2 * NVM Express device driver
3 * Copyright (c) 2011, 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
14 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
19 #include <linux/nvme.h>
20 #include <linux/bio.h>
21 #include <linux/bitops.h>
22 #include <linux/blkdev.h>
23 #include <linux/delay.h>
24 #include <linux/errno.h>
26 #include <linux/genhd.h>
27 #include <linux/idr.h>
28 #include <linux/init.h>
29 #include <linux/interrupt.h>
31 #include <linux/kdev_t.h>
32 #include <linux/kthread.h>
33 #include <linux/kernel.h>
35 #include <linux/module.h>
36 #include <linux/moduleparam.h>
37 #include <linux/pci.h>
38 #include <linux/poison.h>
39 #include <linux/ptrace.h>
40 #include <linux/sched.h>
41 #include <linux/slab.h>
42 #include <linux/types.h>
44 #include <asm-generic/io-64-nonatomic-lo-hi.h>
46 #define NVME_Q_DEPTH 1024
47 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
48 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
49 #define NVME_MINORS 64
50 #define ADMIN_TIMEOUT (60 * HZ)
52 static int nvme_major;
53 module_param(nvme_major, int, 0);
55 static int use_threaded_interrupts;
56 module_param(use_threaded_interrupts, int, 0);
58 static DEFINE_SPINLOCK(dev_list_lock);
59 static LIST_HEAD(dev_list);
60 static struct task_struct *nvme_thread;
61 static struct workqueue_struct *nvme_workq;
64 * An NVM Express queue. Each device has at least two (one for admin
65 * commands and one for I/O commands).
68 struct device *q_dmadev;
71 struct nvme_command *sq_cmds;
72 volatile struct nvme_completion *cqes;
73 dma_addr_t sq_dma_addr;
74 dma_addr_t cq_dma_addr;
75 wait_queue_head_t sq_full;
76 wait_queue_t sq_cong_wait;
77 struct bio_list sq_cong;
87 unsigned long cmdid_data[];
91 * Check we didin't inadvertently grow the command struct
93 static inline void _nvme_check_size(void)
95 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
96 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
97 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
98 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
99 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
100 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
101 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
102 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
103 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
104 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
105 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
108 typedef void (*nvme_completion_fn)(struct nvme_dev *, void *,
109 struct nvme_completion *);
111 struct nvme_cmd_info {
112 nvme_completion_fn fn;
114 unsigned long timeout;
117 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
119 return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
122 static unsigned nvme_queue_extra(int depth)
124 return DIV_ROUND_UP(depth, 8) + (depth * sizeof(struct nvme_cmd_info));
128 * alloc_cmdid() - Allocate a Command ID
129 * @nvmeq: The queue that will be used for this command
130 * @ctx: A pointer that will be passed to the handler
131 * @handler: The function to call on completion
133 * Allocate a Command ID for a queue. The data passed in will
134 * be passed to the completion handler. This is implemented by using
135 * the bottom two bits of the ctx pointer to store the handler ID.
136 * Passing in a pointer that's not 4-byte aligned will cause a BUG.
137 * We can change this if it becomes a problem.
139 * May be called with local interrupts disabled and the q_lock held,
140 * or with interrupts enabled and no locks held.
142 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx,
143 nvme_completion_fn handler, unsigned timeout)
145 int depth = nvmeq->q_depth - 1;
146 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
150 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
153 } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
155 info[cmdid].fn = handler;
156 info[cmdid].ctx = ctx;
157 info[cmdid].timeout = jiffies + timeout;
161 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
162 nvme_completion_fn handler, unsigned timeout)
165 wait_event_killable(nvmeq->sq_full,
166 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
167 return (cmdid < 0) ? -EINTR : cmdid;
170 /* Special values must be less than 0x1000 */
171 #define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
172 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
173 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
174 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
175 #define CMD_CTX_FLUSH (0x318 + CMD_CTX_BASE)
177 static void special_completion(struct nvme_dev *dev, void *ctx,
178 struct nvme_completion *cqe)
180 if (ctx == CMD_CTX_CANCELLED)
182 if (ctx == CMD_CTX_FLUSH)
184 if (ctx == CMD_CTX_COMPLETED) {
185 dev_warn(&dev->pci_dev->dev,
186 "completed id %d twice on queue %d\n",
187 cqe->command_id, le16_to_cpup(&cqe->sq_id));
190 if (ctx == CMD_CTX_INVALID) {
191 dev_warn(&dev->pci_dev->dev,
192 "invalid id %d completed on queue %d\n",
193 cqe->command_id, le16_to_cpup(&cqe->sq_id));
197 dev_warn(&dev->pci_dev->dev, "Unknown special completion %p\n", ctx);
201 * Called with local interrupts disabled and the q_lock held. May not sleep.
203 static void *free_cmdid(struct nvme_queue *nvmeq, int cmdid,
204 nvme_completion_fn *fn)
207 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
209 if (cmdid >= nvmeq->q_depth) {
210 *fn = special_completion;
211 return CMD_CTX_INVALID;
214 *fn = info[cmdid].fn;
215 ctx = info[cmdid].ctx;
216 info[cmdid].fn = special_completion;
217 info[cmdid].ctx = CMD_CTX_COMPLETED;
218 clear_bit(cmdid, nvmeq->cmdid_data);
219 wake_up(&nvmeq->sq_full);
223 static void *cancel_cmdid(struct nvme_queue *nvmeq, int cmdid,
224 nvme_completion_fn *fn)
227 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
229 *fn = info[cmdid].fn;
230 ctx = info[cmdid].ctx;
231 info[cmdid].fn = special_completion;
232 info[cmdid].ctx = CMD_CTX_CANCELLED;
236 struct nvme_queue *get_nvmeq(struct nvme_dev *dev)
238 return dev->queues[get_cpu() + 1];
241 void put_nvmeq(struct nvme_queue *nvmeq)
247 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
248 * @nvmeq: The queue to use
249 * @cmd: The command to send
251 * Safe to use from interrupt context
253 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
257 spin_lock_irqsave(&nvmeq->q_lock, flags);
258 tail = nvmeq->sq_tail;
259 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
260 if (++tail == nvmeq->q_depth)
262 writel(tail, nvmeq->q_db);
263 nvmeq->sq_tail = tail;
264 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
269 static __le64 **iod_list(struct nvme_iod *iod)
271 return ((void *)iod) + iod->offset;
275 * Will slightly overestimate the number of pages needed. This is OK
276 * as it only leads to a small amount of wasted memory for the lifetime of
279 static int nvme_npages(unsigned size)
281 unsigned nprps = DIV_ROUND_UP(size + PAGE_SIZE, PAGE_SIZE);
282 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
285 static struct nvme_iod *
286 nvme_alloc_iod(unsigned nseg, unsigned nbytes, gfp_t gfp)
288 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
289 sizeof(__le64 *) * nvme_npages(nbytes) +
290 sizeof(struct scatterlist) * nseg, gfp);
293 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
295 iod->length = nbytes;
297 iod->start_time = jiffies;
303 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
305 const int last_prp = PAGE_SIZE / 8 - 1;
307 __le64 **list = iod_list(iod);
308 dma_addr_t prp_dma = iod->first_dma;
310 if (iod->npages == 0)
311 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
