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/bio.h>
17 #include <linux/bitops.h>
18 #include <linux/blkdev.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/percpu.h>
37 #include <linux/poison.h>
38 #include <linux/ptrace.h>
39 #include <linux/sched.h>
40 #include <linux/slab.h>
41 #include <linux/types.h>
43 #include <asm-generic/io-64-nonatomic-lo-hi.h>
45 #include <trace/events/block.h>
47 #define NVME_Q_DEPTH 1024
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)
52 #define IOD_TIMEOUT (retry_time * HZ)
54 static unsigned char admin_timeout = 60;
55 module_param(admin_timeout, byte, 0644);
56 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
58 unsigned char nvme_io_timeout = 30;
59 module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
60 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
62 static unsigned char retry_time = 30;
63 module_param(retry_time, byte, 0644);
64 MODULE_PARM_DESC(retry_time, "time in seconds to retry failed I/O");
66 static unsigned char shutdown_timeout = 5;
67 module_param(shutdown_timeout, byte, 0644);
68 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
70 static int nvme_major;
71 module_param(nvme_major, int, 0);
73 static int use_threaded_interrupts;
74 module_param(use_threaded_interrupts, int, 0);
76 static DEFINE_SPINLOCK(dev_list_lock);
77 static LIST_HEAD(dev_list);
78 static struct task_struct *nvme_thread;
79 static struct workqueue_struct *nvme_workq;
80 static wait_queue_head_t nvme_kthread_wait;
81 static struct notifier_block nvme_nb;
83 static void nvme_reset_failed_dev(struct work_struct *ws);
85 struct async_cmd_info {
86 struct kthread_work work;
87 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;
107 wait_queue_head_t sq_full;
108 wait_queue_t sq_cong_wait;
109 struct bio_list sq_cong;
110 struct list_head iod_bio;
121 cpumask_var_t cpu_mask;
122 struct async_cmd_info cmdinfo;
123 unsigned long cmdid_data[];
127 * Check we didin't inadvertently grow the command struct
129 static inline void _nvme_check_size(void)
131 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
132 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
133 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
134 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
135 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
136 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
137 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
138 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
139 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
140 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
141 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
142 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
145 typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
146 struct nvme_completion *);
148 struct nvme_cmd_info {
149 nvme_completion_fn fn;
151 unsigned long timeout;
155 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
157 return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
160 static unsigned nvme_queue_extra(int depth)
162 return DIV_ROUND_UP(depth, 8) + (depth * sizeof(struct nvme_cmd_info));
166 * alloc_cmdid() - Allocate a Command ID
167 * @nvmeq: The queue that will be used for this command
168 * @ctx: A pointer that will be passed to the handler
169 * @handler: The function to call on completion
171 * Allocate a Command ID for a queue. The data passed in will
172 * be passed to the completion handler. This is implemented by using
173 * the bottom two bits of the ctx pointer to store the handler ID.
174 * Passing in a pointer that's not 4-byte aligned will cause a BUG.
175 * We can change this if it becomes a problem.
177 * May be called with local interrupts disabled and the q_lock held,
178 * or with interrupts enabled and no locks held.
180 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx,
181 nvme_completion_fn handler, unsigned timeout)
183 int depth = nvmeq->q_depth - 1;
184 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
188 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
191 } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
193 info[cmdid].fn = handler;
194 info[cmdid].ctx = ctx;
195 info[cmdid].timeout = jiffies + timeout;
196 info[cmdid].aborted = 0;
200 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
201 nvme_completion_fn handler, unsigned timeout)
204 wait_event_killable(nvmeq->sq_full,
205 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
206 return (cmdid < 0) ? -EINTR : cmdid;
209 /* Special values must be less than 0x1000 */
210 #define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
211 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
212 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
213 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
214 #define CMD_CTX_ABORT (0x318 + CMD_CTX_BASE)
215 #define CMD_CTX_ASYNC (0x31C + CMD_CTX_BASE)
217 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
218 struct nvme_completion *cqe)
220 if (ctx == CMD_CTX_CANCELLED)
222 if (ctx == CMD_CTX_ABORT) {
223 ++nvmeq->dev->abort_limit;
226 if (ctx == CMD_CTX_COMPLETED) {
227 dev_warn(nvmeq->q_dmadev,
228 "completed id %d twice on queue %d\n",
229 cqe->command_id, le16_to_cpup(&cqe->sq_id));
232 if (ctx == CMD_CTX_INVALID) {
233 dev_warn(nvmeq->q_dmadev,
234 "invalid id %d completed on queue %d\n",
235 cqe->command_id, le16_to_cpup(&cqe->sq_id));
238 if (ctx == CMD_CTX_ASYNC) {
239 u32 result = le32_to_cpup(&cqe->result);
240 u16 status = le16_to_cpup(&cqe->status) >> 1;
242 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
243 ++nvmeq->dev->event_limit;
244 if (status == NVME_SC_SUCCESS)
245 dev_warn(nvmeq->q_dmadev,
246 "async event result %08x\n", result);
250 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
253 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
254 struct nvme_completion *cqe)
256 struct async_cmd_info *cmdinfo = ctx;
257 cmdinfo->result = le32_to_cpup(&cqe->result);
258 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
259 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
263 * Called with local interrupts disabled and the q_lock held. May not sleep.
265 static void *free_cmdid(struct nvme_queue *nvmeq, int cmdid,
266 nvme_completion_fn *fn)
269 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
271 if (cmdid >= nvmeq->q_depth || !info[cmdid].fn) {
273 *fn = special_completion;
274 return CMD_CTX_INVALID;
277 *fn = info[cmdid].fn;
278 ctx = info[cmdid].ctx;
279 info[cmdid].fn = special_completion;
280 info[cmdid].ctx = CMD_CTX_COMPLETED;
281 clear_bit(cmdid, nvmeq->cmdid_data);
282 wake_up(&nvmeq->sq_full);
286 static void *cancel_cmdid(struct nvme_queue *nvmeq, int cmdid,
287 nvme_completion_fn *fn)
290 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
292 *fn = info[cmdid].fn;
293 ctx = info[cmdid].ctx;
294 info[cmdid].fn = special_completion;
295 info[cmdid].ctx = CMD_CTX_CANCELLED;
299 static struct nvme_queue *raw_nvmeq(struct nvme_dev *dev, int qid)
301 return rcu_dereference_raw(dev->queues[qid]);
304 static struct nvme_queue *get_nvmeq(struct nvme_dev *dev) __acquires(RCU)
306 struct nvme_queue *nvmeq;
307 unsigned queue_id = get_cpu_var(*dev->io_queue);
310 nvmeq = rcu_dereference(dev->queues[queue_id]);
315 put_cpu_var(*dev->io_queue);
319 static void put_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
322 put_cpu_var(nvmeq->dev->io_queue);
325 static struct nvme_queue *lock_nvmeq(struct nvme_dev *dev, int q_idx)
328 struct nvme_queue *nvmeq;
331 nvmeq = rcu_dereference(dev->queues[q_idx]);
339 static void unlock_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
345 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
346 * @nvmeq: The queue to use
347 * @cmd: The command to send
349 * Safe to use from interrupt context
351 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
355 spin_lock_irqsave(&nvmeq->q_lock, flags);
356 if (nvmeq->q_suspended) {
357 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
360 tail = nvmeq->sq_tail;
361 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
362 if (++tail == nvmeq->q_depth)
364 writel(tail, nvmeq->q_db);
365 nvmeq->sq_tail = tail;
366 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
371 static __le64 **iod_list(struct nvme_iod *iod)
373 return ((void *)iod) + iod->offset;
377 * Will slightly overestimate the number of pages needed. This is OK
378 * as it only leads to a small amount of wasted memory for the lifetime of
381 static int nvme_npages(unsigned size, struct nvme_dev *dev)
383 unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
384 return DIV_ROUND_UP(8 * nprps, dev->page_size - 8);
387 static struct nvme_iod *
388 nvme_alloc_iod(unsigned nseg, unsigned nbytes, struct nvme_dev *dev, gfp_t gfp)
390 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
