projects / linux-2.6-block.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
writeback: don't drain bdi_writeback_congested on bdi destruction
[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/bitops.h>
17#include <linux/blkdev.h>
18#include <linux/blk-mq.h>
19#include <linux/cpu.h>
20#include <linux/delay.h>
21#include <linux/errno.h>
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/list_sort.h>
33#include <linux/mm.h>
34#include <linux/module.h>
35#include <linux/moduleparam.h>
36#include <linux/pci.h>
37#include <linux/poison.h>
38#include <linux/ptrace.h>
39#include <linux/sched.h>
40#include <linux/slab.h>
41#include <linux/t10-pi.h>
42#include <linux/types.h>
43#include <scsi/sg.h>
44#include <asm-generic/io-64-nonatomic-lo-hi.h>
45
46#define NVME_MINORS (1U << MINORBITS)
47#define NVME_Q_DEPTH 1024
48#define NVME_AQ_DEPTH 256
49#define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
50#define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
51#define ADMIN_TIMEOUT (admin_timeout * HZ)
52#define SHUTDOWN_TIMEOUT (shutdown_timeout * HZ)
53
54static unsigned char admin_timeout = 60;
55module_param(admin_timeout, byte, 0644);
56MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
57
58unsigned char nvme_io_timeout = 30;
59module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
60MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
61
62static unsigned char shutdown_timeout = 5;
63module_param(shutdown_timeout, byte, 0644);
64MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
65
66static int nvme_major;
67module_param(nvme_major, int, 0);
68
69static int nvme_char_major;
70module_param(nvme_char_major, int, 0);
71
72static int use_threaded_interrupts;
73module_param(use_threaded_interrupts, int, 0);
74
75static DEFINE_SPINLOCK(dev_list_lock);
76static LIST_HEAD(dev_list);
77static struct task_struct *nvme_thread;
78static struct workqueue_struct *nvme_workq;
79static wait_queue_head_t nvme_kthread_wait;
80
81static struct class *nvme_class;
82
83static void nvme_reset_failed_dev(struct work_struct *ws);
84static int nvme_reset(struct nvme_dev *dev);
85static int nvme_process_cq(struct nvme_queue *nvmeq);
86
87struct async_cmd_info {
88 struct kthread_work work;
89 struct kthread_worker *worker;
90 struct request *req;
91 u32 result;
92 int status;
93 void *ctx;
94};
95
96/*
97 * An NVM Express queue. Each device has at least two (one for admin
98 * commands and one for I/O commands).
99 */
100struct nvme_queue {
101 struct device *q_dmadev;
102 struct nvme_dev *dev;
103 char irqname[24]; /* nvme4294967295-65535\0 */
104 spinlock_t q_lock;
105 struct nvme_command *sq_cmds;
106 volatile struct nvme_completion *cqes;
107 struct blk_mq_tags **tags;
108 dma_addr_t sq_dma_addr;
109 dma_addr_t cq_dma_addr;
110 u32 __iomem *q_db;
111 u16 q_depth;
112 s16 cq_vector;
113 u16 sq_head;
114 u16 sq_tail;
115 u16 cq_head;
116 u16 qid;
117 u8 cq_phase;
118 u8 cqe_seen;
119 struct async_cmd_info cmdinfo;
120};
121
122/*
123 * Check we didin't inadvertently grow the command struct
124 */
125static inline void _nvme_check_size(void)
126{
127 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
128 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
129 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
130 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
131 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
132 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
133 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
134 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
135 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
136 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
137 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
138 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
139}
140
141typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
142 struct nvme_completion *);
143
144struct nvme_cmd_info {
145 nvme_completion_fn fn;
146 void *ctx;
147 int aborted;
148 struct nvme_queue *nvmeq;
149 struct nvme_iod iod[0];
150};
151
152/*
153 * Max size of iod being embedded in the request payload
154 */
155#define NVME_INT_PAGES 2
156#define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->page_size)
157#define NVME_INT_MASK 0x01
158
159/*
160 * Will slightly overestimate the number of pages needed. This is OK
161 * as it only leads to a small amount of wasted memory for the lifetime of
162 * the I/O.
