i = dev->queue_count - 1; i >= lowest; i--) {
1375 struct nvme_queue *nvmeq = dev->queues[i];
1377 dev->queues[i] = NULL;
1378 nvme_free_queue(nvmeq);
1383 * nvme_suspend_queue - put queue into suspended state
1384 * @nvmeq - queue to suspend
1386 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1390 spin_lock_irq(&nvmeq->q_lock);
1391 if (nvmeq->cq_vector == -1) {
1392 spin_unlock_irq(&nvmeq->q_lock);
1395 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1396 nvmeq->dev->online_queues--;
1397 nvmeq->cq_vector = -1;
1398 spin_unlock_irq(&nvmeq->q_lock);
1400 if (!nvmeq->qid && nvmeq->dev->admin_q)
1401 blk_mq_freeze_queue_start(nvmeq->dev->admin_q);
1403 irq_set_affinity_hint(vector, NULL);
1404 free_irq(vector, nvmeq);
1409 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1411 struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
1413 spin_lock_irq(&nvmeq->q_lock);
1414 if (hctx && hctx->tags)
1415 blk_mq_tag_busy_iter(hctx, nvme_cancel_queue_ios, nvmeq);
1416 spin_unlock_irq(&nvmeq->q_lock);
1419 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1421 struct nvme_queue *nvmeq = dev->queues[qid];
1425 if (nvme_suspend_queue(nvmeq))
1428 /* Don't tell the adapter to delete the admin queue.
1429 * Don't tell a removed adapter to delete IO queues. */
1430 if (qid && readl(&dev->bar->csts) != -1) {
1431 adapter_delete_sq(dev, qid);
1432 adapter_delete_cq(dev, qid);
1435 spin_lock_irq(&nvmeq->q_lock);
1436 nvme_process_cq(nvmeq);
1437 spin_unlock_irq(&nvmeq->q_lock);
1440 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1443 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1447 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1448 &nvmeq->cq_dma_addr, GFP_KERNEL);
1452 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1453 &nvmeq->sq_dma_addr, GFP_KERNEL);
1454 if (!nvmeq->sq_cmds)
1457 nvmeq->q_dmadev = dev->dev;
1459 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1460 dev->instance, qid);
1461 spin_lock_init(&nvmeq->q_lock);
1463 nvmeq->cq_phase = 1;
1464 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1465 nvmeq->q_depth = depth;
1468 dev->queues[qid] = nvmeq;
1473 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1474 nvmeq->cq_dma_addr);
1480 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1483 if (use_threaded_interrupts)
1484 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1485 nvme_irq_check, nvme_irq, IRQF_SHARED,
1487 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1488 IRQF_SHARED, name, nvmeq);
1491 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1493 struct nvme_dev *dev = nvmeq->dev;
1495 spin_lock_irq(&nvmeq->q_lock);
1498 nvmeq->cq_phase = 1;
1499 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1500 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1501 dev->online_queues++;
1502 spin_unlock_irq(&nvmeq->q_lock);
1505 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1507 struct nvme_dev *dev = nvmeq->dev;
1510 nvmeq->cq_vector = qid - 1;
1511 result = adapter_alloc_cq(dev, qid, nvmeq);
1515 result = adapter_alloc_sq(dev, qid, nvmeq);
1519 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1523 nvme_init_queue(nvmeq, qid);
1527 adapter_delete_sq(dev, qid);
1529 adapter_delete_cq(dev, qid);
1533 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1535 unsigned long timeout;
1536 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1538 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1540 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1542 if (fatal_signal_pending(current))
1544 if (time_after(jiffies, timeout)) {
1546 "Device not ready; aborting %s\n", enabled ?
1547 "initialisation" : "reset");
1556 * If the device has been passed off to us in an enabled state, just clear
1557 * the enabled bit. The spec says we should set the 'shutdown notification
1558 * bits', but doing so may cause the device to complete commands to the
1559 * admin queue ... and we don't know what memory that might be pointing at!