391 sizeof(__le64 *) * nvme_npages(nbytes, dev) +
392 sizeof(struct scatterlist) * nseg, gfp);
395 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
397 iod->length = nbytes;
399 iod->first_dma = 0ULL;
400 iod->start_time = jiffies;
406 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
408 const int last_prp = dev->page_size / 8 - 1;
410 __le64 **list = iod_list(iod);
411 dma_addr_t prp_dma = iod->first_dma;
413 if (iod->npages == 0)
414 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
415 for (i = 0; i < iod->npages; i++) {
416 __le64 *prp_list = list[i];
417 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
418 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
419 prp_dma = next_prp_dma;
424 static void nvme_start_io_acct(struct bio *bio)
426 struct gendisk *disk = bio->bi_bdev->bd_disk;
427 if (blk_queue_io_stat(disk->queue)) {
428 const int rw = bio_data_dir(bio);
429 int cpu = part_stat_lock();
430 part_round_stats(cpu, &disk->part0);
431 part_stat_inc(cpu, &disk->part0, ios[rw]);
432 part_stat_add(cpu, &disk->part0, sectors[rw],
434 part_inc_in_flight(&disk->part0, rw);
439 static void nvme_end_io_acct(struct bio *bio, unsigned long start_time)
441 struct gendisk *disk = bio->bi_bdev->bd_disk;
442 if (blk_queue_io_stat(disk->queue)) {
443 const int rw = bio_data_dir(bio);
444 unsigned long duration = jiffies - start_time;
445 int cpu = part_stat_lock();
446 part_stat_add(cpu, &disk->part0, ticks[rw], duration);
447 part_round_stats(cpu, &disk->part0);
448 part_dec_in_flight(&disk->part0, rw);
453 static int nvme_error_status(u16 status)
455 switch (status & 0x7ff) {
456 case NVME_SC_SUCCESS:
458 case NVME_SC_CAP_EXCEEDED:
465 static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
466 struct nvme_completion *cqe)
468 struct nvme_iod *iod = ctx;
469 struct bio *bio = iod->private;
470 u16 status = le16_to_cpup(&cqe->status) >> 1;
473 if (unlikely(status)) {
474 if (!(status & NVME_SC_DNR ||
475 bio->bi_rw & REQ_FAILFAST_MASK) &&
476 (jiffies - iod->start_time) < IOD_TIMEOUT) {
477 if (!waitqueue_active(&nvmeq->sq_full))
478 add_wait_queue(&nvmeq->sq_full,
479 &nvmeq->sq_cong_wait);
480 list_add_tail(&iod->node, &nvmeq->iod_bio);
481 wake_up(&nvmeq->sq_full);
484 error = nvme_error_status(status);
487 dma_unmap_sg(nvmeq->q_dmadev, iod->sg, iod->nents,
488 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
489 nvme_end_io_acct(bio, iod->start_time);
491 nvme_free_iod(nvmeq->dev, iod);
493 trace_block_bio_complete(bdev_get_queue(bio->bi_bdev), bio, error);
494 bio_endio(bio, error);
497 /* length is in bytes. gfp flags indicates whether we may sleep. */
498 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod, int total_len,
501 struct dma_pool *pool;
502 int length = total_len;
503 struct scatterlist *sg = iod->sg;
504 int dma_len = sg_dma_len(sg);
505 u64 dma_addr = sg_dma_address(sg);
506 int offset = offset_in_page(dma_addr);
508 __le64 **list = iod_list(iod);
511 u32 page_size = dev->page_size;
513 length -= (page_size - offset);
517 dma_len -= (page_size - offset);
519 dma_addr += (page_size - offset);
522 dma_addr = sg_dma_address(sg);
523 dma_len = sg_dma_len(sg);
526 if (length <= page_size) {
527 iod->first_dma = dma_addr;
531 nprps = DIV_ROUND_UP(length, page_size);
532 if (nprps <= (256 / 8)) {
533 pool = dev->prp_small_pool;
536 pool = dev->prp_page_pool;
540 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
542 iod->first_dma = dma_addr;
544 return (total_len - length) + page_size;
547 iod->first_dma = prp_dma;
550 if (i == page_size >> 3) {
551 __le64 *old_prp_list = prp_list;
552 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
554 return total_len - length;
555 list[iod->npages++] = prp_list;
556 prp_list[0] = old_prp_list[i - 1];
557 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
560 prp_list[i++] = cpu_to_le64(dma_addr);
561 dma_len -= page_size;
562 dma_addr += page_size;
570 dma_addr = sg_dma_address(sg);
571 dma_len = sg_dma_len(sg);
577 static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
580 struct bio *split = bio_split(bio, len >> 9, GFP_ATOMIC, NULL);
584 trace_block_split(bdev_get_queue(bio->bi_bdev), bio,
585 split->bi_iter.bi_sector);
586 bio_chain(split, bio);
588 if (!waitqueue_active(&nvmeq->sq_full))
589 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
590 bio_list_add(&nvmeq->sq_cong, split);
591 bio_list_add(&nvmeq->sq_cong, bio);
592 wake_up(&nvmeq->sq_full);
597 /* NVMe scatterlists require no holes in the virtual address */
598 #define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2) ((vec2)->bv_offset || \
599 (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
601 static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
602 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
604 struct bio_vec bvec, bvprv;
605 struct bvec_iter iter;
606 struct scatterlist *sg = NULL;
607 int length = 0, nsegs = 0, split_len = bio->bi_iter.bi_size;
610 if (nvmeq->dev->stripe_size)
611 split_len = nvmeq->dev->stripe_size -
612 ((bio->bi_iter.bi_sector << 9) &
613 (nvmeq->dev->stripe_size - 1));
615 sg_init_table(iod->sg, psegs);
616 bio_for_each_segment(bvec, bio, iter) {
617 if (!first && BIOVEC_PHYS_MERGEABLE(&bvprv, &bvec)) {
618 sg->length += bvec.bv_len;
620 if (!first && BIOVEC_NOT_VIRT_MERGEABLE(&bvprv, &bvec))
621 return nvme_split_and_submit(bio, nvmeq,
624 sg = sg ? sg + 1 : iod->sg;
625 sg_set_page(sg, bvec.bv_page,
626 bvec.bv_len, bvec.bv_offset);
630 if (split_len - length < bvec.bv_len)
631 return nvme_split_and_submit(bio, nvmeq, split_len);
632 length += bvec.bv_len;
638 if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
641 BUG_ON(length != bio->bi_iter.bi_size);
645 static int nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
646 struct bio *bio, struct nvme_iod *iod, int cmdid)
648 struct nvme_dsm_range *range =
649 (struct nvme_dsm_range *)iod_list(iod)[0];
650 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
652 range->cattr = cpu_to_le32(0);
653 range->nlb = cpu_to_le32(bio->bi_iter.bi_size >> ns->lba_shift);
654 range->slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
656 memset(cmnd, 0, sizeof(*cmnd));
657 cmnd->dsm.opcode = nvme_cmd_dsm;
658 cmnd->dsm.command_id = cmdid;
659 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
660 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
662 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
664 if (++nvmeq->sq_tail == nvmeq->q_depth)
666 writel(nvmeq->sq_tail, nvmeq->q_db);
671 static int nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
674 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
676 memset(cmnd, 0, sizeof(*cmnd));
677 cmnd->common.opcode = nvme_cmd_flush;
678 cmnd->common.command_id = cmdid;
679 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
681 if (++nvmeq->sq_tail == nvmeq->q_depth)
683 writel(nvmeq->sq_tail, nvmeq->q_db);
688 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod)
690 struct bio *bio = iod->private;
691 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
692 struct nvme_command *cmnd;
697 cmdid = alloc_cmdid(nvmeq, iod, bio_completion, NVME_IO_TIMEOUT);
698 if (unlikely(cmdid < 0))
701 if (bio->bi_rw & REQ_DISCARD)
702 return nvme_submit_discard(nvmeq, ns, bio, iod, cmdid);
703 if (bio->bi_rw & REQ_FLUSH)
704 return nvme_submit_flush(nvmeq, ns, cmdid);
707 if (bio->bi_rw & REQ_FUA)
708 control |= NVME_RW_FUA;
709 if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
710 control |= NVME_RW_LR;
713 if (bio->bi_rw & REQ_RAHEAD)
714 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
716 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
717 memset(cmnd, 0, sizeof(*cmnd));
719 cmnd->rw.opcode = bio_data_dir(bio) ? nvme_cmd_write : nvme_cmd_read;
720 cmnd->rw.command_id = cmdid;
721 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
722 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
723 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
724 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
726 cpu_to_le16((bio->bi_iter.bi_size >> ns->lba_shift) - 1);
727 cmnd->rw.control = cpu_to_le16(control);
728 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
730 if (++nvmeq->sq_tail == nvmeq->q_depth)
732 writel(nvmeq->sq_tail, nvmeq->q_db);
737 static int nvme_split_flush_data(struct nvme_queue *nvmeq, struct bio *bio)
739 struct bio *split = bio_clone(bio, GFP_ATOMIC);
743 split->bi_iter.bi_size = 0;
744 split->bi_phys_segments = 0;
745 bio->bi_rw &= ~REQ_FLUSH;
746 bio_chain(split, bio);
748 if (!waitqueue_active(&nvmeq->sq_full))
749 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
750 bio_list_add(&nvmeq->sq_cong, split);
751 bio_list_add(&nvmeq->sq_cong, bio);
752 wake_up_process(nvme_thread);
758 * Called with local interrupts disabled and the q_lock held. May not sleep.
