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/types.h>
42 #include <asm-generic/io-64-nonatomic-lo-hi.h>
44 #define NVME_Q_DEPTH 1024
45 #define NVME_AQ_DEPTH 64
46 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
47 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
48 #define ADMIN_TIMEOUT (admin_timeout * HZ)
49 #define SHUTDOWN_TIMEOUT (shutdown_timeout * HZ)
50 #define IOD_TIMEOUT (retry_time * HZ)
52 static unsigned char admin_timeout = 60;
53 module_param(admin_timeout, byte, 0644);
54 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
56 unsigned char nvme_io_timeout = 30;
57 module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
58 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
60 static unsigned char retry_time = 30;
61 module_param(retry_time, byte, 0644);
62 MODULE_PARM_DESC(retry_time, "time in seconds to retry failed I/O");
64 static unsigned char shutdown_timeout = 5;
65 module_param(shutdown_timeout, byte, 0644);
66 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
68 static int nvme_major;
69 module_param(nvme_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;
79 static struct notifier_block nvme_nb;
81 static void nvme_reset_failed_dev(struct work_struct *ws);
82 static int nvme_process_cq(struct nvme_queue *nvmeq);
84 struct async_cmd_info {
85 struct kthread_work work;
86 struct kthread_worker *worker;
94 * An NVM Express queue. Each device has at least two (one for admin
95 * commands and one for I/O commands).
98 struct llist_node node;
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)
157 * Will slightly overestimate the number of pages needed. This is OK
158 * as it only leads to a small amount of wasted memory for the lifetime of
161 static int nvme_npages(unsigned size, struct nvme_dev *dev)
163 unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
164 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
167 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
169 unsigned int ret = sizeof(struct nvme_cmd_info);
171 ret += sizeof(struct nvme_iod);
172 ret += sizeof(__le64 *) * nvme_npages(NVME_INT_BYTES(dev), dev);
173 ret += sizeof(struct scatterlist) * NVME_INT_PAGES;
178 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
179 unsigned int hctx_idx)
181 struct nvme_dev *dev = data;
182 struct nvme_queue *nvmeq = dev->queues[0];
184 WARN_ON(nvmeq->hctx);
186 hctx->driver_data = nvmeq;
190 static int nvme_admin_init_request(void *data, struct request *req,
191 unsigned int hctx_idx, unsigned int rq_idx,
192 unsigned int numa_node)
194 struct nvme_dev *dev = data;
195 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
196 struct nvme_queue *nvmeq = dev->queues[0];
203 static void nvme_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
205 struct nvme_queue *nvmeq = hctx->driver_data;
210 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
211 unsigned int hctx_idx)
213 struct nvme_dev *dev = data;
214 struct nvme_queue *nvmeq = dev->queues[
215 (hctx_idx % dev->queue_count) + 1];
220 /* nvmeq queues are shared between namespaces. We assume here that
221 * blk-mq map the tags so they match up with the nvme queue tags. */
222 WARN_ON(nvmeq->hctx->tags != hctx->tags);
224 hctx->driver_data = nvmeq;
228 static int nvme_init_request(void *data, struct request *req,
229 unsigned int hctx_idx, unsigned int rq_idx,
230 unsigned int numa_node)
232 struct nvme_dev *dev = data;
233 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
234 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
241 static void nvme_set_info(struct nvme_cmd_info *cmd, void *ctx,
242 nvme_completion_fn handler)
247 blk_mq_start_request(blk_mq_rq_from_pdu(cmd));
250 static void *iod_get_private(struct nvme_iod *iod)
252 return (void *) (iod->private & ~0x1UL);
256 * If bit 0 is set, the iod is embedded in the request payload.
258 static bool iod_should_kfree(struct nvme_iod *iod)
260 return (iod->private & 0x01) == 0;
263 /* Special values must be less than 0x1000 */
264 #define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
265 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
266 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
267 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
269 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
270 struct nvme_completion *cqe)
272 if (ctx == CMD_CTX_CANCELLED)
274 if (ctx == CMD_CTX_COMPLETED) {
275 dev_warn(nvmeq->q_dmadev,
276 "completed id %d twice on queue %d\n",
277 cqe->command_id, le16_to_cpup(&cqe->sq_id));
280 if (ctx == CMD_CTX_INVALID) {
281 dev_warn(nvmeq->q_dmadev,
282 "invalid id %d completed on queue %d\n",
283 cqe->command_id, le16_to_cpup(&cqe->sq_id));
286 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
289 static void *cancel_cmd_info(struct nvme_cmd_info *cmd, nvme_completion_fn *fn)
296 cmd->fn = special_completion;
297 cmd->ctx = CMD_CTX_CANCELLED;
301 static void async_req_completion(struct nvme_queue *nvmeq, void *ctx,
302 struct nvme_completion *cqe)
304 struct request *req = ctx;
306 u32 result = le32_to_cpup(&cqe->result);
307 u16 status = le16_to_cpup(&cqe->status) >> 1;
309 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
310 ++nvmeq->dev->event_limit;
311 if (status == NVME_SC_SUCCESS)
312 dev_warn(nvmeq->q_dmadev,
313 "async event result %08x\n", result);
315 blk_mq_free_hctx_request(nvmeq->hctx, req);
318 static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
319 struct nvme_completion *cqe)
321 struct request *req = ctx;
323 u16 status = le16_to_cpup(&cqe->status) >> 1;
324 u32 result = le32_to_cpup(&cqe->result);
326 blk_mq_free_hctx_request(nvmeq->hctx, req);
328 dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
329 ++nvmeq->dev->abort_limit;
332 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
333 struct nvme_completion *cqe)
335 struct async_cmd_info *cmdinfo = ctx;
336 cmdinfo->result = le32_to_cpup(&cqe->result);
337 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
338 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
339 blk_mq_free_hctx_request(nvmeq->hctx, cmdinfo->req);
342 static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
345 struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
346 struct request *req = blk_mq_tag_to_rq(hctx->tags, tag);
348 return blk_mq_rq_to_pdu(req);
352 * Called with local interrupts disabled and the q_lock held. May not sleep.
