projects / linux-2.6-block.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
NVMe: Update namespace and controller identify structures to the 1.1a spec
[linux-2.6-block.git] / drivers / block / nvme-core.c
... / ...
CommitLineData
1/*
2 * NVM Express device driver
3 * Copyright (c) 2011-2014, Intel Corporation.
4 *
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.
8 *
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
12 * more details.
13 */
14
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>
22#include <linux/fs.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>
28#include <linux/io.h>
29#include <linux/kdev_t.h>
30#include <linux/kthread.h>
31#include <linux/kernel.h>
32#include <linux/mm.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>
42#include <scsi/sg.h>
43#include <asm-generic/io-64-nonatomic-lo-hi.h>
44
45#include <trace/events/block.h>
46
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 (60 * HZ)
51#define IOD_TIMEOUT (4 * NVME_IO_TIMEOUT)
52
53unsigned char io_timeout = 30;
54module_param(io_timeout, byte, 0644);
55MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
56
57static int nvme_major;
58module_param(nvme_major, int, 0);
59
60static int use_threaded_interrupts;
61module_param(use_threaded_interrupts, int, 0);
62
63static DEFINE_SPINLOCK(dev_list_lock);
64static LIST_HEAD(dev_list);
65static struct task_struct *nvme_thread;
66static struct workqueue_struct *nvme_workq;
67static wait_queue_head_t nvme_kthread_wait;
68
69static void nvme_reset_failed_dev(struct work_struct *ws);
70
71struct async_cmd_info {
72 struct kthread_work work;
73 struct kthread_worker *worker;
74 u32 result;
75 int status;
76 void *ctx;
77};
78
79/*
80 * An NVM Express queue. Each device has at least two (one for admin
81 * commands and one for I/O commands).
82 */
83struct nvme_queue {
84 struct rcu_head r_head;
85 struct device *q_dmadev;
86 struct nvme_dev *dev;
87 char irqname[24]; /* nvme4294967295-65535\0 */
88 spinlock_t q_lock;
89 struct nvme_command *sq_cmds;
90 volatile struct nvme_completion *cqes;
91 dma_addr_t sq_dma_addr;
92 dma_addr_t cq_dma_addr;
93 wait_queue_head_t sq_full;
94 wait_queue_t sq_cong_wait;
95 struct bio_list sq_cong;
96 struct list_head iod_bio;
97 u32 __iomem *q_db;
98 u16 q_depth;
99 u16 cq_vector;
100 u16 sq_head;
101 u16 sq_tail;
102 u16 cq_head;
103 u16 qid;
104 u8 cq_phase;
105 u8 cqe_seen;
106 u8 q_suspended;
107 cpumask_var_t cpu_mask;
108 struct async_cmd_info cmdinfo;
109 unsigned long cmdid_data[];
110};
111
112/*
113 * Check we didin't inadvertently grow the command struct
114 */
115static inline void _nvme_check_size(void)
116{
117 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
118 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
119 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
120 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
121 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
122 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
123 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
124 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
125 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
126 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
127 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
128 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
129}
130
131typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
132 struct nvme_completion *);
133
134struct nvme_cmd_info {
135 nvme_completion_fn fn;
136 void *ctx;
137 unsigned long timeout;
138 int aborted;
139};
140
141static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
142{
143 return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
144}
145
146static unsigned nvme_queue_extra(int depth)
147{
148 return DIV_ROUND_UP(depth, 8) + (depth * sizeof(struct nvme_cmd_info));
149}
150
151/**
152 * alloc_cmdid() - Allocate a Command ID
153 * @nvmeq: The queue that will be used for this command
154 * @ctx: A pointer that will be passed to the handler
155 * @handler: The function to call on completion
156 *
157 * Allocate a Command ID for a queue. The data passed in will
158 * be passed to the completion handler. This is implemented by using
159 * the bottom two bits of the ctx pointer to store the handler ID.
160 * Passing in a pointer that's not 4-byte aligned will cause a BUG.
161 * We can change this if it becomes a problem.
162 *
163 * May be called with local interrupts disabled and the q_lock held,
164 * or with interrupts enabled and no locks held.
165 */
166static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx,
167 nvme_completion_fn handler, unsigned timeout)
168{
169 int depth = nvmeq->q_depth - 1;
170 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
171 int cmdid;
172
173 do {
174 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
175 if (cmdid >= depth)
176 return -EBUSY;
177 } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
178
179 info[cmdid].fn = handler;
180 info[cmdid].ctx = ctx;
181 info[cmdid].timeout = jiffies + timeout;
182 info[cmdid].aborted = 0;
183 return cmdid;
184}
185
186static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
187 nvme_completion_fn handler, unsigned timeout)
188{
189 int cmdid;
190 wait_event_killable(nvmeq->sq_full,
191 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
192 return (cmdid < 0) ? -EINTR : cmdid;
193}
194
195/* Special values must be less than 0x1000 */
196#define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
197#define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
198#define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
199#define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
200#define CMD_CTX_ABORT (0x318 + CMD_CTX_BASE)
201
202static void special_completion(struct nvme_queue *nvmeq, void *ctx,
203 struct nvme_completion *cqe)
204{
205 if (ctx == CMD_CTX_CANCELLED)
206 return;
207 if (ctx == CMD_CTX_ABORT) {
208 ++nvmeq->dev->abort_limit;
209 return;
210 }
211 if (ctx == CMD_CTX_COMPLETED) {
212 dev_warn(nvmeq->q_dmadev,
213 "completed id %d twice on queue %d\n",
214 cqe->command_id, le16_to_cpup(&cqe->sq_id));
215 return;
216 }
217 if (ctx == CMD_CTX_INVALID) {
218 dev_warn(nvmeq->q_dmadev,
219 "invalid id %d completed on queue %d\n",
220 cqe->command_id, le16_to_cpup(&cqe->sq_id));
221 return;
222 }
223
224 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
225}
226
227static void async_completion(struct nvme_queue *nvmeq, void *ctx,
228 struct nvme_completion *cqe)
229{
230 struct async_cmd_info *cmdinfo = ctx;
231 cmdinfo->result = le32_to_cpup(&cqe->result);
232 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
233 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
234}
235
236/*
237 * Called with local interrupts disabled and the q_lock held. May not sleep.
238 */
239static void *free_cmdid(struct nvme_queue *nvmeq, int cmdid,
240 nvme_completion_fn *fn)
241{
242 void *ctx;
243 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
244
245 if (cmdid >= nvmeq->q_depth || !info[cmdid].fn) {
246 if (fn)
247 *fn = special_completion;
248 return CMD_CTX_INVALID;
249 }
250 if (fn)
251 *fn = info[cmdid].fn;
252 ctx = info[cmdid].ctx;
253 info[cmdid].fn = special_completion;
254 info[cmdid].ctx = CMD_CTX_COMPLETED;
255 clear_bit(cmdid, nvmeq->cmdid_data);
256 wake_up(&nvmeq->sq_full);
257 return ctx;
258}
259
260static void *cancel_cmdid(struct nvme_queue *nvmeq, int cmdid,
261 nvme_completion_fn *fn)
262{
263 void *ctx;
264 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
265 if (fn)
266 *fn = info[cmdid].fn;
267 ctx = info[cmdid].ctx;
268 info[cmdid].fn = special_completion;
269 info[cmdid].ctx = CMD_CTX_CANCELLED;
270 return ctx;
271}
272
273static struct nvme_queue *raw_nvmeq(struct nvme_dev *dev, int qid)
274{
275 return rcu_dereference_raw(dev->queues[qid]);
276}
277
278static struct nvme_queue *get_nvmeq(struct nvme_dev *dev) __acquires(RCU)
279{
280 unsigned queue_id = get_cpu_var(*dev->io_queue);
281 rcu_read_lock();
282 return rcu_dereference(dev->queues[queue_id]);
283}
284
285static void put_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
286{
287 rcu_read_unlock();
288 put_cpu_var(nvmeq->dev->io_queue);
289}
290
291static struct nvme_queue *lock_nvmeq(struct nvme_dev *dev, int q_idx)
292 __acquires(RCU)
293{
294 rcu_read_lock();
295 return rcu_dereference(dev->queues[q_idx]);
296}
297
298static void unlock_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
299{
300 rcu_read_unlock();
301}
302
303/**
304 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
305 * @nvmeq: The queue to use
306 * @cmd: The command to send
307 *
308 * Safe to use from interrupt context
309 */
310static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
311{
312 unsigned long flags;
313 u16 tail;
314 spin_lock_irqsave(&nvmeq->q_lock, flags);
315 if (nvmeq->q_suspended) {
316 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
317 return -EBUSY;
318 }
319 tail = nvmeq->sq_tail;
320 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
321 if (++tail == nvmeq->q_depth)
322 tail = 0;
323 writel(tail, nvmeq->q_db);
324 nvmeq->sq_tail = tail;
325 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
326
327 return 0;
328}
329
330static __le64 **iod_list(struct nvme_iod *iod)
331{
332 return ((void *)iod) + iod->offset;
333}
334
335/*
336 * Will slightly overestimate the number of pages needed. This is OK
337 * as it only leads to a small amount of wasted memory for the lifetime of
338 * the I/O.