1561 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1563 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1564 dev->ctrl_config &= ~NVME_CC_ENABLE;
1565 writel(dev->ctrl_config, &dev->bar->cc);
1567 return nvme_wait_ready(dev, cap, false);
1570 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1572 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1573 dev->ctrl_config |= NVME_CC_ENABLE;
1574 writel(dev->ctrl_config, &dev->bar->cc);
1576 return nvme_wait_ready(dev, cap, true);
1579 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1581 unsigned long timeout;
1583 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1584 dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1586 writel(dev->ctrl_config, &dev->bar->cc);
1588 timeout = SHUTDOWN_TIMEOUT + jiffies;
1589 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1590 NVME_CSTS_SHST_CMPLT) {
1592 if (fatal_signal_pending(current))
1594 if (time_after(jiffies, timeout)) {
1596 "Device shutdown incomplete; abort shutdown\n");
1604 static struct blk_mq_ops nvme_mq_admin_ops = {
1605 .queue_rq = nvme_queue_rq,
1606 .map_queue = blk_mq_map_queue,
1607 .init_hctx = nvme_admin_init_hctx,
1608 .exit_hctx = nvme_exit_hctx,
1609 .init_request = nvme_admin_init_request,
1610 .timeout = nvme_timeout,
1613 static struct blk_mq_ops nvme_mq_ops = {
1614 .queue_rq = nvme_queue_rq,
1615 .map_queue = blk_mq_map_queue,
1616 .init_hctx = nvme_init_hctx,
1617 .exit_hctx = nvme_exit_hctx,
1618 .init_request = nvme_init_request,
1619 .timeout = nvme_timeout,
1622 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1624 if (dev->admin_q && !blk_queue_dying(dev->admin_q)) {
1625 blk_cleanup_queue(dev->admin_q);
1626 blk_mq_free_tag_set(&dev->admin_tagset);
1630 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1632 if (!dev->admin_q) {
1633 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1634 dev->admin_tagset.nr_hw_queues = 1;
1635 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1636 dev->admin_tagset.reserved_tags = 1;
1637 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1638 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1639 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1640 dev->admin_tagset.driver_data = dev;
1642 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1645 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1646 if (IS_ERR(dev->admin_q)) {
1647 blk_mq_free_tag_set(&dev->admin_tagset);
1650 if (!blk_get_queue(dev->admin_q)) {
1651 nvme_dev_remove_admin(dev);
1655 blk_mq_unfreeze_queue(dev->admin_q);
1660 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1664 u64 cap = readq(&dev->bar->cap);
1665 struct nvme_queue *nvmeq;
1666 unsigned page_shift = PAGE_SHIFT;
1667 unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1668 unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1670 if (page_shift < dev_page_min) {
1672 "Minimum device page size (%u) too large for "
1673 "host (%u)\n", 1 << dev_page_min,
1677 if (page_shift > dev_page_max) {
1679 "Device maximum page size (%u) smaller than "
1680 "host (%u); enabling work-around\n",
1681 1 << dev_page_max, 1 << page_shift);
1682 page_shift = dev_page_max;
1685 result = nvme_disable_ctrl(dev, cap);
1689 nvmeq = dev->queues[0];
1691 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1696 aqa = nvmeq->q_depth - 1;
1699 dev->page_size = 1 << page_shift;
1701 dev->ctrl_config = NVME_CC_CSS_NVM;
1702 dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1703 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1704 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1706 writel(aqa, &dev->bar->aqa);
1707 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1708 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1710 result = nvme_enable_ctrl(dev, cap);
1714 nvmeq->cq_vector = 0;
1715 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1722 nvme_free_queues(dev, 0);
1726 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1728 struct nvme_dev *dev = ns->dev;
1729 struct nvme_user_io io;
1730 struct nvme_command c;
1731 unsigned length, meta_len;
1733 dma_addr_t meta_dma = 0;
1736 if (copy_from_user(&io, uio, sizeof(io)))
1739 switch (io.opcode) {
1740 case nvme_cmd_write:
1742 case nvme_cmd_compare:
1748 length = (io.nblocks + 1) << ns->lba_shift;
1749 meta_len = (io.nblocks + 1) * ns->ms;
1750 write = io.opcode & 1;
1753 if (((io.metadata & 3) || !io.metadata) && !ns->ext)
1761 meta = dma_alloc_coherent(dev->dev, meta_len,
1762 &meta_dma, GFP_KERNEL);
1768 if (copy_from_user(meta, (void __user *)io.metadata,
1776 memset(&c, 0, sizeof(c));
1777 c.rw.opcode = io.opcode;
1778 c.rw.flags = io.flags;
1779 c.rw.nsid = cpu_to_le32(ns->ns_id);
1780 c.rw.slba = cpu_to_le64(io.slba);
1781 c.rw.length = cpu_to_le16(io.nblocks);
1782 c.rw.control = cpu_to_le16(io.control);
1783 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1784 c.rw.reftag = cpu_to_le32(io.reftag);
1785 c.rw.apptag = cpu_to_le16(io.apptag);
1786 c.rw.appmask = cpu_to_le16(io.appmask);
1787 c.rw.metadata = cpu_to_le64(meta_dma);
1789 status = __nvme_submit_sync_cmd(ns->queue, &c, NULL,