760 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
763 struct nvme_iod *iod;
764 int psegs = bio_phys_segments(ns->queue, bio);
767 if ((bio->bi_rw & REQ_FLUSH) && psegs)
768 return nvme_split_flush_data(nvmeq, bio);
770 iod = nvme_alloc_iod(psegs, bio->bi_iter.bi_size, ns->dev, GFP_ATOMIC);
775 if (bio->bi_rw & REQ_DISCARD) {
778 * We reuse the small pool to allocate the 16-byte range here
779 * as it is not worth having a special pool for these or
780 * additional cases to handle freeing the iod.
782 range = dma_pool_alloc(nvmeq->dev->prp_small_pool,
789 iod_list(iod)[0] = (__le64 *)range;
792 result = nvme_map_bio(nvmeq, iod, bio,
793 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE,
797 if (nvme_setup_prps(nvmeq->dev, iod, result, GFP_ATOMIC) !=
802 nvme_start_io_acct(bio);
804 if (unlikely(nvme_submit_iod(nvmeq, iod))) {
805 if (!waitqueue_active(&nvmeq->sq_full))
806 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
807 list_add_tail(&iod->node, &nvmeq->iod_bio);
812 nvme_free_iod(nvmeq->dev, iod);
816 static int nvme_process_cq(struct nvme_queue *nvmeq)
820 head = nvmeq->cq_head;
821 phase = nvmeq->cq_phase;
825 nvme_completion_fn fn;
826 struct nvme_completion cqe = nvmeq->cqes[head];
827 if ((le16_to_cpu(cqe.status) & 1) != phase)
829 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
830 if (++head == nvmeq->q_depth) {
835 ctx = free_cmdid(nvmeq, cqe.command_id, &fn);
836 fn(nvmeq, ctx, &cqe);
839 /* If the controller ignores the cq head doorbell and continuously
840 * writes to the queue, it is theoretically possible to wrap around
841 * the queue twice and mistakenly return IRQ_NONE. Linux only
842 * requires that 0.1% of your interrupts are handled, so this isn't
845 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
848 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
849 nvmeq->cq_head = head;
850 nvmeq->cq_phase = phase;
856 static void nvme_make_request(struct request_queue *q, struct bio *bio)
858 struct nvme_ns *ns = q->queuedata;
859 struct nvme_queue *nvmeq = get_nvmeq(ns->dev);
863 bio_endio(bio, -EIO);
867 spin_lock_irq(&nvmeq->q_lock);
868 if (!nvmeq->q_suspended && bio_list_empty(&nvmeq->sq_cong))
869 result = nvme_submit_bio_queue(nvmeq, ns, bio);
870 if (unlikely(result)) {
871 if (!waitqueue_active(&nvmeq->sq_full))
872 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
873 bio_list_add(&nvmeq->sq_cong, bio);
876 nvme_process_cq(nvmeq);
877 spin_unlock_irq(&nvmeq->q_lock);
881 static irqreturn_t nvme_irq(int irq, void *data)
884 struct nvme_queue *nvmeq = data;
885 spin_lock(&nvmeq->q_lock);
886 nvme_process_cq(nvmeq);
887 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
889 spin_unlock(&nvmeq->q_lock);
893 static irqreturn_t nvme_irq_check(int irq, void *data)
895 struct nvme_queue *nvmeq = data;
896 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
897 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
899 return IRQ_WAKE_THREAD;
902 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
904 spin_lock_irq(&nvmeq->q_lock);
905 cancel_cmdid(nvmeq, cmdid, NULL);
906 spin_unlock_irq(&nvmeq->q_lock);
909 struct sync_cmd_info {
910 struct task_struct *task;
915 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
916 struct nvme_completion *cqe)
918 struct sync_cmd_info *cmdinfo = ctx;
919 cmdinfo->result = le32_to_cpup(&cqe->result);
920 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
921 wake_up_process(cmdinfo->task);
925 * Returns 0 on success. If the result is negative, it's a Linux error code;
926 * if the result is positive, it's an NVM Express status code
928 static int nvme_submit_sync_cmd(struct nvme_dev *dev, int q_idx,
929 struct nvme_command *cmd,
930 u32 *result, unsigned timeout)
933 struct sync_cmd_info cmdinfo;
934 struct nvme_queue *nvmeq;
936 nvmeq = lock_nvmeq(dev, q_idx);
940 cmdinfo.task = current;
941 cmdinfo.status = -EINTR;
943 cmdid = alloc_cmdid(nvmeq, &cmdinfo, sync_completion, timeout);
948 cmd->common.command_id = cmdid;
950 set_current_state(TASK_KILLABLE);
951 ret = nvme_submit_cmd(nvmeq, cmd);
953 free_cmdid(nvmeq, cmdid, NULL);
955 set_current_state(TASK_RUNNING);
959 schedule_timeout(timeout);
961 if (cmdinfo.status == -EINTR) {
962 nvmeq = lock_nvmeq(dev, q_idx);
964 nvme_abort_command(nvmeq, cmdid);
971 *result = cmdinfo.result;
973 return cmdinfo.status;
976 static int nvme_submit_async_cmd(struct nvme_queue *nvmeq,
977 struct nvme_command *cmd,
978 struct async_cmd_info *cmdinfo, unsigned timeout)
982 cmdid = alloc_cmdid_killable(nvmeq, cmdinfo, async_completion, timeout);
985 cmdinfo->status = -EINTR;
986 cmd->common.command_id = cmdid;
987 return nvme_submit_cmd(nvmeq, cmd);
990 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
993 return nvme_submit_sync_cmd(dev, 0, cmd, result, ADMIN_TIMEOUT);
996 int nvme_submit_io_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
999 return nvme_submit_sync_cmd(dev, this_cpu_read(*dev->io_queue), cmd,
1000 result, NVME_IO_TIMEOUT);
1003 static int nvme_submit_admin_cmd_async(struct nvme_dev *dev,
1004 struct nvme_command *cmd, struct async_cmd_info *cmdinfo)
1006 return nvme_submit_async_cmd(raw_nvmeq(dev, 0), cmd, cmdinfo,
1010 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1013 struct nvme_command c;
1015 memset(&c, 0, sizeof(c));
1016 c.delete_queue.opcode = opcode;
1017 c.delete_queue.qid = cpu_to_le16(id);
1019 status = nvme_submit_admin_cmd(dev, &c, NULL);
1025 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1026 struct nvme_queue *nvmeq)
1029 struct nvme_command c;
1030 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1032 memset(&c, 0, sizeof(c));
1033 c.create_cq.opcode = nvme_admin_create_cq;
1034 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1035 c.create_cq.cqid = cpu_to_le16(qid);
1036 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1037 c.create_cq.cq_flags = cpu_to_le16(flags);
1038 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1040 status = nvme_submit_admin_cmd(dev, &c, NULL);
1046 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1047 struct nvme_queue *nvmeq)
1050 struct nvme_command c;
1051 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1053 memset(&c, 0, sizeof(c));
1054 c.create_sq.opcode = nvme_admin_create_sq;
1055 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1056 c.create_sq.sqid = cpu_to_le16(qid);
1057 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1058 c.create_sq.sq_flags = cpu_to_le16(flags);
1059 c.create_sq.cqid = cpu_to_le16(qid);
1061 status = nvme_submit_admin_cmd(dev, &c, NULL);
1067 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1069 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1072 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1074 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1077 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
1078 dma_addr_t dma_addr)
1080 struct nvme_command c;
1082 memset(&c, 0, sizeof(c));
1083 c.identify.opcode = nvme_admin_identify;
1084 c.identify.nsid = cpu_to_le32(nsid);
1085 c.identify.prp1 = cpu_to_le64(dma_addr);
1086 c.identify.cns = cpu_to_le32(cns);
1088 return nvme_submit_admin_cmd(dev, &c, NULL);
1091 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1092 dma_addr_t dma_addr, u32 *result)
1094 struct nvme_command c;
1096 memset(&c, 0, sizeof(c));
1097 c.features.opcode = nvme_admin_get_features;
1098 c.features.nsid = cpu_to_le32(nsid);
1099 c.features.prp1 = cpu_to_le64(dma_addr);
1100 c.features.fid = cpu_to_le32(fid);
1102 return nvme_submit_admin_cmd(dev, &c, result);
1105 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1106 dma_addr_t dma_addr, u32 *result)
1108 struct nvme_command c;
1110 memset(&c, 0, sizeof(c));
1111 c.features.opcode = nvme_admin_set_features;
1112 c.features.prp1 = cpu_to_le64(dma_addr);
1113 c.features.fid = cpu_to_le32(fid);
1114 c.features.dword11 = cpu_to_le32(dword11);
1116 return nvme_submit_admin_cmd(dev, &c, result);
1120 * nvme_abort_cmd - Attempt aborting a command
1121 * @cmdid: Command id of a timed out IO
1122 * @queue: The queue with timed out IO
1124 * Schedule controller reset if the command was already aborted once before and
1125 * still hasn't been returned to the driver, or if this is the admin queue.