163 */
164static int nvme_npages(unsigned size, struct nvme_dev *dev)
165{
166 unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
167 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
168}
169
170static unsigned int nvme_cmd_size(struct nvme_dev *dev)
171{
172 unsigned int ret = sizeof(struct nvme_cmd_info);
173
174 ret += sizeof(struct nvme_iod);
175 ret += sizeof(__le64 *) * nvme_npages(NVME_INT_BYTES(dev), dev);
176 ret += sizeof(struct scatterlist) * NVME_INT_PAGES;
177
178 return ret;
179}
180
181static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
182 unsigned int hctx_idx)
183{
184 struct nvme_dev *dev = data;
185 struct nvme_queue *nvmeq = dev->queues[0];
186
187 WARN_ON(hctx_idx != 0);
188 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
189 WARN_ON(nvmeq->tags);
190
191 hctx->driver_data = nvmeq;
192 nvmeq->tags = &dev->admin_tagset.tags[0];
193 return 0;
194}
195
196static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
197{
198 struct nvme_queue *nvmeq = hctx->driver_data;
199
200 nvmeq->tags = NULL;
201}
202
203static int nvme_admin_init_request(void *data, struct request *req,
204 unsigned int hctx_idx, unsigned int rq_idx,
205 unsigned int numa_node)
206{
207 struct nvme_dev *dev = data;
208 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
209 struct nvme_queue *nvmeq = dev->queues[0];
210
211 BUG_ON(!nvmeq);
212 cmd->nvmeq = nvmeq;
213 return 0;
214}
215
216static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
217 unsigned int hctx_idx)
218{
219 struct nvme_dev *dev = data;
220 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
221
222 if (!nvmeq->tags)
223 nvmeq->tags = &dev->tagset.tags[hctx_idx];
224
225 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
226 hctx->driver_data = nvmeq;
227 return 0;
228}
229
230static int nvme_init_request(void *data, struct request *req,
231 unsigned int hctx_idx, unsigned int rq_idx,
232 unsigned int numa_node)
233{
234 struct nvme_dev *dev = data;
235 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
236 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
237
238 BUG_ON(!nvmeq);
239 cmd->nvmeq = nvmeq;
240 return 0;
241}
242
243static void nvme_set_info(struct nvme_cmd_info *cmd, void *ctx,
244 nvme_completion_fn handler)
245{
246 cmd->fn = handler;
247 cmd->ctx = ctx;
248 cmd->aborted = 0;
249 blk_mq_start_request(blk_mq_rq_from_pdu(cmd));
250}
251
252static void *iod_get_private(struct nvme_iod *iod)
253{
254 return (void *) (iod->private & ~0x1UL);
255}
256
257/*
258 * If bit 0 is set, the iod is embedded in the request payload.
259 */
260static bool iod_should_kfree(struct nvme_iod *iod)
261{
262 return (iod->private & NVME_INT_MASK) == 0;
263}
264
265/* Special values must be less than 0x1000 */
266#define CMD_CTX_BASE ((void *)POISON_POINTER_DELTA)
267#define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
268#define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
269#define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
270
271static void special_completion(struct nvme_queue *nvmeq, void *ctx,
272 struct nvme_completion *cqe)
273{
274 if (ctx == CMD_CTX_CANCELLED)
275 return;
276 if (ctx == CMD_CTX_COMPLETED) {
277 dev_warn(nvmeq->q_dmadev,
278 "completed id %d twice on queue %d\n",
279 cqe->command_id, le16_to_cpup(&cqe->sq_id));
280 return;
281 }
282 if (ctx == CMD_CTX_INVALID) {
283 dev_warn(nvmeq->q_dmadev,
284 "invalid id %d completed on queue %d\n",
285 cqe->command_id, le16_to_cpup(&cqe->sq_id));
286 return;
287 }
288 dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
289}
290
291static void *cancel_cmd_info(struct nvme_cmd_info *cmd, nvme_completion_fn *fn)
292{
293 void *ctx;
294
295 if (fn)
296 *fn = cmd->fn;
297 ctx = cmd->ctx;
298 cmd->fn = special_completion;
299 cmd->ctx = CMD_CTX_CANCELLED;
300 return ctx;
301}
302
303static void async_req_completion(struct nvme_queue *nvmeq, void *ctx,
304 struct nvme_completion *cqe)
305{
306 u32 result = le32_to_cpup(&cqe->result);
307 u16 status = le16_to_cpup(&cqe->status) >> 1;
308
309 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
310 ++nvmeq->dev->event_limit;
311 if (status != NVME_SC_SUCCESS)
312 return;
313
314 switch (result & 0xff07) {
315 case NVME_AER_NOTICE_NS_CHANGED:
316 dev_info(nvmeq->q_dmadev, "rescanning\n");
317 schedule_work(&nvmeq->dev->scan_work);
318 default:
319 dev_warn(nvmeq->q_dmadev, "async event result %08x\n", result);
320 }
321}
322
323static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
324 struct nvme_completion *cqe)
325{
326 struct request *req = ctx;
327
328 u16 status = le16_to_cpup(&cqe->status) >> 1;
329 u32 result = le32_to_cpup(&cqe->result);
330
331 blk_mq_free_request(req);
332
333 dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
334 ++nvmeq->dev->abort_limit;
335}
336
337static void async_completion(struct nvme_queue *nvmeq, void *ctx,
338 struct nvme_completion *cqe)
339{
340 struct async_cmd_info *cmdinfo = ctx;
341 cmdinfo->result = le32_to_cpup(&cqe->result);
342 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
343 queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
344 blk_mq_free_request(cmdinfo->req);
345}
346
347static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
348 unsigned int tag)
349{
350 struct request *req = blk_mq_tag_to_rq(*nvmeq->tags, tag);
351
352 return blk_mq_rq_to_pdu(req);
353}
354
355/*
356 * Called with local interrupts disabled and the q_lock held. May not sleep.