354 static void *nvme_finish_cmd(struct nvme_queue *nvmeq, int tag,
355 nvme_completion_fn *fn)
357 struct nvme_cmd_info *cmd = get_cmd_from_tag(nvmeq, tag);
359 if (tag >= nvmeq->q_depth) {
360 *fn = special_completion;
361 return CMD_CTX_INVALID;
366 cmd->fn = special_completion;
367 cmd->ctx = CMD_CTX_COMPLETED;
372 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
373 * @nvmeq: The queue to use
374 * @cmd: The command to send
376 * Safe to use from interrupt context
378 static int __nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
380 u16 tail = nvmeq->sq_tail;
382 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
383 if (++tail == nvmeq->q_depth)
385 writel(tail, nvmeq->q_db);
386 nvmeq->sq_tail = tail;
391 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
395 spin_lock_irqsave(&nvmeq->q_lock, flags);
396 ret = __nvme_submit_cmd(nvmeq, cmd);
397 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
401 static __le64 **iod_list(struct nvme_iod *iod)
403 return ((void *)iod) + iod->offset;
406 static inline void iod_init(struct nvme_iod *iod, unsigned nbytes,
407 unsigned nseg, unsigned long private)
409 iod->private = private;
410 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
412 iod->length = nbytes;
416 static struct nvme_iod *
417 __nvme_alloc_iod(unsigned nseg, unsigned bytes, struct nvme_dev *dev,
418 unsigned long priv, gfp_t gfp)
420 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
421 sizeof(__le64 *) * nvme_npages(bytes, dev) +
422 sizeof(struct scatterlist) * nseg, gfp);
425 iod_init(iod, bytes, nseg, priv);
430 static struct nvme_iod *nvme_alloc_iod(struct request *rq, struct nvme_dev *dev,
433 unsigned size = !(rq->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(rq) :
434 sizeof(struct nvme_dsm_range);
435 unsigned long mask = 0;
436 struct nvme_iod *iod;
438 if (rq->nr_phys_segments <= NVME_INT_PAGES &&
439 size <= NVME_INT_BYTES(dev)) {
440 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(rq);
444 iod_init(iod, size, rq->nr_phys_segments,
445 (unsigned long) rq | 0x01);
449 return __nvme_alloc_iod(rq->nr_phys_segments, size, dev,
450 (unsigned long) rq, gfp);
453 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
455 const int last_prp = dev->page_size / 8 - 1;
457 __le64 **list = iod_list(iod);
458 dma_addr_t prp_dma = iod->first_dma;
460 if (iod->npages == 0)
461 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
462 for (i = 0; i < iod->npages; i++) {
463 __le64 *prp_list = list[i];
464 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
465 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
466 prp_dma = next_prp_dma;
469 if (iod_should_kfree(iod))
473 static int nvme_error_status(u16 status)
475 switch (status & 0x7ff) {
476 case NVME_SC_SUCCESS:
478 case NVME_SC_CAP_EXCEEDED:
485 static void req_completion(struct nvme_queue *nvmeq, void *ctx,
486 struct nvme_completion *cqe)
488 struct nvme_iod *iod = ctx;
489 struct request *req = iod_get_private(iod);
490 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
492 u16 status = le16_to_cpup(&cqe->status) >> 1;
494 if (unlikely(status)) {
495 if (!(status & NVME_SC_DNR || blk_noretry_request(req))
496 && (jiffies - req->start_time) < req->timeout) {
499 blk_mq_requeue_request(req);
500 spin_lock_irqsave(req->q->queue_lock, flags);
501 if (!blk_queue_stopped(req->q))
502 blk_mq_kick_requeue_list(req->q);
503 spin_unlock_irqrestore(req->q->queue_lock, flags);
506 req->errors = nvme_error_status(status);
511 dev_warn(&nvmeq->dev->pci_dev->dev,
512 "completing aborted command with status:%04x\n",
516 dma_unmap_sg(&nvmeq->dev->pci_dev->dev, iod->sg, iod->nents,
517 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
518 nvme_free_iod(nvmeq->dev, iod);
520 blk_mq_complete_request(req);
523 /* length is in bytes. gfp flags indicates whether we may sleep. */
524 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod, int total_len,
527 struct dma_pool *pool;
528 int length = total_len;
529 struct scatterlist *sg = iod->sg;
530 int dma_len = sg_dma_len(sg);
531 u64 dma_addr = sg_dma_address(sg);
532 int offset = offset_in_page(dma_addr);
534 __le64 **list = iod_list(iod);
537 u32 page_size = dev->page_size;
539 length -= (page_size - offset);
543 dma_len -= (page_size - offset);
545 dma_addr += (page_size - offset);
548 dma_addr = sg_dma_address(sg);
549 dma_len = sg_dma_len(sg);
552 if (length <= page_size) {
553 iod->first_dma = dma_addr;
557 nprps = DIV_ROUND_UP(length, page_size);
558 if (nprps <= (256 / 8)) {
559 pool = dev->prp_small_pool;
562 pool = dev->prp_page_pool;
566 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
568 iod->first_dma = dma_addr;
570 return (total_len - length) + page_size;
573 iod->first_dma = prp_dma;
576 if (i == page_size >> 3) {
577 __le64 *old_prp_list = prp_list;
578 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
580 return total_len - length;
581 list[iod->npages++] = prp_list;
582 prp_list[0] = old_prp_list[i - 1];
583 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
586 prp_list[i++] = cpu_to_le64(dma_addr);
587 dma_len -= page_size;
588 dma_addr += page_size;
596 dma_addr = sg_dma_address(sg);
597 dma_len = sg_dma_len(sg);
604 * We reuse the small pool to allocate the 16-byte range here as it is not
605 * worth having a special pool for these or additional cases to handle freeing
608 static void nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
609 struct request *req, struct nvme_iod *iod)
611 struct nvme_dsm_range *range =
612 (struct nvme_dsm_range *)iod_list(iod)[0];
613 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
615 range->cattr = cpu_to_le32(0);
616 range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift);
617 range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
619 memset(cmnd, 0, sizeof(*cmnd));
620 cmnd->dsm.opcode = nvme_cmd_dsm;
621 cmnd->dsm.command_id = req->tag;
622 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
623 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
625 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
627 if (++nvmeq->sq_tail == nvmeq->q_depth)
629 writel(nvmeq->sq_tail, nvmeq->q_db);
632 static void nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
635 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
637 memset(cmnd, 0, sizeof(*cmnd));
638 cmnd->common.opcode = nvme_cmd_flush;
639 cmnd->common.command_id = cmdid;
640 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
642 if (++nvmeq->sq_tail == nvmeq->q_depth)
644 writel(nvmeq->sq_tail, nvmeq->q_db);
647 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod,
650 struct request *req = iod_get_private(iod);
651 struct nvme_command *cmnd;
655 if (req->cmd_flags & REQ_FUA)
656 control |= NVME_RW_FUA;
657 if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
658 control |= NVME_RW_LR;
660 if (req->cmd_flags & REQ_RAHEAD)
661 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
663 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
664 memset(cmnd, 0, sizeof(*cmnd));
666 cmnd->rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
667 cmnd->rw.command_id = req->tag;
668 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
669 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
670 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
671 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
672 cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
673 cmnd->rw.control = cpu_to_le16(control);
674 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
676 if (++nvmeq->sq_tail == nvmeq->q_depth)
678 writel(nvmeq->sq_tail, nvmeq->q_db);
683 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
684 const struct blk_mq_queue_data *bd)
686 struct nvme_ns *ns = hctx->queue->queuedata;
687 struct nvme_queue *nvmeq = hctx->driver_data;
688 struct request *req = bd->rq;
689 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
690 struct nvme_iod *iod;
691 enum dma_data_direction dma_dir;
693 iod = nvme_alloc_iod(req, ns->dev, GFP_ATOMIC);
695 return BLK_MQ_RQ_QUEUE_BUSY;
697 if (req->cmd_flags & REQ_DISCARD) {
700 * We reuse the small pool to allocate the 16-byte range here
701 * as it is not worth having a special pool for these or
702 * additional cases to handle freeing the iod.
704 range = dma_pool_alloc(nvmeq->dev->prp_small_pool,
709 iod_list(iod)[0] = (__le64 *)range;
711 } else if (req->nr_phys_segments) {
712 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
714 sg_init_table(iod->sg, req->nr_phys_segments);
715 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
719 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
722 if (blk_rq_bytes(req) !=
723 nvme_setup_prps(nvmeq->dev, iod, blk_rq_bytes(req), GFP_ATOMIC)) {
724 dma_unmap_sg(&nvmeq->dev->pci_dev->dev, iod->sg,
725 iod->nents, dma_dir);
730 nvme_set_info(cmd, iod, req_completion);
731 spin_lock_irq(&nvmeq->q_lock);
732 if (req->cmd_flags & REQ_DISCARD)
733 nvme_submit_discard(nvmeq, ns, req, iod);
734 else if (req->cmd_flags & REQ_FLUSH)
735 nvme_submit_flush(nvmeq, ns, req->tag);
737 nvme_submit_iod(nvmeq, iod, ns);
739 nvme_process_cq(nvmeq);
740 spin_unlock_irq(&nvmeq->q_lock);
741 return BLK_MQ_RQ_QUEUE_OK;
744 nvme_free_iod(nvmeq->dev, iod);
745 return BLK_MQ_RQ_QUEUE_ERROR;
747 nvme_free_iod(nvmeq->dev, iod);
748 return BLK_MQ_RQ_QUEUE_BUSY;
751 static int nvme_process_cq(struct nvme_queue *nvmeq)
755 head = nvmeq->cq_head;
756 phase = nvmeq->cq_phase;
760 nvme_completion_fn fn;
761 struct nvme_completion cqe = nvmeq->cqes[head];
762 if ((le16_to_cpu(cqe.status) & 1) != phase)
764 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
765 if (++head == nvmeq->q_depth) {
769 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
770 fn(nvmeq, ctx, &cqe);