1127 static void nvme_abort_cmd(int cmdid, struct nvme_queue *nvmeq)
1130 struct nvme_command cmd;
1131 struct nvme_dev *dev = nvmeq->dev;
1132 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1133 struct nvme_queue *adminq;
1135 if (!nvmeq->qid || info[cmdid].aborted) {
1136 if (work_busy(&dev->reset_work))
1138 list_del_init(&dev->node);
1139 dev_warn(&dev->pci_dev->dev,
1140 "I/O %d QID %d timeout, reset controller\n", cmdid,
1142 dev->reset_workfn = nvme_reset_failed_dev;
1143 queue_work(nvme_workq, &dev->reset_work);
1147 if (!dev->abort_limit)
1150 adminq = rcu_dereference(dev->queues[0]);
1151 a_cmdid = alloc_cmdid(adminq, CMD_CTX_ABORT, special_completion,
1156 memset(&cmd, 0, sizeof(cmd));
1157 cmd.abort.opcode = nvme_admin_abort_cmd;
1158 cmd.abort.cid = cmdid;
1159 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1160 cmd.abort.command_id = a_cmdid;
1163 info[cmdid].aborted = 1;
1164 info[cmdid].timeout = jiffies + ADMIN_TIMEOUT;
1166 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", cmdid,
1168 nvme_submit_cmd(adminq, &cmd);
1172 * nvme_cancel_ios - Cancel outstanding I/Os
1173 * @queue: The queue to cancel I/Os on
1174 * @timeout: True to only cancel I/Os which have timed out
1176 static void nvme_cancel_ios(struct nvme_queue *nvmeq, bool timeout)
1178 int depth = nvmeq->q_depth - 1;
1179 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1180 unsigned long now = jiffies;
1183 for_each_set_bit(cmdid, nvmeq->cmdid_data, depth) {
1185 nvme_completion_fn fn;
1186 static struct nvme_completion cqe = {
1187 .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1190 if (timeout && !time_after(now, info[cmdid].timeout))
1192 if (info[cmdid].ctx == CMD_CTX_CANCELLED)
1194 if (timeout && info[cmdid].ctx == CMD_CTX_ASYNC)
1196 if (timeout && nvmeq->dev->initialized) {
1197 nvme_abort_cmd(cmdid, nvmeq);
1200 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n", cmdid,
1202 ctx = cancel_cmdid(nvmeq, cmdid, &fn);
1203 fn(nvmeq, ctx, &cqe);
1207 static void nvme_free_queue(struct nvme_queue *nvmeq)
1209 spin_lock_irq(&nvmeq->q_lock);
1210 while (bio_list_peek(&nvmeq->sq_cong)) {
1211 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1212 bio_endio(bio, -EIO);
1214 while (!list_empty(&nvmeq->iod_bio)) {
1215 static struct nvme_completion cqe = {
1216 .status = cpu_to_le16(
1217 (NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1),
1219 struct nvme_iod *iod = list_first_entry(&nvmeq->iod_bio,
1222 list_del(&iod->node);
1223 bio_completion(nvmeq, iod, &cqe);
1225 spin_unlock_irq(&nvmeq->q_lock);
1227 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1228 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1229 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1230 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1232 free_cpumask_var(nvmeq->cpu_mask);
1236 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1239 struct nvme_queue *nvmeq, *next;
1240 struct llist_node *entry;
1243 for (i = dev->queue_count - 1; i >= lowest; i--) {
1244 nvmeq = raw_nvmeq(dev, i);
1245 RCU_INIT_POINTER(dev->queues[i], NULL);
1246 llist_add(&nvmeq->node, &q_list);
1250 entry = llist_del_all(&q_list);
1251 llist_for_each_entry_safe(nvmeq, next, entry, node)
1252 nvme_free_queue(nvmeq);
1256 * nvme_suspend_queue - put queue into suspended state
1257 * @nvmeq - queue to suspend
1259 * Returns 1 if already suspended, 0 otherwise.
1261 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1263 int vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1265 spin_lock_irq(&nvmeq->q_lock);
1266 if (nvmeq->q_suspended) {
1267 spin_unlock_irq(&nvmeq->q_lock);
1270 nvmeq->q_suspended = 1;
1271 nvmeq->dev->online_queues--;
1272 spin_unlock_irq(&nvmeq->q_lock);
1274 irq_set_affinity_hint(vector, NULL);
1275 free_irq(vector, nvmeq);
1280 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1282 spin_lock_irq(&nvmeq->q_lock);
1283 nvme_process_cq(nvmeq);
1284 nvme_cancel_ios(nvmeq, false);
1285 spin_unlock_irq(&nvmeq->q_lock);
1288 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1290 struct nvme_queue *nvmeq = raw_nvmeq(dev, qid);
1294 if (nvme_suspend_queue(nvmeq))
1297 /* Don't tell the adapter to delete the admin queue.
1298 * Don't tell a removed adapter to delete IO queues. */
1299 if (qid && readl(&dev->bar->csts) != -1) {
1300 adapter_delete_sq(dev, qid);
1301 adapter_delete_cq(dev, qid);
1303 nvme_clear_queue(nvmeq);
1306 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1307 int depth, int vector)
1309 struct device *dmadev = &dev->pci_dev->dev;
1310 unsigned extra = nvme_queue_extra(depth);
1311 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
1315 nvmeq->cqes = dma_zalloc_coherent(dmadev, CQ_SIZE(depth),
1316 &nvmeq->cq_dma_addr, GFP_KERNEL);
1320 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1321 &nvmeq->sq_dma_addr, GFP_KERNEL);
1322 if (!nvmeq->sq_cmds)
1325 if (qid && !zalloc_cpumask_var(&nvmeq->cpu_mask, GFP_KERNEL))
1328 nvmeq->q_dmadev = dmadev;
1330 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1331 dev->instance, qid);
1332 spin_lock_init(&nvmeq->q_lock);
1334 nvmeq->cq_phase = 1;
1335 init_waitqueue_head(&nvmeq->sq_full);
1336 bio_list_init(&nvmeq->sq_cong);
1337 INIT_LIST_HEAD(&nvmeq->iod_bio);
1338 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1339 nvmeq->q_depth = depth;
1340 nvmeq->cq_vector = vector;
1342 nvmeq->q_suspended = 1;
1344 rcu_assign_pointer(dev->queues[qid], nvmeq);
1349 dma_free_coherent(dmadev, SQ_SIZE(depth), (void *)nvmeq->sq_cmds,
1350 nvmeq->sq_dma_addr);
1352 dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1353 nvmeq->cq_dma_addr);
1359 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1362 if (use_threaded_interrupts)
1363 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1364 nvme_irq_check, nvme_irq, IRQF_SHARED,
1366 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1367 IRQF_SHARED, name, nvmeq);
1370 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1372 struct nvme_dev *dev = nvmeq->dev;
1373 unsigned extra = nvme_queue_extra(nvmeq->q_depth);
1375 spin_lock_irq(&nvmeq->q_lock);
1376 init_waitqueue_entry(&nvmeq->sq_cong_wait, nvme_thread);
1379 nvmeq->cq_phase = 1;
1380 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1381 memset(nvmeq->cmdid_data, 0, extra);
1382 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1383 nvme_cancel_ios(nvmeq, false);
1384 nvmeq->q_suspended = 0;
1385 dev->online_queues++;
1386 spin_unlock_irq(&nvmeq->q_lock);
1389 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1391 struct nvme_dev *dev = nvmeq->dev;
1394 result = adapter_alloc_cq(dev, qid, nvmeq);
1398 result = adapter_alloc_sq(dev, qid, nvmeq);
1402 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1406 nvme_init_queue(nvmeq, qid);
1410 adapter_delete_sq(dev, qid);
1412 adapter_delete_cq(dev, qid);
1416 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1418 unsigned long timeout;
1419 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1421 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1423 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1425 if (fatal_signal_pending(current))
1427 if (time_after(jiffies, timeout)) {
1428 dev_err(&dev->pci_dev->dev,
1429 "Device not ready; aborting %s\n", enabled ?