357 */
358static void *nvme_finish_cmd(struct nvme_queue *nvmeq, int tag,
359 nvme_completion_fn *fn)
360{
361 struct nvme_cmd_info *cmd = get_cmd_from_tag(nvmeq, tag);
362 void *ctx;
363 if (tag >= nvmeq->q_depth) {
364 *fn = special_completion;
365 return CMD_CTX_INVALID;
366 }
367 if (fn)
368 *fn = cmd->fn;
369 ctx = cmd->ctx;
370 cmd->fn = special_completion;
371 cmd->ctx = CMD_CTX_COMPLETED;
372 return ctx;
373}
374
375/**
376 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
377 * @nvmeq: The queue to use
378 * @cmd: The command to send
379 *
380 * Safe to use from interrupt context
381 */
382static int __nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
383{
384 u16 tail = nvmeq->sq_tail;
385
386 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
387 if (++tail == nvmeq->q_depth)
388 tail = 0;
389 writel(tail, nvmeq->q_db);
390 nvmeq->sq_tail = tail;
391
392 return 0;
393}
394
395static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
396{
397 unsigned long flags;
398 int ret;
399 spin_lock_irqsave(&nvmeq->q_lock, flags);
400 ret = __nvme_submit_cmd(nvmeq, cmd);
401 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
402 return ret;
403}
404
405static __le64 **iod_list(struct nvme_iod *iod)
406{
407 return ((void *)iod) + iod->offset;
408}
409
410static inline void iod_init(struct nvme_iod *iod, unsigned nbytes,
411 unsigned nseg, unsigned long private)
412{
413 iod->private = private;
414 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
415 iod->npages = -1;
416 iod->length = nbytes;
417 iod->nents = 0;
418}
419
420static struct nvme_iod *
421__nvme_alloc_iod(unsigned nseg, unsigned bytes, struct nvme_dev *dev,
422 unsigned long priv, gfp_t gfp)
423{
424 struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
425 sizeof(__le64 *) * nvme_npages(bytes, dev) +
426 sizeof(struct scatterlist) * nseg, gfp);
427
428 if (iod)
429 iod_init(iod, bytes, nseg, priv);
430
431 return iod;
432}
433
434static struct nvme_iod *nvme_alloc_iod(struct request *rq, struct nvme_dev *dev,
435 gfp_t gfp)
436{
437 unsigned size = !(rq->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(rq) :
438 sizeof(struct nvme_dsm_range);
439 struct nvme_iod *iod;
440
441 if (rq->nr_phys_segments <= NVME_INT_PAGES &&
442 size <= NVME_INT_BYTES(dev)) {
443 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(rq);
444
445 iod = cmd->iod;
446 iod_init(iod, size, rq->nr_phys_segments,
447 (unsigned long) rq | NVME_INT_MASK);
448 return iod;
449 }
450
451 return __nvme_alloc_iod(rq->nr_phys_segments, size, dev,
452 (unsigned long) rq, gfp);
453}
454
455static void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
456{
457 const int last_prp = dev->page_size / 8 - 1;
458 int i;
459 __le64 **list = iod_list(iod);
460 dma_addr_t prp_dma = iod->first_dma;
461
462 if (iod->npages == 0)
463 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
464 for (i = 0; i < iod->npages; i++) {
465 __le64 *prp_list = list[i];
466 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
467 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
468 prp_dma = next_prp_dma;
469 }
470
471 if (iod_should_kfree(iod))
472 kfree(iod);
473}
474
475static int nvme_error_status(u16 status)
476{
477 switch (status & 0x7ff) {
478 case NVME_SC_SUCCESS:
479 return 0;
480 case NVME_SC_CAP_EXCEEDED:
481 return -ENOSPC;
482 default:
483 return -EIO;
484 }
485}
486
487#ifdef CONFIG_BLK_DEV_INTEGRITY
488static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
489{
490 if (be32_to_cpu(pi->ref_tag) == v)
491 pi->ref_tag = cpu_to_be32(p);
492}
493
494static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
495{
496 if (be32_to_cpu(pi->ref_tag) == p)
497 pi->ref_tag = cpu_to_be32(v);
498}
499
500/**
501 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
502 *
503 * The virtual start sector is the one that was originally submitted by the
504 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
505 * start sector may be different. Remap protection information to match the
506 * physical LBA on writes, and back to the original seed on reads.