312 for (i = 0; i < iod->npages; i++) {
313 __le64 *prp_list = list[i];
314 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
315 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
316 prp_dma = next_prp_dma;
321 static void nvme_start_io_acct(struct bio *bio)
323 struct gendisk *disk = bio->bi_bdev->bd_disk;
324 const int rw = bio_data_dir(bio);
325 int cpu = part_stat_lock();
326 part_round_stats(cpu, &disk->part0);
327 part_stat_inc(cpu, &disk->part0, ios[rw]);
328 part_stat_add(cpu, &disk->part0, sectors[rw], bio_sectors(bio));
329 part_inc_in_flight(&disk->part0, rw);
333 static void nvme_end_io_acct(struct bio *bio, unsigned long start_time)
335 struct gendisk *disk = bio->bi_bdev->bd_disk;
336 const int rw = bio_data_dir(bio);
337 unsigned long duration = jiffies - start_time;
338 int cpu = part_stat_lock();
339 part_stat_add(cpu, &disk->part0, ticks[rw], duration);
340 part_round_stats(cpu, &disk->part0);
341 part_dec_in_flight(&disk->part0, rw);
345 static void bio_completion(struct nvme_dev *dev, void *ctx,
346 struct nvme_completion *cqe)
348 struct nvme_iod *iod = ctx;
349 struct bio *bio = iod->private;
350 u16 status = le16_to_cpup(&cqe->status) >> 1;
353 dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
354 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
355 nvme_end_io_acct(bio, iod->start_time);
357 nvme_free_iod(dev, iod);
359 bio_endio(bio, -EIO);
364 /* length is in bytes. gfp flags indicates whether we may sleep. */
365 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_common_command *cmd,
366 struct nvme_iod *iod, int total_len, gfp_t gfp)
368 struct dma_pool *pool;
369 int length = total_len;
370 struct scatterlist *sg = iod->sg;
371 int dma_len = sg_dma_len(sg);
372 u64 dma_addr = sg_dma_address(sg);
373 int offset = offset_in_page(dma_addr);
375 __le64 **list = iod_list(iod);
379 cmd->prp1 = cpu_to_le64(dma_addr);
380 length -= (PAGE_SIZE - offset);
384 dma_len -= (PAGE_SIZE - offset);
386 dma_addr += (PAGE_SIZE - offset);
389 dma_addr = sg_dma_address(sg);
390 dma_len = sg_dma_len(sg);
393 if (length <= PAGE_SIZE) {
394 cmd->prp2 = cpu_to_le64(dma_addr);
398 nprps = DIV_ROUND_UP(length, PAGE_SIZE);
399 if (nprps <= (256 / 8)) {
400 pool = dev->prp_small_pool;
403 pool = dev->prp_page_pool;
407 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
409 cmd->prp2 = cpu_to_le64(dma_addr);
411 return (total_len - length) + PAGE_SIZE;
414 iod->first_dma = prp_dma;
415 cmd->prp2 = cpu_to_le64(prp_dma);
418 if (i == PAGE_SIZE / 8) {
419 __le64 *old_prp_list = prp_list;
420 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
422 return total_len - length;
423 list[iod->npages++] = prp_list;
424 prp_list[0] = old_prp_list[i - 1];
425 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
428 prp_list[i++] = cpu_to_le64(dma_addr);
429 dma_len -= PAGE_SIZE;
430 dma_addr += PAGE_SIZE;
438 dma_addr = sg_dma_address(sg);
439 dma_len = sg_dma_len(sg);
445 struct nvme_bio_pair {
446 struct bio b1, b2, *parent;
447 struct bio_vec *bv1, *bv2;
452 static void nvme_bio_pair_endio(struct bio *bio, int err)
454 struct nvme_bio_pair *bp = bio->bi_private;
459 if (atomic_dec_and_test(&bp->cnt)) {
460 bio_endio(bp->parent, bp->err);
467 static struct nvme_bio_pair *nvme_bio_split(struct bio *bio, int idx,
470 struct nvme_bio_pair *bp;
472 BUG_ON(len > bio->bi_size);
473 BUG_ON(idx > bio->bi_vcnt);
475 bp = kmalloc(sizeof(*bp), GFP_ATOMIC);
483 bp->b1.bi_size = len;
484 bp->b2.bi_size -= len;
485 bp->b1.bi_vcnt = idx;
487 bp->b2.bi_sector += len >> 9;
490 bp->bv1 = kmalloc(bio->bi_max_vecs * sizeof(struct bio_vec),
495 bp->bv2 = kmalloc(bio->bi_max_vecs * sizeof(struct bio_vec),
500 memcpy(bp->bv1, bio->bi_io_vec,
501 bio->bi_max_vecs * sizeof(struct bio_vec));
502 memcpy(bp->bv2, bio->bi_io_vec,
503 bio->bi_max_vecs * sizeof(struct bio_vec));
505 bp->b1.bi_io_vec = bp->bv1;
506 bp->b2.bi_io_vec = bp->bv2;
507 bp->b2.bi_io_vec[idx].bv_offset += offset;
508 bp->b2.bi_io_vec[idx].bv_len -= offset;
509 bp->b1.bi_io_vec[idx].bv_len = offset;
512 bp->bv1 = bp->bv2 = NULL;
514 bp->b1.bi_private = bp;
515 bp->b2.bi_private = bp;
517 bp->b1.bi_end_io = nvme_bio_pair_endio;
518 bp->b2.bi_end_io = nvme_bio_pair_endio;
521 atomic_set(&bp->cnt, 2);
532 static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
533 int idx, int len, int offset)
535 struct nvme_bio_pair *bp = nvme_bio_split(bio, idx, len, offset);
539 if (bio_list_empty(&nvmeq->sq_cong))
540 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
541 bio_list_add(&nvmeq->sq_cong, &bp->b1);
542 bio_list_add(&nvmeq->sq_cong, &bp->b2);
547 /* NVMe scatterlists require no holes in the virtual address */
548 #define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2) ((vec2)->bv_offset || \
549 (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
551 static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
552 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
554 struct bio_vec *bvec, *bvprv = NULL;
555 struct scatterlist *sg = NULL;
556 int i, length = 0, nsegs = 0, split_len = bio->bi_size;
558 if (nvmeq->dev->stripe_size)
559 split_len = nvmeq->dev->stripe_size -
560 ((bio->bi_sector << 9) & (nvmeq->dev->stripe_size - 1));
562 sg_init_table(iod->sg, psegs);
563 bio_for_each_segment(bvec, bio, i) {
564 if (bvprv && BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) {
565 sg->length += bvec->bv_len;
567 if (bvprv && BIOVEC_NOT_VIRT_MERGEABLE(bvprv, bvec))
568 return nvme_split_and_submit(bio, nvmeq, i,
571 sg = sg ? sg + 1 : iod->sg;
572 sg_set_page(sg, bvec->bv_page, bvec->bv_len,
577 if (split_len - length < bvec->bv_len)
578 return nvme_split_and_submit(bio, nvmeq, i, split_len,
580 length += bvec->bv_len;
585 if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
588 BUG_ON(length != bio->bi_size);
593 * We reuse the small pool to allocate the 16-byte range here as it is not
594 * worth having a special pool for these or additional cases to handle freeing
597 static int nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
598 struct bio *bio, struct nvme_iod *iod, int cmdid)
600 struct nvme_dsm_range *range;
601 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
603 range = dma_pool_alloc(nvmeq->dev->prp_small_pool, GFP_ATOMIC,
608 iod_list(iod)[0] = (__le64 *)range;
611 range->cattr = cpu_to_le32(0);
612 range->nlb = cpu_to_le32(bio->bi_size >> ns->lba_shift);
613 range->slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_sector));
615 memset(cmnd, 0, sizeof(*cmnd));
616 cmnd->dsm.opcode = nvme_cmd_dsm;
617 cmnd->dsm.command_id = cmdid;
618 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
619 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
621 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
623 if (++nvmeq->sq_tail == nvmeq->q_depth)
625 writel(nvmeq->sq_tail, nvmeq->q_db);
630 static int nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
633 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
635 memset(cmnd, 0, sizeof(*cmnd));
636 cmnd->common.opcode = nvme_cmd_flush;
637 cmnd->common.command_id = cmdid;
638 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
640 if (++nvmeq->sq_tail == nvmeq->q_depth)
642 writel(nvmeq->sq_tail, nvmeq->q_db);
647 int nvme_submit_flush_data(struct nvme_queue *nvmeq, struct nvme_ns *ns)
649 int cmdid = alloc_cmdid(nvmeq, (void *)CMD_CTX_FLUSH,
650 special_completion, NVME_IO_TIMEOUT);
651 if (unlikely(cmdid < 0))
654 return nvme_submit_flush(nvmeq, ns, cmdid);
658 * Called with local interrupts disabled and the q_lock held. May not sleep.