1430 "initialisation" : "reset");
1439 * If the device has been passed off to us in an enabled state, just clear
1440 * the enabled bit. The spec says we should set the 'shutdown notification
1441 * bits', but doing so may cause the device to complete commands to the
1442 * admin queue ... and we don't know what memory that might be pointing at!
1444 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1446 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1447 dev->ctrl_config &= ~NVME_CC_ENABLE;
1448 writel(dev->ctrl_config, &dev->bar->cc);
1450 return nvme_wait_ready(dev, cap, false);
1453 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1455 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1456 dev->ctrl_config |= NVME_CC_ENABLE;
1457 writel(dev->ctrl_config, &dev->bar->cc);
1459 return nvme_wait_ready(dev, cap, true);
1462 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1464 unsigned long timeout;
1466 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1467 dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1469 writel(dev->ctrl_config, &dev->bar->cc);
1471 timeout = SHUTDOWN_TIMEOUT + jiffies;
1472 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1473 NVME_CSTS_SHST_CMPLT) {
1475 if (fatal_signal_pending(current))
1477 if (time_after(jiffies, timeout)) {
1478 dev_err(&dev->pci_dev->dev,
1479 "Device shutdown incomplete; abort shutdown\n");
1487 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1491 u64 cap = readq(&dev->bar->cap);
1492 struct nvme_queue *nvmeq;
1493 unsigned page_shift = PAGE_SHIFT;
1494 unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1495 unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1497 if (page_shift < dev_page_min) {
1498 dev_err(&dev->pci_dev->dev,
1499 "Minimum device page size (%u) too large for "
1500 "host (%u)\n", 1 << dev_page_min,
1504 if (page_shift > dev_page_max) {
1505 dev_info(&dev->pci_dev->dev,
1506 "Device maximum page size (%u) smaller than "
1507 "host (%u); enabling work-around\n",
1508 1 << dev_page_max, 1 << page_shift);
1509 page_shift = dev_page_max;
1512 result = nvme_disable_ctrl(dev, cap);
1516 nvmeq = raw_nvmeq(dev, 0);
1518 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
1523 aqa = nvmeq->q_depth - 1;
1526 dev->page_size = 1 << page_shift;
1528 dev->ctrl_config = NVME_CC_CSS_NVM;
1529 dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1530 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1531 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1533 writel(aqa, &dev->bar->aqa);
1534 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1535 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1537 result = nvme_enable_ctrl(dev, cap);
1541 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
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, 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, &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_admin_cmd(struct nvme_dev *dev,
1736 struct nvme_admin_cmd __user *ucmd)
1738 struct nvme_admin_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) :
1775 if (length != cmd.data_len)
1778 status = nvme_submit_sync_cmd(dev, 0, &c, &cmd.result, timeout);
1781 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1782 nvme_free_iod(dev, iod);
1785 if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1786 sizeof(cmd.result)))
1792 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1795 struct nvme_ns *ns = bdev->bd_disk->private_data;
1799 force_successful_syscall_return();
1801 case NVME_IOCTL_ADMIN_CMD:
1802 return nvme_user_admin_cmd(ns->dev, (void __user *)arg);
1803 case NVME_IOCTL_SUBMIT_IO:
1804 return nvme_submit_io(ns, (void __user *)arg);
1805 case SG_GET_VERSION_NUM:
1806 return nvme_sg_get_version_num((void __user *)arg);
1808 return nvme_sg_io(ns, (void __user *)arg);
1814 #ifdef CONFIG_COMPAT
1815 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1816 unsigned int cmd, unsigned long arg)
1820 return -ENOIOCTLCMD;
1822 return nvme_ioctl(bdev, mode, cmd, arg);
1825 #define nvme_compat_ioctl NULL
1828 static int nvme_open(struct block_device *bdev, fmode_t mode)
1830 struct nvme_ns *ns = bdev->bd_disk->private_data;
1831 struct nvme_dev *dev = ns->dev;
1833 kref_get(&dev->kref);
1837 static void nvme_free_dev(struct kref *kref);
1839 static void nvme_release(struct gendisk *disk, fmode_t mode)
1841 struct nvme_ns *ns = disk->private_data;
1842 struct nvme_dev *dev = ns->dev;
1844 kref_put(&dev->kref, nvme_free_dev);
1847 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1849 /* some standard values */
1850 geo->heads = 1 << 6;
1851 geo->sectors = 1 << 5;
1852 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1856 static int nvme_revalidate_disk(struct gendisk *disk)
1858 struct nvme_ns *ns = disk->private_data;
1859 struct nvme_dev *dev = ns->dev;
1860 struct nvme_id_ns *id;
1861 dma_addr_t dma_addr;
1864 id = dma_alloc_coherent(&dev->pci_dev->dev, 4096, &dma_addr,
1867 dev_warn(&dev->pci_dev->dev, "%s: Memory alocation failure\n",
1872 if (nvme_identify(dev, ns->ns_id, 0, dma_addr))
1875 lbaf = id->flbas & 0xf;
1876 ns->lba_shift = id->lbaf[lbaf].ds;
1878 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1879 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1881 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1885 static const struct block_device_operations nvme_fops = {
1886 .owner = THIS_MODULE,
1887 .ioctl = nvme_ioctl,
1888 .compat_ioctl = nvme_compat_ioctl,
1890 .release = nvme_release,
1891 .getgeo = nvme_getgeo,
1892 .revalidate_disk= nvme_revalidate_disk,
1895 static void nvme_resubmit_iods(struct nvme_queue *nvmeq)
1897 struct nvme_iod *iod, *next;
1899 list_for_each_entry_safe(iod, next, &nvmeq->iod_bio, node) {
1900 if (unlikely(nvme_submit_iod(nvmeq, iod)))
1902 list_del(&iod->node);
1903 if (bio_list_empty(&nvmeq->sq_cong) &&
1904 list_empty(&nvmeq->iod_bio))
1905 remove_wait_queue(&nvmeq->sq_full,
1906 &nvmeq->sq_cong_wait);
1910 static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
1912 while (bio_list_peek(&nvmeq->sq_cong)) {
1913 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1914 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
1916 if (bio_list_empty(&nvmeq->sq_cong) &&
1917 list_empty(&nvmeq->iod_bio))
1918 remove_wait_queue(&nvmeq->sq_full,
1919 &nvmeq->sq_cong_wait);
1920 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
1921 if (!waitqueue_active(&nvmeq->sq_full))
1922 add_wait_queue(&nvmeq->sq_full,
1923 &nvmeq->sq_cong_wait);
1924 bio_list_add_head(&nvmeq->sq_cong, bio);
1930 static int nvme_submit_async_req(struct nvme_queue *nvmeq)
1932 struct nvme_command *c;
1935 cmdid = alloc_cmdid(nvmeq, CMD_CTX_ASYNC, special_completion, 0);
1939 c = &nvmeq->sq_cmds[nvmeq->sq_tail];
1940 memset(c, 0, sizeof(*c));
1941 c->common.opcode = nvme_admin_async_event;
1942 c->common.command_id = cmdid;
1944 if (++nvmeq->sq_tail == nvmeq->q_depth)
1946 writel(nvmeq->sq_tail, nvmeq->q_db);
1951 static int nvme_kthread(void *data)
1953 struct nvme_dev *dev, *next;
1955 while (!kthread_should_stop()) {
1956 set_current_state(TASK_INTERRUPTIBLE);
1957 spin_lock(&dev_list_lock);
1958 list_for_each_entry_safe(dev, next, &dev_list, node) {
1960 if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1962 if (work_busy(&dev->reset_work))
1964 list_del_init(&dev->node);
1965 dev_warn(&dev->pci_dev->dev,
1966 "Failed status, reset controller\n");
1967 dev->reset_workfn = nvme_reset_failed_dev;
1968 queue_work(nvme_workq, &dev->reset_work);
1972 for (i = 0; i < dev->queue_count; i++) {
1973 struct nvme_queue *nvmeq =
1974 rcu_dereference(dev->queues[i]);
1977 spin_lock_irq(&nvmeq->q_lock);
1978 if (nvmeq->q_suspended)
1980 nvme_process_cq(nvmeq);
1981 nvme_cancel_ios(nvmeq, true);
1982 nvme_resubmit_bios(nvmeq);
1983 nvme_resubmit_iods(nvmeq);
1985 while ((i == 0) && (dev->event_limit > 0)) {