773 /* If the controller ignores the cq head doorbell and continuously
774 * writes to the queue, it is theoretically possible to wrap around
775 * the queue twice and mistakenly return IRQ_NONE. Linux only
776 * requires that 0.1% of your interrupts are handled, so this isn't
779 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
782 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
783 nvmeq->cq_head = head;
784 nvmeq->cq_phase = phase;
790 /* Admin queue isn't initialized as a request queue. If at some point this
791 * happens anyway, make sure to notify the user */
792 static int nvme_admin_queue_rq(struct blk_mq_hw_ctx *hctx,
793 const struct blk_mq_queue_data *bd)
796 return BLK_MQ_RQ_QUEUE_ERROR;
799 static irqreturn_t nvme_irq(int irq, void *data)
802 struct nvme_queue *nvmeq = data;
803 spin_lock(&nvmeq->q_lock);
804 nvme_process_cq(nvmeq);
805 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
807 spin_unlock(&nvmeq->q_lock);
811 static irqreturn_t nvme_irq_check(int irq, void *data)
813 struct nvme_queue *nvmeq = data;
814 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
815 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
817 return IRQ_WAKE_THREAD;
820 static void nvme_abort_cmd_info(struct nvme_queue *nvmeq, struct nvme_cmd_info *
823 spin_lock_irq(&nvmeq->q_lock);
824 cancel_cmd_info(cmd_info, NULL);
825 spin_unlock_irq(&nvmeq->q_lock);
828 struct sync_cmd_info {
829 struct task_struct *task;
834 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
835 struct nvme_completion *cqe)
837 struct sync_cmd_info *cmdinfo = ctx;
838 cmdinfo->result = le32_to_cpup(&cqe->result);
839 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
840 wake_up_process(cmdinfo->task);
844 * Returns 0 on success. If the result is negative, it's a Linux error code;
845 * if the result is positive, it's an NVM Express status code
847 static int nvme_submit_sync_cmd(struct request *req, struct nvme_command *cmd,
848 u32 *result, unsigned timeout)
851 struct sync_cmd_info cmdinfo;
852 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
853 struct nvme_queue *nvmeq = cmd_rq->nvmeq;
855 cmdinfo.task = current;
856 cmdinfo.status = -EINTR;
858 cmd->common.command_id = req->tag;
860 nvme_set_info(cmd_rq, &cmdinfo, sync_completion);
862 set_current_state(TASK_KILLABLE);
863 ret = nvme_submit_cmd(nvmeq, cmd);
865 nvme_finish_cmd(nvmeq, req->tag, NULL);
866 set_current_state(TASK_RUNNING);
868 ret = schedule_timeout(timeout);
871 * Ensure that sync_completion has either run, or that it will
874 nvme_abort_cmd_info(nvmeq, blk_mq_rq_to_pdu(req));
877 * We never got the completion
879 if (cmdinfo.status == -EINTR)
883 *result = cmdinfo.result;
885 return cmdinfo.status;
888 static int nvme_submit_async_admin_req(struct nvme_dev *dev)
890 struct nvme_queue *nvmeq = dev->queues[0];
891 struct nvme_command c;
892 struct nvme_cmd_info *cmd_info;
895 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, false);
899 req->cmd_flags |= REQ_NO_TIMEOUT;
900 cmd_info = blk_mq_rq_to_pdu(req);
901 nvme_set_info(cmd_info, req, async_req_completion);
903 memset(&c, 0, sizeof(c));
904 c.common.opcode = nvme_admin_async_event;
905 c.common.command_id = req->tag;
907 return __nvme_submit_cmd(nvmeq, &c);
910 static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
911 struct nvme_command *cmd,
912 struct async_cmd_info *cmdinfo, unsigned timeout)
914 struct nvme_queue *nvmeq = dev->queues[0];
916 struct nvme_cmd_info *cmd_rq;
918 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
922 req->timeout = timeout;
923 cmd_rq = blk_mq_rq_to_pdu(req);
925 nvme_set_info(cmd_rq, cmdinfo, async_completion);
926 cmdinfo->status = -EINTR;
928 cmd->common.command_id = req->tag;
930 return nvme_submit_cmd(nvmeq, cmd);
933 static int __nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
934 u32 *result, unsigned timeout)
939 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
942 res = nvme_submit_sync_cmd(req, cmd, result, timeout);
943 blk_mq_free_request(req);
947 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
950 return __nvme_submit_admin_cmd(dev, cmd, result, ADMIN_TIMEOUT);
953 int nvme_submit_io_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
954 struct nvme_command *cmd, u32 *result)
959 req = blk_mq_alloc_request(ns->queue, WRITE, (GFP_KERNEL|__GFP_WAIT),
963 res = nvme_submit_sync_cmd(req, cmd, result, NVME_IO_TIMEOUT);
964 blk_mq_free_request(req);
968 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
970 struct nvme_command c;
972 memset(&c, 0, sizeof(c));
973 c.delete_queue.opcode = opcode;
974 c.delete_queue.qid = cpu_to_le16(id);
976 return nvme_submit_admin_cmd(dev, &c, NULL);
979 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
980 struct nvme_queue *nvmeq)
982 struct nvme_command c;
983 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
985 memset(&c, 0, sizeof(c));
986 c.create_cq.opcode = nvme_admin_create_cq;
987 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
988 c.create_cq.cqid = cpu_to_le16(qid);
989 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
990 c.create_cq.cq_flags = cpu_to_le16(flags);
991 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
993 return nvme_submit_admin_cmd(dev, &c, NULL);
996 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
997 struct nvme_queue *nvmeq)
999 struct nvme_command c;
1000 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1002 memset(&c, 0, sizeof(c));
1003 c.create_sq.opcode = nvme_admin_create_sq;
1004 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1005 c.create_sq.sqid = cpu_to_le16(qid);
1006 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1007 c.create_sq.sq_flags = cpu_to_le16(flags);
1008 c.create_sq.cqid = cpu_to_le16(qid);
1010 return nvme_submit_admin_cmd(dev, &c, NULL);
1013 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1015 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1018 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1020 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1023 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
1024 dma_addr_t dma_addr)
1026 struct nvme_command c;
1028 memset(&c, 0, sizeof(c));
1029 c.identify.opcode = nvme_admin_identify;
1030 c.identify.nsid = cpu_to_le32(nsid);
1031 c.identify.prp1 = cpu_to_le64(dma_addr);
1032 c.identify.cns = cpu_to_le32(cns);
1034 return nvme_submit_admin_cmd(dev, &c, NULL);
1037 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1038 dma_addr_t dma_addr, u32 *result)
1040 struct nvme_command c;
1042 memset(&c, 0, sizeof(c));
1043 c.features.opcode = nvme_admin_get_features;
1044 c.features.nsid = cpu_to_le32(nsid);
1045 c.features.prp1 = cpu_to_le64(dma_addr);
1046 c.features.fid = cpu_to_le32(fid);
1048 return nvme_submit_admin_cmd(dev, &c, result);
1051 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1052 dma_addr_t dma_addr, u32 *result)
1054 struct nvme_command c;
1056 memset(&c, 0, sizeof(c));
1057 c.features.opcode = nvme_admin_set_features;
1058 c.features.prp1 = cpu_to_le64(dma_addr);
1059 c.features.fid = cpu_to_le32(fid);
1060 c.features.dword11 = cpu_to_le32(dword11);
1062 return nvme_submit_admin_cmd(dev, &c, result);
1066 * nvme_abort_req - Attempt aborting a request
1068 * Schedule controller reset if the command was already aborted once before and
1069 * still hasn't been returned to the driver, or if this is the admin queue.
1071 static void nvme_abort_req(struct request *req)
1073 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1074 struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1075 struct nvme_dev *dev = nvmeq->dev;
1076 struct request *abort_req;
1077 struct nvme_cmd_info *abort_cmd;
1078 struct nvme_command cmd;
1080 if (!nvmeq->qid || cmd_rq->aborted) {
1081 unsigned long flags;
1083 spin_lock_irqsave(&dev_list_lock, flags);
1084 if (work_busy(&dev->reset_work))
1086 list_del_init(&dev->node);
1087 dev_warn(&dev->pci_dev->dev,
1088 "I/O %d QID %d timeout, reset controller\n",
1089 req->tag, nvmeq->qid);
1090 dev->reset_workfn = nvme_reset_failed_dev;
1091 queue_work(nvme_workq, &dev->reset_work);
1093 spin_unlock_irqrestore(&dev_list_lock, flags);
1097 if (!dev->abort_limit)
1100 abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1102 if (IS_ERR(abort_req))
1105 abort_cmd = blk_mq_rq_to_pdu(abort_req);
1106 nvme_set_info(abort_cmd, abort_req, abort_completion);
1108 memset(&cmd, 0, sizeof(cmd));
1109 cmd.abort.opcode = nvme_admin_abort_cmd;
1110 cmd.abort.cid = req->tag;
1111 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1112 cmd.abort.command_id = abort_req->tag;
1115 cmd_rq->aborted = 1;
1117 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1119 if (nvme_submit_cmd(dev->queues[0], &cmd) < 0) {
1120 dev_warn(nvmeq->q_dmadev,
1121 "Could not abort I/O %d QID %d",
1122 req->tag, nvmeq->qid);
1123 blk_mq_free_request(abort_req);
1127 static void nvme_cancel_queue_ios(struct blk_mq_hw_ctx *hctx,
1128 struct request *req, void *data, bool reserved)
1130 struct nvme_queue *nvmeq = data;
1132 nvme_completion_fn fn;
1133 struct nvme_cmd_info *cmd;
1134 struct nvme_completion cqe;
1136 if (!blk_mq_request_started(req))
1139 cmd = blk_mq_rq_to_pdu(req);
1141 if (cmd->ctx == CMD_CTX_CANCELLED)
1144 if (blk_queue_dying(req->q))
1145 cqe.status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1147 cqe.status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1150 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1151 req->tag, nvmeq->qid);
1152 ctx = cancel_cmd_info(cmd, &fn);
1153 fn(nvmeq, ctx, &cqe);
1156 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1158 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1159 struct nvme_queue *nvmeq = cmd->nvmeq;
1162 * The aborted req will be completed on receiving the abort req.