660 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
663 struct nvme_command *cmnd;
664 struct nvme_iod *iod;
665 enum dma_data_direction dma_dir;
666 int cmdid, length, result;
669 int psegs = bio_phys_segments(ns->queue, bio);
671 if ((bio->bi_rw & REQ_FLUSH) && psegs) {
672 result = nvme_submit_flush_data(nvmeq, ns);
678 iod = nvme_alloc_iod(psegs, bio->bi_size, GFP_ATOMIC);
684 cmdid = alloc_cmdid(nvmeq, iod, bio_completion, NVME_IO_TIMEOUT);
685 if (unlikely(cmdid < 0))
688 if (bio->bi_rw & REQ_DISCARD) {
689 result = nvme_submit_discard(nvmeq, ns, bio, iod, cmdid);
694 if ((bio->bi_rw & REQ_FLUSH) && !psegs)
695 return nvme_submit_flush(nvmeq, ns, cmdid);
698 if (bio->bi_rw & REQ_FUA)
699 control |= NVME_RW_FUA;
700 if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
701 control |= NVME_RW_LR;
704 if (bio->bi_rw & REQ_RAHEAD)
705 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
707 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
709 memset(cmnd, 0, sizeof(*cmnd));
710 if (bio_data_dir(bio)) {
711 cmnd->rw.opcode = nvme_cmd_write;
712 dma_dir = DMA_TO_DEVICE;
714 cmnd->rw.opcode = nvme_cmd_read;
715 dma_dir = DMA_FROM_DEVICE;
718 result = nvme_map_bio(nvmeq, iod, bio, dma_dir, psegs);
723 cmnd->rw.command_id = cmdid;
724 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
725 length = nvme_setup_prps(nvmeq->dev, &cmnd->common, iod, length,
727 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_sector));
728 cmnd->rw.length = cpu_to_le16((length >> ns->lba_shift) - 1);
729 cmnd->rw.control = cpu_to_le16(control);
730 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
732 nvme_start_io_acct(bio);
733 if (++nvmeq->sq_tail == nvmeq->q_depth)
735 writel(nvmeq->sq_tail, nvmeq->q_db);
740 free_cmdid(nvmeq, cmdid, NULL);
742 nvme_free_iod(nvmeq->dev, iod);
747 static int nvme_process_cq(struct nvme_queue *nvmeq)
751 head = nvmeq->cq_head;
752 phase = nvmeq->cq_phase;
756 nvme_completion_fn fn;
757 struct nvme_completion cqe = nvmeq->cqes[head];
758 if ((le16_to_cpu(cqe.status) & 1) != phase)
760 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
761 if (++head == nvmeq->q_depth) {
766 ctx = free_cmdid(nvmeq, cqe.command_id, &fn);
767 fn(nvmeq->dev, ctx, &cqe);
770 /* If the controller ignores the cq head doorbell and continuously
771 * writes to the queue, it is theoretically possible to wrap around
772 * the queue twice and mistakenly return IRQ_NONE. Linux only
773 * requires that 0.1% of your interrupts are handled, so this isn't
776 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
779 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
780 nvmeq->cq_head = head;
781 nvmeq->cq_phase = phase;
787 static void nvme_make_request(struct request_queue *q, struct bio *bio)
789 struct nvme_ns *ns = q->queuedata;
790 struct nvme_queue *nvmeq = get_nvmeq(ns->dev);
795 bio_endio(bio, -EIO);
799 spin_lock_irq(&nvmeq->q_lock);
800 if (!nvmeq->q_suspended && bio_list_empty(&nvmeq->sq_cong))
801 result = nvme_submit_bio_queue(nvmeq, ns, bio);
802 if (unlikely(result)) {
803 if (bio_list_empty(&nvmeq->sq_cong))
804 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
805 bio_list_add(&nvmeq->sq_cong, bio);
808 nvme_process_cq(nvmeq);
809 spin_unlock_irq(&nvmeq->q_lock);
813 static irqreturn_t nvme_irq(int irq, void *data)
816 struct nvme_queue *nvmeq = data;
817 spin_lock(&nvmeq->q_lock);
818 nvme_process_cq(nvmeq);
819 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
821 spin_unlock(&nvmeq->q_lock);
825 static irqreturn_t nvme_irq_check(int irq, void *data)
827 struct nvme_queue *nvmeq = data;
828 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
829 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
831 return IRQ_WAKE_THREAD;
834 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
836 spin_lock_irq(&nvmeq->q_lock);
837 cancel_cmdid(nvmeq, cmdid, NULL);
838 spin_unlock_irq(&nvmeq->q_lock);
841 struct sync_cmd_info {
842 struct task_struct *task;
847 static void sync_completion(struct nvme_dev *dev, void *ctx,
848 struct nvme_completion *cqe)
850 struct sync_cmd_info *cmdinfo = ctx;
851 cmdinfo->result = le32_to_cpup(&cqe->result);
852 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
853 wake_up_process(cmdinfo->task);
857 * Returns 0 on success. If the result is negative, it's a Linux error code;
858 * if the result is positive, it's an NVM Express status code
860 int nvme_submit_sync_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
861 u32 *result, unsigned timeout)
864 struct sync_cmd_info cmdinfo;
866 cmdinfo.task = current;
867 cmdinfo.status = -EINTR;
869 cmdid = alloc_cmdid_killable(nvmeq, &cmdinfo, sync_completion,
873 cmd->common.command_id = cmdid;
875 set_current_state(TASK_KILLABLE);
876 nvme_submit_cmd(nvmeq, cmd);
877 schedule_timeout(timeout);
879 if (cmdinfo.status == -EINTR) {
880 nvme_abort_command(nvmeq, cmdid);
885 *result = cmdinfo.result;
887 return cmdinfo.status;
890 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
893 return nvme_submit_sync_cmd(dev->queues[0], cmd, result, ADMIN_TIMEOUT);
896 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
899 struct nvme_command c;
901 memset(&c, 0, sizeof(c));
902 c.delete_queue.opcode = opcode;
903 c.delete_queue.qid = cpu_to_le16(id);
905 status = nvme_submit_admin_cmd(dev, &c, NULL);
911 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
912 struct nvme_queue *nvmeq)
915 struct nvme_command c;
916 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
918 memset(&c, 0, sizeof(c));
919 c.create_cq.opcode = nvme_admin_create_cq;
920 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
921 c.create_cq.cqid = cpu_to_le16(qid);
922 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
923 c.create_cq.cq_flags = cpu_to_le16(flags);
924 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
926 status = nvme_submit_admin_cmd(dev, &c, NULL);
932 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
933 struct nvme_queue *nvmeq)
936 struct nvme_command c;
937 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
939 memset(&c, 0, sizeof(c));
940 c.create_sq.opcode = nvme_admin_create_sq;
941 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
942 c.create_sq.sqid = cpu_to_le16(qid);
943 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
944 c.create_sq.sq_flags = cpu_to_le16(flags);
945 c.create_sq.cqid = cpu_to_le16(qid);