1986 if (nvme_submit_async_req(nvmeq))
1991 spin_unlock_irq(&nvmeq->q_lock);
1995 spin_unlock(&dev_list_lock);
1996 schedule_timeout(round_jiffies_relative(HZ));
2001 static void nvme_config_discard(struct nvme_ns *ns)
2003 u32 logical_block_size = queue_logical_block_size(ns->queue);
2004 ns->queue->limits.discard_zeroes_data = 0;
2005 ns->queue->limits.discard_alignment = logical_block_size;
2006 ns->queue->limits.discard_granularity = logical_block_size;
2007 ns->queue->limits.max_discard_sectors = 0xffffffff;
2008 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
2011 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
2012 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
2015 struct gendisk *disk;
2018 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
2021 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
2024 ns->queue = blk_alloc_queue(GFP_KERNEL);
2027 ns->queue->queue_flags = QUEUE_FLAG_DEFAULT;
2028 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
2029 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
2030 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, ns->queue);
2031 blk_queue_make_request(ns->queue, nvme_make_request);
2033 ns->queue->queuedata = ns;
2035 disk = alloc_disk(0);
2037 goto out_free_queue;
2040 lbaf = id->flbas & 0xf;
2041 ns->lba_shift = id->lbaf[lbaf].ds;
2042 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
2043 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2044 if (dev->max_hw_sectors)
2045 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2046 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2047 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2049 disk->major = nvme_major;
2050 disk->first_minor = 0;
2051 disk->fops = &nvme_fops;
2052 disk->private_data = ns;
2053 disk->queue = ns->queue;
2054 disk->driverfs_dev = &dev->pci_dev->dev;
2055 disk->flags = GENHD_FL_EXT_DEVT;
2056 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2057 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
2059 if (dev->oncs & NVME_CTRL_ONCS_DSM)
2060 nvme_config_discard(ns);
2065 blk_cleanup_queue(ns->queue);
2071 static int nvme_find_closest_node(int node)
2073 int n, val, min_val = INT_MAX, best_node = node;
2075 for_each_online_node(n) {
2078 val = node_distance(node, n);
2079 if (val < min_val) {
2087 static void nvme_set_queue_cpus(cpumask_t *qmask, struct nvme_queue *nvmeq,
2091 for_each_cpu(cpu, qmask) {
2092 if (cpumask_weight(nvmeq->cpu_mask) >= count)
2094 if (!cpumask_test_and_set_cpu(cpu, nvmeq->cpu_mask))
2095 *per_cpu_ptr(nvmeq->dev->io_queue, cpu) = nvmeq->qid;
2099 static void nvme_add_cpus(cpumask_t *mask, const cpumask_t *unassigned_cpus,
2100 const cpumask_t *new_mask, struct nvme_queue *nvmeq, int cpus_per_queue)
2103 for_each_cpu(next_cpu, new_mask) {
2104 cpumask_or(mask, mask, get_cpu_mask(next_cpu));
2105 cpumask_or(mask, mask, topology_thread_cpumask(next_cpu));
2106 cpumask_and(mask, mask, unassigned_cpus);
2107 nvme_set_queue_cpus(mask, nvmeq, cpus_per_queue);
2111 static void nvme_create_io_queues(struct nvme_dev *dev)
2115 max = min(dev->max_qid, num_online_cpus());
2116 for (i = dev->queue_count; i <= max; i++)
2117 if (!nvme_alloc_queue(dev, i, dev->q_depth, i - 1))
2120 max = min(dev->queue_count - 1, num_online_cpus());
2121 for (i = dev->online_queues; i <= max; i++)
2122 if (nvme_create_queue(raw_nvmeq(dev, i), i))
2127 * If there are fewer queues than online cpus, this will try to optimally
2128 * assign a queue to multiple cpus by grouping cpus that are "close" together:
2129 * thread siblings, core, socket, closest node, then whatever else is
2132 static void nvme_assign_io_queues(struct nvme_dev *dev)
2134 unsigned cpu, cpus_per_queue, queues, remainder, i;
2135 cpumask_var_t unassigned_cpus;
2137 nvme_create_io_queues(dev);
2139 queues = min(dev->online_queues - 1, num_online_cpus());
2143 cpus_per_queue = num_online_cpus() / queues;
2144 remainder = queues - (num_online_cpus() - queues * cpus_per_queue);
2146 if (!alloc_cpumask_var(&unassigned_cpus, GFP_KERNEL))
2149 cpumask_copy(unassigned_cpus, cpu_online_mask);
2150 cpu = cpumask_first(unassigned_cpus);
2151 for (i = 1; i <= queues; i++) {
2152 struct nvme_queue *nvmeq = lock_nvmeq(dev, i);
2155 cpumask_clear(nvmeq->cpu_mask);
2156 if (!cpumask_weight(unassigned_cpus)) {
2157 unlock_nvmeq(nvmeq);
2161 mask = *get_cpu_mask(cpu);
2162 nvme_set_queue_cpus(&mask, nvmeq, cpus_per_queue);
2163 if (cpus_weight(mask) < cpus_per_queue)
2164 nvme_add_cpus(&mask, unassigned_cpus,
2165 topology_thread_cpumask(cpu),
2166 nvmeq, cpus_per_queue);
2167 if (cpus_weight(mask) < cpus_per_queue)
2168 nvme_add_cpus(&mask, unassigned_cpus,
2169 topology_core_cpumask(cpu),
2170 nvmeq, cpus_per_queue);
2171 if (cpus_weight(mask) < cpus_per_queue)
2172 nvme_add_cpus(&mask, unassigned_cpus,
2173 cpumask_of_node(cpu_to_node(cpu)),
2174 nvmeq, cpus_per_queue);
2175 if (cpus_weight(mask) < cpus_per_queue)
2176 nvme_add_cpus(&mask, unassigned_cpus,
2178 nvme_find_closest_node(
2180 nvmeq, cpus_per_queue);
2181 if (cpus_weight(mask) < cpus_per_queue)
2182 nvme_add_cpus(&mask, unassigned_cpus,
2184 nvmeq, cpus_per_queue);
2186 WARN(cpumask_weight(nvmeq->cpu_mask) != cpus_per_queue,
2187 "nvme%d qid:%d mis-matched queue-to-cpu assignment\n",
2190 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2192 cpumask_andnot(unassigned_cpus, unassigned_cpus,
2194 cpu = cpumask_next(cpu, unassigned_cpus);
2195 if (remainder && !--remainder)
2197 unlock_nvmeq(nvmeq);
2199 WARN(cpumask_weight(unassigned_cpus), "nvme%d unassigned online cpus\n",
2202 cpumask_andnot(unassigned_cpus, cpu_possible_mask, cpu_online_mask);
2203 for_each_cpu(cpu, unassigned_cpus)
2204 *per_cpu_ptr(dev->io_queue, cpu) = (i++ % queues) + 1;
2205 free_cpumask_var(unassigned_cpus);
2208 static int set_queue_count(struct nvme_dev *dev, int count)
2212 u32 q_count = (count - 1) | ((count - 1) << 16);
2214 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2219 dev_err(&dev->pci_dev->dev, "Could not set queue count (%d)\n",
2223 return min(result & 0xffff, result >> 16) + 1;
2226 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2228 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2231 static void nvme_cpu_workfn(struct work_struct *work)
2233 struct nvme_dev *dev = container_of(work, struct nvme_dev, cpu_work);
2234 if (dev->initialized)
2235 nvme_assign_io_queues(dev);
2238 static int nvme_cpu_notify(struct notifier_block *self,
2239 unsigned long action, void *hcpu)
2241 struct nvme_dev *dev;
2246 spin_lock(&dev_list_lock);
2247 list_for_each_entry(dev, &dev_list, node)
2248 schedule_work(&dev->cpu_work);
2249 spin_unlock(&dev_list_lock);
2255 static int nvme_setup_io_queues(struct nvme_dev *dev)
2257 struct nvme_queue *adminq = raw_nvmeq(dev, 0);
2258 struct pci_dev *pdev = dev->pci_dev;
2259 int result, i, vecs, nr_io_queues, size;
2261 nr_io_queues = num_possible_cpus();
2262 result = set_queue_count(dev, nr_io_queues);
2265 if (result < nr_io_queues)
2266 nr_io_queues = result;
2268 size = db_bar_size(dev, nr_io_queues);
2272 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2275 if (!--nr_io_queues)
2277 size = db_bar_size(dev, nr_io_queues);
2279 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2280 adminq->q_db = dev->dbs;
2283 /* Deregister the admin queue's interrupt */
2284 free_irq(dev->entry[0].vector, adminq);
2286 for (i = 0; i < nr_io_queues; i++)
2287 dev->entry[i].entry = i;
2288 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2290 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2294 for (i = 0; i < vecs; i++)
2295 dev->entry[i].vector = i + pdev->irq;
2300 * Should investigate if there's a performance win from allocating
2301 * more queues than interrupt vectors; it might allow the submission
2302 * path to scale better, even if the receive path is limited by the
2303 * number of interrupts.