339 */
340static int nvme_npages(unsigned size)
341{
342 unsigned nprps = DIV_ROUND_UP(size + PAGE_SIZE, PAGE_SIZE);
343 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
344}
345
346static struct nvme_iod *
347nvme_alloc_iod(unsigned nseg, unsigned nbytes, gfp_t gfp)
348{
349 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
350 sizeof(__le64 *) * nvme_npages(nbytes) +
351 sizeof(struct scatterlist) * nseg, gfp);
352
353 if (iod) {
354 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
355 iod->npages = -1;
356 iod->length = nbytes;
357 iod->nents = 0;
358 iod->first_dma = 0ULL;
359 iod->start_time = jiffies;
360 }
361
362 return iod;
363}
364
365void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
366{
367 const int last_prp = PAGE_SIZE / 8 - 1;
368 int i;
369 __le64 **list = iod_list(iod);
370 dma_addr_t prp_dma = iod->first_dma;
371
372 if (iod->npages == 0)
373 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
374 for (i = 0; i < iod->npages; i++) {
375 __le64 *prp_list = list[i];
376 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
377 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
378 prp_dma = next_prp_dma;
379 }
380 kfree(iod);
381}
382
383static void nvme_start_io_acct(struct bio *bio)
384{
385 struct gendisk *disk = bio->bi_bdev->bd_disk;
386 const int rw = bio_data_dir(bio);
387 int cpu = part_stat_lock();
388 part_round_stats(cpu, &disk->part0);
389 part_stat_inc(cpu, &disk->part0, ios[rw]);
390 part_stat_add(cpu, &disk->part0, sectors[rw], bio_sectors(bio));
391 part_inc_in_flight(&disk->part0, rw);
392 part_stat_unlock();
393}
394
395static void nvme_end_io_acct(struct bio *bio, unsigned long start_time)
396{
397 struct gendisk *disk = bio->bi_bdev->bd_disk;
398 const int rw = bio_data_dir(bio);
399 unsigned long duration = jiffies - start_time;
400 int cpu = part_stat_lock();
401 part_stat_add(cpu, &disk->part0, ticks[rw], duration);
402 part_round_stats(cpu, &disk->part0);
403 part_dec_in_flight(&disk->part0, rw);
404 part_stat_unlock();
405}
406
407static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
408 struct nvme_completion *cqe)
409{
410 struct nvme_iod *iod = ctx;
411 struct bio *bio = iod->private;
412 u16 status = le16_to_cpup(&cqe->status) >> 1;
413 int error = 0;
414
415 if (unlikely(status)) {
416 if (!(status & NVME_SC_DNR ||
417 bio->bi_rw & REQ_FAILFAST_MASK) &&
418 (jiffies - iod->start_time) < IOD_TIMEOUT) {
419 if (!waitqueue_active(&nvmeq->sq_full))
420 add_wait_queue(&nvmeq->sq_full,
421 &nvmeq->sq_cong_wait);
422 list_add_tail(&iod->node, &nvmeq->iod_bio);
423 wake_up(&nvmeq->sq_full);
424 return;
425 }
426 error = -EIO;
427 }
428 if (iod->nents) {
429 dma_unmap_sg(nvmeq->q_dmadev, iod->sg, iod->nents,
430 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
431 nvme_end_io_acct(bio, iod->start_time);
432 }
433 nvme_free_iod(nvmeq->dev, iod);
434
435 trace_block_bio_complete(bdev_get_queue(bio->bi_bdev), bio, error);
436 bio_endio(bio, error);
437}
438
439/* length is in bytes. gfp flags indicates whether we may sleep. */
440int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod, int total_len,
441 gfp_t gfp)
442{
443 struct dma_pool *pool;
444 int length = total_len;
445 struct scatterlist *sg = iod->sg;
446 int dma_len = sg_dma_len(sg);
447 u64 dma_addr = sg_dma_address(sg);
448 int offset = offset_in_page(dma_addr);
449 __le64 *prp_list;
450 __le64 **list = iod_list(iod);
451 dma_addr_t prp_dma;
452 int nprps, i;
453
454 length -= (PAGE_SIZE - offset);
455 if (length <= 0)
456 return total_len;
457
458 dma_len -= (PAGE_SIZE - offset);
459 if (dma_len) {
460 dma_addr += (PAGE_SIZE - offset);
461 } else {
462 sg = sg_next(sg);
463 dma_addr = sg_dma_address(sg);
464 dma_len = sg_dma_len(sg);
465 }
466
467 if (length <= PAGE_SIZE) {
468 iod->first_dma = dma_addr;
469 return total_len;
470 }
471
472 nprps = DIV_ROUND_UP(length, PAGE_SIZE);
473 if (nprps <= (256 / 8)) {
474 pool = dev->prp_small_pool;
475 iod->npages = 0;
476 } else {
477 pool = dev->prp_page_pool;
478 iod->npages = 1;
479 }
480
481 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
482 if (!prp_list) {
483 iod->first_dma = dma_addr;
484 iod->npages = -1;
485 return (total_len - length) + PAGE_SIZE;
486 }
487 list[0] = prp_list;
488 iod->first_dma = prp_dma;
489 i = 0;
490 for (;;) {
491 if (i == PAGE_SIZE / 8) {
492 __le64 *old_prp_list = prp_list;
493 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
494 if (!prp_list)
495 return total_len - length;
496 list[iod->npages++] = prp_list;
497 prp_list[0] = old_prp_list[i - 1];
498 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
499 i = 1;
500 }
501 prp_list[i++] = cpu_to_le64(dma_addr);
502 dma_len -= PAGE_SIZE;
503 dma_addr += PAGE_SIZE;
504 length -= PAGE_SIZE;
505 if (length <= 0)
506 break;
507 if (dma_len > 0)
508 continue;
509 BUG_ON(dma_len < 0);
510 sg = sg_next(sg);
511 dma_addr = sg_dma_address(sg);
512 dma_len = sg_dma_len(sg);
513 }
514
515 return total_len;
516}
517
518static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
519 int len)
520{
521 struct bio *split = bio_split(bio, len >> 9, GFP_ATOMIC, NULL);
522 if (!split)
523 return -ENOMEM;
524
525 trace_block_split(bdev_get_queue(bio->bi_bdev), bio,
526 split->bi_iter.bi_sector);
527 bio_chain(split, bio);
528
529 if (!waitqueue_active(&nvmeq->sq_full))
530 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
531 bio_list_add(&nvmeq->sq_cong, split);
532 bio_list_add(&nvmeq->sq_cong, bio);
533 wake_up(&nvmeq->sq_full);
534
535 return 0;
536}
537
538/* NVMe scatterlists require no holes in the virtual address */
539#define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2) ((vec2)->bv_offset || \
540 (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
541
542static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
543 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
544{
545 struct bio_vec bvec, bvprv;
546 struct bvec_iter iter;
547 struct scatterlist *sg = NULL;
548 int length = 0, nsegs = 0, split_len = bio->bi_iter.bi_size;
549 int first = 1;
550
551 if (nvmeq->dev->stripe_size)
552 split_len = nvmeq->dev->stripe_size -
553 ((bio->bi_iter.bi_sector << 9) &
554 (nvmeq->dev->stripe_size - 1));
555
556 sg_init_table(iod->sg, psegs);
557 bio_for_each_segment(bvec, bio, iter) {
558 if (!first && BIOVEC_PHYS_MERGEABLE(&bvprv, &bvec)) {
559 sg->length += bvec.bv_len;
560 } else {
561 if (!first && BIOVEC_NOT_VIRT_MERGEABLE(&bvprv, &bvec))
562 return nvme_split_and_submit(bio, nvmeq,
563 length);
564
565 sg = sg ? sg + 1 : iod->sg;
566 sg_set_page(sg, bvec.bv_page,
567 bvec.bv_len, bvec.bv_offset);
568 nsegs++;
569 }
570
571 if (split_len - length < bvec.bv_len)
572 return nvme_split_and_submit(bio, nvmeq, split_len);
573 length += bvec.bv_len;
574 bvprv = bvec;
575 first = 0;
576 }
577 iod->nents = nsegs;
578 sg_mark_end(sg);
579 if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
580 return -ENOMEM;
581
582 BUG_ON(length != bio->bi_iter.bi_size);
583 return length;
584}
585
586static int nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
587 struct bio *bio, struct nvme_iod *iod, int cmdid)
588{
589 struct nvme_dsm_range *range =
590 (struct nvme_dsm_range *)iod_list(iod)[0];
591 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
592
593 range->cattr = cpu_to_le32(0);
594 range->nlb = cpu_to_le32(bio->bi_iter.bi_size >> ns->lba_shift);
595 range->slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
596
597 memset(cmnd, 0, sizeof(*cmnd));
598 cmnd->dsm.opcode = nvme_cmd_dsm;
599 cmnd->dsm.command_id = cmdid;
600 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
601 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
602 cmnd->dsm.nr = 0;
603 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
604
605 if (++nvmeq->sq_tail == nvmeq->q_depth)
606 nvmeq->sq_tail = 0;
607 writel(nvmeq->sq_tail, nvmeq->q_db);
608
609 return 0;
610}
611
612static int nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
613 int cmdid)
614{
615 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
616
617 memset(cmnd, 0, sizeof(*cmnd));
618 cmnd->common.opcode = nvme_cmd_flush;
619 cmnd->common.command_id = cmdid;
620 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
621
622 if (++nvmeq->sq_tail == nvmeq->q_depth)
623 nvmeq->sq_tail = 0;
624 writel(nvmeq->sq_tail, nvmeq->q_db);
625
626 return 0;
627}
628
629static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod)
630{
631 struct bio *bio = iod->private;
632 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
633 struct nvme_command *cmnd;
634 int cmdid;
635 u16 control;
636 u32 dsmgmt;
637
638 cmdid = alloc_cmdid(nvmeq, iod, bio_completion, NVME_IO_TIMEOUT);
639 if (unlikely(cmdid < 0))
640 return cmdid;
641
642 if (bio->bi_rw & REQ_DISCARD)
643 return nvme_submit_discard(nvmeq, ns, bio, iod, cmdid);
644 if (bio->bi_rw & REQ_FLUSH)
645 return nvme_submit_flush(nvmeq, ns, cmdid);
646
647 control = 0;
648 if (bio->bi_rw & REQ_FUA)
649 control |= NVME_RW_FUA;
650 if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
651 control |= NVME_RW_LR;
652
653 dsmgmt = 0;
654 if (bio->bi_rw & REQ_RAHEAD)
655 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
656
657 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
658 memset(cmnd, 0, sizeof(*cmnd));
659
660 cmnd->rw.opcode = bio_data_dir(bio) ? nvme_cmd_write : nvme_cmd_read;
661 cmnd->rw.command_id = cmdid;
662 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
663 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
664 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
665 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
666 cmnd->rw.length =
667 cpu_to_le16((bio->bi_iter.bi_size >> ns->lba_shift) - 1);
668 cmnd->rw.control = cpu_to_le16(control);
669 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
670
671 if (++nvmeq->sq_tail == nvmeq->q_depth)
672 nvmeq->sq_tail = 0;
673 writel(nvmeq->sq_tail, nvmeq->q_db);
674
675 return 0;
676}
677
678static int nvme_split_flush_data(struct nvme_queue *nvmeq, struct bio *bio)
679{
680 struct bio *split = bio_clone(bio, GFP_ATOMIC);
681 if (!split)
682 return -ENOMEM;
683
684 split->bi_iter.bi_size = 0;
685 split->bi_phys_segments = 0;
686 bio->bi_rw &= ~REQ_FLUSH;
687 bio_chain(split, bio);
688
689 if (!waitqueue_active(&nvmeq->sq_full))
690 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
691 bio_list_add(&nvmeq->sq_cong, split);
692 bio_list_add(&nvmeq->sq_cong, bio);
693 wake_up_process(nvme_thread);
694
695 return 0;
696}
697
698/*
699 * Called with local interrupts disabled and the q_lock held. May not sleep.
700 */
701static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
702 struct bio *bio)
703{
704 struct nvme_iod *iod;
705 int psegs = bio_phys_segments(ns->queue, bio);
706 int result;
707
708 if ((bio->bi_rw & REQ_FLUSH) && psegs)
709 return nvme_split_flush_data(nvmeq, bio);
710
711 iod = nvme_alloc_iod(psegs, bio->bi_iter.bi_size, GFP_ATOMIC);
712 if (!iod)
713 return -ENOMEM;
714
715 iod->private = bio;
716 if (bio->bi_rw & REQ_DISCARD) {
717 void *range;
718 /*
719 * We reuse the small pool to allocate the 16-byte range here
720 * as it is not worth having a special pool for these or
721 * additional cases to handle freeing the iod.
722 */
723 range = dma_pool_alloc(nvmeq->dev->prp_small_pool,
724 GFP_ATOMIC,
725 &iod->first_dma);
726 if (!range) {
727 result = -ENOMEM;
728 goto free_iod;
729 }
730 iod_list(iod)[0] = (__le64 *)range;
731 iod->npages = 0;
732 } else if (psegs) {
733 result = nvme_map_bio(nvmeq, iod, bio,
734 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE,
735 psegs);
736 if (result <= 0)
737 goto free_iod;
738 if (nvme_setup_prps(nvmeq->dev, iod, result, GFP_ATOMIC) !=
739 result) {
740 result = -ENOMEM;
741 goto free_iod;
742 }
743 nvme_start_io_acct(bio);
744 }
745 if (unlikely(nvme_submit_iod(nvmeq, iod))) {
746 if (!waitqueue_active(&nvmeq->sq_full))
747 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
748 list_add_tail(&iod->node, &nvmeq->iod_bio);
749 }
750 return 0;
751
752 free_iod:
753 nvme_free_iod(nvmeq->dev, iod);
754 return result;
755}
756
757static int nvme_process_cq(struct nvme_queue *nvmeq)
758{
759 u16 head, phase;
760
761 head = nvmeq->cq_head;
762 phase = nvmeq->cq_phase;
763
764 for (;;) {
765 void *ctx;
766 nvme_completion_fn fn;
767 struct nvme_completion cqe = nvmeq->cqes[head];
768 if ((le16_to_cpu(cqe.status) & 1) != phase)
769 break;
770 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
771 if (++head == nvmeq->q_depth) {
772 head = 0;
773 phase = !phase;
774 }
775
776 ctx = free_cmdid(nvmeq, cqe.command_id, &fn);
777 fn(nvmeq, ctx, &cqe);
778 }
779
780 /* If the controller ignores the cq head doorbell and continuously
781 * writes to the queue, it is theoretically possible to wrap around
782 * the queue twice and mistakenly return IRQ_NONE. Linux only
783 * requires that 0.1% of your interrupts are handled, so this isn't
784 * a big problem.