1163 * We enable the timer again. If hit twice, it'll cause a device reset,
1164 * as the device then is in a faulty state.
1166 int ret = BLK_EH_RESET_TIMER;
1168 dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1171 spin_lock_irq(&nvmeq->q_lock);
1172 if (!nvmeq->dev->initialized) {
1174 * Force cancelled command frees the request, which requires we
1175 * return BLK_EH_NOT_HANDLED.
1177 nvme_cancel_queue_ios(nvmeq->hctx, req, nvmeq, reserved);
1178 ret = BLK_EH_NOT_HANDLED;
1180 nvme_abort_req(req);
1181 spin_unlock_irq(&nvmeq->q_lock);
1186 static void nvme_free_queue(struct nvme_queue *nvmeq)
1188 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1189 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1190 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1191 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1195 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1199 for (i = dev->queue_count - 1; i >= lowest; i--) {
1200 struct nvme_queue *nvmeq = dev->queues[i];
1202 dev->queues[i] = NULL;
1203 nvme_free_queue(nvmeq);
1208 * nvme_suspend_queue - put queue into suspended state
1209 * @nvmeq - queue to suspend
1211 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1215 spin_lock_irq(&nvmeq->q_lock);
1216 if (nvmeq->cq_vector == -1) {
1217 spin_unlock_irq(&nvmeq->q_lock);
1220 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1221 nvmeq->dev->online_queues--;
1222 nvmeq->cq_vector = -1;
1223 spin_unlock_irq(&nvmeq->q_lock);
1225 irq_set_affinity_hint(vector, NULL);
1226 free_irq(vector, nvmeq);
1231 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1233 struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
1235 spin_lock_irq(&nvmeq->q_lock);
1236 nvme_process_cq(nvmeq);
1237 if (hctx && hctx->tags)
1238 blk_mq_tag_busy_iter(hctx, nvme_cancel_queue_ios, nvmeq);
1239 spin_unlock_irq(&nvmeq->q_lock);
1242 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1244 struct nvme_queue *nvmeq = dev->queues[qid];
1248 if (nvme_suspend_queue(nvmeq))
1251 /* Don't tell the adapter to delete the admin queue.
1252 * Don't tell a removed adapter to delete IO queues. */
1253 if (qid && readl(&dev->bar->csts) != -1) {
1254 adapter_delete_sq(dev, qid);
1255 adapter_delete_cq(dev, qid);
1257 if (!qid && dev->admin_q)
1258 blk_mq_freeze_queue_start(dev->admin_q);
1259 nvme_clear_queue(nvmeq);
1262 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1265 struct device *dmadev = &dev->pci_dev->dev;
1266 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1270 nvmeq->cqes = dma_zalloc_coherent(dmadev, CQ_SIZE(depth),
1271 &nvmeq->cq_dma_addr, GFP_KERNEL);
1275 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1276 &nvmeq->sq_dma_addr, GFP_KERNEL);
1277 if (!nvmeq->sq_cmds)
1280 nvmeq->q_dmadev = dmadev;
1282 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1283 dev->instance, qid);
1284 spin_lock_init(&nvmeq->q_lock);
1286 nvmeq->cq_phase = 1;
1287 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1288 nvmeq->q_depth = depth;
1291 dev->queues[qid] = nvmeq;
1296 dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1297 nvmeq->cq_dma_addr);
1303 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1306 if (use_threaded_interrupts)
1307 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1308 nvme_irq_check, nvme_irq, IRQF_SHARED,
1310 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1311 IRQF_SHARED, name, nvmeq);
1314 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1316 struct nvme_dev *dev = nvmeq->dev;
1318 spin_lock_irq(&nvmeq->q_lock);
1321 nvmeq->cq_phase = 1;
1322 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1323 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1324 dev->online_queues++;
1325 spin_unlock_irq(&nvmeq->q_lock);
1328 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1330 struct nvme_dev *dev = nvmeq->dev;
1333 nvmeq->cq_vector = qid - 1;
1334 result = adapter_alloc_cq(dev, qid, nvmeq);
1338 result = adapter_alloc_sq(dev, qid, nvmeq);
1342 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1346 nvme_init_queue(nvmeq, qid);
1350 adapter_delete_sq(dev, qid);
1352 adapter_delete_cq(dev, qid);
1356 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1358 unsigned long timeout;
1359 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1361 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1363 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1365 if (fatal_signal_pending(current))
1367 if (time_after(jiffies, timeout)) {
1368 dev_err(&dev->pci_dev->dev,
1369 "Device not ready; aborting %s\n", enabled ?
1370 "initialisation" : "reset");
1379 * If the device has been passed off to us in an enabled state, just clear
1380 * the enabled bit. The spec says we should set the 'shutdown notification
1381 * bits', but doing so may cause the device to complete commands to the
1382 * admin queue ... and we don't know what memory that might be pointing at!
1384 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1386 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1387 dev->ctrl_config &= ~NVME_CC_ENABLE;
1388 writel(dev->ctrl_config, &dev->bar->cc);
1390 return nvme_wait_ready(dev, cap, false);
1393 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1395 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1396 dev->ctrl_config |= NVME_CC_ENABLE;
1397 writel(dev->ctrl_config, &dev->bar->cc);
1399 return nvme_wait_ready(dev, cap, true);
1402 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1404 unsigned long timeout;
1406 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1407 dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1409 writel(dev->ctrl_config, &dev->bar->cc);
1411 timeout = SHUTDOWN_TIMEOUT + jiffies;
1412 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1413 NVME_CSTS_SHST_CMPLT) {
1415 if (fatal_signal_pending(current))
1417 if (time_after(jiffies, timeout)) {
1418 dev_err(&dev->pci_dev->dev,
1419 "Device shutdown incomplete; abort shutdown\n");
1427 static struct blk_mq_ops nvme_mq_admin_ops = {
1428 .queue_rq = nvme_admin_queue_rq,
1429 .map_queue = blk_mq_map_queue,
1430 .init_hctx = nvme_admin_init_hctx,
1431 .exit_hctx = nvme_exit_hctx,
1432 .init_request = nvme_admin_init_request,
1433 .timeout = nvme_timeout,
1436 static struct blk_mq_ops nvme_mq_ops = {
1437 .queue_rq = nvme_queue_rq,
1438 .map_queue = blk_mq_map_queue,
1439 .init_hctx = nvme_init_hctx,
1440 .exit_hctx = nvme_exit_hctx,
1441 .init_request = nvme_init_request,
1442 .timeout = nvme_timeout,
1445 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1447 if (dev->admin_q && !blk_queue_dying(dev->admin_q)) {
1448 blk_cleanup_queue(dev->admin_q);
1449 blk_mq_free_tag_set(&dev->admin_tagset);
1453 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1455 if (!dev->admin_q) {
1456 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1457 dev->admin_tagset.nr_hw_queues = 1;
1458 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1459 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1460 dev->admin_tagset.numa_node = dev_to_node(&dev->pci_dev->dev);
1461 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1462 dev->admin_tagset.driver_data = dev;
1464 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1467 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1468 if (IS_ERR(dev->admin_q)) {
1469 blk_mq_free_tag_set(&dev->admin_tagset);
1472 if (!blk_get_queue(dev->admin_q)) {
1473 nvme_dev_remove_admin(dev);
1477 blk_mq_unfreeze_queue(dev->admin_q);
1482 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1486 u64 cap = readq(&dev->bar->cap);
1487 struct nvme_queue *nvmeq;
1488 unsigned page_shift = PAGE_SHIFT;
1489 unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1490 unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1492 if (page_shift < dev_page_min) {
1493 dev_err(&dev->pci_dev->dev,
1494 "Minimum device page size (%u) too large for "
1495 "host (%u)\n", 1 << dev_page_min,
1499 if (page_shift > dev_page_max) {
1500 dev_info(&dev->pci_dev->dev,
1501 "Device maximum page size (%u) smaller than "
1502 "host (%u); enabling work-around\n",
1503 1 << dev_page_max, 1 << page_shift);
1504 page_shift = dev_page_max;
1507 result = nvme_disable_ctrl(dev, cap);
1511 nvmeq = dev->queues[0];
1513 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1518 aqa = nvmeq->q_depth - 1;
1521 dev->page_size = 1 << page_shift;
1523 dev->ctrl_config = NVME_CC_CSS_NVM;