2305 nr_io_queues = vecs;
2306 dev->max_qid = nr_io_queues;
2308 result = queue_request_irq(dev, adminq, adminq->irqname);
2310 adminq->q_suspended = 1;
2314 /* Free previously allocated queues that are no longer usable */
2315 nvme_free_queues(dev, nr_io_queues + 1);
2316 nvme_assign_io_queues(dev);
2321 nvme_free_queues(dev, 1);
2326 * Return: error value if an error occurred setting up the queues or calling
2327 * Identify Device. 0 if these succeeded, even if adding some of the
2328 * namespaces failed. At the moment, these failures are silent. TBD which
2329 * failures should be reported.
2331 static int nvme_dev_add(struct nvme_dev *dev)
2333 struct pci_dev *pdev = dev->pci_dev;
2337 struct nvme_id_ctrl *ctrl;
2338 struct nvme_id_ns *id_ns;
2340 dma_addr_t dma_addr;
2341 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2343 mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
2347 res = nvme_identify(dev, 0, 1, dma_addr);
2349 dev_err(&pdev->dev, "Identify Controller failed (%d)\n", res);
2355 nn = le32_to_cpup(&ctrl->nn);
2356 dev->oncs = le16_to_cpup(&ctrl->oncs);
2357 dev->abort_limit = ctrl->acl + 1;
2358 dev->vwc = ctrl->vwc;
2359 dev->event_limit = min(ctrl->aerl + 1, 8);
2360 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2361 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2362 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2364 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2365 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2366 (pdev->device == 0x0953) && ctrl->vs[3])
2367 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2370 for (i = 1; i <= nn; i++) {
2371 res = nvme_identify(dev, i, 0, dma_addr);
2375 if (id_ns->ncap == 0)
2378 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
2379 dma_addr + 4096, NULL);
2381 memset(mem + 4096, 0, 4096);
2383 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
2385 list_add_tail(&ns->list, &dev->namespaces);
2387 list_for_each_entry(ns, &dev->namespaces, list)
2392 dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
2396 static int nvme_dev_map(struct nvme_dev *dev)
2399 int bars, result = -ENOMEM;
2400 struct pci_dev *pdev = dev->pci_dev;
2402 if (pci_enable_device_mem(pdev))
2405 dev->entry[0].vector = pdev->irq;
2406 pci_set_master(pdev);
2407 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2408 if (pci_request_selected_regions(pdev, bars, "nvme"))
2411 if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2412 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2415 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2418 if (readl(&dev->bar->csts) == -1) {
2422 cap = readq(&dev->bar->cap);
2423 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2424 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2425 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2433 pci_release_regions(pdev);
2435 pci_disable_device(pdev);
2439 static void nvme_dev_unmap(struct nvme_dev *dev)
2441 if (dev->pci_dev->msi_enabled)
2442 pci_disable_msi(dev->pci_dev);
2443 else if (dev->pci_dev->msix_enabled)
2444 pci_disable_msix(dev->pci_dev);
2449 pci_release_regions(dev->pci_dev);
2452 if (pci_is_enabled(dev->pci_dev))
2453 pci_disable_device(dev->pci_dev);
2456 struct nvme_delq_ctx {
2457 struct task_struct *waiter;
2458 struct kthread_worker *worker;
2462 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2464 dq->waiter = current;
2468 set_current_state(TASK_KILLABLE);
2469 if (!atomic_read(&dq->refcount))
2471 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2472 fatal_signal_pending(current)) {
2473 set_current_state(TASK_RUNNING);
2475 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2476 nvme_disable_queue(dev, 0);
2478 send_sig(SIGKILL, dq->worker->task, 1);
2479 flush_kthread_worker(dq->worker);
2483 set_current_state(TASK_RUNNING);
2486 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2488 atomic_dec(&dq->refcount);
2490 wake_up_process(dq->waiter);
2493 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2495 atomic_inc(&dq->refcount);
2499 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2501 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2503 nvme_clear_queue(nvmeq);
2507 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2508 kthread_work_func_t fn)
2510 struct nvme_command c;
2512 memset(&c, 0, sizeof(c));
2513 c.delete_queue.opcode = opcode;
2514 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2516 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2517 return nvme_submit_admin_cmd_async(nvmeq->dev, &c, &nvmeq->cmdinfo);
2520 static void nvme_del_cq_work_handler(struct kthread_work *work)
2522 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2524 nvme_del_queue_end(nvmeq);
2527 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2529 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2530 nvme_del_cq_work_handler);
2533 static void nvme_del_sq_work_handler(struct kthread_work *work)
2535 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2537 int status = nvmeq->cmdinfo.status;
2540 status = nvme_delete_cq(nvmeq);
2542 nvme_del_queue_end(nvmeq);
2545 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2547 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2548 nvme_del_sq_work_handler);
2551 static void nvme_del_queue_start(struct kthread_work *work)
2553 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2555 allow_signal(SIGKILL);
2556 if (nvme_delete_sq(nvmeq))
2557 nvme_del_queue_end(nvmeq);
2560 static void nvme_disable_io_queues(struct nvme_dev *dev)
2563 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2564 struct nvme_delq_ctx dq;
2565 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2566 &worker, "nvme%d", dev->instance);
2568 if (IS_ERR(kworker_task)) {
2569 dev_err(&dev->pci_dev->dev,
2570 "Failed to create queue del task\n");
2571 for (i = dev->queue_count - 1; i > 0; i--)
2572 nvme_disable_queue(dev, i);
2577 atomic_set(&dq.refcount, 0);
2578 dq.worker = &worker;
2579 for (i = dev->queue_count - 1; i > 0; i--) {
2580 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2582 if (nvme_suspend_queue(nvmeq))
2584 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2585 nvmeq->cmdinfo.worker = dq.worker;
2586 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2587 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2589 nvme_wait_dq(&dq, dev);
2590 kthread_stop(kworker_task);
2594 * Remove the node from the device list and check
2595 * for whether or not we need to stop the nvme_thread.
2597 static void nvme_dev_list_remove(struct nvme_dev *dev)
2599 struct task_struct *tmp = NULL;
2601 spin_lock(&dev_list_lock);
2602 list_del_init(&dev->node);
2603 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2607 spin_unlock(&dev_list_lock);
2613 static void nvme_dev_shutdown(struct nvme_dev *dev)
2618 dev->initialized = 0;
2619 nvme_dev_list_remove(dev);
2622 csts = readl(&dev->bar->csts);
2623 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2624 for (i = dev->queue_count - 1; i >= 0; i--) {
2625 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2626 nvme_suspend_queue(nvmeq);
2627 nvme_clear_queue(nvmeq);
2630 nvme_disable_io_queues(dev);
2631 nvme_shutdown_ctrl(dev);
2632 nvme_disable_queue(dev, 0);
2634 nvme_dev_unmap(dev);
2637 static void nvme_dev_remove(struct nvme_dev *dev)
2641 list_for_each_entry(ns, &dev->namespaces, list) {
2642 if (ns->disk->flags & GENHD_FL_UP)
2643 del_gendisk(ns->disk);
2644 if (!blk_queue_dying(ns->queue))
2645 blk_cleanup_queue(ns->queue);
2649 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2651 struct device *dmadev = &dev->pci_dev->dev;
2652 dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2653 PAGE_SIZE, PAGE_SIZE, 0);
2654 if (!dev->prp_page_pool)
2657 /* Optimisation for I/Os between 4k and 128k */
2658 dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2660 if (!dev->prp_small_pool) {
2661 dma_pool_destroy(dev->prp_page_pool);
2667 static void nvme_release_prp_pools(struct nvme_dev *dev)
2669 dma_pool_destroy(dev->prp_page_pool);
2670 dma_pool_destroy(dev->prp_small_pool);
2673 static DEFINE_IDA(nvme_instance_ida);
2675 static int nvme_set_instance(struct nvme_dev *dev)
2677 int instance, error;
2680 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2683 spin_lock(&dev_list_lock);
2684 error = ida_get_new(&nvme_instance_ida, &instance);
2685 spin_unlock(&dev_list_lock);
2686 } while (error == -EAGAIN);