1790 (void __user *)io.addr, length, NULL, 0);
1793 if (status == NVME_SC_SUCCESS && !write) {
1794 if (copy_to_user((void __user *)io.metadata, meta,
1798 dma_free_coherent(dev->dev, meta_len, meta, meta_dma);
1803 static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1804 struct nvme_passthru_cmd __user *ucmd)
1806 struct nvme_passthru_cmd cmd;
1807 struct nvme_command c;
1808 unsigned timeout = 0;
1811 if (!capable(CAP_SYS_ADMIN))
1813 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1816 memset(&c, 0, sizeof(c));
1817 c.common.opcode = cmd.opcode;
1818 c.common.flags = cmd.flags;
1819 c.common.nsid = cpu_to_le32(cmd.nsid);
1820 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1821 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1822 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1823 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1824 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1825 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1826 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1827 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1830 timeout = msecs_to_jiffies(cmd.timeout_ms);
1832 status = __nvme_submit_sync_cmd(ns ? ns->queue : dev->admin_q, &c,
1833 NULL, (void __user *)cmd.addr, cmd.data_len,
1834 &cmd.result, timeout);
1836 if (put_user(cmd.result, &ucmd->result))
1843 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1846 struct nvme_ns *ns = bdev->bd_disk->private_data;
1850 force_successful_syscall_return();
1852 case NVME_IOCTL_ADMIN_CMD:
1853 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1854 case NVME_IOCTL_IO_CMD:
1855 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1856 case NVME_IOCTL_SUBMIT_IO:
1857 return nvme_submit_io(ns, (void __user *)arg);
1858 case SG_GET_VERSION_NUM:
1859 return nvme_sg_get_version_num((void __user *)arg);
1861 return nvme_sg_io(ns, (void __user *)arg);
1867 #ifdef CONFIG_COMPAT
1868 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1869 unsigned int cmd, unsigned long arg)
1873 return -ENOIOCTLCMD;
1875 return nvme_ioctl(bdev, mode, cmd, arg);
1878 #define nvme_compat_ioctl NULL
1881 static int nvme_open(struct block_device *bdev, fmode_t mode)
1886 spin_lock(&dev_list_lock);
1887 ns = bdev->bd_disk->private_data;
1890 else if (!kref_get_unless_zero(&ns->dev->kref))
1892 spin_unlock(&dev_list_lock);
1897 static void nvme_free_dev(struct kref *kref);
1899 static void nvme_release(struct gendisk *disk, fmode_t mode)
1901 struct nvme_ns *ns = disk->private_data;
1902 struct nvme_dev *dev = ns->dev;
1904 kref_put(&dev->kref, nvme_free_dev);
1907 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1909 /* some standard values */
1910 geo->heads = 1 << 6;
1911 geo->sectors = 1 << 5;
1912 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1916 static void nvme_config_discard(struct nvme_ns *ns)
1918 u32 logical_block_size = queue_logical_block_size(ns->queue);
1919 ns->queue->limits.discard_zeroes_data = 0;
1920 ns->queue->limits.discard_alignment = logical_block_size;
1921 ns->queue->limits.discard_granularity = logical_block_size;
1922 ns->queue->limits.max_discard_sectors = 0xffffffff;
1923 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1926 static int nvme_revalidate_disk(struct gendisk *disk)
1928 struct nvme_ns *ns = disk->private_data;
1929 struct nvme_dev *dev = ns->dev;
1930 struct nvme_id_ns *id;
1935 if (nvme_identify_ns(dev, ns->ns_id, &id)) {
1936 dev_warn(dev->dev, "%s: Identify failure\n", __func__);
1941 lbaf = id->flbas & NVME_NS_FLBAS_LBA_MASK;
1942 ns->lba_shift = id->lbaf[lbaf].ds;
1943 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1944 ns->ext = ns->ms && (id->flbas & NVME_NS_FLBAS_META_EXT);
1947 * If identify namespace failed, use default 512 byte block size so
1948 * block layer can use before failing read/write for 0 capacity.
1950 if (ns->lba_shift == 0)
1952 bs = 1 << ns->lba_shift;
1954 /* XXX: PI implementation requires metadata equal t10 pi tuple size */
1955 pi_type = ns->ms == sizeof(struct t10_pi_tuple) ?
1956 id->dps & NVME_NS_DPS_PI_MASK : 0;
1958 if (blk_get_integrity(disk) && (ns->pi_type != pi_type ||
1960 bs != queue_logical_block_size(disk->queue) ||
1961 (ns->ms && ns->ext)))
1962 blk_integrity_unregister(disk);
1964 ns->pi_type = pi_type;
1965 blk_queue_logical_block_size(ns->queue, bs);
1967 if (ns->ms && !blk_get_integrity(disk) && (disk->flags & GENHD_FL_UP) &&
1969 nvme_init_integrity(ns);
1971 if (id->ncap == 0 || (ns->ms && !blk_get_integrity(disk)))
1972 set_capacity(disk, 0);
1974 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1976 if (dev->oncs & NVME_CTRL_ONCS_DSM)
1977 nvme_config_discard(ns);
1983 static const struct block_device_operations nvme_fops = {
1984 .owner = THIS_MODULE,
1985 .ioctl = nvme_ioctl,
1986 .compat_ioctl = nvme_compat_ioctl,
1988 .release = nvme_release,
1989 .getgeo = nvme_getgeo,
1990 .revalidate_disk= nvme_revalidate_disk,
1993 static int nvme_kthread(void *data)
1995 struct nvme_dev *dev, *next;
1997 while (!kthread_should_stop()) {