785 */
786 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
787 return 0;
788
789 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
790 nvmeq->cq_head = head;
791 nvmeq->cq_phase = phase;
792
793 nvmeq->cqe_seen = 1;
794 return 1;
795}
796
797static void nvme_make_request(struct request_queue *q, struct bio *bio)
798{
799 struct nvme_ns *ns = q->queuedata;
800 struct nvme_queue *nvmeq = get_nvmeq(ns->dev);
801 int result = -EBUSY;
802
803 if (!nvmeq) {
804 put_nvmeq(NULL);
805 bio_endio(bio, -EIO);
806 return;
807 }
808
809 spin_lock_irq(&nvmeq->q_lock);
810 if (!nvmeq->q_suspended && bio_list_empty(&nvmeq->sq_cong))
811 result = nvme_submit_bio_queue(nvmeq, ns, bio);
812 if (unlikely(result)) {
813 if (!waitqueue_active(&nvmeq->sq_full))
814 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
815 bio_list_add(&nvmeq->sq_cong, bio);
816 }
817
818 nvme_process_cq(nvmeq);
819 spin_unlock_irq(&nvmeq->q_lock);
820 put_nvmeq(nvmeq);
821}
822
823static irqreturn_t nvme_irq(int irq, void *data)
824{
825 irqreturn_t result;
826 struct nvme_queue *nvmeq = data;
827 spin_lock(&nvmeq->q_lock);
828 nvme_process_cq(nvmeq);
829 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
830 nvmeq->cqe_seen = 0;
831 spin_unlock(&nvmeq->q_lock);
832 return result;
833}
834
835static irqreturn_t nvme_irq_check(int irq, void *data)
836{
837 struct nvme_queue *nvmeq = data;
838 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
839 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
840 return IRQ_NONE;
841 return IRQ_WAKE_THREAD;
842}
843
844static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
845{
846 spin_lock_irq(&nvmeq->q_lock);
847 cancel_cmdid(nvmeq, cmdid, NULL);
848 spin_unlock_irq(&nvmeq->q_lock);
849}
850
851struct sync_cmd_info {
852 struct task_struct *task;
853 u32 result;
854 int status;
855};
856
857static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
858 struct nvme_completion *cqe)
859{
860 struct sync_cmd_info *cmdinfo = ctx;
861 cmdinfo->result = le32_to_cpup(&cqe->result);
862 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
863 wake_up_process(cmdinfo->task);
864}
865
866/*
867 * Returns 0 on success. If the result is negative, it's a Linux error code;
868 * if the result is positive, it's an NVM Express status code
869 */
870static int nvme_submit_sync_cmd(struct nvme_dev *dev, int q_idx,
871 struct nvme_command *cmd,
872 u32 *result, unsigned timeout)
873{
874 int cmdid, ret;
875 struct sync_cmd_info cmdinfo;
876 struct nvme_queue *nvmeq;
877
878 nvmeq = lock_nvmeq(dev, q_idx);
879 if (!nvmeq) {
880 unlock_nvmeq(nvmeq);
881 return -ENODEV;
882 }
883
884 cmdinfo.task = current;
885 cmdinfo.status = -EINTR;
886
887 cmdid = alloc_cmdid(nvmeq, &cmdinfo, sync_completion, timeout);
888 if (cmdid < 0) {
889 unlock_nvmeq(nvmeq);
890 return cmdid;
891 }
892 cmd->common.command_id = cmdid;
893
894 set_current_state(TASK_KILLABLE);
895 ret = nvme_submit_cmd(nvmeq, cmd);
896 if (ret) {
897 free_cmdid(nvmeq, cmdid, NULL);
898 unlock_nvmeq(nvmeq);
899 set_current_state(TASK_RUNNING);
900 return ret;
901 }
902 unlock_nvmeq(nvmeq);
903 schedule_timeout(timeout);
904
905 if (cmdinfo.status == -EINTR) {
906 nvmeq = lock_nvmeq(dev, q_idx);
907 if (nvmeq)
908 nvme_abort_command(nvmeq, cmdid);
909 unlock_nvmeq(nvmeq);
910 return -EINTR;
911 }
912
913 if (result)
914 *result = cmdinfo.result;
915
916 return cmdinfo.status;
917}
918
919static int nvme_submit_async_cmd(struct nvme_queue *nvmeq,
920 struct nvme_command *cmd,
921 struct async_cmd_info *cmdinfo, unsigned timeout)
922{
923 int cmdid;
924
925 cmdid = alloc_cmdid_killable(nvmeq, cmdinfo, async_completion, timeout);
926 if (cmdid < 0)
927 return cmdid;
928 cmdinfo->status = -EINTR;
929 cmd->common.command_id = cmdid;
930 return nvme_submit_cmd(nvmeq, cmd);
931}
932
933int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
934 u32 *result)
935{
936 return nvme_submit_sync_cmd(dev, 0, cmd, result, ADMIN_TIMEOUT);
937}
938
939int nvme_submit_io_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
940 u32 *result)
941{
942 return nvme_submit_sync_cmd(dev, smp_processor_id() + 1, cmd, result,
943 NVME_IO_TIMEOUT);
944}
945
946static int nvme_submit_admin_cmd_async(struct nvme_dev *dev,
947 struct nvme_command *cmd, struct async_cmd_info *cmdinfo)
948{
949 return nvme_submit_async_cmd(raw_nvmeq(dev, 0), cmd, cmdinfo,
950 ADMIN_TIMEOUT);
951}
952
953static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
954{
955 int status;
956 struct nvme_command c;
957
958 memset(&c, 0, sizeof(c));
959 c.delete_queue.opcode = opcode;
960 c.delete_queue.qid = cpu_to_le16(id);
961
962 status = nvme_submit_admin_cmd(dev, &c, NULL);
963 if (status)
964 return -EIO;
965 return 0;
966}
967
968static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
969 struct nvme_queue *nvmeq)
970{
971 int status;
972 struct nvme_command c;
973 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
974
975 memset(&c, 0, sizeof(c));
976 c.create_cq.opcode = nvme_admin_create_cq;
977 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
978 c.create_cq.cqid = cpu_to_le16(qid);
979 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
980 c.create_cq.cq_flags = cpu_to_le16(flags);
981 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
982
983 status = nvme_submit_admin_cmd(dev, &c, NULL);
984 if (status)
985 return -EIO;
986 return 0;
987}
988
989static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
990 struct nvme_queue *nvmeq)
991{
992 int status;
993 struct nvme_command c;
994 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
995
996 memset(&c, 0, sizeof(c));
997 c.create_sq.opcode = nvme_admin_create_sq;
998 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
999 c.create_sq.sqid = cpu_to_le16(qid);
1000 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1001 c.create_sq.sq_flags = cpu_to_le16(flags);
1002 c.create_sq.cqid = cpu_to_le16(qid);
1003
1004 status = nvme_submit_admin_cmd(dev, &c, NULL);
1005 if (status)
1006 return -EIO;
1007 return 0;
1008}
1009
1010static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1011{
1012 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1013}
1014
1015static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1016{
1017 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1018}
1019
1020int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
1021 dma_addr_t dma_addr)
1022{
1023 struct nvme_command c;
1024
1025 memset(&c, 0, sizeof(c));
1026 c.identify.opcode = nvme_admin_identify;
1027 c.identify.nsid = cpu_to_le32(nsid);
1028 c.identify.prp1 = cpu_to_le64(dma_addr);
1029 c.identify.cns = cpu_to_le32(cns);
1030
1031 return nvme_submit_admin_cmd(dev, &c, NULL);
1032}
1033
1034int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1035 dma_addr_t dma_addr, u32 *result)
1036{
1037 struct nvme_command c;
1038
1039 memset(&c, 0, sizeof(c));
1040 c.features.opcode = nvme_admin_get_features;
1041 c.features.nsid = cpu_to_le32(nsid);
1042 c.features.prp1 = cpu_to_le64(dma_addr);
1043 c.features.fid = cpu_to_le32(fid);
1044
1045 return nvme_submit_admin_cmd(dev, &c, result);
1046}
1047
1048int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1049 dma_addr_t dma_addr, u32 *result)
1050{
1051 struct nvme_command c;
1052
1053 memset(&c, 0, sizeof(c));
1054 c.features.opcode = nvme_admin_set_features;
1055 c.features.prp1 = cpu_to_le64(dma_addr);
1056 c.features.fid = cpu_to_le32(fid);
1057 c.features.dword11 = cpu_to_le32(dword11);
1058
1059 return nvme_submit_admin_cmd(dev, &c, result);
1060}
1061
1062/**
1063 * nvme_abort_cmd - Attempt aborting a command
1064 * @cmdid: Command id of a timed out IO
1065 * @queue: The queue with timed out IO
1066 *
1067 * Schedule controller reset if the command was already aborted once before and
1068 * still hasn't been returned to the driver, or if this is the admin queue.