947 status = nvme_submit_admin_cmd(dev, &c, NULL);
953 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
955 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
958 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
960 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
963 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
966 struct nvme_command c;
968 memset(&c, 0, sizeof(c));
969 c.identify.opcode = nvme_admin_identify;
970 c.identify.nsid = cpu_to_le32(nsid);
971 c.identify.prp1 = cpu_to_le64(dma_addr);
972 c.identify.cns = cpu_to_le32(cns);
974 return nvme_submit_admin_cmd(dev, &c, NULL);
977 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
978 dma_addr_t dma_addr, u32 *result)
980 struct nvme_command c;
982 memset(&c, 0, sizeof(c));
983 c.features.opcode = nvme_admin_get_features;
984 c.features.nsid = cpu_to_le32(nsid);
985 c.features.prp1 = cpu_to_le64(dma_addr);
986 c.features.fid = cpu_to_le32(fid);
988 return nvme_submit_admin_cmd(dev, &c, result);
991 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
992 dma_addr_t dma_addr, u32 *result)
994 struct nvme_command c;
996 memset(&c, 0, sizeof(c));
997 c.features.opcode = nvme_admin_set_features;
998 c.features.prp1 = cpu_to_le64(dma_addr);
999 c.features.fid = cpu_to_le32(fid);
1000 c.features.dword11 = cpu_to_le32(dword11);
1002 return nvme_submit_admin_cmd(dev, &c, result);
1006 * nvme_cancel_ios - Cancel outstanding I/Os
1007 * @queue: The queue to cancel I/Os on
1008 * @timeout: True to only cancel I/Os which have timed out
1010 static void nvme_cancel_ios(struct nvme_queue *nvmeq, bool timeout)
1012 int depth = nvmeq->q_depth - 1;
1013 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1014 unsigned long now = jiffies;
1017 for_each_set_bit(cmdid, nvmeq->cmdid_data, depth) {
1019 nvme_completion_fn fn;
1020 static struct nvme_completion cqe = {
1021 .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1024 if (timeout && !time_after(now, info[cmdid].timeout))
1026 if (info[cmdid].ctx == CMD_CTX_CANCELLED)
1028 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d\n", cmdid);
1029 ctx = cancel_cmdid(nvmeq, cmdid, &fn);
1030 fn(nvmeq->dev, ctx, &cqe);
1034 static void nvme_free_queue(struct nvme_queue *nvmeq)
1036 spin_lock_irq(&nvmeq->q_lock);
1037 while (bio_list_peek(&nvmeq->sq_cong)) {
1038 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1039 bio_endio(bio, -EIO);
1041 spin_unlock_irq(&nvmeq->q_lock);
1043 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1044 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1045 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1046 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1050 static void nvme_free_queues(struct nvme_dev *dev)
1054 for (i = dev->queue_count - 1; i >= 0; i--) {
1055 nvme_free_queue(dev->queues[i]);
1057 dev->queues[i] = NULL;
1061 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1063 struct nvme_queue *nvmeq = dev->queues[qid];
1064 int vector = dev->entry[nvmeq->cq_vector].vector;
1066 spin_lock_irq(&nvmeq->q_lock);
1067 if (nvmeq->q_suspended) {
1068 spin_unlock_irq(&nvmeq->q_lock);
1071 nvmeq->q_suspended = 1;
1072 spin_unlock_irq(&nvmeq->q_lock);
1074 irq_set_affinity_hint(vector, NULL);
1075 free_irq(vector, nvmeq);
1077 /* Don't tell the adapter to delete the admin queue */
1079 adapter_delete_sq(dev, qid);
1080 adapter_delete_cq(dev, qid);
1083 spin_lock_irq(&nvmeq->q_lock);
1084 nvme_process_cq(nvmeq);
1085 nvme_cancel_ios(nvmeq, false);
1086 spin_unlock_irq(&nvmeq->q_lock);
1089 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1090 int depth, int vector)
1092 struct device *dmadev = &dev->pci_dev->dev;
1093 unsigned extra = nvme_queue_extra(depth);
1094 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
1098 nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
1099 &nvmeq->cq_dma_addr, GFP_KERNEL);
1102 memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
1104 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1105 &nvmeq->sq_dma_addr, GFP_KERNEL);
1106 if (!nvmeq->sq_cmds)
1109 nvmeq->q_dmadev = dmadev;
1111 spin_lock_init(&nvmeq->q_lock);
1113 nvmeq->cq_phase = 1;
1114 init_waitqueue_head(&nvmeq->sq_full);
1115 init_waitqueue_entry(&nvmeq->sq_cong_wait, nvme_thread);
1116 bio_list_init(&nvmeq->sq_cong);
1117 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1118 nvmeq->q_depth = depth;
1119 nvmeq->cq_vector = vector;
1120 nvmeq->q_suspended = 1;
1126 dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1127 nvmeq->cq_dma_addr);
1133 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1136 if (use_threaded_interrupts)
1137 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1138 nvme_irq_check, nvme_irq, IRQF_SHARED,
1140 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1141 IRQF_SHARED, name, nvmeq);
1144 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1146 struct nvme_dev *dev = nvmeq->dev;
1147 unsigned extra = nvme_queue_extra(nvmeq->q_depth);
1151 nvmeq->cq_phase = 1;
1152 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1153 memset(nvmeq->cmdid_data, 0, extra);
1154 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1155 nvme_cancel_ios(nvmeq, false);
1156 nvmeq->q_suspended = 0;
1159 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1161 struct nvme_dev *dev = nvmeq->dev;
1164 result = adapter_alloc_cq(dev, qid, nvmeq);
1168 result = adapter_alloc_sq(dev, qid, nvmeq);
1172 result = queue_request_irq(dev, nvmeq, "nvme");
1176 spin_lock_irq(&nvmeq->q_lock);
1177 nvme_init_queue(nvmeq, qid);
1178 spin_unlock_irq(&nvmeq->q_lock);
1183 adapter_delete_sq(dev, qid);
1185 adapter_delete_cq(dev, qid);
1189 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1191 unsigned long timeout;
1192 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1194 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1196 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1198 if (fatal_signal_pending(current))
1200 if (time_after(jiffies, timeout)) {
1201 dev_err(&dev->pci_dev->dev,
1202 "Device not ready; aborting initialisation\n");
1211 * If the device has been passed off to us in an enabled state, just clear
1212 * the enabled bit. The spec says we should set the 'shutdown notification
1213 * bits', but doing so may cause the device to complete commands to the
1214 * admin queue ... and we don't know what memory that might be pointing at!