1998 set_current_state(TASK_INTERRUPTIBLE);
1999 spin_lock(&dev_list_lock);
2000 list_for_each_entry_safe(dev, next, &dev_list, node) {
2002 if (readl(&dev->bar->csts) & NVME_CSTS_CFS) {
2003 if (work_busy(&dev->reset_work))
2005 list_del_init(&dev->node);
2007 "Failed status: %x, reset controller\n",
2008 readl(&dev->bar->csts));
2009 dev->reset_workfn = nvme_reset_failed_dev;
2010 queue_work(nvme_workq, &dev->reset_work);
2013 for (i = 0; i < dev->queue_count; i++) {
2014 struct nvme_queue *nvmeq = dev->queues[i];
2017 spin_lock_irq(&nvmeq->q_lock);
2018 nvme_process_cq(nvmeq);
2020 while ((i == 0) && (dev->event_limit > 0)) {
2021 if (nvme_submit_async_admin_req(dev))
2025 spin_unlock_irq(&nvmeq->q_lock);
2028 spin_unlock(&dev_list_lock);
2029 schedule_timeout(round_jiffies_relative(HZ));
2034 static void nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid)
2037 struct gendisk *disk;
2038 int node = dev_to_node(dev->dev);
2040 ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
2044 ns->queue = blk_mq_init_queue(&dev->tagset);
2045 if (IS_ERR(ns->queue))
2047 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
2048 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
2049 queue_flag_set_unlocked(QUEUE_FLAG_SG_GAPS, ns->queue);
2051 ns->queue->queuedata = ns;
2053 disk = alloc_disk_node(0, node);
2055 goto out_free_queue;
2059 ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
2060 list_add_tail(&ns->list, &dev->namespaces);
2062 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2063 if (dev->max_hw_sectors)
2064 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2065 if (dev->stripe_size)
2066 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
2067 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2068 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2070 disk->major = nvme_major;
2071 disk->first_minor = 0;
2072 disk->fops = &nvme_fops;
2073 disk->private_data = ns;
2074 disk->queue = ns->queue;
2075 disk->driverfs_dev = dev->device;
2076 disk->flags = GENHD_FL_EXT_DEVT;
2077 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2080 * Initialize capacity to 0 until we establish the namespace format and
2081 * setup integrity extentions if necessary. The revalidate_disk after
2082 * add_disk allows the driver to register with integrity if the format
2085 set_capacity(disk, 0);
2086 nvme_revalidate_disk(ns->disk);
2089 revalidate_disk(ns->disk);
2092 blk_cleanup_queue(ns->queue);
2097 static void nvme_create_io_queues(struct nvme_dev *dev)
2101 for (i = dev->queue_count; i <= dev->max_qid; i++)
2102 if (!nvme_alloc_queue(dev, i, dev->q_depth))
2105 for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
2106 if (nvme_create_queue(dev->queues[i], i))
2110 static int set_queue_count(struct nvme_dev *dev, int count)
2114 u32 q_count = (count - 1) | ((count - 1) << 16);
2116 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2121 dev_err(dev->dev, "Could not set queue count (%d)\n", status);
2124 return min(result & 0xffff, result >> 16) + 1;
2127 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2129 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2132 static int nvme_setup_io_queues(struct nvme_dev *dev)
2134 struct nvme_queue *adminq = dev->queues[0];
2135 struct pci_dev *pdev = to_pci_dev(dev->dev);
2136 int result, i, vecs, nr_io_queues, size;
2138 nr_io_queues = num_possible_cpus();
2139 result = set_queue_count(dev, nr_io_queues);
2142 if (result < nr_io_queues)
2143 nr_io_queues = result;
2145 size = db_bar_size(dev, nr_io_queues);
2149 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2152 if (!--nr_io_queues)
2154 size = db_bar_size(dev, nr_io_queues);
2156 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2157 adminq->q_db = dev->dbs;
2160 /* Deregister the admin queue's interrupt */
2161 free_irq(dev->entry[0].vector, adminq);
2164 * If we enable msix early due to not intx, disable it again before
2165 * setting up the full range we need.
2168 pci_disable_msix(pdev);
2170 for (i = 0; i < nr_io_queues; i++)
2171 dev->entry[i].entry = i;
2172 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2174 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2178 for (i = 0; i < vecs; i++)
2179 dev->entry[i].vector = i + pdev->irq;
2184 * Should investigate if there's a performance win from allocating
2185 * more queues than interrupt vectors; it might allow the submission
2186 * path to scale better, even if the receive path is limited by the
2187 * number of interrupts.
2189 nr_io_queues = vecs;
2190 dev->max_qid = nr_io_queues;
2192 result = queue_request_irq(dev, adminq, adminq->irqname);
2196 /* Free previously allocated queues that are no longer usable */
2197 nvme_free_queues(dev, nr_io_queues + 1);
2198 nvme_create_io_queues(dev);
2203 nvme_free_queues(dev, 1);
2208 * Return: error value if an error occurred setting up the queues or calling
2209 * Identify Device. 0 if these succeeded, even if adding some of the
2210 * namespaces failed. At the moment, these failures are silent. TBD which
2211 * failures should be reported.