1069 */
1070static void nvme_abort_cmd(int cmdid, struct nvme_queue *nvmeq)
1071{
1072 int a_cmdid;
1073 struct nvme_command cmd;
1074 struct nvme_dev *dev = nvmeq->dev;
1075 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1076 struct nvme_queue *adminq;
1077
1078 if (!nvmeq->qid || info[cmdid].aborted) {
1079 if (work_busy(&dev->reset_work))
1080 return;
1081 list_del_init(&dev->node);
1082 dev_warn(&dev->pci_dev->dev,
1083 "I/O %d QID %d timeout, reset controller\n", cmdid,
1084 nvmeq->qid);
1085 dev->reset_workfn = nvme_reset_failed_dev;
1086 queue_work(nvme_workq, &dev->reset_work);
1087 return;
1088 }
1089
1090 if (!dev->abort_limit)
1091 return;
1092
1093 adminq = rcu_dereference(dev->queues[0]);
1094 a_cmdid = alloc_cmdid(adminq, CMD_CTX_ABORT, special_completion,
1095 ADMIN_TIMEOUT);
1096 if (a_cmdid < 0)
1097 return;
1098
1099 memset(&cmd, 0, sizeof(cmd));
1100 cmd.abort.opcode = nvme_admin_abort_cmd;
1101 cmd.abort.cid = cmdid;
1102 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1103 cmd.abort.command_id = a_cmdid;
1104
1105 --dev->abort_limit;
1106 info[cmdid].aborted = 1;
1107 info[cmdid].timeout = jiffies + ADMIN_TIMEOUT;
1108
1109 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", cmdid,
1110 nvmeq->qid);
1111 nvme_submit_cmd(adminq, &cmd);
1112}
1113
1114/**
1115 * nvme_cancel_ios - Cancel outstanding I/Os
1116 * @queue: The queue to cancel I/Os on
1117 * @timeout: True to only cancel I/Os which have timed out
1118 */
1119static void nvme_cancel_ios(struct nvme_queue *nvmeq, bool timeout)
1120{
1121 int depth = nvmeq->q_depth - 1;
1122 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1123 unsigned long now = jiffies;
1124 int cmdid;
1125
1126 for_each_set_bit(cmdid, nvmeq->cmdid_data, depth) {
1127 void *ctx;
1128 nvme_completion_fn fn;
1129 static struct nvme_completion cqe = {
1130 .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1131 };
1132
1133 if (timeout && !time_after(now, info[cmdid].timeout))
1134 continue;
1135 if (info[cmdid].ctx == CMD_CTX_CANCELLED)
1136 continue;
1137 if (timeout && nvmeq->dev->initialized) {
1138 nvme_abort_cmd(cmdid, nvmeq);
1139 continue;
1140 }
1141 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n", cmdid,
1142 nvmeq->qid);
1143 ctx = cancel_cmdid(nvmeq, cmdid, &fn);
1144 fn(nvmeq, ctx, &cqe);
1145 }
1146}
1147
1148static void nvme_free_queue(struct rcu_head *r)
1149{
1150 struct nvme_queue *nvmeq = container_of(r, struct nvme_queue, r_head);
1151
1152 spin_lock_irq(&nvmeq->q_lock);
1153 while (bio_list_peek(&nvmeq->sq_cong)) {
1154 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1155 bio_endio(bio, -EIO);
1156 }
1157 while (!list_empty(&nvmeq->iod_bio)) {
1158 static struct nvme_completion cqe = {
1159 .status = cpu_to_le16(
1160 (NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1),
1161 };
1162 struct nvme_iod *iod = list_first_entry(&nvmeq->iod_bio,
1163 struct nvme_iod,
1164 node);
1165 list_del(&iod->node);
1166 bio_completion(nvmeq, iod, &cqe);
1167 }
1168 spin_unlock_irq(&nvmeq->q_lock);
1169
1170 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1171 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1172 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1173 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1174 if (nvmeq->qid)
1175 free_cpumask_var(nvmeq->cpu_mask);
1176 kfree(nvmeq);
1177}
1178
1179static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1180{
1181 int i;
1182
1183 for (i = dev->queue_count - 1; i >= lowest; i--) {
1184 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
1185 rcu_assign_pointer(dev->queues[i], NULL);
1186 call_rcu(&nvmeq->r_head, nvme_free_queue);
1187 dev->queue_count--;
1188 }
1189}
1190
1191/**
1192 * nvme_suspend_queue - put queue into suspended state
1193 * @nvmeq - queue to suspend
1194 *
1195 * Returns 1 if already suspended, 0 otherwise.
1196 */
1197static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1198{
1199 int vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1200
1201 spin_lock_irq(&nvmeq->q_lock);
1202 if (nvmeq->q_suspended) {
1203 spin_unlock_irq(&nvmeq->q_lock);
1204 return 1;
1205 }
1206 nvmeq->q_suspended = 1;
1207 nvmeq->dev->online_queues--;
1208 spin_unlock_irq(&nvmeq->q_lock);
1209
1210 irq_set_affinity_hint(vector, NULL);
1211 free_irq(vector, nvmeq);
1212
1213 return 0;
1214}
1215
1216static void nvme_clear_queue(struct nvme_queue *nvmeq)
1217{
1218 spin_lock_irq(&nvmeq->q_lock);
1219 nvme_process_cq(nvmeq);
1220 nvme_cancel_ios(nvmeq, false);
1221 spin_unlock_irq(&nvmeq->q_lock);
1222}
1223
1224static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1225{
1226 struct nvme_queue *nvmeq = raw_nvmeq(dev, qid);
1227
1228 if (!nvmeq)
1229 return;
1230 if (nvme_suspend_queue(nvmeq))
1231 return;
1232
1233 /* Don't tell the adapter to delete the admin queue.
1234 * Don't tell a removed adapter to delete IO queues. */
1235 if (qid && readl(&dev->bar->csts) != -1) {
1236 adapter_delete_sq(dev, qid);
1237 adapter_delete_cq(dev, qid);
1238 }
1239 nvme_clear_queue(nvmeq);
1240}
1241
1242static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1243 int depth, int vector)
1244{
1245 struct device *dmadev = &dev->pci_dev->dev;
1246 unsigned extra = nvme_queue_extra(depth);
1247 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
1248 if (!nvmeq)
1249 return NULL;
1250
1251 nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
1252 &nvmeq->cq_dma_addr, GFP_KERNEL);
1253 if (!nvmeq->cqes)
1254 goto free_nvmeq;
1255 memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
1256
1257 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1258 &nvmeq->sq_dma_addr, GFP_KERNEL);
1259 if (!nvmeq->sq_cmds)
1260 goto free_cqdma;
1261
1262 if (qid && !zalloc_cpumask_var(&nvmeq->cpu_mask, GFP_KERNEL))
1263 goto free_sqdma;
1264
1265 nvmeq->q_dmadev = dmadev;
1266 nvmeq->dev = dev;
1267 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1268 dev->instance, qid);
1269 spin_lock_init(&nvmeq->q_lock);
1270 nvmeq->cq_head = 0;
1271 nvmeq->cq_phase = 1;
1272 init_waitqueue_head(&nvmeq->sq_full);
1273 init_waitqueue_entry(&nvmeq->sq_cong_wait, nvme_thread);
1274 bio_list_init(&nvmeq->sq_cong);
1275 INIT_LIST_HEAD(&nvmeq->iod_bio);
1276 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1277 nvmeq->q_depth = depth;
1278 nvmeq->cq_vector = vector;
1279 nvmeq->qid = qid;
1280 nvmeq->q_suspended = 1;
1281 dev->queue_count++;
1282 rcu_assign_pointer(dev->queues[qid], nvmeq);
1283
1284 return nvmeq;
1285
1286 free_sqdma:
1287 dma_free_coherent(dmadev, SQ_SIZE(depth), (void *)nvmeq->sq_cmds,
1288 nvmeq->sq_dma_addr);
1289 free_cqdma:
1290 dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1291 nvmeq->cq_dma_addr);
1292 free_nvmeq:
1293 kfree(nvmeq);
1294 return NULL;
1295}
1296
1297static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1298 const char *name)
1299{
1300 if (use_threaded_interrupts)
1301 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1302 nvme_irq_check, nvme_irq, IRQF_SHARED,
1303 name, nvmeq);
1304 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1305 IRQF_SHARED, name, nvmeq);
1306}
1307
1308static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1309{
1310 struct nvme_dev *dev = nvmeq->dev;
1311 unsigned extra = nvme_queue_extra(nvmeq->q_depth);
1312
1313 nvmeq->sq_tail = 0;
1314 nvmeq->cq_head = 0;
1315 nvmeq->cq_phase = 1;
1316 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1317 memset(nvmeq->cmdid_data, 0, extra);
1318 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1319 nvme_cancel_ios(nvmeq, false);
1320 nvmeq->q_suspended = 0;
1321 dev->online_queues++;
1322}
1323
1324static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1325{
1326 struct nvme_dev *dev = nvmeq->dev;
1327 int result;
1328
1329 result = adapter_alloc_cq(dev, qid, nvmeq);
1330 if (result < 0)
1331 return result;
1332
1333 result = adapter_alloc_sq(dev, qid, nvmeq);
1334 if (result < 0)
1335 goto release_cq;
1336
1337 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1338 if (result < 0)
1339 goto release_sq;
1340
1341 spin_lock_irq(&nvmeq->q_lock);
1342 nvme_init_queue(nvmeq, qid);
1343 spin_unlock_irq(&nvmeq->q_lock);
1344
1345 return result;
1346
1347 release_sq:
1348 adapter_delete_sq(dev, qid);
1349 release_cq:
1350 adapter_delete_cq(dev, qid);
1351 return result;
1352}
1353
1354static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1355{
1356 unsigned long timeout;
1357 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1358
1359 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1360
1361 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1362 msleep(100);
1363 if (fatal_signal_pending(current))
1364 return -EINTR;
1365 if (time_after(jiffies, timeout)) {
1366 dev_err(&dev->pci_dev->dev,
1367 "Device not ready; aborting %s\n", enabled ?
1368 "initialisation" : "reset");
1369 return -ENODEV;
1370 }
1371 }
1372
1373 return 0;
1374}
1375
1376/*
1377 * If the device has been passed off to us in an enabled state, just clear
1378 * the enabled bit. The spec says we should set the 'shutdown notification
1379 * bits', but doing so may cause the device to complete commands to the
1380 * admin queue ... and we don't know what memory that might be pointing at!