507 *
508 * Type 0 and 3 do not have a ref tag, so no remapping required.
509 */
510static void nvme_dif_remap(struct request *req,
511 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
512{
513 struct nvme_ns *ns = req->rq_disk->private_data;
514 struct bio_integrity_payload *bip;
515 struct t10_pi_tuple *pi;
516 void *p, *pmap;
517 u32 i, nlb, ts, phys, virt;
518
519 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
520 return;
521
522 bip = bio_integrity(req->bio);
523 if (!bip)
524 return;
525
526 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
527
528 p = pmap;
529 virt = bip_get_seed(bip);
530 phys = nvme_block_nr(ns, blk_rq_pos(req));
531 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
532 ts = ns->disk->integrity->tuple_size;
533
534 for (i = 0; i < nlb; i++, virt++, phys++) {
535 pi = (struct t10_pi_tuple *)p;
536 dif_swap(phys, virt, pi);
537 p += ts;
538 }
539 kunmap_atomic(pmap);
540}
541
542static int nvme_noop_verify(struct blk_integrity_iter *iter)
543{
544 return 0;
545}
546
547static int nvme_noop_generate(struct blk_integrity_iter *iter)
548{
549 return 0;
550}
551
552struct blk_integrity nvme_meta_noop = {
553 .name = "NVME_META_NOOP",
554 .generate_fn = nvme_noop_generate,
555 .verify_fn = nvme_noop_verify,
556};
557
558static void nvme_init_integrity(struct nvme_ns *ns)
559{
560 struct blk_integrity integrity;
561
562 switch (ns->pi_type) {
563 case NVME_NS_DPS_PI_TYPE3:
564 integrity = t10_pi_type3_crc;
565 break;
566 case NVME_NS_DPS_PI_TYPE1:
567 case NVME_NS_DPS_PI_TYPE2:
568 integrity = t10_pi_type1_crc;
569 break;
570 default:
571 integrity = nvme_meta_noop;
572 break;
573 }
574 integrity.tuple_size = ns->ms;
575 blk_integrity_register(ns->disk, &integrity);
576 blk_queue_max_integrity_segments(ns->queue, 1);
577}
578#else /* CONFIG_BLK_DEV_INTEGRITY */
579static void nvme_dif_remap(struct request *req,
580 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
581{
582}
583static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
584{
585}
586static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
587{
588}
589static void nvme_init_integrity(struct nvme_ns *ns)
590{
591}
592#endif
593
594static void req_completion(struct nvme_queue *nvmeq, void *ctx,
595 struct nvme_completion *cqe)
596{
597 struct nvme_iod *iod = ctx;
598 struct request *req = iod_get_private(iod);
599 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
600
601 u16 status = le16_to_cpup(&cqe->status) >> 1;
602
603 if (unlikely(status)) {
604 if (!(status & NVME_SC_DNR || blk_noretry_request(req))
605 && (jiffies - req->start_time) < req->timeout) {
606 unsigned long flags;
607
608 blk_mq_requeue_request(req);
609 spin_lock_irqsave(req->q->queue_lock, flags);
610 if (!blk_queue_stopped(req->q))
611 blk_mq_kick_requeue_list(req->q);
612 spin_unlock_irqrestore(req->q->queue_lock, flags);
613 return;
614 }
615 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
616 if (cmd_rq->ctx == CMD_CTX_CANCELLED)
617 req->errors = -EINTR;
618 else
619 req->errors = status;
620 } else {
621 req->errors = nvme_error_status(status);
622 }
623 } else
624 req->errors = 0;
625 if (req->cmd_type == REQ_TYPE_DRV_PRIV) {
626 u32 result = le32_to_cpup(&cqe->result);
627 req->special = (void *)(uintptr_t)result;
628 }
629
630 if (cmd_rq->aborted)