2691 dev->instance = instance;
2695 static void nvme_release_instance(struct nvme_dev *dev)
2697 spin_lock(&dev_list_lock);
2698 ida_remove(&nvme_instance_ida, dev->instance);
2699 spin_unlock(&dev_list_lock);
2702 static void nvme_free_namespaces(struct nvme_dev *dev)
2704 struct nvme_ns *ns, *next;
2706 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2707 list_del(&ns->list);
2713 static void nvme_free_dev(struct kref *kref)
2715 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2717 pci_dev_put(dev->pci_dev);
2718 nvme_free_namespaces(dev);
2719 free_percpu(dev->io_queue);
2725 static int nvme_dev_open(struct inode *inode, struct file *f)
2727 struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2729 kref_get(&dev->kref);
2730 f->private_data = dev;
2734 static int nvme_dev_release(struct inode *inode, struct file *f)
2736 struct nvme_dev *dev = f->private_data;
2737 kref_put(&dev->kref, nvme_free_dev);
2741 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2743 struct nvme_dev *dev = f->private_data;
2745 case NVME_IOCTL_ADMIN_CMD:
2746 return nvme_user_admin_cmd(dev, (void __user *)arg);
2752 static const struct file_operations nvme_dev_fops = {
2753 .owner = THIS_MODULE,
2754 .open = nvme_dev_open,
2755 .release = nvme_dev_release,
2756 .unlocked_ioctl = nvme_dev_ioctl,
2757 .compat_ioctl = nvme_dev_ioctl,
2760 static int nvme_dev_start(struct nvme_dev *dev)
2763 bool start_thread = false;
2765 result = nvme_dev_map(dev);
2769 result = nvme_configure_admin_queue(dev);
2773 spin_lock(&dev_list_lock);
2774 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2775 start_thread = true;
2778 list_add(&dev->node, &dev_list);
2779 spin_unlock(&dev_list_lock);
2782 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2783 wake_up(&nvme_kthread_wait);
2785 wait_event_killable(nvme_kthread_wait, nvme_thread);
2787 if (IS_ERR_OR_NULL(nvme_thread)) {
2788 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2791 nvme_init_queue(raw_nvmeq(dev, 0), 0);
2793 result = nvme_setup_io_queues(dev);
2800 nvme_disable_queue(dev, 0);
2801 nvme_dev_list_remove(dev);
2803 nvme_dev_unmap(dev);
2807 static int nvme_remove_dead_ctrl(void *arg)
2809 struct nvme_dev *dev = (struct nvme_dev *)arg;
2810 struct pci_dev *pdev = dev->pci_dev;
2812 if (pci_get_drvdata(pdev))
2813 pci_stop_and_remove_bus_device_locked(pdev);
2814 kref_put(&dev->kref, nvme_free_dev);
2818 static void nvme_remove_disks(struct work_struct *ws)
2820 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2822 nvme_free_queues(dev, 1);
2823 nvme_dev_remove(dev);
2826 static int nvme_dev_resume(struct nvme_dev *dev)
2830 ret = nvme_dev_start(dev);
2833 if (dev->online_queues < 2) {
2834 spin_lock(&dev_list_lock);
2835 dev->reset_workfn = nvme_remove_disks;
2836 queue_work(nvme_workq, &dev->reset_work);
2837 spin_unlock(&dev_list_lock);
2839 dev->initialized = 1;
2843 static void nvme_dev_reset(struct nvme_dev *dev)
2845 nvme_dev_shutdown(dev);
2846 if (nvme_dev_resume(dev)) {
2847 dev_err(&dev->pci_dev->dev, "Device failed to resume\n");
2848 kref_get(&dev->kref);
2849 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2851 dev_err(&dev->pci_dev->dev,
2852 "Failed to start controller remove task\n");
2853 kref_put(&dev->kref, nvme_free_dev);
2858 static void nvme_reset_failed_dev(struct work_struct *ws)
2860 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2861 nvme_dev_reset(dev);
2864 static void nvme_reset_workfn(struct work_struct *work)
2866 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2867 dev->reset_workfn(work);
2870 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2872 int result = -ENOMEM;
2873 struct nvme_dev *dev;
2875 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2878 dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
2882 dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
2886 dev->io_queue = alloc_percpu(unsigned short);
2890 INIT_LIST_HEAD(&dev->namespaces);
2891 dev->reset_workfn = nvme_reset_failed_dev;
2892 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
2893 INIT_WORK(&dev->cpu_work, nvme_cpu_workfn);
2894 dev->pci_dev = pci_dev_get(pdev);
2895 pci_set_drvdata(pdev, dev);
2896 result = nvme_set_instance(dev);
2900 result = nvme_setup_prp_pools(dev);
2904 kref_init(&dev->kref);
2905 result = nvme_dev_start(dev);
2909 if (dev->online_queues > 1)
2910 result = nvme_dev_add(dev);
2914 scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2915 dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2916 dev->miscdev.parent = &pdev->dev;
2917 dev->miscdev.name = dev->name;
2918 dev->miscdev.fops = &nvme_dev_fops;
2919 result = misc_register(&dev->miscdev);
2923 dev->initialized = 1;
2927 nvme_dev_remove(dev);
2928 nvme_free_namespaces(dev);
2930 nvme_dev_shutdown(dev);
2932 nvme_free_queues(dev, 0);
2933 nvme_release_prp_pools(dev);
2935 nvme_release_instance(dev);
2937 pci_dev_put(dev->pci_dev);
2939 free_percpu(dev->io_queue);
2946 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2948 struct nvme_dev *dev = pci_get_drvdata(pdev);
2951 nvme_dev_shutdown(dev);
2953 nvme_dev_resume(dev);
2956 static void nvme_shutdown(struct pci_dev *pdev)
2958 struct nvme_dev *dev = pci_get_drvdata(pdev);
2959 nvme_dev_shutdown(dev);
2962 static void nvme_remove(struct pci_dev *pdev)
2964 struct nvme_dev *dev = pci_get_drvdata(pdev);
2966 spin_lock(&dev_list_lock);
2967 list_del_init(&dev->node);
2968 spin_unlock(&dev_list_lock);
2970 pci_set_drvdata(pdev, NULL);
2971 flush_work(&dev->reset_work);
2972 flush_work(&dev->cpu_work);
2973 misc_deregister(&dev->miscdev);
2974 nvme_dev_shutdown(dev);
2975 nvme_free_queues(dev, 0);
2976 nvme_dev_remove(dev);
2977 nvme_release_instance(dev);
2978 nvme_release_prp_pools(dev);
2979 kref_put(&dev->kref, nvme_free_dev);
2982 /* These functions are yet to be implemented */
2983 #define nvme_error_detected NULL
2984 #define nvme_dump_registers NULL
2985 #define nvme_link_reset NULL
2986 #define nvme_slot_reset NULL
2987 #define nvme_error_resume NULL
2989 #ifdef CONFIG_PM_SLEEP
2990 static int nvme_suspend(struct device *dev)
2992 struct pci_dev *pdev = to_pci_dev(dev);
2993 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2995 nvme_dev_shutdown(ndev);
2999 static int nvme_resume(struct device *dev)
3001 struct pci_dev *pdev = to_pci_dev(dev);
3002 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3004 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
3005 ndev->reset_workfn = nvme_reset_failed_dev;
3006 queue_work(nvme_workq, &ndev->reset_work);
3012 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
3014 static const struct pci_error_handlers nvme_err_handler = {
3015 .error_detected = nvme_error_detected,
3016 .mmio_enabled = nvme_dump_registers,
3017 .link_reset = nvme_link_reset,
3018 .slot_reset = nvme_slot_reset,
3019 .resume = nvme_error_resume,
3020 .reset_notify = nvme_reset_notify,
3023 /* Move to pci_ids.h later */
3024 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
3026 static const struct pci_device_id nvme_id_table[] = {
3027 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3030 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3032 static struct pci_driver nvme_driver = {
3034 .id_table = nvme_id_table,
3035 .probe = nvme_probe,
3036 .remove = nvme_remove,
3037 .shutdown = nvme_shutdown,
3039 .pm = &nvme_dev_pm_ops,
3041 .err_handler = &nvme_err_handler,
3044 static int __init nvme_init(void)
3048 init_waitqueue_head(&nvme_kthread_wait);
3050 nvme_workq = create_singlethread_workqueue("nvme");
3054 result = register_blkdev(nvme_major, "nvme");
3057 else if (result > 0)
3058 nvme_major = result;
3060 nvme_nb.notifier_call = &nvme_cpu_notify;
3061 result = register_hotcpu_notifier(&nvme_nb);
3063 goto unregister_blkdev;
3065 result = pci_register_driver(&nvme_driver);
3067 goto unregister_hotcpu;
3071 unregister_hotcpu_notifier(&nvme_nb);
3073 unregister_blkdev(nvme_major, "nvme");
3075 destroy_workqueue(nvme_workq);
3079 static void __exit nvme_exit(void)
3081 pci_unregister_driver(&nvme_driver);
3082 unregister_hotcpu_notifier(&nvme_nb);
3083 unregister_blkdev(nvme_major, "nvme");
3084 destroy_workqueue(nvme_workq);
3085 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3089 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3090 MODULE_LICENSE("GPL");
3091 MODULE_VERSION("0.9");
3092 module_init(nvme_init);
3093 module_exit(nvme_exit);