1381 */
1382static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1383{
1384 u32 cc = readl(&dev->bar->cc);
1385
1386 if (cc & NVME_CC_ENABLE)
1387 writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
1388 return nvme_wait_ready(dev, cap, false);
1389}
1390
1391static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1392{
1393 return nvme_wait_ready(dev, cap, true);
1394}
1395
1396static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1397{
1398 unsigned long timeout;
1399 u32 cc;
1400
1401 cc = (readl(&dev->bar->cc) & ~NVME_CC_SHN_MASK) | NVME_CC_SHN_NORMAL;
1402 writel(cc, &dev->bar->cc);
1403
1404 timeout = 2 * HZ + jiffies;
1405 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1406 NVME_CSTS_SHST_CMPLT) {
1407 msleep(100);
1408 if (fatal_signal_pending(current))
1409 return -EINTR;
1410 if (time_after(jiffies, timeout)) {
1411 dev_err(&dev->pci_dev->dev,
1412 "Device shutdown incomplete; abort shutdown\n");
1413 return -ENODEV;
1414 }
1415 }
1416
1417 return 0;
1418}
1419
1420static int nvme_configure_admin_queue(struct nvme_dev *dev)
1421{
1422 int result;
1423 u32 aqa;
1424 u64 cap = readq(&dev->bar->cap);
1425 struct nvme_queue *nvmeq;
1426
1427 result = nvme_disable_ctrl(dev, cap);
1428 if (result < 0)
1429 return result;
1430
1431 nvmeq = raw_nvmeq(dev, 0);
1432 if (!nvmeq) {
1433 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
1434 if (!nvmeq)
1435 return -ENOMEM;
1436 }
1437
1438 aqa = nvmeq->q_depth - 1;
1439 aqa |= aqa << 16;
1440
1441 dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
1442 dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
1443 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1444 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1445
1446 writel(aqa, &dev->bar->aqa);
1447 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1448 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1449 writel(dev->ctrl_config, &dev->bar->cc);
1450
1451 result = nvme_enable_ctrl(dev, cap);
1452 if (result)
1453 return result;
1454
1455 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1456 if (result)
1457 return result;
1458
1459 spin_lock_irq(&nvmeq->q_lock);
1460 nvme_init_queue(nvmeq, 0);
1461 spin_unlock_irq(&nvmeq->q_lock);
1462 return result;
1463}
1464
1465struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1466 unsigned long addr, unsigned length)
1467{
1468 int i, err, count, nents, offset;
1469 struct scatterlist *sg;
1470 struct page **pages;
1471 struct nvme_iod *iod;
1472
1473 if (addr & 3)
1474 return ERR_PTR(-EINVAL);
1475 if (!length || length > INT_MAX - PAGE_SIZE)
1476 return ERR_PTR(-EINVAL);
1477
1478 offset = offset_in_page(addr);
1479 count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1480 pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1481 if (!pages)
1482 return ERR_PTR(-ENOMEM);
1483
1484 err = get_user_pages_fast(addr, count, 1, pages);
1485 if (err < count) {
1486 count = err;
1487 err = -EFAULT;
1488 goto put_pages;
1489 }
1490
1491 iod = nvme_alloc_iod(count, length, GFP_KERNEL);
1492 sg = iod->sg;
1493 sg_init_table(sg, count);
1494 for (i = 0; i < count; i++) {
1495 sg_set_page(&sg[i], pages[i],
1496 min_t(unsigned, length, PAGE_SIZE - offset),
1497 offset);
1498 length -= (PAGE_SIZE - offset);
1499 offset = 0;
1500 }
1501 sg_mark_end(&sg[i - 1]);
1502 iod->nents = count;
1503
1504 err = -ENOMEM;
1505 nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1506 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1507 if (!nents)
1508 goto free_iod;
1509
1510 kfree(pages);
1511 return iod;
1512
1513 free_iod:
1514 kfree(iod);
1515 put_pages:
1516 for (i = 0; i < count; i++)
1517 put_page(pages[i]);
1518 kfree(pages);
1519 return ERR_PTR(err);
1520}
1521
1522void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1523 struct nvme_iod *iod)
1524{
1525 int i;
1526
1527 dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1528 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1529
1530 for (i = 0; i < iod->nents; i++)
1531 put_page(sg_page(&iod->sg[i]));
1532}
1533
1534static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1535{
1536 struct nvme_dev *dev = ns->dev;
1537 struct nvme_user_io io;
1538 struct nvme_command c;
1539 unsigned length, meta_len;
1540 int status, i;
1541 struct nvme_iod *iod, *meta_iod = NULL;
1542 dma_addr_t meta_dma_addr;
1543 void *meta, *uninitialized_var(meta_mem);
1544
1545 if (copy_from_user(&io, uio, sizeof(io)))
1546 return -EFAULT;
1547 length = (io.nblocks + 1) << ns->lba_shift;
1548 meta_len = (io.nblocks + 1) * ns->ms;
1549
1550 if (meta_len && ((io.metadata & 3) || !io.metadata))
1551 return -EINVAL;
1552
1553 switch (io.opcode) {
1554 case nvme_cmd_write:
1555 case nvme_cmd_read:
1556 case nvme_cmd_compare:
1557 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1558 break;
1559 default:
1560 return -EINVAL;
1561 }
1562
1563 if (IS_ERR(iod))
1564 return PTR_ERR(iod);
1565
1566 memset(&c, 0, sizeof(c));
1567 c.rw.opcode = io.opcode;
1568 c.rw.flags = io.flags;
1569 c.rw.nsid = cpu_to_le32(ns->ns_id);
1570 c.rw.slba = cpu_to_le64(io.slba);
1571 c.rw.length = cpu_to_le16(io.nblocks);
1572 c.rw.control = cpu_to_le16(io.control);
1573 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1574 c.rw.reftag = cpu_to_le32(io.reftag);
1575 c.rw.apptag = cpu_to_le16(io.apptag);
1576 c.rw.appmask = cpu_to_le16(io.appmask);
1577
1578 if (meta_len) {
1579 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1580 meta_len);
1581 if (IS_ERR(meta_iod)) {
1582 status = PTR_ERR(meta_iod);
1583 meta_iod = NULL;
1584 goto unmap;
1585 }
1586
1587 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1588 &meta_dma_addr, GFP_KERNEL);
1589 if (!meta_mem) {
1590 status = -ENOMEM;
1591 goto unmap;
1592 }
1593
1594 if (io.opcode & 1) {
1595 int meta_offset = 0;
1596
1597 for (i = 0; i < meta_iod->nents; i++) {
1598 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1599 meta_iod->sg[i].offset;
1600 memcpy(meta_mem + meta_offset, meta,
1601 meta_iod->sg[i].length);
1602 kunmap_atomic(meta);
1603 meta_offset += meta_iod->sg[i].length;
1604 }
1605 }
1606
1607 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1608 }
1609
1610 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1611 c.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1612 c.rw.prp2 = cpu_to_le64(iod->first_dma);
1613
1614 if (length != (io.nblocks + 1) << ns->lba_shift)
1615 status = -ENOMEM;
1616 else
1617 status = nvme_submit_io_cmd(dev, &c, NULL);
1618
1619 if (meta_len) {
1620 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1621 int meta_offset = 0;
1622
1623 for (i = 0; i < meta_iod->nents; i++) {
1624 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1625 meta_iod->sg[i].offset;
1626 memcpy(meta, meta_mem + meta_offset,
1627 meta_iod->sg[i].length);
1628 kunmap_atomic(meta);
1629 meta_offset += meta_iod->sg[i].length;
1630 }
1631 }
1632
1633 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1634 meta_dma_addr);
1635 }
1636
1637 unmap:
1638 nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1639 nvme_free_iod(dev, iod);
1640
1641 if (meta_iod) {
1642 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1643 nvme_free_iod(dev, meta_iod);
1644 }
1645
1646 return status;
1647}
1648
1649static int nvme_user_admin_cmd(struct nvme_dev *dev,
1650 struct nvme_admin_cmd __user *ucmd)
1651{
1652 struct nvme_admin_cmd cmd;
1653 struct nvme_command c;
1654 int status, length;
1655 struct nvme_iod *uninitialized_var(iod);
1656 unsigned timeout;
1657
1658 if (!capable(CAP_SYS_ADMIN))
1659 return -EACCES;
1660 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1661 return -EFAULT;
1662
1663 memset(&c, 0, sizeof(c));
1664 c.common.opcode = cmd.opcode;
1665 c.common.flags = cmd.flags;
1666 c.common.nsid = cpu_to_le32(cmd.nsid);
1667 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1668 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1669 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1670 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1671 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1672 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1673 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1674 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1675
1676 length = cmd.data_len;
1677 if (cmd.data_len) {
1678 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1679 length);
1680 if (IS_ERR(iod))
1681 return PTR_ERR(iod);
1682 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1683 c.common.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1684 c.common.prp2 = cpu_to_le64(iod->first_dma);
1685 }
1686
1687 timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1688 ADMIN_TIMEOUT;
1689 if (length != cmd.data_len)
1690 status = -ENOMEM;
1691 else
1692 status = nvme_submit_sync_cmd(dev, 0, &c, &cmd.result, timeout);
1693
1694 if (cmd.data_len) {
1695 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1696 nvme_free_iod(dev, iod);
1697 }
1698
1699 if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1700 sizeof(cmd.result)))
1701 status = -EFAULT;
1702
1703 return status;
1704}
1705
1706static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1707 unsigned long arg)
1708{
1709 struct nvme_ns *ns = bdev->bd_disk->private_data;
1710
1711 switch (cmd) {
1712 case NVME_IOCTL_ID:
1713 force_successful_syscall_return();
1714 return ns->ns_id;
1715 case NVME_IOCTL_ADMIN_CMD:
1716 return nvme_user_admin_cmd(ns->dev, (void __user *)arg);
1717 case NVME_IOCTL_SUBMIT_IO:
1718 return nvme_submit_io(ns, (void __user *)arg);
1719 case SG_GET_VERSION_NUM:
1720 return nvme_sg_get_version_num((void __user *)arg);
1721 case SG_IO:
1722 return nvme_sg_io(ns, (void __user *)arg);
1723 default:
1724 return -ENOTTY;
1725 }
1726}
1727
1728#ifdef CONFIG_COMPAT
1729static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1730 unsigned int cmd, unsigned long arg)
1731{
1732 struct nvme_ns *ns = bdev->bd_disk->private_data;
1733
1734 switch (cmd) {
1735 case SG_IO:
1736 return nvme_sg_io32(ns, arg);
1737 }
1738 return nvme_ioctl(bdev, mode, cmd, arg);
1739}
1740#else
1741#define nvme_compat_ioctl NULL
1742#endif
1743
1744static int nvme_open(struct block_device *bdev, fmode_t mode)
1745{
1746 struct nvme_ns *ns = bdev->bd_disk->private_data;
1747 struct nvme_dev *dev = ns->dev;
1748
1749 kref_get(&dev->kref);
1750 return 0;
1751}
1752
1753static void nvme_free_dev(struct kref *kref);
1754
1755static void nvme_release(struct gendisk *disk, fmode_t mode)
1756{
1757 struct nvme_ns *ns = disk->private_data;
1758 struct nvme_dev *dev = ns->dev;
1759
1760 kref_put(&dev->kref, nvme_free_dev);
1761}
1762
1763static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1764{
1765 /* some standard values */
1766 geo->heads = 1 << 6;
1767 geo->sectors = 1 << 5;
1768 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1769 return 0;
1770}
1771
1772static const struct block_device_operations nvme_fops = {
1773 .owner = THIS_MODULE,
1774 .ioctl = nvme_ioctl,
1775 .compat_ioctl = nvme_compat_ioctl,
1776 .open = nvme_open,
1777 .release = nvme_release,
1778 .getgeo = nvme_getgeo,
1779};
1780
1781static void nvme_resubmit_iods(struct nvme_queue *nvmeq)
1782{
1783 struct nvme_iod *iod, *next;
1784