1524 dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1525 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1526 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1528 writel(aqa, &dev->bar->aqa);
1529 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1530 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1532 result = nvme_enable_ctrl(dev, cap);
1536 nvmeq->cq_vector = 0;
1537 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1544 nvme_free_queues(dev, 0);
1548 struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1549 unsigned long addr, unsigned length)
1551 int i, err, count, nents, offset;
1552 struct scatterlist *sg;
1553 struct page **pages;
1554 struct nvme_iod *iod;
1557 return ERR_PTR(-EINVAL);
1558 if (!length || length > INT_MAX - PAGE_SIZE)
1559 return ERR_PTR(-EINVAL);
1561 offset = offset_in_page(addr);
1562 count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1563 pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1565 return ERR_PTR(-ENOMEM);
1567 err = get_user_pages_fast(addr, count, 1, pages);
1575 iod = __nvme_alloc_iod(count, length, dev, 0, GFP_KERNEL);
1580 sg_init_table(sg, count);
1581 for (i = 0; i < count; i++) {
1582 sg_set_page(&sg[i], pages[i],
1583 min_t(unsigned, length, PAGE_SIZE - offset),
1585 length -= (PAGE_SIZE - offset);
1588 sg_mark_end(&sg[i - 1]);
1591 nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1592 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1602 for (i = 0; i < count; i++)
1605 return ERR_PTR(err);
1608 void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1609 struct nvme_iod *iod)
1613 dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1614 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1616 for (i = 0; i < iod->nents; i++)
1617 put_page(sg_page(&iod->sg[i]));
1620 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1622 struct nvme_dev *dev = ns->dev;
1623 struct nvme_user_io io;
1624 struct nvme_command c;
1625 unsigned length, meta_len;
1627 struct nvme_iod *iod, *meta_iod = NULL;
1628 dma_addr_t meta_dma_addr;
1629 void *meta, *uninitialized_var(meta_mem);
1631 if (copy_from_user(&io, uio, sizeof(io)))
1633 length = (io.nblocks + 1) << ns->lba_shift;
1634 meta_len = (io.nblocks + 1) * ns->ms;
1636 if (meta_len && ((io.metadata & 3) || !io.metadata))
1639 switch (io.opcode) {
1640 case nvme_cmd_write:
1642 case nvme_cmd_compare:
1643 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1650 return PTR_ERR(iod);
1652 memset(&c, 0, sizeof(c));
1653 c.rw.opcode = io.opcode;
1654 c.rw.flags = io.flags;
1655 c.rw.nsid = cpu_to_le32(ns->ns_id);
1656 c.rw.slba = cpu_to_le64(io.slba);
1657 c.rw.length = cpu_to_le16(io.nblocks);
1658 c.rw.control = cpu_to_le16(io.control);
1659 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1660 c.rw.reftag = cpu_to_le32(io.reftag);
1661 c.rw.apptag = cpu_to_le16(io.apptag);
1662 c.rw.appmask = cpu_to_le16(io.appmask);
1665 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1667 if (IS_ERR(meta_iod)) {
1668 status = PTR_ERR(meta_iod);
1673 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1674 &meta_dma_addr, GFP_KERNEL);
1680 if (io.opcode & 1) {
1681 int meta_offset = 0;
1683 for (i = 0; i < meta_iod->nents; i++) {
1684 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1685 meta_iod->sg[i].offset;
1686 memcpy(meta_mem + meta_offset, meta,
1687 meta_iod->sg[i].length);
1688 kunmap_atomic(meta);
1689 meta_offset += meta_iod->sg[i].length;
1693 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1696 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1697 c.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1698 c.rw.prp2 = cpu_to_le64(iod->first_dma);
1700 if (length != (io.nblocks + 1) << ns->lba_shift)
1703 status = nvme_submit_io_cmd(dev, ns, &c, NULL);
1706 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1707 int meta_offset = 0;
1709 for (i = 0; i < meta_iod->nents; i++) {
1710 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1711 meta_iod->sg[i].offset;
1712 memcpy(meta, meta_mem + meta_offset,
1713 meta_iod->sg[i].length);
1714 kunmap_atomic(meta);
1715 meta_offset += meta_iod->sg[i].length;
1719 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1724 nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1725 nvme_free_iod(dev, iod);
1728 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1729 nvme_free_iod(dev, meta_iod);
1735 static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1736 struct nvme_passthru_cmd __user *ucmd)
1738 struct nvme_passthru_cmd cmd;
1739 struct nvme_command c;
1741 struct nvme_iod *uninitialized_var(iod);
1744 if (!capable(CAP_SYS_ADMIN))
1746 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1749 memset(&c, 0, sizeof(c));
1750 c.common.opcode = cmd.opcode;
1751 c.common.flags = cmd.flags;
1752 c.common.nsid = cpu_to_le32(cmd.nsid);
1753 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1754 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1755 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1756 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1757 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1758 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1759 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1760 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1762 length = cmd.data_len;
1764 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1767 return PTR_ERR(iod);
1768 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1769 c.common.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1770 c.common.prp2 = cpu_to_le64(iod->first_dma);
1773 timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1776 if (length != cmd.data_len)
1779 struct request *req;
1781 req = blk_mq_alloc_request(ns->queue, WRITE,
1782 (GFP_KERNEL|__GFP_WAIT), false);
1784 status = PTR_ERR(req);
1786 status = nvme_submit_sync_cmd(req, &c, &cmd.result,
1788 blk_mq_free_request(req);
1791 status = __nvme_submit_admin_cmd(dev, &c, &cmd.result, timeout);
1794 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1795 nvme_free_iod(dev, iod);
1798 if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1799 sizeof(cmd.result)))
1805 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1808 struct nvme_ns *ns = bdev->bd_disk->private_data;
1812 force_successful_syscall_return();
1814 case NVME_IOCTL_ADMIN_CMD:
1815 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1816 case NVME_IOCTL_IO_CMD:
1817 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1818 case NVME_IOCTL_SUBMIT_IO:
1819 return nvme_submit_io(ns, (void __user *)arg);
1820 case SG_GET_VERSION_NUM:
1821 return nvme_sg_get_version_num((void __user *)arg);
1823 return nvme_sg_io(ns, (void __user *)arg);
1829 #ifdef CONFIG_COMPAT
1830 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1831 unsigned int cmd, unsigned long arg)
1835 return -ENOIOCTLCMD;
1837 return nvme_ioctl(bdev, mode, cmd, arg);
1840 #define nvme_compat_ioctl NULL
1843 static int nvme_open(struct block_device *bdev, fmode_t mode)
1848 spin_lock(&dev_list_lock);
1849 ns = bdev->bd_disk->private_data;
1852 else if (!kref_get_unless_zero(&ns->dev->kref))
1854 spin_unlock(&dev_list_lock);
1859 static void nvme_free_dev(struct kref *kref);
1861 static void nvme_release(struct gendisk *disk, fmode_t mode)
1863 struct nvme_ns *ns = disk->private_data;
1864 struct nvme_dev *dev = ns->dev;
1866 kref_put(&dev->kref, nvme_free_dev);
1869 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1871 /* some standard values */
1872 geo->heads = 1 << 6;
1873 geo->sectors = 1 << 5;
1874 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1878 static int nvme_revalidate_disk(struct gendisk *disk)
1880 struct nvme_ns *ns = disk->private_data;
1881 struct nvme_dev *dev = ns->dev;
1882 struct nvme_id_ns *id;
1883 dma_addr_t dma_addr;
1886 id = dma_alloc_coherent(&dev->pci_dev->dev, 4096, &dma_addr,
1889 dev_warn(&dev->pci_dev->dev, "%s: Memory alocation failure\n",
1894 if (nvme_identify(dev, ns->ns_id, 0, dma_addr))
1897 lbaf = id->flbas & 0xf;
1898 ns->lba_shift = id->lbaf[lbaf].ds;
1900 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1901 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1903 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1907 static const struct block_device_operations nvme_fops = {
1908 .owner = THIS_MODULE,
1909 .ioctl = nvme_ioctl,
1910 .compat_ioctl = nvme_compat_ioctl,
1912 .release = nvme_release,
1913 .getgeo = nvme_getgeo,
1914 .revalidate_disk= nvme_revalidate_disk,