2213 static int nvme_dev_add(struct nvme_dev *dev)
2215 struct pci_dev *pdev = to_pci_dev(dev->dev);
2218 struct nvme_id_ctrl *ctrl;
2219 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2221 res = nvme_identify_ctrl(dev, &ctrl);
2223 dev_err(dev->dev, "Identify Controller failed (%d)\n", res);
2227 nn = le32_to_cpup(&ctrl->nn);
2228 dev->oncs = le16_to_cpup(&ctrl->oncs);
2229 dev->abort_limit = ctrl->acl + 1;
2230 dev->vwc = ctrl->vwc;
2231 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2232 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2233 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2235 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2236 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2237 (pdev->device == 0x0953) && ctrl->vs[3]) {
2238 unsigned int max_hw_sectors;
2240 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2241 max_hw_sectors = dev->stripe_size >> (shift - 9);
2242 if (dev->max_hw_sectors) {
2243 dev->max_hw_sectors = min(max_hw_sectors,
2244 dev->max_hw_sectors);
2246 dev->max_hw_sectors = max_hw_sectors;
2250 dev->tagset.ops = &nvme_mq_ops;
2251 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2252 dev->tagset.timeout = NVME_IO_TIMEOUT;
2253 dev->tagset.numa_node = dev_to_node(dev->dev);
2254 dev->tagset.queue_depth =
2255 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2256 dev->tagset.cmd_size = nvme_cmd_size(dev);
2257 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2258 dev->tagset.driver_data = dev;
2260 if (blk_mq_alloc_tag_set(&dev->tagset))
2263 for (i = 1; i <= nn; i++)
2264 nvme_alloc_ns(dev, i);
2269 static int nvme_dev_map(struct nvme_dev *dev)
2272 int bars, result = -ENOMEM;
2273 struct pci_dev *pdev = to_pci_dev(dev->dev);
2275 if (pci_enable_device_mem(pdev))
2278 dev->entry[0].vector = pdev->irq;
2279 pci_set_master(pdev);
2280 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2284 if (pci_request_selected_regions(pdev, bars, "nvme"))
2287 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
2288 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
2291 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2295 if (readl(&dev->bar->csts) == -1) {
2301 * Some devices don't advertse INTx interrupts, pre-enable a single
2302 * MSIX vec for setup. We'll adjust this later.
2305 result = pci_enable_msix(pdev, dev->entry, 1);
2310 cap = readq(&dev->bar->cap);
2311 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2312 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2313 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2321 pci_release_regions(pdev);
2323 pci_disable_device(pdev);
2327 static void nvme_dev_unmap(struct nvme_dev *dev)
2329 struct pci_dev *pdev = to_pci_dev(dev->dev);
2331 if (pdev->msi_enabled)
2332 pci_disable_msi(pdev);
2333 else if (pdev->msix_enabled)
2334 pci_disable_msix(pdev);
2339 pci_release_regions(pdev);
2342 if (pci_is_enabled(pdev))
2343 pci_disable_device(pdev);
2346 struct nvme_delq_ctx {
2347 struct task_struct *waiter;
2348 struct kthread_worker *worker;
2352 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2354 dq->waiter = current;
2358 set_current_state(TASK_KILLABLE);
2359 if (!atomic_read(&dq->refcount))
2361 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2362 fatal_signal_pending(current)) {
2364 * Disable the controller first since we can't trust it
2365 * at this point, but leave the admin queue enabled
2366 * until all queue deletion requests are flushed.
2367 * FIXME: This may take a while if there are more h/w
2368 * queues than admin tags.
2370 set_current_state(TASK_RUNNING);
2371 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2372 nvme_clear_queue(dev->queues[0]);
2373 flush_kthread_worker(dq->worker);
2374 nvme_disable_queue(dev, 0);
2378 set_current_state(TASK_RUNNING);
2381 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2383 atomic_dec(&dq->refcount);
2385 wake_up_process(dq->waiter);
2388 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2390 atomic_inc(&dq->refcount);
2394 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2396 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2400 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2401 kthread_work_func_t fn)
2403 struct nvme_command c;
2405 memset(&c, 0, sizeof(c));
2406 c.delete_queue.opcode = opcode;
2407 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2409 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2410 return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2414 static void nvme_del_cq_work_handler(struct kthread_work *work)
2416 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2418 nvme_del_queue_end(nvmeq);
2421 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2423 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2424 nvme_del_cq_work_handler);
2427 static void nvme_del_sq_work_handler(struct kthread_work *work)
2429 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2431 int status = nvmeq->cmdinfo.status;
2434 status = nvme_delete_cq(nvmeq);
2436 nvme_del_queue_end(nvmeq);
2439 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2441 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2442 nvme_del_sq_work_handler);
2445 static void nvme_del_queue_start(struct kthread_work *work)
2447 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2449 if (nvme_delete_sq(nvmeq))
2450 nvme_del_queue_end(nvmeq);
2453 static void nvme_disable_io_queues(struct nvme_dev *dev)
2456 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2457 struct nvme_delq_ctx dq;
2458 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2459 &worker, "nvme%d", dev->instance);
2461 if (IS_ERR(kworker_task)) {
2463 "Failed to create queue del task\n");
2464 for (i = dev->queue_count - 1; i > 0; i--)
2465 nvme_disable_queue(dev, i);
2470 atomic_set(&dq.refcount, 0);
2471 dq.worker = &worker;
2472 for (i = dev->queue_count - 1; i > 0; i--) {
2473 struct nvme_queue *nvmeq = dev->queues[i];
2475 if (nvme_suspend_queue(nvmeq))
2477 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2478 nvmeq->cmdinfo.worker = dq.worker;
2479 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2480 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2482 nvme_wait_dq(&dq, dev);
2483 kthread_stop(kworker_task);
2487 * Remove the node from the device list and check
2488 * for whether or not we need to stop the nvme_thread.