1785 list_for_each_entry_safe(iod, next, &nvmeq->iod_bio, node) {
1786 if (unlikely(nvme_submit_iod(nvmeq, iod)))
1787 break;
1788 list_del(&iod->node);
1789 if (bio_list_empty(&nvmeq->sq_cong) &&
1790 list_empty(&nvmeq->iod_bio))
1791 remove_wait_queue(&nvmeq->sq_full,
1792 &nvmeq->sq_cong_wait);
1793 }
1794}
1795
1796static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
1797{
1798 while (bio_list_peek(&nvmeq->sq_cong)) {
1799 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1800 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
1801
1802 if (bio_list_empty(&nvmeq->sq_cong) &&
1803 list_empty(&nvmeq->iod_bio))
1804 remove_wait_queue(&nvmeq->sq_full,
1805 &nvmeq->sq_cong_wait);
1806 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
1807 if (!waitqueue_active(&nvmeq->sq_full))
1808 add_wait_queue(&nvmeq->sq_full,
1809 &nvmeq->sq_cong_wait);
1810 bio_list_add_head(&nvmeq->sq_cong, bio);
1811 break;
1812 }
1813 }
1814}
1815
1816static int nvme_kthread(void *data)
1817{
1818 struct nvme_dev *dev, *next;
1819
1820 while (!kthread_should_stop()) {
1821 set_current_state(TASK_INTERRUPTIBLE);
1822 spin_lock(&dev_list_lock);
1823 list_for_each_entry_safe(dev, next, &dev_list, node) {
1824 int i;
1825 if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1826 dev->initialized) {
1827 if (work_busy(&dev->reset_work))
1828 continue;
1829 list_del_init(&dev->node);
1830 dev_warn(&dev->pci_dev->dev,
1831 "Failed status, reset controller\n");
1832 dev->reset_workfn = nvme_reset_failed_dev;
1833 queue_work(nvme_workq, &dev->reset_work);
1834 continue;
1835 }
1836 rcu_read_lock();
1837 for (i = 0; i < dev->queue_count; i++) {
1838 struct nvme_queue *nvmeq =
1839 rcu_dereference(dev->queues[i]);
1840 if (!nvmeq)
1841 continue;
1842 spin_lock_irq(&nvmeq->q_lock);
1843 if (nvmeq->q_suspended)
1844 goto unlock;
1845 nvme_process_cq(nvmeq);
1846 nvme_cancel_ios(nvmeq, true);
1847 nvme_resubmit_bios(nvmeq);
1848 nvme_resubmit_iods(nvmeq);
1849 unlock:
1850 spin_unlock_irq(&nvmeq->q_lock);
1851 }
1852 rcu_read_unlock();
1853 }
1854 spin_unlock(&dev_list_lock);
1855 schedule_timeout(round_jiffies_relative(HZ));
1856 }
1857 return 0;
1858}
1859
1860static void nvme_config_discard(struct nvme_ns *ns)
1861{
1862 u32 logical_block_size = queue_logical_block_size(ns->queue);
1863 ns->queue->limits.discard_zeroes_data = 0;
1864 ns->queue->limits.discard_alignment = logical_block_size;
1865 ns->queue->limits.discard_granularity = logical_block_size;
1866 ns->queue->limits.max_discard_sectors = 0xffffffff;
1867 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1868}
1869
1870static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1871 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1872{
1873 struct nvme_ns *ns;
1874 struct gendisk *disk;
1875 int lbaf;
1876
1877 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1878 return NULL;
1879
1880 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1881 if (!ns)
1882 return NULL;
1883 ns->queue = blk_alloc_queue(GFP_KERNEL);
1884 if (!ns->queue)
1885 goto out_free_ns;
1886 ns->queue->queue_flags = QUEUE_FLAG_DEFAULT;
1887 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1888 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1889 blk_queue_make_request(ns->queue, nvme_make_request);
1890 ns->dev = dev;
1891 ns->queue->queuedata = ns;
1892
1893 disk = alloc_disk(0);
1894 if (!disk)
1895 goto out_free_queue;
1896 ns->ns_id = nsid;
1897 ns->disk = disk;
1898 lbaf = id->flbas & 0xf;
1899 ns->lba_shift = id->lbaf[lbaf].ds;
1900 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1901 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1902 if (dev->max_hw_sectors)
1903 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
1904 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
1905 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
1906
1907 disk->major = nvme_major;
1908 disk->first_minor = 0;
1909 disk->fops = &nvme_fops;
1910 disk->private_data = ns;
1911 disk->queue = ns->queue;
1912 disk->driverfs_dev = &dev->pci_dev->dev;
1913 disk->flags = GENHD_FL_EXT_DEVT;
1914 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
1915 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1916
1917 if (dev->oncs & NVME_CTRL_ONCS_DSM)
1918 nvme_config_discard(ns);
1919
1920 return ns;
1921
1922 out_free_queue:
1923 blk_cleanup_queue(ns->queue);
1924 out_free_ns:
1925 kfree(ns);
1926 return NULL;
1927}
1928
1929static int nvme_find_closest_node(int node)
1930{
1931 int n, val, min_val = INT_MAX, best_node = node;
1932
1933 for_each_online_node(n) {
1934 if (n == node)
1935 continue;
1936 val = node_distance(node, n);
1937 if (val < min_val) {
1938 min_val = val;
1939 best_node = n;
1940 }
1941 }
1942 return best_node;
1943}
1944
1945static void nvme_set_queue_cpus(cpumask_t *qmask, struct nvme_queue *nvmeq,
1946 int count)
1947{
1948 int cpu;
1949 for_each_cpu(cpu, qmask) {
1950 if (cpumask_weight(nvmeq->cpu_mask) >= count)
1951 break;
1952 if (!cpumask_test_and_set_cpu(cpu, nvmeq->cpu_mask))
1953 *per_cpu_ptr(nvmeq->dev->io_queue, cpu) = nvmeq->qid;
1954 }
1955}
1956
1957static void nvme_add_cpus(cpumask_t *mask, const cpumask_t *unassigned_cpus,
1958 const cpumask_t *new_mask, struct nvme_queue *nvmeq, int cpus_per_queue)
1959{
1960 int next_cpu;
1961 for_each_cpu(next_cpu, new_mask) {
1962 cpumask_or(mask, mask, get_cpu_mask(next_cpu));
1963 cpumask_or(mask, mask, topology_thread_cpumask(next_cpu));
1964 cpumask_and(mask, mask, unassigned_cpus);
1965 nvme_set_queue_cpus(mask, nvmeq, cpus_per_queue);
1966 }
1967}
1968
1969static void nvme_create_io_queues(struct nvme_dev *dev)
1970{
1971 unsigned i, max;
1972
1973 max = min(dev->max_qid, num_online_cpus());
1974 for (i = dev->queue_count; i <= max; i++)
1975 if (!nvme_alloc_queue(dev, i, dev->q_depth, i - 1))
1976 break;
1977
1978 max = min(dev->queue_count - 1, num_online_cpus());
1979 for (i = dev->online_queues; i <= max; i++)
1980 if (nvme_create_queue(raw_nvmeq(dev, i), i))
1981 break;
1982}
1983
1984/*
1985 * If there are fewer queues than online cpus, this will try to optimally
1986 * assign a queue to multiple cpus by grouping cpus that are "close" together:
1987 * thread siblings, core, socket, closest node, then whatever else is
1988 * available.
1989 */
1990static void nvme_assign_io_queues(struct nvme_dev *dev)
1991{
1992 unsigned cpu, cpus_per_queue, queues, remainder, i;
1993 cpumask_var_t unassigned_cpus;
1994
1995 nvme_create_io_queues(dev);
1996
1997 queues = min(dev->online_queues - 1, num_online_cpus());
1998 if (!queues)
1999 return;
2000
2001 cpus_per_queue = num_online_cpus() / queues;
2002 remainder = queues - (num_online_cpus() - queues * cpus_per_queue);
2003
2004 if (!alloc_cpumask_var(&unassigned_cpus, GFP_KERNEL))
2005 return;
2006
2007 cpumask_copy(unassigned_cpus, cpu_online_mask);
2008 cpu = cpumask_first(unassigned_cpus);
2009 for (i = 1; i <= queues; i++) {
2010 struct nvme_queue *nvmeq = lock_nvmeq(dev, i);
2011 cpumask_t mask;
2012
2013 cpumask_clear(nvmeq->cpu_mask);
2014 if (!cpumask_weight(unassigned_cpus)) {
2015 unlock_nvmeq(nvmeq);
2016 break;
2017 }
2018
2019 mask = *get_cpu_mask(cpu);
2020 nvme_set_queue_cpus(&mask, nvmeq, cpus_per_queue);
2021 if (cpus_weight(mask) < cpus_per_queue)
2022 nvme_add_cpus(&mask, unassigned_cpus,
2023 topology_thread_cpumask(cpu),
2024 nvmeq, cpus_per_queue);
2025 if (cpus_weight(mask) < cpus_per_queue)
2026 nvme_add_cpus(&mask, unassigned_cpus,
2027 topology_core_cpumask(cpu),
2028 nvmeq, cpus_per_queue);
2029 if (cpus_weight(mask) < cpus_per_queue)
2030 nvme_add_cpus(&mask, unassigned_cpus,
2031 cpumask_of_node(cpu_to_node(cpu)),
2032 nvmeq, cpus_per_queue);
2033 if (cpus_weight(mask) < cpus_per_queue)
2034 nvme_add_cpus(&mask, unassigned_cpus,
2035 cpumask_of_node(
2036 nvme_find_closest_node(
2037 cpu_to_node(cpu))),
2038 nvmeq, cpus_per_queue);
2039 if (cpus_weight(mask) < cpus_per_queue)
2040 nvme_add_cpus(&mask, unassigned_cpus,
2041 unassigned_cpus,
2042 nvmeq, cpus_per_queue);
2043
2044 WARN(cpumask_weight(nvmeq->cpu_mask) != cpus_per_queue,
2045 "nvme%d qid:%d mis-matched queue-to-cpu assignment\n",
2046 dev->instance, i);
2047
2048 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2049 nvmeq->cpu_mask);
2050 cpumask_andnot(unassigned_cpus, unassigned_cpus,
2051 nvmeq->cpu_mask);
2052 cpu = cpumask_next(cpu, unassigned_cpus);
2053 if (remainder && !--remainder)
2054 cpus_per_queue++;
2055 unlock_nvmeq(nvmeq);
2056 }
2057 WARN(cpumask_weight(unassigned_cpus), "nvme%d unassigned online cpus\n",
2058 dev->instance);
2059 i = 0;
2060 cpumask_andnot(unassigned_cpus, cpu_possible_mask, cpu_online_mask);
2061 for_each_cpu(cpu, unassigned_cpus)
2062 *per_cpu_ptr(dev->io_queue, cpu) = (i++ % queues) + 1;
2063 free_cpumask_var(unassigned_cpus);
2064}
2065
2066static int set_queue_count(struct nvme_dev *dev, int count)
2067{
2068 int status;
2069 u32 result;
2070 u32 q_count = (count - 1) | ((count - 1) << 16);
2071
2072 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2073 &result);
2074 if (status < 0)
2075 return status;
2076 if (status > 0) {
2077 dev_err(&dev->pci_dev->dev, "Could not set queue count (%d)\n",
2078 status);
2079 return -EBUSY;
2080 }
2081 return min(result & 0xffff, result >> 16) + 1;
2082}
2083
2084static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2085{
2086 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2087}
2088
2089static int nvme_cpu_notify(struct notifier_block *self,
2090 unsigned long action, void *hcpu)
2091{
2092 struct nvme_dev *dev = container_of(self, struct nvme_dev, nb);
2093 switch (action) {
2094 case CPU_ONLINE:
2095 case CPU_DEAD:
2096 nvme_assign_io_queues(dev);
2097 break;
2098 }
2099 return NOTIFY_OK;
2100}
2101
2102static int nvme_setup_io_queues(struct nvme_dev *dev)
2103{
2104 struct nvme_queue *adminq = raw_nvmeq(dev, 0);
2105 struct pci_dev *pdev = dev->pci_dev;
2106 int result, i, vecs, nr_io_queues, size;
2107
2108 nr_io_queues = num_possible_cpus();
2109 result = set_queue_count(dev, nr_io_queues);
2110 if (result < 0)
2111 return result;
2112 if (result < nr_io_queues)
2113 nr_io_queues = result;
2114
2115 size = db_bar_size(dev, nr_io_queues);
2116 if (size > 8192) {
2117 iounmap(dev->bar);
2118 do {
2119 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2120 if (dev->bar)
2121 break;
2122 if (!--nr_io_queues)
2123 return -ENOMEM;
2124 size = db_bar_size(dev, nr_io_queues);
2125 } while (1);
2126 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2127 adminq->q_db = dev->dbs;
2128 }
2129
2130 /* Deregister the admin queue's interrupt */
2131 free_irq(dev->entry[0].vector, adminq);
2132
2133 for (i = 0; i < nr_io_queues; i++)
2134 dev->entry[i].entry = i;
2135 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2136 if (vecs < 0) {
2137 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2138 if (vecs < 0) {
2139 vecs = 1;
2140 } else {
2141 for (i = 0; i < vecs; i++)
2142 dev->entry[i].vector = i + pdev->irq;
2143 }
2144 }
2145
2146 /*
2147 * Should investigate if there's a performance win from allocating
2148 * more queues than interrupt vectors; it might allow the submission
2149 * path to scale better, even if the receive path is limited by the
2150 * number of interrupts.