631 dev_warn(nvmeq->dev->dev,
632 "completing aborted command with status:%04x\n",
633 status);
634
635 if (iod->nents) {
636 dma_unmap_sg(nvmeq->dev->dev, iod->sg, iod->nents,
637 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
638 if (blk_integrity_rq(req)) {
639 if (!rq_data_dir(req))
640 nvme_dif_remap(req, nvme_dif_complete);
641 dma_unmap_sg(nvmeq->dev->dev, iod->meta_sg, 1,
642 rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
643 }
644 }
645 nvme_free_iod(nvmeq->dev, iod);
646
647 blk_mq_complete_request(req);
648}
649
650/* length is in bytes. gfp flags indicates whether we may sleep. */
651static int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
652 int total_len, gfp_t gfp)
653{
654 struct dma_pool *pool;
655 int length = total_len;
656 struct scatterlist *sg = iod->sg;
657 int dma_len = sg_dma_len(sg);
658 u64 dma_addr = sg_dma_address(sg);
659 u32 page_size = dev->page_size;
660 int offset = dma_addr & (page_size - 1);
661 __le64 *prp_list;
662 __le64 **list = iod_list(iod);
663 dma_addr_t prp_dma;
664 int nprps, i;
665
666 length -= (page_size - offset);
667 if (length <= 0)
668 return total_len;
669
670 dma_len -= (page_size - offset);
671 if (dma_len) {
672 dma_addr += (page_size - offset);
673 } else {
674 sg = sg_next(sg);
675 dma_addr = sg_dma_address(sg);
676 dma_len = sg_dma_len(sg);
677 }
678
679 if (length <= page_size) {
680 iod->first_dma = dma_addr;
681 return total_len;
682 }
683
684 nprps = DIV_ROUND_UP(length, page_size);
685 if (nprps <= (256 / 8)) {
686 pool = dev->prp_small_pool;
687 iod->npages = 0;
688 } else {
689 pool = dev->prp_page_pool;
690 iod->npages = 1;
691 }
692
693 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
694 if (!prp_list) {
695 iod->first_dma = dma_addr;
696 iod->npages = -1;
697 return (total_len - length) + page_size;
698 }
699 list[0] = prp_list;
700 iod->first_dma = prp_dma;
701 i = 0;
702 for (;;) {
703 if (i == page_size >> 3) {
704 __le64 *old_prp_list = prp_list;
705 prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
706 if (!prp_list)
707 return total_len - length;
708 list[iod->npages++] = prp_list;
709 prp_list[0] = old_prp_list[i - 1];
710 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
711 i = 1;
712 }
713 prp_list[i++] = cpu_to_le64(dma_addr);
714 dma_len -= page_size;
715 dma_addr += page_size;
716 length -= page_size;
717 if (length <= 0)
718 break;
719 if (dma_len > 0)
720 continue;
721 BUG_ON(dma_len < 0);
722 sg = sg_next(sg);
723 dma_addr = sg_dma_address(sg);
724 dma_len = sg_dma_len(sg);
725 }
726
727 return total_len;
728}
729
730static void nvme_submit_priv(struct nvme_queue *nvmeq, struct request *req,
731 struct nvme_iod *iod)
732{
733 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
734
735 memcpy(cmnd, req->cmd, sizeof(struct nvme_command));
736 cmnd->rw.command_id = req->tag;
737 if (req->nr_phys_segments) {
738 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
739 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
740 }
741
742 if (++nvmeq->sq_tail == nvmeq->q_depth)
743 nvmeq->sq_tail = 0;
744 writel(nvmeq->sq_tail, nvmeq->q_db);
745}
746
747/*
748 * We reuse the small pool to allocate the 16-byte range here as it is not
749 * worth having a special pool for these or additional cases to handle freeing
750 * the iod.