1917 static int nvme_kthread(void *data)
1919 struct nvme_dev *dev, *next;
1921 while (!kthread_should_stop()) {
1922 set_current_state(TASK_INTERRUPTIBLE);
1923 spin_lock(&dev_list_lock);
1924 list_for_each_entry_safe(dev, next, &dev_list, node) {
1926 if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1928 if (work_busy(&dev->reset_work))
1930 list_del_init(&dev->node);
1931 dev_warn(&dev->pci_dev->dev,
1932 "Failed status: %x, reset controller\n",
1933 readl(&dev->bar->csts));
1934 dev->reset_workfn = nvme_reset_failed_dev;
1935 queue_work(nvme_workq, &dev->reset_work);
1938 for (i = 0; i < dev->queue_count; i++) {
1939 struct nvme_queue *nvmeq = dev->queues[i];
1942 spin_lock_irq(&nvmeq->q_lock);
1943 nvme_process_cq(nvmeq);
1945 while ((i == 0) && (dev->event_limit > 0)) {
1946 if (nvme_submit_async_admin_req(dev))
1950 spin_unlock_irq(&nvmeq->q_lock);
1953 spin_unlock(&dev_list_lock);
1954 schedule_timeout(round_jiffies_relative(HZ));
1959 static void nvme_config_discard(struct nvme_ns *ns)
1961 u32 logical_block_size = queue_logical_block_size(ns->queue);
1962 ns->queue->limits.discard_zeroes_data = 0;
1963 ns->queue->limits.discard_alignment = logical_block_size;
1964 ns->queue->limits.discard_granularity = logical_block_size;
1965 ns->queue->limits.max_discard_sectors = 0xffffffff;
1966 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1969 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1970 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1973 struct gendisk *disk;
1974 int node = dev_to_node(&dev->pci_dev->dev);
1977 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1980 ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
1983 ns->queue = blk_mq_init_queue(&dev->tagset);
1984 if (IS_ERR(ns->queue))
1986 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1987 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1988 queue_flag_set_unlocked(QUEUE_FLAG_SG_GAPS, ns->queue);
1990 ns->queue->queuedata = ns;
1992 disk = alloc_disk_node(0, node);
1994 goto out_free_queue;
1998 lbaf = id->flbas & 0xf;
1999 ns->lba_shift = id->lbaf[lbaf].ds;
2000 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
2001 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2002 if (dev->max_hw_sectors)
2003 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2004 if (dev->stripe_size)
2005 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
2006 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2007 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2009 disk->major = nvme_major;
2010 disk->first_minor = 0;
2011 disk->fops = &nvme_fops;
2012 disk->private_data = ns;
2013 disk->queue = ns->queue;
2014 disk->driverfs_dev = &dev->pci_dev->dev;
2015 disk->flags = GENHD_FL_EXT_DEVT;
2016 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2017 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
2019 if (dev->oncs & NVME_CTRL_ONCS_DSM)
2020 nvme_config_discard(ns);
2025 blk_cleanup_queue(ns->queue);
2031 static void nvme_create_io_queues(struct nvme_dev *dev)
2035 for (i = dev->queue_count; i <= dev->max_qid; i++)
2036 if (!nvme_alloc_queue(dev, i, dev->q_depth))
2039 for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
2040 if (nvme_create_queue(dev->queues[i], i))
2044 static int set_queue_count(struct nvme_dev *dev, int count)
2048 u32 q_count = (count - 1) | ((count - 1) << 16);
2050 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2055 dev_err(&dev->pci_dev->dev, "Could not set queue count (%d)\n",
2059 return min(result & 0xffff, result >> 16) + 1;
2062 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2064 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2067 static int nvme_setup_io_queues(struct nvme_dev *dev)
2069 struct nvme_queue *adminq = dev->queues[0];
2070 struct pci_dev *pdev = dev->pci_dev;
2071 int result, i, vecs, nr_io_queues, size;
2073 nr_io_queues = num_possible_cpus();
2074 result = set_queue_count(dev, nr_io_queues);
2077 if (result < nr_io_queues)
2078 nr_io_queues = result;
2080 size = db_bar_size(dev, nr_io_queues);
2084 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2087 if (!--nr_io_queues)
2089 size = db_bar_size(dev, nr_io_queues);
2091 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2092 adminq->q_db = dev->dbs;
2095 /* Deregister the admin queue's interrupt */
2096 free_irq(dev->entry[0].vector, adminq);
2099 * If we enable msix early due to not intx, disable it again before
2100 * setting up the full range we need.
2103 pci_disable_msix(pdev);
2105 for (i = 0; i < nr_io_queues; i++)
2106 dev->entry[i].entry = i;
2107 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2109 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2113 for (i = 0; i < vecs; i++)
2114 dev->entry[i].vector = i + pdev->irq;
2119 * Should investigate if there's a performance win from allocating
2120 * more queues than interrupt vectors; it might allow the submission
2121 * path to scale better, even if the receive path is limited by the
2122 * number of interrupts.
2124 nr_io_queues = vecs;
2125 dev->max_qid = nr_io_queues;
2127 result = queue_request_irq(dev, adminq, adminq->irqname);
2131 /* Free previously allocated queues that are no longer usable */
2132 nvme_free_queues(dev, nr_io_queues + 1);
2133 nvme_create_io_queues(dev);
2138 nvme_free_queues(dev, 1);
2143 * Return: error value if an error occurred setting up the queues or calling
2144 * Identify Device. 0 if these succeeded, even if adding some of the
2145 * namespaces failed. At the moment, these failures are silent. TBD which
2146 * failures should be reported.
2148 static int nvme_dev_add(struct nvme_dev *dev)
2150 struct pci_dev *pdev = dev->pci_dev;
2154 struct nvme_id_ctrl *ctrl;
2155 struct nvme_id_ns *id_ns;
2157 dma_addr_t dma_addr;
2158 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2160 mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
2164 res = nvme_identify(dev, 0, 1, dma_addr);
2166 dev_err(&pdev->dev, "Identify Controller failed (%d)\n", res);
2172 nn = le32_to_cpup(&ctrl->nn);
2173 dev->oncs = le16_to_cpup(&ctrl->oncs);
2174 dev->abort_limit = ctrl->acl + 1;
2175 dev->vwc = ctrl->vwc;
2176 dev->event_limit = min(ctrl->aerl + 1, 8);
2177 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2178 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2179 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2181 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2182 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2183 (pdev->device == 0x0953) && ctrl->vs[3]) {
2184 unsigned int max_hw_sectors;
2186 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2187 max_hw_sectors = dev->stripe_size >> (shift - 9);
2188 if (dev->max_hw_sectors) {
2189 dev->max_hw_sectors = min(max_hw_sectors,
2190 dev->max_hw_sectors);
2192 dev->max_hw_sectors = max_hw_sectors;
2195 dev->tagset.ops = &nvme_mq_ops;
2196 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2197 dev->tagset.timeout = NVME_IO_TIMEOUT;
2198 dev->tagset.numa_node = dev_to_node(&dev->pci_dev->dev);
2199 dev->tagset.queue_depth =
2200 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2201 dev->tagset.cmd_size = nvme_cmd_size(dev);
2202 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2203 dev->tagset.driver_data = dev;
2205 if (blk_mq_alloc_tag_set(&dev->tagset))
2209 for (i = 1; i <= nn; i++) {
2210 res = nvme_identify(dev, i, 0, dma_addr);
2214 if (id_ns->ncap == 0)
2217 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
2218 dma_addr + 4096, NULL);
2220 memset(mem + 4096, 0, 4096);
2222 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
2224 list_add_tail(&ns->list, &dev->namespaces);
2226 list_for_each_entry(ns, &dev->namespaces, list)
2231 dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
2235 static int nvme_dev_map(struct nvme_dev *dev)
2238 int bars, result = -ENOMEM;
2239 struct pci_dev *pdev = dev->pci_dev;
2241 if (pci_enable_device_mem(pdev))
2244 dev->entry[0].vector = pdev->irq;
2245 pci_set_master(pdev);
2246 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2250 if (pci_request_selected_regions(pdev, bars, "nvme"))
2253 if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2254 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2257 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2261 if (readl(&dev->bar->csts) == -1) {
2267 * Some devices don't advertse INTx interrupts, pre-enable a single
2268 * MSIX vec for setup. We'll adjust this later.