2151 */
2152 nr_io_queues = vecs;
2153 dev->max_qid = nr_io_queues;
2154
2155 result = queue_request_irq(dev, adminq, adminq->irqname);
2156 if (result) {
2157 adminq->q_suspended = 1;
2158 goto free_queues;
2159 }
2160
2161 /* Free previously allocated queues that are no longer usable */
2162 nvme_free_queues(dev, nr_io_queues + 1);
2163 nvme_assign_io_queues(dev);
2164
2165 dev->nb.notifier_call = &nvme_cpu_notify;
2166 result = register_hotcpu_notifier(&dev->nb);
2167 if (result)
2168 goto free_queues;
2169
2170 return 0;
2171
2172 free_queues:
2173 nvme_free_queues(dev, 1);
2174 return result;
2175}
2176
2177/*
2178 * Return: error value if an error occurred setting up the queues or calling
2179 * Identify Device. 0 if these succeeded, even if adding some of the
2180 * namespaces failed. At the moment, these failures are silent. TBD which
2181 * failures should be reported.
2182 */
2183static int nvme_dev_add(struct nvme_dev *dev)
2184{
2185 struct pci_dev *pdev = dev->pci_dev;
2186 int res;
2187 unsigned nn, i;
2188 struct nvme_ns *ns;
2189 struct nvme_id_ctrl *ctrl;
2190 struct nvme_id_ns *id_ns;
2191 void *mem;
2192 dma_addr_t dma_addr;
2193 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2194
2195 mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
2196 if (!mem)
2197 return -ENOMEM;
2198
2199 res = nvme_identify(dev, 0, 1, dma_addr);
2200 if (res) {
2201 dev_err(&pdev->dev, "Identify Controller failed (%d)\n", res);
2202 res = -EIO;
2203 goto out;
2204 }
2205
2206 ctrl = mem;
2207 nn = le32_to_cpup(&ctrl->nn);
2208 dev->oncs = le16_to_cpup(&ctrl->oncs);
2209 dev->abort_limit = ctrl->acl + 1;
2210 dev->vwc = ctrl->vwc;
2211 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2212 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2213 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2214 if (ctrl->mdts)
2215 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2216 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2217 (pdev->device == 0x0953) && ctrl->vs[3])
2218 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2219
2220 id_ns = mem;
2221 for (i = 1; i <= nn; i++) {
2222 res = nvme_identify(dev, i, 0, dma_addr);
2223 if (res)
2224 continue;
2225
2226 if (id_ns->ncap == 0)
2227 continue;
2228
2229 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
2230 dma_addr + 4096, NULL);
2231 if (res)
2232 memset(mem + 4096, 0, 4096);
2233
2234 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
2235 if (ns)
2236 list_add_tail(&ns->list, &dev->namespaces);
2237 }
2238 list_for_each_entry(ns, &dev->namespaces, list)
2239 add_disk(ns->disk);
2240 res = 0;
2241
2242 out:
2243 dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
2244 return res;
2245}
2246
2247static int nvme_dev_map(struct nvme_dev *dev)
2248{
2249 u64 cap;
2250 int bars, result = -ENOMEM;
2251 struct pci_dev *pdev = dev->pci_dev;
2252
2253 if (pci_enable_device_mem(pdev))
2254 return result;
2255
2256 dev->entry[0].vector = pdev->irq;
2257 pci_set_master(pdev);
2258 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2259 if (pci_request_selected_regions(pdev, bars, "nvme"))
2260 goto disable_pci;
2261
2262 if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2263 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2264 goto disable;
2265
2266 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2267 if (!dev->bar)
2268 goto disable;
2269 if (readl(&dev->bar->csts) == -1) {
2270 result = -ENODEV;
2271 goto unmap;
2272 }
2273 cap = readq(&dev->bar->cap);
2274 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2275 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2276 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2277
2278 return 0;
2279
2280 unmap:
2281 iounmap(dev->bar);
2282 dev->bar = NULL;
2283 disable:
2284 pci_release_regions(pdev);
2285 disable_pci:
2286 pci_disable_device(pdev);
2287 return result;
2288}
2289
2290static void nvme_dev_unmap(struct nvme_dev *dev)
2291{
2292 if (dev->pci_dev->msi_enabled)
2293 pci_disable_msi(dev->pci_dev);
2294 else if (dev->pci_dev->msix_enabled)
2295 pci_disable_msix(dev->pci_dev);
2296
2297 if (dev->bar) {
2298 iounmap(dev->bar);
2299 dev->bar = NULL;
2300 pci_release_regions(dev->pci_dev);
2301 }
2302
2303 if (pci_is_enabled(dev->pci_dev))
2304 pci_disable_device(dev->pci_dev);
2305}
2306
2307struct nvme_delq_ctx {
2308 struct task_struct *waiter;
2309 struct kthread_worker *worker;
2310 atomic_t refcount;
2311};
2312
2313static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2314{
2315 dq->waiter = current;
2316 mb();
2317
2318 for (;;) {
2319 set_current_state(TASK_KILLABLE);
2320 if (!atomic_read(&dq->refcount))
2321 break;
2322 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2323 fatal_signal_pending(current)) {
2324 set_current_state(TASK_RUNNING);
2325
2326 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2327 nvme_disable_queue(dev, 0);
2328
2329 send_sig(SIGKILL, dq->worker->task, 1);
2330 flush_kthread_worker(dq->worker);
2331 return;
2332 }
2333 }
2334 set_current_state(TASK_RUNNING);
2335}
2336
2337static void nvme_put_dq(struct nvme_delq_ctx *dq)
2338{
2339 atomic_dec(&dq->refcount);
2340 if (dq->waiter)
2341 wake_up_process(dq->waiter);
2342}
2343
2344static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2345{
2346 atomic_inc(&dq->refcount);
2347 return dq;
2348}
2349
2350static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2351{
2352 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2353
2354 nvme_clear_queue(nvmeq);
2355 nvme_put_dq(dq);
2356}
2357
2358static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2359 kthread_work_func_t fn)
2360{
2361 struct nvme_command c;
2362
2363 memset(&c, 0, sizeof(c));
2364 c.delete_queue.opcode = opcode;
2365 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2366
2367 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2368 return nvme_submit_admin_cmd_async(nvmeq->dev, &c, &nvmeq->cmdinfo);
2369}
2370
2371static void nvme_del_cq_work_handler(struct kthread_work *work)
2372{
2373 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2374 cmdinfo.work);
2375 nvme_del_queue_end(nvmeq);
2376}
2377
2378static int nvme_delete_cq(struct nvme_queue *nvmeq)
2379{
2380 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2381 nvme_del_cq_work_handler);
2382}
2383
2384static void nvme_del_sq_work_handler(struct kthread_work *work)
2385{
2386 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2387 cmdinfo.work);
2388 int status = nvmeq->cmdinfo.status;
2389
2390 if (!status)
2391 status = nvme_delete_cq(nvmeq);
2392 if (status)
2393 nvme_del_queue_end(nvmeq);
2394}
2395
2396static int nvme_delete_sq(struct nvme_queue *nvmeq)
2397{
2398 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2399 nvme_del_sq_work_handler);
2400}
2401
2402static void nvme_del_queue_start(struct kthread_work *work)
2403{
2404 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2405 cmdinfo.work);
2406 allow_signal(SIGKILL);
2407 if (nvme_delete_sq(nvmeq))
2408 nvme_del_queue_end(nvmeq);
2409}
2410
2411static void nvme_disable_io_queues(struct nvme_dev *dev)
2412{
2413 int i;
2414 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2415 struct nvme_delq_ctx dq;
2416 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2417 &worker, "nvme%d", dev->instance);
2418
2419 if (IS_ERR(kworker_task)) {
2420 dev_err(&dev->pci_dev->dev,
2421 "Failed to create queue del task\n");
2422 for (i = dev->queue_count - 1; i > 0; i--)
2423 nvme_disable_queue(dev, i);
2424 return;
2425 }
2426
2427 dq.waiter = NULL;
2428 atomic_set(&dq.refcount, 0);
2429 dq.worker = &worker;
2430 for (i = dev->queue_count - 1; i > 0; i--) {
2431 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2432
2433 if (nvme_suspend_queue(nvmeq))
2434 continue;
2435 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2436 nvmeq->cmdinfo.worker = dq.worker;
2437 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2438 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2439 }
2440 nvme_wait_dq(&dq, dev);
2441 kthread_stop(kworker_task);
2442}
2443
2444/*
2445* Remove the node from the device list and check
2446* for whether or not we need to stop the nvme_thread.