2490 static void nvme_dev_list_remove(struct nvme_dev *dev)
2492 struct task_struct *tmp = NULL;
2494 spin_lock(&dev_list_lock);
2495 list_del_init(&dev->node);
2496 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2500 spin_unlock(&dev_list_lock);
2506 static void nvme_freeze_queues(struct nvme_dev *dev)
2510 list_for_each_entry(ns, &dev->namespaces, list) {
2511 blk_mq_freeze_queue_start(ns->queue);
2513 spin_lock_irq(ns->queue->queue_lock);
2514 queue_flag_set(QUEUE_FLAG_STOPPED, ns->queue);
2515 spin_unlock_irq(ns->queue->queue_lock);
2517 blk_mq_cancel_requeue_work(ns->queue);
2518 blk_mq_stop_hw_queues(ns->queue);
2522 static void nvme_unfreeze_queues(struct nvme_dev *dev)
2526 list_for_each_entry(ns, &dev->namespaces, list) {
2527 queue_flag_clear_unlocked(QUEUE_FLAG_STOPPED, ns->queue);
2528 blk_mq_unfreeze_queue(ns->queue);
2529 blk_mq_start_stopped_hw_queues(ns->queue, true);
2530 blk_mq_kick_requeue_list(ns->queue);
2534 static void nvme_dev_shutdown(struct nvme_dev *dev)
2539 nvme_dev_list_remove(dev);
2542 nvme_freeze_queues(dev);
2543 csts = readl(&dev->bar->csts);
2545 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2546 for (i = dev->queue_count - 1; i >= 0; i--) {
2547 struct nvme_queue *nvmeq = dev->queues[i];
2548 nvme_suspend_queue(nvmeq);
2551 nvme_disable_io_queues(dev);
2552 nvme_shutdown_ctrl(dev);
2553 nvme_disable_queue(dev, 0);
2555 nvme_dev_unmap(dev);
2557 for (i = dev->queue_count - 1; i >= 0; i--)
2558 nvme_clear_queue(dev->queues[i]);
2561 static void nvme_dev_remove(struct nvme_dev *dev)
2565 list_for_each_entry(ns, &dev->namespaces, list) {
2566 if (ns->disk->flags & GENHD_FL_UP) {
2567 if (blk_get_integrity(ns->disk))
2568 blk_integrity_unregister(ns->disk);
2569 del_gendisk(ns->disk);
2571 if (!blk_queue_dying(ns->queue)) {
2572 blk_mq_abort_requeue_list(ns->queue);
2573 blk_cleanup_queue(ns->queue);
2578 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2580 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2581 PAGE_SIZE, PAGE_SIZE, 0);
2582 if (!dev->prp_page_pool)
2585 /* Optimisation for I/Os between 4k and 128k */
2586 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2588 if (!dev->prp_small_pool) {
2589 dma_pool_destroy(dev->prp_page_pool);
2595 static void nvme_release_prp_pools(struct nvme_dev *dev)
2597 dma_pool_destroy(dev->prp_page_pool);
2598 dma_pool_destroy(dev->prp_small_pool);
2601 static DEFINE_IDA(nvme_instance_ida);
2603 static int nvme_set_instance(struct nvme_dev *dev)
2605 int instance, error;
2608 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2611 spin_lock(&dev_list_lock);
2612 error = ida_get_new(&nvme_instance_ida, &instance);
2613 spin_unlock(&dev_list_lock);
2614 } while (error == -EAGAIN);
2619 dev->instance = instance;
2623 static void nvme_release_instance(struct nvme_dev *dev)
2625 spin_lock(&dev_list_lock);
2626 ida_remove(&nvme_instance_ida, dev->instance);
2627 spin_unlock(&dev_list_lock);
2630 static void nvme_free_namespaces(struct nvme_dev *dev)
2632 struct nvme_ns *ns, *next;
2634 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2635 list_del(&ns->list);
2637 spin_lock(&dev_list_lock);
2638 ns->disk->private_data = NULL;
2639 spin_unlock(&dev_list_lock);
2646 static void nvme_free_dev(struct kref *kref)
2648 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2650 put_device(dev->dev);
2651 put_device(dev->device);
2652 nvme_free_namespaces(dev);
2653 nvme_release_instance(dev);
2654 blk_mq_free_tag_set(&dev->tagset);
2655 blk_put_queue(dev->admin_q);
2661 static int nvme_dev_open(struct inode *inode, struct file *f)
2663 struct nvme_dev *dev;
2664 int instance = iminor(inode);
2667 spin_lock(&dev_list_lock);
2668 list_for_each_entry(dev, &dev_list, node) {
2669 if (dev->instance == instance) {
2670 if (!dev->admin_q) {
2674 if (!kref_get_unless_zero(&dev->kref))
2676 f->private_data = dev;
2681 spin_unlock(&dev_list_lock);
2686 static int nvme_dev_release(struct inode *inode, struct file *f)
2688 struct nvme_dev *dev = f->private_data;
2689 kref_put(&dev->kref, nvme_free_dev);
2693 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2695 struct nvme_dev *dev = f->private_data;
2699 case NVME_IOCTL_ADMIN_CMD:
2700 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2701 case NVME_IOCTL_IO_CMD:
2702 if (list_empty(&dev->namespaces))
2704 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2705 return nvme_user_cmd(dev, ns, (void __user *)arg);
2711 static const struct file_operations nvme_dev_fops = {
2712 .owner = THIS_MODULE,
2713 .open = nvme_dev_open,
2714 .release = nvme_dev_release,
2715 .unlocked_ioctl = nvme_dev_ioctl,
2716 .compat_ioctl = nvme_dev_ioctl,
2719 static void nvme_set_irq_hints(struct nvme_dev *dev)
2721 struct nvme_queue *nvmeq;
2724 for (i = 0; i < dev->online_queues; i++) {
2725 nvmeq = dev->queues[i];
2730 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2731 nvmeq->hctx->cpumask);
2735 static int nvme_dev_start(struct nvme_dev *dev)
2738 bool start_thread = false;
2740 result = nvme_dev_map(dev);
2744 result = nvme_configure_admin_queue(dev);
2748 spin_lock(&dev_list_lock);
2749 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2750 start_thread = true;
2753 list_add(&dev->node, &dev_list);
2754 spin_unlock(&dev_list_lock);
2757 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2758 wake_up_all(&nvme_kthread_wait);
2760 wait_event_killable(nvme_kthread_wait, nvme_thread);
2762 if (IS_ERR_OR_NULL(nvme_thread)) {
2763 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2767 nvme_init_queue(dev->queues[0], 0);
2768 result = nvme_alloc_admin_tags(dev);
2772 result = nvme_setup_io_queues(dev);
2776 nvme_set_irq_hints(dev);
2778 dev->event_limit = 1;
2782 nvme_dev_remove_admin(dev);
2784 nvme_disable_queue(dev, 0);
2785 nvme_dev_list_remove(dev);
2787 nvme_dev_unmap(dev);
2791 static int nvme_remove_dead_ctrl(void *arg)
2793 struct nvme_dev *dev = (struct nvme_dev *)arg;
2794 struct pci_dev *pdev = to_pci_dev(dev->dev);
2796 if (pci_get_drvdata(pdev))
2797 pci_stop_and_remove_bus_device_locked(pdev);
2798 kref_put(&dev->kref, nvme_free_dev);
2802 static void nvme_remove_disks(struct work_struct *ws)
2804 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2806 nvme_free_queues(dev, 1);
2807 nvme_dev_remove(dev);
2810 static int nvme_dev_resume(struct nvme_dev *dev)
2814 ret = nvme_dev_start(dev);
2817 if (dev->online_queues < 2) {
2818 spin_lock(&dev_list_lock);
2819 dev->reset_workfn = nvme_remove_disks;
2820 queue_work(nvme_workq, &dev->reset_work);
2821 spin_unlock(&dev_list_lock);
2823 nvme_unfreeze_queues(dev);
2824 nvme_set_irq_hints(dev);
2829 static void nvme_dev_reset(struct nvme_dev *dev)
2831 nvme_dev_shutdown(dev);
2832 if (nvme_dev_resume(dev)) {
2833 dev_warn(dev->dev, "Device failed to resume\n");
2834 kref_get(&dev->kref);
2835 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2838 "Failed to start controller remove task\n");
2839 kref_put(&dev->kref, nvme_free_dev);
2844 static void nvme_reset_failed_dev(struct work_struct *ws)
2846 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2847 nvme_dev_reset(dev);
2850 static void nvme_reset_workfn(struct work_struct *work)
2852 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2853 dev->reset_workfn(work);
2856 static void nvme_async_probe(struct work_struct *work);
2857 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2859 int node, result = -ENOMEM;
2860 struct nvme_dev *dev;
2862 node = dev_to_node(&pdev->dev);
2863 if (node == NUMA_NO_NODE)
2864 set_dev_node(&pdev->dev, 0);
2866 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2869 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
2873 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2878 INIT_LIST_HEAD(&dev->namespaces);
2879 dev->reset_workfn = nvme_reset_failed_dev;
2880 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
2881 dev->dev = get_device(&pdev->dev);
2882 pci_set_drvdata(pdev, dev);
2883 result = nvme_set_instance(dev);
2887 result = nvme_setup_prp_pools(dev);
2891 kref_init(&dev->kref);
2892 dev->device = device_create(nvme_class, &pdev->dev,
2893 MKDEV(nvme_char_major, dev->instance),
2894 dev, "nvme%d", dev->instance);
2895 if (IS_ERR(dev->device)) {
2896 result = PTR_ERR(dev->device);
2899 get_device(dev->device);
2901 INIT_LIST_HEAD(&dev->node);
2902 INIT_WORK(&dev->probe_work, nvme_async_probe);
2903 schedule_work(&dev->probe_work);
2907 nvme_release_prp_pools(dev);