2271 result = pci_enable_msix(pdev, dev->entry, 1);
2276 cap = readq(&dev->bar->cap);
2277 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2278 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2279 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2287 pci_release_regions(pdev);
2289 pci_disable_device(pdev);
2293 static void nvme_dev_unmap(struct nvme_dev *dev)
2295 if (dev->pci_dev->msi_enabled)
2296 pci_disable_msi(dev->pci_dev);
2297 else if (dev->pci_dev->msix_enabled)
2298 pci_disable_msix(dev->pci_dev);
2303 pci_release_regions(dev->pci_dev);
2306 if (pci_is_enabled(dev->pci_dev))
2307 pci_disable_device(dev->pci_dev);
2310 struct nvme_delq_ctx {
2311 struct task_struct *waiter;
2312 struct kthread_worker *worker;
2316 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2318 dq->waiter = current;
2322 set_current_state(TASK_KILLABLE);
2323 if (!atomic_read(&dq->refcount))
2325 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2326 fatal_signal_pending(current)) {
2328 * Disable the controller first since we can't trust it
2329 * at this point, but leave the admin queue enabled
2330 * until all queue deletion requests are flushed.
2331 * FIXME: This may take a while if there are more h/w
2332 * queues than admin tags.
2334 set_current_state(TASK_RUNNING);
2335 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2336 nvme_clear_queue(dev->queues[0]);
2337 flush_kthread_worker(dq->worker);
2338 nvme_disable_queue(dev, 0);
2342 set_current_state(TASK_RUNNING);
2345 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2347 atomic_dec(&dq->refcount);
2349 wake_up_process(dq->waiter);
2352 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2354 atomic_inc(&dq->refcount);
2358 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2360 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2362 nvme_clear_queue(nvmeq);
2366 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2367 kthread_work_func_t fn)
2369 struct nvme_command c;
2371 memset(&c, 0, sizeof(c));
2372 c.delete_queue.opcode = opcode;
2373 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2375 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2376 return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2380 static void nvme_del_cq_work_handler(struct kthread_work *work)
2382 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2384 nvme_del_queue_end(nvmeq);
2387 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2389 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2390 nvme_del_cq_work_handler);
2393 static void nvme_del_sq_work_handler(struct kthread_work *work)
2395 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2397 int status = nvmeq->cmdinfo.status;
2400 status = nvme_delete_cq(nvmeq);
2402 nvme_del_queue_end(nvmeq);
2405 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2407 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2408 nvme_del_sq_work_handler);
2411 static void nvme_del_queue_start(struct kthread_work *work)
2413 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2415 if (nvme_delete_sq(nvmeq))
2416 nvme_del_queue_end(nvmeq);
2419 static void nvme_disable_io_queues(struct nvme_dev *dev)
2422 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2423 struct nvme_delq_ctx dq;
2424 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2425 &worker, "nvme%d", dev->instance);
2427 if (IS_ERR(kworker_task)) {
2428 dev_err(&dev->pci_dev->dev,
2429 "Failed to create queue del task\n");
2430 for (i = dev->queue_count - 1; i > 0; i--)
2431 nvme_disable_queue(dev, i);
2436 atomic_set(&dq.refcount, 0);
2437 dq.worker = &worker;
2438 for (i = dev->queue_count - 1; i > 0; i--) {
2439 struct nvme_queue *nvmeq = dev->queues[i];
2441 if (nvme_suspend_queue(nvmeq))
2443 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2444 nvmeq->cmdinfo.worker = dq.worker;
2445 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2446 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2448 nvme_wait_dq(&dq, dev);
2449 kthread_stop(kworker_task);
2453 * Remove the node from the device list and check
2454 * for whether or not we need to stop the nvme_thread.
2456 static void nvme_dev_list_remove(struct nvme_dev *dev)
2458 struct task_struct *tmp = NULL;
2460 spin_lock(&dev_list_lock);
2461 list_del_init(&dev->node);
2462 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2466 spin_unlock(&dev_list_lock);
2472 static void nvme_freeze_queues(struct nvme_dev *dev)
2476 list_for_each_entry(ns, &dev->namespaces, list) {
2477 blk_mq_freeze_queue_start(ns->queue);
2479 spin_lock(ns->queue->queue_lock);
2480 queue_flag_set(QUEUE_FLAG_STOPPED, ns->queue);
2481 spin_unlock(ns->queue->queue_lock);
2483 blk_mq_cancel_requeue_work(ns->queue);
2484 blk_mq_stop_hw_queues(ns->queue);
2488 static void nvme_unfreeze_queues(struct nvme_dev *dev)
2492 list_for_each_entry(ns, &dev->namespaces, list) {
2493 queue_flag_clear_unlocked(QUEUE_FLAG_STOPPED, ns->queue);
2494 blk_mq_unfreeze_queue(ns->queue);
2495 blk_mq_start_stopped_hw_queues(ns->queue, true);
2496 blk_mq_kick_requeue_list(ns->queue);
2500 static void nvme_dev_shutdown(struct nvme_dev *dev)
2505 dev->initialized = 0;
2506 nvme_dev_list_remove(dev);
2509 nvme_freeze_queues(dev);
2510 csts = readl(&dev->bar->csts);
2512 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2513 for (i = dev->queue_count - 1; i >= 0; i--) {
2514 struct nvme_queue *nvmeq = dev->queues[i];
2515 nvme_suspend_queue(nvmeq);
2516 nvme_clear_queue(nvmeq);
2519 nvme_disable_io_queues(dev);
2520 nvme_shutdown_ctrl(dev);
2521 nvme_disable_queue(dev, 0);
2523 nvme_dev_unmap(dev);
2526 static void nvme_dev_remove(struct nvme_dev *dev)
2530 list_for_each_entry(ns, &dev->namespaces, list) {
2531 if (ns->disk->flags & GENHD_FL_UP)
2532 del_gendisk(ns->disk);
2533 if (!blk_queue_dying(ns->queue)) {
2534 blk_mq_abort_requeue_list(ns->queue);
2535 blk_cleanup_queue(ns->queue);
2540 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2542 struct device *dmadev = &dev->pci_dev->dev;
2543 dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2544 PAGE_SIZE, PAGE_SIZE, 0);
2545 if (!dev->prp_page_pool)
2548 /* Optimisation for I/Os between 4k and 128k */
2549 dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2551 if (!dev->prp_small_pool) {
2552 dma_pool_destroy(dev->prp_page_pool);
2558 static void nvme_release_prp_pools(struct nvme_dev *dev)
2560 dma_pool_destroy(dev->prp_page_pool);
2561 dma_pool_destroy(dev->prp_small_pool);
2564 static DEFINE_IDA(nvme_instance_ida);
2566 static int nvme_set_instance(struct nvme_dev *dev)
2568 int instance, error;
2571 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2574 spin_lock(&dev_list_lock);
2575 error = ida_get_new(&nvme_instance_ida, &instance);
2576 spin_unlock(&dev_list_lock);
2577 } while (error == -EAGAIN);
2582 dev->instance = instance;
2586 static void nvme_release_instance(struct nvme_dev *dev)
2588 spin_lock(&dev_list_lock);
2589 ida_remove(&nvme_instance_ida, dev->instance);
2590 spin_unlock(&dev_list_lock);
2593 static void nvme_free_namespaces(struct nvme_dev *dev)
2595 struct nvme_ns *ns, *next;
2597 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2598 list_del(&ns->list);
2600 spin_lock(&dev_list_lock);
2601 ns->disk->private_data = NULL;
2602 spin_unlock(&dev_list_lock);
2609 static void nvme_free_dev(struct kref *kref)
2611 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2613 pci_dev_put(dev->pci_dev);
2614 nvme_free_namespaces(dev);
2615 nvme_release_instance(dev);
2616 blk_mq_free_tag_set(&dev->tagset);
2617 blk_put_queue(dev->admin_q);
2623 static int nvme_dev_open(struct inode *inode, struct file *f)
2625 struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2627 kref_get(&dev->kref);
2628 f->private_data = dev;
2632 static int nvme_dev_release(struct inode *inode, struct file *f)
2634 struct nvme_dev *dev = f->private_data;
2635 kref_put(&dev->kref, nvme_free_dev);