1216 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1218 u32 cc = readl(&dev->bar->cc);
1220 if (cc & NVME_CC_ENABLE)
1221 writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
1222 return nvme_wait_ready(dev, cap, false);
1225 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1227 return nvme_wait_ready(dev, cap, true);
1230 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1232 unsigned long timeout;
1235 cc = (readl(&dev->bar->cc) & ~NVME_CC_SHN_MASK) | NVME_CC_SHN_NORMAL;
1236 writel(cc, &dev->bar->cc);
1238 timeout = 2 * HZ + jiffies;
1239 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1240 NVME_CSTS_SHST_CMPLT) {
1242 if (fatal_signal_pending(current))
1244 if (time_after(jiffies, timeout)) {
1245 dev_err(&dev->pci_dev->dev,
1246 "Device shutdown incomplete; abort shutdown\n");
1254 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1258 u64 cap = readq(&dev->bar->cap);
1259 struct nvme_queue *nvmeq;
1261 result = nvme_disable_ctrl(dev, cap);
1265 nvmeq = dev->queues[0];
1267 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
1270 dev->queues[0] = nvmeq;
1273 aqa = nvmeq->q_depth - 1;
1276 dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
1277 dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
1278 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1279 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1281 writel(aqa, &dev->bar->aqa);
1282 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1283 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1284 writel(dev->ctrl_config, &dev->bar->cc);
1286 result = nvme_enable_ctrl(dev, cap);
1290 result = queue_request_irq(dev, nvmeq, "nvme admin");
1294 spin_lock_irq(&nvmeq->q_lock);
1295 nvme_init_queue(nvmeq, 0);
1296 spin_unlock_irq(&nvmeq->q_lock);
1300 struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1301 unsigned long addr, unsigned length)
1303 int i, err, count, nents, offset;
1304 struct scatterlist *sg;
1305 struct page **pages;
1306 struct nvme_iod *iod;
1309 return ERR_PTR(-EINVAL);
1310 if (!length || length > INT_MAX - PAGE_SIZE)
1311 return ERR_PTR(-EINVAL);
1313 offset = offset_in_page(addr);
1314 count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1315 pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1317 return ERR_PTR(-ENOMEM);
1319 err = get_user_pages_fast(addr, count, 1, pages);
1326 iod = nvme_alloc_iod(count, length, GFP_KERNEL);
1328 sg_init_table(sg, count);
1329 for (i = 0; i < count; i++) {
1330 sg_set_page(&sg[i], pages[i],
1331 min_t(unsigned, length, PAGE_SIZE - offset),
1333 length -= (PAGE_SIZE - offset);
1336 sg_mark_end(&sg[i - 1]);
1340 nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1341 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1351 for (i = 0; i < count; i++)
1354 return ERR_PTR(err);
1357 void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1358 struct nvme_iod *iod)
1362 dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1363 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1365 for (i = 0; i < iod->nents; i++)
1366 put_page(sg_page(&iod->sg[i]));
1369 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1371 struct nvme_dev *dev = ns->dev;
1372 struct nvme_queue *nvmeq;
1373 struct nvme_user_io io;
1374 struct nvme_command c;
1375 unsigned length, meta_len;
1377 struct nvme_iod *iod, *meta_iod = NULL;
1378 dma_addr_t meta_dma_addr;
1379 void *meta, *uninitialized_var(meta_mem);
1381 if (copy_from_user(&io, uio, sizeof(io)))
1383 length = (io.nblocks + 1) << ns->lba_shift;
1384 meta_len = (io.nblocks + 1) * ns->ms;
1386 if (meta_len && ((io.metadata & 3) || !io.metadata))
1389 switch (io.opcode) {
1390 case nvme_cmd_write:
1392 case nvme_cmd_compare:
1393 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1400 return PTR_ERR(iod);
1402 memset(&c, 0, sizeof(c));
1403 c.rw.opcode = io.opcode;
1404 c.rw.flags = io.flags;
1405 c.rw.nsid = cpu_to_le32(ns->ns_id);
1406 c.rw.slba = cpu_to_le64(io.slba);
1407 c.rw.length = cpu_to_le16(io.nblocks);
1408 c.rw.control = cpu_to_le16(io.control);
1409 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1410 c.rw.reftag = cpu_to_le32(io.reftag);
1411 c.rw.apptag = cpu_to_le16(io.apptag);
1412 c.rw.appmask = cpu_to_le16(io.appmask);
1415 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1417 if (IS_ERR(meta_iod)) {
1418 status = PTR_ERR(meta_iod);
1423 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1424 &meta_dma_addr, GFP_KERNEL);
1430 if (io.opcode & 1) {
1431 int meta_offset = 0;
1433 for (i = 0; i < meta_iod->nents; i++) {
1434 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1435 meta_iod->sg[i].offset;
1436 memcpy(meta_mem + meta_offset, meta,
1437 meta_iod->sg[i].length);
1438 kunmap_atomic(meta);
1439 meta_offset += meta_iod->sg[i].length;
1443 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1446 length = nvme_setup_prps(dev, &c.common, iod, length, GFP_KERNEL);
1448 nvmeq = get_nvmeq(dev);
1450 * Since nvme_submit_sync_cmd sleeps, we can't keep preemption
1451 * disabled. We may be preempted at any point, and be rescheduled
1452 * to a different CPU. That will cause cacheline bouncing, but no
1453 * additional races since q_lock already protects against other CPUs.