2909 nvme_release_instance(dev);
2911 put_device(dev->dev);
2919 static void nvme_async_probe(struct work_struct *work)
2921 struct nvme_dev *dev = container_of(work, struct nvme_dev, probe_work);
2924 result = nvme_dev_start(dev);
2928 if (dev->online_queues > 1)
2929 result = nvme_dev_add(dev);
2933 nvme_set_irq_hints(dev);
2936 if (!work_busy(&dev->reset_work)) {
2937 dev->reset_workfn = nvme_reset_failed_dev;
2938 queue_work(nvme_workq, &dev->reset_work);
2942 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2944 struct nvme_dev *dev = pci_get_drvdata(pdev);
2947 nvme_dev_shutdown(dev);
2949 nvme_dev_resume(dev);
2952 static void nvme_shutdown(struct pci_dev *pdev)
2954 struct nvme_dev *dev = pci_get_drvdata(pdev);
2955 nvme_dev_shutdown(dev);
2958 static void nvme_remove(struct pci_dev *pdev)
2960 struct nvme_dev *dev = pci_get_drvdata(pdev);
2962 spin_lock(&dev_list_lock);
2963 list_del_init(&dev->node);
2964 spin_unlock(&dev_list_lock);
2966 pci_set_drvdata(pdev, NULL);
2967 flush_work(&dev->probe_work);
2968 flush_work(&dev->reset_work);
2969 nvme_dev_shutdown(dev);
2970 nvme_dev_remove(dev);
2971 nvme_dev_remove_admin(dev);
2972 device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
2973 nvme_free_queues(dev, 0);
2974 nvme_release_prp_pools(dev);
2975 kref_put(&dev->kref, nvme_free_dev);
2978 /* These functions are yet to be implemented */
2979 #define nvme_error_detected NULL
2980 #define nvme_dump_registers NULL
2981 #define nvme_link_reset NULL
2982 #define nvme_slot_reset NULL
2983 #define nvme_error_resume NULL
2985 #ifdef CONFIG_PM_SLEEP
2986 static int nvme_suspend(struct device *dev)
2988 struct pci_dev *pdev = to_pci_dev(dev);
2989 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2991 nvme_dev_shutdown(ndev);
2995 static int nvme_resume(struct device *dev)
2997 struct pci_dev *pdev = to_pci_dev(dev);
2998 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3000 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
3001 ndev->reset_workfn = nvme_reset_failed_dev;
3002 queue_work(nvme_workq, &ndev->reset_work);
3008 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
3010 static const struct pci_error_handlers nvme_err_handler = {
3011 .error_detected = nvme_error_detected,
3012 .mmio_enabled = nvme_dump_registers,
3013 .link_reset = nvme_link_reset,
3014 .slot_reset = nvme_slot_reset,
3015 .resume = nvme_error_resume,
3016 .reset_notify = nvme_reset_notify,
3019 /* Move to pci_ids.h later */
3020 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
3022 static const struct pci_device_id nvme_id_table[] = {
3023 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3026 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3028 static struct pci_driver nvme_driver = {
3030 .id_table = nvme_id_table,
3031 .probe = nvme_probe,
3032 .remove = nvme_remove,
3033 .shutdown = nvme_shutdown,
3035 .pm = &nvme_dev_pm_ops,
3037 .err_handler = &nvme_err_handler,
3040 static int __init nvme_init(void)
3044 init_waitqueue_head(&nvme_kthread_wait);
3046 nvme_workq = create_singlethread_workqueue("nvme");
3050 result = register_blkdev(nvme_major, "nvme");
3053 else if (result > 0)
3054 nvme_major = result;
3056 result = __register_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme",
3059 goto unregister_blkdev;
3060 else if (result > 0)
3061 nvme_char_major = result;
3063 nvme_class = class_create(THIS_MODULE, "nvme");
3064 if (IS_ERR(nvme_class)) {
3065 result = PTR_ERR(nvme_class);
3066 goto unregister_chrdev;
3069 result = pci_register_driver(&nvme_driver);
3075 class_destroy(nvme_class);
3077 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3079 unregister_blkdev(nvme_major, "nvme");
3081 destroy_workqueue(nvme_workq);
3085 static void __exit nvme_exit(void)
3087 pci_unregister_driver(&nvme_driver);
3088 unregister_blkdev(nvme_major, "nvme");
3089 destroy_workqueue(nvme_workq);
3090 class_destroy(nvme_class);
3091 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3092 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3096 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3097 MODULE_LICENSE("GPL");
3098 MODULE_VERSION("1.0");
3099 module_init(nvme_init);
3100 module_exit(nvme_exit);