2447*/
2448static void nvme_dev_list_remove(struct nvme_dev *dev)
2449{
2450 struct task_struct *tmp = NULL;
2451
2452 spin_lock(&dev_list_lock);
2453 list_del_init(&dev->node);
2454 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2455 tmp = nvme_thread;
2456 nvme_thread = NULL;
2457 }
2458 spin_unlock(&dev_list_lock);
2459
2460 if (tmp)
2461 kthread_stop(tmp);
2462}
2463
2464static void nvme_dev_shutdown(struct nvme_dev *dev)
2465{
2466 int i;
2467
2468 dev->initialized = 0;
2469 unregister_hotcpu_notifier(&dev->nb);
2470
2471 nvme_dev_list_remove(dev);
2472
2473 if (!dev->bar || (dev->bar && readl(&dev->bar->csts) == -1)) {
2474 for (i = dev->queue_count - 1; i >= 0; i--) {
2475 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2476 nvme_suspend_queue(nvmeq);
2477 nvme_clear_queue(nvmeq);
2478 }
2479 } else {
2480 nvme_disable_io_queues(dev);
2481 nvme_shutdown_ctrl(dev);
2482 nvme_disable_queue(dev, 0);
2483 }
2484 nvme_dev_unmap(dev);
2485}
2486
2487static void nvme_dev_remove(struct nvme_dev *dev)
2488{
2489 struct nvme_ns *ns;
2490
2491 list_for_each_entry(ns, &dev->namespaces, list) {
2492 if (ns->disk->flags & GENHD_FL_UP)
2493 del_gendisk(ns->disk);
2494 if (!blk_queue_dying(ns->queue))
2495 blk_cleanup_queue(ns->queue);
2496 }
2497}
2498
2499static int nvme_setup_prp_pools(struct nvme_dev *dev)
2500{
2501 struct device *dmadev = &dev->pci_dev->dev;
2502 dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2503 PAGE_SIZE, PAGE_SIZE, 0);
2504 if (!dev->prp_page_pool)
2505 return -ENOMEM;
2506
2507 /* Optimisation for I/Os between 4k and 128k */
2508 dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2509 256, 256, 0);
2510 if (!dev->prp_small_pool) {
2511 dma_pool_destroy(dev->prp_page_pool);
2512 return -ENOMEM;
2513 }
2514 return 0;
2515}
2516
2517static void nvme_release_prp_pools(struct nvme_dev *dev)
2518{
2519 dma_pool_destroy(dev->prp_page_pool);
2520 dma_pool_destroy(dev->prp_small_pool);
2521}
2522
2523static DEFINE_IDA(nvme_instance_ida);
2524
2525static int nvme_set_instance(struct nvme_dev *dev)
2526{
2527 int instance, error;
2528
2529 do {
2530 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2531 return -ENODEV;
2532
2533 spin_lock(&dev_list_lock);
2534 error = ida_get_new(&nvme_instance_ida, &instance);
2535 spin_unlock(&dev_list_lock);
2536 } while (error == -EAGAIN);
2537
2538 if (error)
2539 return -ENODEV;
2540
2541 dev->instance = instance;
2542 return 0;
2543}
2544
2545static void nvme_release_instance(struct nvme_dev *dev)
2546{
2547 spin_lock(&dev_list_lock);
2548 ida_remove(&nvme_instance_ida, dev->instance);
2549 spin_unlock(&dev_list_lock);
2550}
2551
2552static void nvme_free_namespaces(struct nvme_dev *dev)
2553{
2554 struct nvme_ns *ns, *next;
2555
2556 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2557 list_del(&ns->list);
2558 put_disk(ns->disk);
2559 kfree(ns);
2560 }
2561}
2562
2563static void nvme_free_dev(struct kref *kref)
2564{
2565 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2566
2567 nvme_free_namespaces(dev);
2568 free_percpu(dev->io_queue);
2569 kfree(dev->queues);
2570 kfree(dev->entry);
2571 kfree(dev);
2572}
2573
2574static int nvme_dev_open(struct inode *inode, struct file *f)
2575{
2576 struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2577 miscdev);
2578 kref_get(&dev->kref);
2579 f->private_data = dev;
2580 return 0;
2581}
2582
2583static int nvme_dev_release(struct inode *inode, struct file *f)
2584{
2585 struct nvme_dev *dev = f->private_data;
2586 kref_put(&dev->kref, nvme_free_dev);
2587 return 0;
2588}
2589
2590static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2591{
2592 struct nvme_dev *dev = f->private_data;
2593 switch (cmd) {
2594 case NVME_IOCTL_ADMIN_CMD:
2595 return nvme_user_admin_cmd(dev, (void __user *)arg);
2596 default:
2597 return -ENOTTY;
2598 }
2599}
2600
2601static const struct file_operations nvme_dev_fops = {
2602 .owner = THIS_MODULE,
2603 .open = nvme_dev_open,
2604 .release = nvme_dev_release,
2605 .unlocked_ioctl = nvme_dev_ioctl,
2606 .compat_ioctl = nvme_dev_ioctl,
2607};
2608
2609static int nvme_dev_start(struct nvme_dev *dev)
2610{
2611 int result;
2612 bool start_thread = false;
2613
2614 result = nvme_dev_map(dev);
2615 if (result)
2616 return result;
2617
2618 result = nvme_configure_admin_queue(dev);
2619 if (result)
2620 goto unmap;
2621
2622 spin_lock(&dev_list_lock);
2623 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2624 start_thread = true;
2625 nvme_thread = NULL;
2626 }
2627 list_add(&dev->node, &dev_list);
2628 spin_unlock(&dev_list_lock);
2629
2630 if (start_thread) {
2631 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2632 wake_up(&nvme_kthread_wait);
2633 } else
2634 wait_event_killable(nvme_kthread_wait, nvme_thread);
2635
2636 if (IS_ERR_OR_NULL(nvme_thread)) {
2637 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2638 goto disable;
2639 }
2640
2641 result = nvme_setup_io_queues(dev);
2642 if (result && result != -EBUSY)
2643 goto disable;
2644
2645 return result;
2646
2647 disable:
2648 nvme_disable_queue(dev, 0);
2649 nvme_dev_list_remove(dev);
2650 unmap:
2651 nvme_dev_unmap(dev);
2652 return result;
2653}
2654
2655static int nvme_remove_dead_ctrl(void *arg)
2656{
2657 struct nvme_dev *dev = (struct nvme_dev *)arg;
2658 struct pci_dev *pdev = dev->pci_dev;
2659
2660 if (pci_get_drvdata(pdev))
2661 pci_stop_and_remove_bus_device(pdev);
2662 kref_put(&dev->kref, nvme_free_dev);
2663 return 0;
2664}
2665
2666static void nvme_remove_disks(struct work_struct *ws)
2667{
2668 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2669
2670 nvme_dev_remove(dev);
2671 nvme_free_queues(dev, 1);
2672}
2673
2674static int nvme_dev_resume(struct nvme_dev *dev)
2675{
2676 int ret;
2677
2678 ret = nvme_dev_start(dev);
2679 if (ret && ret != -EBUSY)
2680 return ret;
2681 if (ret == -EBUSY) {
2682 spin_lock(&dev_list_lock);
2683 dev->reset_workfn = nvme_remove_disks;
2684 queue_work(nvme_workq, &dev->reset_work);
2685 spin_unlock(&dev_list_lock);
2686 }
2687 dev->initialized = 1;
2688 return 0;
2689}
2690
2691static void nvme_dev_reset(struct nvme_dev *dev)
2692{
2693 nvme_dev_shutdown(dev);
2694 if (nvme_dev_resume(dev)) {
2695 dev_err(&dev->pci_dev->dev, "Device failed to resume\n");
2696 kref_get(&dev->kref);
2697 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2698 dev->instance))) {
2699 dev_err(&dev->pci_dev->dev,
2700 "Failed to start controller remove task\n");
2701 kref_put(&dev->kref, nvme_free_dev);
2702 }
2703 }
2704}
2705
2706static void nvme_reset_failed_dev(struct work_struct *ws)
2707{
2708 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2709 nvme_dev_reset(dev);
2710}
2711
2712static void nvme_reset_workfn(struct work_struct *work)
2713{
2714 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2715 dev->reset_workfn(work);
2716}
2717
2718static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2719{
2720 int result = -ENOMEM;
2721 struct nvme_dev *dev;
2722
2723 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2724 if (!dev)
2725 return -ENOMEM;
2726 dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
2727 GFP_KERNEL);
2728 if (!dev->entry)
2729 goto free;
2730 dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
2731 GFP_KERNEL);
2732 if (!dev->queues)
2733 goto free;
2734 dev->io_queue = alloc_percpu(unsigned short);
2735 if (!dev->io_queue)
2736 goto free;
2737
2738 INIT_LIST_HEAD(&dev->namespaces);
2739 dev->reset_workfn = nvme_reset_failed_dev;
2740 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
2741 dev->pci_dev = pdev;
2742 pci_set_drvdata(pdev, dev);
2743 result = nvme_set_instance(dev);
2744 if (result)
2745 goto free;
2746
2747 result = nvme_setup_prp_pools(dev);
2748 if (result)
2749 goto release;
2750
2751 kref_init(&dev->kref);
2752 result = nvme_dev_start(dev);
2753 if (result) {
2754 if (result == -EBUSY)
2755 goto create_cdev;
2756 goto release_pools;
2757 }
2758
2759 result = nvme_dev_add(dev);
2760 if (result)
2761 goto shutdown;
2762
2763 create_cdev:
2764 scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2765 dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2766 dev->miscdev.parent = &pdev->dev;
2767 dev->miscdev.name = dev->name;
2768 dev->miscdev.fops = &nvme_dev_fops;
2769 result = misc_register(&dev->miscdev);
2770 if (result)
2771 goto remove;
2772
2773 dev->initialized = 1;
2774 return 0;
2775
2776 remove:
2777 nvme_dev_remove(dev);
2778 nvme_free_namespaces(dev);
2779 shutdown:
2780 nvme_dev_shutdown(dev);
2781 release_pools:
2782 nvme_free_queues(dev, 0);
2783 nvme_release_prp_pools(dev);
2784 release:
2785 nvme_release_instance(dev);
2786 free:
2787 free_percpu(dev->io_queue);
2788 kfree(dev->queues);
2789 kfree(dev->entry);
2790 kfree(dev);
2791 return result;
2792}
2793
2794static void nvme_shutdown(struct pci_dev *pdev)
2795{
2796 struct nvme_dev *dev = pci_get_drvdata(pdev);
2797 nvme_dev_shutdown(dev);
2798}
2799
2800static void nvme_remove(struct pci_dev *pdev)
2801{
2802 struct nvme_dev *dev = pci_get_drvdata(pdev);
2803
2804 spin_lock(&dev_list_lock);
2805 list_del_init(&dev->node);
2806 spin_unlock(&dev_list_lock);
2807
2808 pci_set_drvdata(pdev, NULL);
2809 flush_work(&dev->reset_work);
2810 misc_deregister(&dev->miscdev);
2811 nvme_dev_remove(dev);
2812 nvme_dev_shutdown(dev);
2813 nvme_free_queues(dev, 0);
2814 rcu_barrier();
2815 nvme_release_instance(dev);
2816 nvme_release_prp_pools(dev);
2817 kref_put(&dev->kref, nvme_free_dev);
2818}
2819
2820/* These functions are yet to be implemented */
2821#define nvme_error_detected NULL
2822#define nvme_dump_registers NULL
2823#define nvme_link_reset NULL
2824#define nvme_slot_reset NULL
2825#define nvme_error_resume NULL
2826
2827#ifdef CONFIG_PM_SLEEP
2828static int nvme_suspend(struct device *dev)
2829{
2830 struct pci_dev *pdev = to_pci_dev(dev);
2831 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2832
2833 nvme_dev_shutdown(ndev);
2834 return 0;
2835}
2836
2837static int nvme_resume(struct device *dev)
2838{
2839 struct pci_dev *pdev = to_pci_dev(dev);
2840 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2841
2842 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2843 ndev->reset_workfn = nvme_reset_failed_dev;
2844 queue_work(nvme_workq, &ndev->reset_work);
2845 }
2846 return 0;
2847}
2848#endif
2849
2850static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2851
2852static const struct pci_error_handlers nvme_err_handler = {
2853 .error_detected = nvme_error_detected,
2854 .mmio_enabled = nvme_dump_registers,
2855 .link_reset = nvme_link_reset,
2856 .slot_reset = nvme_slot_reset,
2857 .resume = nvme_error_resume,
2858};
2859
2860/* Move to pci_ids.h later */
2861#define PCI_CLASS_STORAGE_EXPRESS 0x010802
2862
2863static const struct pci_device_id nvme_id_table[] = {
2864 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2865 { 0, }
2866};
2867MODULE_DEVICE_TABLE(pci, nvme_id_table);
2868
2869static struct pci_driver nvme_driver = {
2870 .name = "nvme",
2871 .id_table = nvme_id_table,
2872 .probe = nvme_probe,
2873 .remove = nvme_remove,
2874 .shutdown = nvme_shutdown,
2875 .driver = {
2876 .pm = &nvme_dev_pm_ops,
2877 },
2878 .err_handler = &nvme_err_handler,
2879};
2880
2881static int __init nvme_init(void)
2882{
2883 int result;
2884
2885 init_waitqueue_head(&nvme_kthread_wait);
2886
2887 nvme_workq = create_singlethread_workqueue("nvme");
2888 if (!nvme_workq)
2889 return -ENOMEM;
2890
2891 result = register_blkdev(nvme_major, "nvme");
2892 if (result < 0)
2893 goto kill_workq;
2894 else if (result > 0)
2895 nvme_major = result;
2896
2897 result = pci_register_driver(&nvme_driver);
2898 if (result)
2899 goto unregister_blkdev;
2900 return 0;
2901
2902 unregister_blkdev:
2903 unregister_blkdev(nvme_major, "nvme");
2904 kill_workq:
2905 destroy_workqueue(nvme_workq);
2906 return result;
2907}
2908
2909static void __exit nvme_exit(void)
2910{
2911 pci_unregister_driver(&nvme_driver);
2912 unregister_blkdev(nvme_major, "nvme");
2913 destroy_workqueue(nvme_workq);
2914 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
2915}
2916
2917MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2918MODULE_LICENSE("GPL");
2919MODULE_VERSION("0.9");
2920module_init(nvme_init);
2921module_exit(nvme_exit);
Linux 6.x block layer and io_uring tree(s)