1456 if (length != (io.nblocks + 1) << ns->lba_shift)
1458 else if (!nvmeq || nvmeq->q_suspended)
1461 status = nvme_submit_sync_cmd(nvmeq, &c, NULL, NVME_IO_TIMEOUT);
1464 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1465 int meta_offset = 0;
1467 for (i = 0; i < meta_iod->nents; i++) {
1468 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1469 meta_iod->sg[i].offset;
1470 memcpy(meta, meta_mem + meta_offset,
1471 meta_iod->sg[i].length);
1472 kunmap_atomic(meta);
1473 meta_offset += meta_iod->sg[i].length;
1477 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1482 nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1483 nvme_free_iod(dev, iod);
1486 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1487 nvme_free_iod(dev, meta_iod);
1493 static int nvme_user_admin_cmd(struct nvme_dev *dev,
1494 struct nvme_admin_cmd __user *ucmd)
1496 struct nvme_admin_cmd cmd;
1497 struct nvme_command c;
1499 struct nvme_iod *uninitialized_var(iod);
1502 if (!capable(CAP_SYS_ADMIN))
1504 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1507 memset(&c, 0, sizeof(c));
1508 c.common.opcode = cmd.opcode;
1509 c.common.flags = cmd.flags;
1510 c.common.nsid = cpu_to_le32(cmd.nsid);
1511 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1512 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1513 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1514 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1515 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1516 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1517 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1518 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1520 length = cmd.data_len;
1522 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1525 return PTR_ERR(iod);
1526 length = nvme_setup_prps(dev, &c.common, iod, length,
1530 timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1532 if (length != cmd.data_len)
1535 status = nvme_submit_sync_cmd(dev->queues[0], &c, &cmd.result,
1539 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1540 nvme_free_iod(dev, iod);
1543 if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1544 sizeof(cmd.result)))
1550 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1553 struct nvme_ns *ns = bdev->bd_disk->private_data;
1557 force_successful_syscall_return();
1559 case NVME_IOCTL_ADMIN_CMD:
1560 return nvme_user_admin_cmd(ns->dev, (void __user *)arg);
1561 case NVME_IOCTL_SUBMIT_IO:
1562 return nvme_submit_io(ns, (void __user *)arg);
1563 case SG_GET_VERSION_NUM:
1564 return nvme_sg_get_version_num((void __user *)arg);
1566 return nvme_sg_io(ns, (void __user *)arg);
1572 #ifdef CONFIG_COMPAT
1573 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1574 unsigned int cmd, unsigned long arg)
1576 struct nvme_ns *ns = bdev->bd_disk->private_data;
1580 return nvme_sg_io32(ns, arg);
1582 return nvme_ioctl(bdev, mode, cmd, arg);
1585 #define nvme_compat_ioctl NULL
1588 static const struct block_device_operations nvme_fops = {
1589 .owner = THIS_MODULE,
1590 .ioctl = nvme_ioctl,
1591 .compat_ioctl = nvme_compat_ioctl,
1594 static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
1596 while (bio_list_peek(&nvmeq->sq_cong)) {
1597 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1598 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
1600 if (bio_list_empty(&nvmeq->sq_cong))
1601 remove_wait_queue(&nvmeq->sq_full,
1602 &nvmeq->sq_cong_wait);
1603 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
1604 if (bio_list_empty(&nvmeq->sq_cong))
1605 add_wait_queue(&nvmeq->sq_full,
1606 &nvmeq->sq_cong_wait);
1607 bio_list_add_head(&nvmeq->sq_cong, bio);
1613 static int nvme_kthread(void *data)
1615 struct nvme_dev *dev;
1617 while (!kthread_should_stop()) {
1618 set_current_state(TASK_INTERRUPTIBLE);
1619 spin_lock(&dev_list_lock);
1620 list_for_each_entry(dev, &dev_list, node) {
1622 for (i = 0; i < dev->queue_count; i++) {
1623 struct nvme_queue *nvmeq = dev->queues[i];
1626 spin_lock_irq(&nvmeq->q_lock);
1627 if (nvmeq->q_suspended)
1629 nvme_process_cq(nvmeq);
1630 nvme_cancel_ios(nvmeq, true);
1631 nvme_resubmit_bios(nvmeq);
1633 spin_unlock_irq(&nvmeq->q_lock);
1636 spin_unlock(&dev_list_lock);
1637 schedule_timeout(round_jiffies_relative(HZ));
1642 static DEFINE_IDA(nvme_index_ida);
1644 static int nvme_get_ns_idx(void)
1649 if (!ida_pre_get(&nvme_index_ida, GFP_KERNEL))
1652 spin_lock(&dev_list_lock);
1653 error = ida_get_new(&nvme_index_ida, &index);
1654 spin_unlock(&dev_list_lock);
1655 } while (error == -EAGAIN);
1662 static void nvme_put_ns_idx(int index)
1664 spin_lock(&dev_list_lock);
1665 ida_remove(&nvme_index_ida, index);
1666 spin_unlock(&dev_list_lock);
1669 static void nvme_config_discard(struct nvme_ns *ns)
1671 u32 logical_block_size = queue_logical_block_size(ns->queue);
1672 ns->queue->limits.discard_zeroes_data = 0;
1673 ns->queue->limits.discard_alignment = logical_block_size;
1674 ns->queue->limits.discard_granularity = logical_block_size;
1675 ns->queue->limits.max_discard_sectors = 0xffffffff;
1676 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1679 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1680 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1683 struct gendisk *disk;
1686 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1689 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1692 ns->queue = blk_alloc_queue(GFP_KERNEL);
1695 ns->queue->queue_flags = QUEUE_FLAG_DEFAULT;
1696 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1697 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1698 blk_queue_make_request(ns->queue, nvme_make_request);
1700 ns->queue->queuedata = ns;
1702 disk = alloc_disk(NVME_MINORS);
1704 goto out_free_queue;
1707 lbaf = id->flbas & 0xf;
1708 ns->lba_shift = id->lbaf[lbaf].ds;
1709 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1710 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1711 if (dev->max_hw_sectors)
1712 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
1714 disk->major = nvme_major;
1715 disk->minors = NVME_MINORS;
1716 disk->first_minor = NVME_MINORS * nvme_get_ns_idx();
1717 disk->fops = &nvme_fops;
1718 disk->private_data = ns;
1719 disk->queue = ns->queue;
1720 disk->driverfs_dev = &dev->pci_dev->dev;
1721 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
1722 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1724 if (dev->oncs & NVME_CTRL_ONCS_DSM)
1725 nvme_config_discard(ns);
1730 blk_cleanup_queue(ns->queue);
1736 static void nvme_ns_free(struct nvme_ns *ns)
1738 int index = ns->disk->first_minor / NVME_MINORS;
1740 nvme_put_ns_idx(index);
1741 blk_cleanup_queue(ns->queue);
1745 static int set_queue_count(struct nvme_dev *dev, int count)
1749 u32 q_count = (count - 1) | ((count - 1) << 16);
1751 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
1754 return status < 0 ? -EIO : -EBUSY;
1755 return min(result & 0xffff, result >> 16) + 1;
1758 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1760 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1763 static int nvme_setup_io_queues(struct nvme_dev *dev)
1765 struct pci_dev *pdev = dev->pci_dev;
1766 int result, cpu, i, vecs, nr_io_queues, size, q_depth;
1768 nr_io_queues = num_online_cpus();
1769 result = set_queue_count(dev, nr_io_queues);
1772 if (result < nr_io_queues)
1773 nr_io_queues = result;
1775 size = db_bar_size(dev, nr_io_queues);
1779 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1782 if (!--nr_io_queues)
1784 size = db_bar_size(dev, nr_io_queues);
1786 dev->dbs = ((void __iomem *)dev->bar) + 4096;
1787 dev->queues[0]->q_db = dev->dbs;
1790 /* Deregister the admin queue's interrupt */
1791 free_irq(dev->entry[0].vector, dev->queues[0]);
1793 vecs = nr_io_queues;
1794 for (i = 0; i < vecs; i++)
1795 dev->entry[i].entry = i;
1797 result = pci_enable_msix(pdev, dev->entry, vecs);
1804 vecs = nr_io_queues;
1808 result = pci_enable_msi_block(pdev, vecs);
1810 for (i = 0; i < vecs; i++)
1811 dev->entry[i].vector = i + pdev->irq;
1813 } else if (result < 0) {
1822 * Should investigate if there's a performance win from allocating
1823 * more queues than interrupt vectors; it might allow the submission
1824 * path to scale better, even if the receive path is limited by the
1825 * number of interrupts.