2639 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2641 struct nvme_dev *dev = f->private_data;
2645 case NVME_IOCTL_ADMIN_CMD:
2646 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2647 case NVME_IOCTL_IO_CMD:
2648 if (list_empty(&dev->namespaces))
2650 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2651 return nvme_user_cmd(dev, ns, (void __user *)arg);
2657 static const struct file_operations nvme_dev_fops = {
2658 .owner = THIS_MODULE,
2659 .open = nvme_dev_open,
2660 .release = nvme_dev_release,
2661 .unlocked_ioctl = nvme_dev_ioctl,
2662 .compat_ioctl = nvme_dev_ioctl,
2665 static void nvme_set_irq_hints(struct nvme_dev *dev)
2667 struct nvme_queue *nvmeq;
2670 for (i = 0; i < dev->online_queues; i++) {
2671 nvmeq = dev->queues[i];
2676 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2677 nvmeq->hctx->cpumask);
2681 static int nvme_dev_start(struct nvme_dev *dev)
2684 bool start_thread = false;
2686 result = nvme_dev_map(dev);
2690 result = nvme_configure_admin_queue(dev);
2694 spin_lock(&dev_list_lock);
2695 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2696 start_thread = true;
2699 list_add(&dev->node, &dev_list);
2700 spin_unlock(&dev_list_lock);
2703 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2704 wake_up_all(&nvme_kthread_wait);
2706 wait_event_killable(nvme_kthread_wait, nvme_thread);
2708 if (IS_ERR_OR_NULL(nvme_thread)) {
2709 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2713 nvme_init_queue(dev->queues[0], 0);
2714 result = nvme_alloc_admin_tags(dev);
2718 result = nvme_setup_io_queues(dev);
2722 nvme_set_irq_hints(dev);
2727 nvme_dev_remove_admin(dev);
2729 nvme_disable_queue(dev, 0);
2730 nvme_dev_list_remove(dev);
2732 nvme_dev_unmap(dev);
2736 static int nvme_remove_dead_ctrl(void *arg)
2738 struct nvme_dev *dev = (struct nvme_dev *)arg;
2739 struct pci_dev *pdev = dev->pci_dev;
2741 if (pci_get_drvdata(pdev))
2742 pci_stop_and_remove_bus_device_locked(pdev);
2743 kref_put(&dev->kref, nvme_free_dev);
2747 static void nvme_remove_disks(struct work_struct *ws)
2749 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2751 nvme_free_queues(dev, 1);
2752 nvme_dev_remove(dev);
2755 static int nvme_dev_resume(struct nvme_dev *dev)
2759 ret = nvme_dev_start(dev);
2762 if (dev->online_queues < 2) {
2763 spin_lock(&dev_list_lock);
2764 dev->reset_workfn = nvme_remove_disks;
2765 queue_work(nvme_workq, &dev->reset_work);
2766 spin_unlock(&dev_list_lock);
2768 nvme_unfreeze_queues(dev);
2769 nvme_set_irq_hints(dev);
2771 dev->initialized = 1;
2775 static void nvme_dev_reset(struct nvme_dev *dev)
2777 nvme_dev_shutdown(dev);
2778 if (nvme_dev_resume(dev)) {
2779 dev_warn(&dev->pci_dev->dev, "Device failed to resume\n");
2780 kref_get(&dev->kref);
2781 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2783 dev_err(&dev->pci_dev->dev,
2784 "Failed to start controller remove task\n");
2785 kref_put(&dev->kref, nvme_free_dev);
2790 static void nvme_reset_failed_dev(struct work_struct *ws)
2792 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2793 nvme_dev_reset(dev);
2796 static void nvme_reset_workfn(struct work_struct *work)
2798 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2799 dev->reset_workfn(work);
2802 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2804 int node, result = -ENOMEM;
2805 struct nvme_dev *dev;
2807 node = dev_to_node(&pdev->dev);
2808 if (node == NUMA_NO_NODE)
2809 set_dev_node(&pdev->dev, 0);
2811 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2814 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
2818 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2823 INIT_LIST_HEAD(&dev->namespaces);
2824 dev->reset_workfn = nvme_reset_failed_dev;
2825 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
2826 dev->pci_dev = pci_dev_get(pdev);
2827 pci_set_drvdata(pdev, dev);
2828 result = nvme_set_instance(dev);
2832 result = nvme_setup_prp_pools(dev);
2836 kref_init(&dev->kref);
2837 result = nvme_dev_start(dev);
2841 if (dev->online_queues > 1)
2842 result = nvme_dev_add(dev);
2846 scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2847 dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2848 dev->miscdev.parent = &pdev->dev;
2849 dev->miscdev.name = dev->name;
2850 dev->miscdev.fops = &nvme_dev_fops;
2851 result = misc_register(&dev->miscdev);
2855 nvme_set_irq_hints(dev);
2857 dev->initialized = 1;
2861 nvme_dev_remove(dev);
2862 nvme_dev_remove_admin(dev);
2863 nvme_free_namespaces(dev);
2865 nvme_dev_shutdown(dev);
2867 nvme_free_queues(dev, 0);
2868 nvme_release_prp_pools(dev);
2870 nvme_release_instance(dev);
2872 pci_dev_put(dev->pci_dev);
2880 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2882 struct nvme_dev *dev = pci_get_drvdata(pdev);
2885 nvme_dev_shutdown(dev);
2887 nvme_dev_resume(dev);
2890 static void nvme_shutdown(struct pci_dev *pdev)
2892 struct nvme_dev *dev = pci_get_drvdata(pdev);
2893 nvme_dev_shutdown(dev);
2896 static void nvme_remove(struct pci_dev *pdev)
2898 struct nvme_dev *dev = pci_get_drvdata(pdev);
2900 spin_lock(&dev_list_lock);
2901 list_del_init(&dev->node);
2902 spin_unlock(&dev_list_lock);
2904 pci_set_drvdata(pdev, NULL);
2905 flush_work(&dev->reset_work);
2906 misc_deregister(&dev->miscdev);
2907 nvme_dev_shutdown(dev);
2908 nvme_dev_remove(dev);
2909 nvme_dev_remove_admin(dev);
2910 nvme_free_queues(dev, 0);
2911 nvme_release_prp_pools(dev);
2912 kref_put(&dev->kref, nvme_free_dev);
2915 /* These functions are yet to be implemented */
2916 #define nvme_error_detected NULL
2917 #define nvme_dump_registers NULL
2918 #define nvme_link_reset NULL
2919 #define nvme_slot_reset NULL
2920 #define nvme_error_resume NULL
2922 #ifdef CONFIG_PM_SLEEP
2923 static int nvme_suspend(struct device *dev)
2925 struct pci_dev *pdev = to_pci_dev(dev);
2926 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2928 nvme_dev_shutdown(ndev);
2932 static int nvme_resume(struct device *dev)
2934 struct pci_dev *pdev = to_pci_dev(dev);
2935 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2937 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2938 ndev->reset_workfn = nvme_reset_failed_dev;
2939 queue_work(nvme_workq, &ndev->reset_work);
2945 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2947 static const struct pci_error_handlers nvme_err_handler = {
2948 .error_detected = nvme_error_detected,
2949 .mmio_enabled = nvme_dump_registers,
2950 .link_reset = nvme_link_reset,
2951 .slot_reset = nvme_slot_reset,
2952 .resume = nvme_error_resume,
2953 .reset_notify = nvme_reset_notify,
2956 /* Move to pci_ids.h later */
2957 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
2959 static const struct pci_device_id nvme_id_table[] = {
2960 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2963 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2965 static struct pci_driver nvme_driver = {
2967 .id_table = nvme_id_table,
2968 .probe = nvme_probe,
2969 .remove = nvme_remove,
2970 .shutdown = nvme_shutdown,
2972 .pm = &nvme_dev_pm_ops,
2974 .err_handler = &nvme_err_handler,
2977 static int __init nvme_init(void)
2981 init_waitqueue_head(&nvme_kthread_wait);
2983 nvme_workq = create_singlethread_workqueue("nvme");
2987 result = register_blkdev(nvme_major, "nvme");
2990 else if (result > 0)
2991 nvme_major = result;
2993 result = pci_register_driver(&nvme_driver);
2995 goto unregister_blkdev;
2999 unregister_blkdev(nvme_major, "nvme");
3001 destroy_workqueue(nvme_workq);
3005 static void __exit nvme_exit(void)
3007 pci_unregister_driver(&nvme_driver);
3008 unregister_hotcpu_notifier(&nvme_nb);
3009 unregister_blkdev(nvme_major, "nvme");
3010 destroy_workqueue(nvme_workq);
3011 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3015 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3016 MODULE_LICENSE("GPL");
3017 MODULE_VERSION("1.0");
3018 module_init(nvme_init);
3019 module_exit(nvme_exit);