1827 nr_io_queues = vecs;
1829 result = queue_request_irq(dev, dev->queues[0], "nvme admin");
1831 dev->queues[0]->q_suspended = 1;
1835 /* Free previously allocated queues that are no longer usable */
1836 spin_lock(&dev_list_lock);
1837 for (i = dev->queue_count - 1; i > nr_io_queues; i--) {
1838 struct nvme_queue *nvmeq = dev->queues[i];
1840 spin_lock_irq(&nvmeq->q_lock);
1841 nvme_cancel_ios(nvmeq, false);
1842 spin_unlock_irq(&nvmeq->q_lock);
1844 nvme_free_queue(nvmeq);
1846 dev->queues[i] = NULL;
1848 spin_unlock(&dev_list_lock);
1850 cpu = cpumask_first(cpu_online_mask);
1851 for (i = 0; i < nr_io_queues; i++) {
1852 irq_set_affinity_hint(dev->entry[i].vector, get_cpu_mask(cpu));
1853 cpu = cpumask_next(cpu, cpu_online_mask);
1856 q_depth = min_t(int, NVME_CAP_MQES(readq(&dev->bar->cap)) + 1,
1858 for (i = dev->queue_count - 1; i < nr_io_queues; i++) {
1859 dev->queues[i + 1] = nvme_alloc_queue(dev, i + 1, q_depth, i);
1860 if (!dev->queues[i + 1]) {
1866 for (; i < num_possible_cpus(); i++) {
1867 int target = i % rounddown_pow_of_two(dev->queue_count - 1);
1868 dev->queues[i + 1] = dev->queues[target + 1];
1871 for (i = 1; i < dev->queue_count; i++) {
1872 result = nvme_create_queue(dev->queues[i], i);
1874 for (--i; i > 0; i--)
1875 nvme_disable_queue(dev, i);
1883 nvme_free_queues(dev);
1888 * Return: error value if an error occurred setting up the queues or calling
1889 * Identify Device. 0 if these succeeded, even if adding some of the
1890 * namespaces failed. At the moment, these failures are silent. TBD which
1891 * failures should be reported.
1893 static int nvme_dev_add(struct nvme_dev *dev)
1895 struct pci_dev *pdev = dev->pci_dev;
1899 struct nvme_id_ctrl *ctrl;
1900 struct nvme_id_ns *id_ns;
1902 dma_addr_t dma_addr;
1903 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
1905 mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
1909 res = nvme_identify(dev, 0, 1, dma_addr);
1916 nn = le32_to_cpup(&ctrl->nn);
1917 dev->oncs = le16_to_cpup(&ctrl->oncs);
1918 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
1919 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
1920 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
1922 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
1923 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
1924 (pdev->device == 0x0953) && ctrl->vs[3])
1925 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
1928 for (i = 1; i <= nn; i++) {
1929 res = nvme_identify(dev, i, 0, dma_addr);
1933 if (id_ns->ncap == 0)
1936 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
1937 dma_addr + 4096, NULL);
1939 memset(mem + 4096, 0, 4096);
1941 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
1943 list_add_tail(&ns->list, &dev->namespaces);
1945 list_for_each_entry(ns, &dev->namespaces, list)
1950 dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
1954 static int nvme_dev_map(struct nvme_dev *dev)
1956 int bars, result = -ENOMEM;
1957 struct pci_dev *pdev = dev->pci_dev;
1959 if (pci_enable_device_mem(pdev))
1962 dev->entry[0].vector = pdev->irq;
1963 pci_set_master(pdev);
1964 bars = pci_select_bars(pdev, IORESOURCE_MEM);
1965 if (pci_request_selected_regions(pdev, bars, "nvme"))
1968 if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
1969 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
1972 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1976 dev->db_stride = 1 << NVME_CAP_STRIDE(readq(&dev->bar->cap));
1977 dev->dbs = ((void __iomem *)dev->bar) + 4096;
1982 pci_release_regions(pdev);
1984 pci_disable_device(pdev);
1988 static void nvme_dev_unmap(struct nvme_dev *dev)
1990 if (dev->pci_dev->msi_enabled)
1991 pci_disable_msi(dev->pci_dev);
1992 else if (dev->pci_dev->msix_enabled)
1993 pci_disable_msix(dev->pci_dev);
1998 pci_release_regions(dev->pci_dev);
2001 if (pci_is_enabled(dev->pci_dev))
2002 pci_disable_device(dev->pci_dev);
2005 static void nvme_dev_shutdown(struct nvme_dev *dev)
2009 for (i = dev->queue_count - 1; i >= 0; i--)
2010 nvme_disable_queue(dev, i);
2012 spin_lock(&dev_list_lock);
2013 list_del_init(&dev->node);
2014 spin_unlock(&dev_list_lock);
2017 nvme_shutdown_ctrl(dev);
2018 nvme_dev_unmap(dev);
2021 static void nvme_dev_remove(struct nvme_dev *dev)
2023 struct nvme_ns *ns, *next;
2025 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2026 list_del(&ns->list);
2027 del_gendisk(ns->disk);
2032 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2034 struct device *dmadev = &dev->pci_dev->dev;
2035 dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2036 PAGE_SIZE, PAGE_SIZE, 0);
2037 if (!dev->prp_page_pool)
2040 /* Optimisation for I/Os between 4k and 128k */
2041 dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2043 if (!dev->prp_small_pool) {
2044 dma_pool_destroy(dev->prp_page_pool);
2050 static void nvme_release_prp_pools(struct nvme_dev *dev)
2052 dma_pool_destroy(dev->prp_page_pool);
2053 dma_pool_destroy(dev->prp_small_pool);
2056 static DEFINE_IDA(nvme_instance_ida);
2058 static int nvme_set_instance(struct nvme_dev *dev)
2060 int instance, error;
2063 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2066 spin_lock(&dev_list_lock);
2067 error = ida_get_new(&nvme_instance_ida, &instance);
2068 spin_unlock(&dev_list_lock);
2069 } while (error == -EAGAIN);
2074 dev->instance = instance;
2078 static void nvme_release_instance(struct nvme_dev *dev)
2080 spin_lock(&dev_list_lock);
2081 ida_remove(&nvme_instance_ida, dev->instance);
2082 spin_unlock(&dev_list_lock);
2085 static void nvme_free_dev(struct kref *kref)
2087 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2093 static int nvme_dev_open(struct inode *inode, struct file *f)
2095 struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2097 kref_get(&dev->kref);
2098 f->private_data = dev;
2102 static int nvme_dev_release(struct inode *inode, struct file *f)
2104 struct nvme_dev *dev = f->private_data;
2105 kref_put(&dev->kref, nvme_free_dev);
2109 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2111 struct nvme_dev *dev = f->private_data;
2113 case NVME_IOCTL_ADMIN_CMD:
2114 return nvme_user_admin_cmd(dev, (void __user *)arg);
2120 static const struct file_operations nvme_dev_fops = {
2121 .owner = THIS_MODULE,
2122 .open = nvme_dev_open,
2123 .release = nvme_dev_release,
2124 .unlocked_ioctl = nvme_dev_ioctl,
2125 .compat_ioctl = nvme_dev_ioctl,
2128 static int nvme_dev_start(struct nvme_dev *dev)
2132 result = nvme_dev_map(dev);
2136 result = nvme_configure_admin_queue(dev);
2140 spin_lock(&dev_list_lock);
2141 list_add(&dev->node, &dev_list);
2142 spin_unlock(&dev_list_lock);
2144 result = nvme_setup_io_queues(dev);
2145 if (result && result != -EBUSY)
2151 spin_lock(&dev_list_lock);
2152 list_del_init(&dev->node);
2153 spin_unlock(&dev_list_lock);
2155 nvme_dev_unmap(dev);
2159 static int nvme_remove_dead_ctrl(void *arg)
2161 struct nvme_dev *dev = (struct nvme_dev *)arg;
2162 struct pci_dev *pdev = dev->pci_dev;
2164 if (pci_get_drvdata(pdev))
2165 pci_stop_and_remove_bus_device(pdev);
2166 kref_put(&dev->kref, nvme_free_dev);
2170 static void nvme_remove_disks(struct work_struct *ws)
2173 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2175 nvme_dev_remove(dev);
2176 spin_lock(&dev_list_lock);
2177 for (i = dev->queue_count - 1; i > 0; i--) {
2178 BUG_ON(!dev->queues[i] || !dev->queues[i]->q_suspended);
2179 nvme_free_queue(dev->queues[i]);
2181 dev->queues[i] = NULL;
2183 spin_unlock(&dev_list_lock);
2186 static int nvme_dev_resume(struct nvme_dev *dev)
2190 ret = nvme_dev_start(dev);
2191 if (ret && ret != -EBUSY)
2193 if (ret == -EBUSY) {
2194 spin_lock(&dev_list_lock);
2195 INIT_WORK(&dev->reset_work, nvme_remove_disks);
2196 queue_work(nvme_workq, &dev->reset_work);
2197 spin_unlock(&dev_list_lock);
2202 static void nvme_dev_reset(struct nvme_dev *dev)
2204 nvme_dev_shutdown(dev);
2205 if (nvme_dev_resume(dev)) {
2206 dev_err(&dev->pci_dev->dev, "Device failed to resume\n");
2207 kref_get(&dev->kref);
2208 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2210 dev_err(&dev->pci_dev->dev,
2211 "Failed to start controller remove task\n");
2212 kref_put(&dev->kref, nvme_free_dev);
2217 static void nvme_reset_failed_dev(struct work_struct *ws)
2219 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2220 nvme_dev_reset(dev);
2223 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2225 int result = -ENOMEM;
2226 struct nvme_dev *dev;
2228 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2231 dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
2235 dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
2240 INIT_LIST_HEAD(&dev->namespaces);
2241 dev->pci_dev = pdev;
2242 pci_set_drvdata(pdev, dev);
2243 result = nvme_set_instance(dev);
2247 result = nvme_setup_prp_pools(dev);
2251 result = nvme_dev_start(dev);
2253 if (result == -EBUSY)
2258 result = nvme_dev_add(dev);
2263 scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2264 dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2265 dev->miscdev.parent = &pdev->dev;
2266 dev->miscdev.name = dev->name;
2267 dev->miscdev.fops = &nvme_dev_fops;
2268 result = misc_register(&dev->miscdev);
2272 kref_init(&dev->kref);
2276 nvme_dev_remove(dev);
2278 nvme_dev_shutdown(dev);
2280 nvme_free_queues(dev);
2281 nvme_release_prp_pools(dev);
2283 nvme_release_instance(dev);
2291 static void nvme_remove(struct pci_dev *pdev)
2293 struct nvme_dev *dev = pci_get_drvdata(pdev);
2295 spin_lock(&dev_list_lock);
2296 list_del_init(&dev->node);
2297 spin_unlock(&dev_list_lock);
2299 pci_set_drvdata(pdev, NULL);
2300 flush_work(&dev->reset_work);
2301 misc_deregister(&dev->miscdev);
2302 nvme_dev_remove(dev);
2303 nvme_dev_shutdown(dev);
2304 nvme_free_queues(dev);
2305 nvme_release_instance(dev);
2306 nvme_release_prp_pools(dev);
2307 kref_put(&dev->kref, nvme_free_dev);
2310 /* These functions are yet to be implemented */
2311 #define nvme_error_detected NULL
2312 #define nvme_dump_registers NULL
2313 #define nvme_link_reset NULL
2314 #define nvme_slot_reset NULL
2315 #define nvme_error_resume NULL
2317 static int nvme_suspend(struct device *dev)
2319 struct pci_dev *pdev = to_pci_dev(dev);
2320 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2322 nvme_dev_shutdown(ndev);
2326 static int nvme_resume(struct device *dev)
2328 struct pci_dev *pdev = to_pci_dev(dev);
2329 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2331 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2332 INIT_WORK(&ndev->reset_work, nvme_reset_failed_dev);
2333 queue_work(nvme_workq, &ndev->reset_work);
2338 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2340 static const struct pci_error_handlers nvme_err_handler = {
2341 .error_detected = nvme_error_detected,
2342 .mmio_enabled = nvme_dump_registers,
2343 .link_reset = nvme_link_reset,
2344 .slot_reset = nvme_slot_reset,
2345 .resume = nvme_error_resume,
2348 /* Move to pci_ids.h later */
2349 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
2351 static DEFINE_PCI_DEVICE_TABLE(nvme_id_table) = {
2352 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2355 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2357 static struct pci_driver nvme_driver = {
2359 .id_table = nvme_id_table,
2360 .probe = nvme_probe,
2361 .remove = nvme_remove,
2363 .pm = &nvme_dev_pm_ops,
2365 .err_handler = &nvme_err_handler,
2368 static int __init nvme_init(void)
2372 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2373 if (IS_ERR(nvme_thread))
2374 return PTR_ERR(nvme_thread);
2377 nvme_workq = create_singlethread_workqueue("nvme");
2381 result = register_blkdev(nvme_major, "nvme");
2384 else if (result > 0)
2385 nvme_major = result;
2387 result = pci_register_driver(&nvme_driver);
2389 goto unregister_blkdev;
2393 unregister_blkdev(nvme_major, "nvme");
2395 destroy_workqueue(nvme_workq);
2397 kthread_stop(nvme_thread);
2401 static void __exit nvme_exit(void)
2403 pci_unregister_driver(&nvme_driver);
2404 unregister_blkdev(nvme_major, "nvme");
2405 destroy_workqueue(nvme_workq);
2406 kthread_stop(nvme_thread);
2409 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2410 MODULE_LICENSE("GPL");
2411 MODULE_VERSION("0.8");
2412 module_init(nvme_init);
2413 module_exit(nvme_exit);