751 */
752static void nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
753 struct request *req, struct nvme_iod *iod)
754{
755 struct nvme_dsm_range *range =
756 (struct nvme_dsm_range *)iod_list(iod)[0];
757 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
758
759 range->cattr = cpu_to_le32(0);
760 range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift);
761 range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
762
763 memset(cmnd, 0, sizeof(*cmnd));
764 cmnd->dsm.opcode = nvme_cmd_dsm;
765 cmnd->dsm.command_id = req->tag;
766 cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
767 cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
768 cmnd->dsm.nr = 0;
769 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
770
771 if (++nvmeq->sq_tail == nvmeq->q_depth)
772 nvmeq->sq_tail = 0;
773 writel(nvmeq->sq_tail, nvmeq->q_db);
774}
775
776static void nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
777 int cmdid)
778{
779 struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
780
781 memset(cmnd, 0, sizeof(*cmnd));
782 cmnd->common.opcode = nvme_cmd_flush;
783 cmnd->common.command_id = cmdid;
784 cmnd->common.nsid = cpu_to_le32(ns->ns_id);
785
786 if (++nvmeq->sq_tail == nvmeq->q_depth)
787 nvmeq->sq_tail = 0;
788 writel(nvmeq->sq_tail, nvmeq->q_db);
789}
790
791static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod,
792 struct nvme_ns *ns)
793{
794 struct request *req = iod_get_private(iod);
795 struct nvme_command *cmnd;
796 u16 control = 0;
797 u32 dsmgmt = 0;
798
799 if (req->cmd_flags & REQ_FUA)
800 control |= NVME_RW_FUA;
801 if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
802 control |= NVME_RW_LR;
803
804 if (req->cmd_flags & REQ_RAHEAD)
805 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
806
807 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
808 memset(cmnd, 0, sizeof(*cmnd));
809
810 cmnd->rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
811 cmnd->rw.command_id = req->tag;
812 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
813 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
814 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
815 cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
816 cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
817
818 if (blk_integrity_rq(req)) {
819 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(iod->meta_sg));
820 switch (ns->pi_type) {
821 case NVME_NS_DPS_PI_TYPE3:
822 control |= NVME_RW_PRINFO_PRCHK_GUARD;
823 break;
824 case NVME_NS_DPS_PI_TYPE1:
825 case NVME_NS_DPS_PI_TYPE2:
826 control |= NVME_RW_PRINFO_PRCHK_GUARD |
827 NVME_RW_PRINFO_PRCHK_REF;
828 cmnd->rw.reftag = cpu_to_le32(
829 nvme_block_nr(ns, blk_rq_pos(req)));
830 break;
831 }
832 } else if (ns->ms)
833 control |= NVME_RW_PRINFO_PRACT;
834
835 cmnd->rw.control = cpu_to_le16(control);
836 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
837
838 if (++nvmeq->sq_tail == nvmeq->q_depth)
839 nvmeq->sq_tail = 0;
840 writel(nvmeq->sq_tail, nvmeq->q_db);
841
842 return 0;
843}
844
845/*
846 * NOTE: ns is NULL when called on the admin queue.
847 */
848static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
849 const struct blk_mq_queue_data *bd)
850{
851 struct nvme_ns *ns = hctx->queue->queuedata;
852 struct nvme_queue *nvmeq = hctx->driver_data;
853 struct nvme_dev *dev = nvmeq->dev;
854 struct request *req = bd->rq;
855 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
856 struct nvme_iod *iod;
857 enum dma_data_direction dma_dir;
858
859 /*
860 * If formated with metadata, require the block layer provide a buffer
861 * unless this namespace is formated such that the metadata can be
862 * stripped/generated by the controller with PRACT=1.
863 */
864 if (ns && ns->ms && !blk_integrity_rq(req)) {
865 if (!(ns->pi_type && ns->ms == 8) &&
866 req->cmd_type != REQ_TYPE_DRV_PRIV) {
867 req->errors = -EFAULT;
868 blk_mq_complete_request(req);
869 return BLK_MQ_RQ_QUEUE_OK;
870 }
871 }
872
873 iod = nvme_alloc_iod(req, dev, GFP_ATOMIC);
874 if (!iod)
875 return BLK_MQ_RQ_QUEUE_BUSY;
876
877 if (req->cmd_flags & REQ_DISCARD) {
878 void *range;
879 /*
880 * We reuse the small pool to allocate the 16-byte range here
881 * as it is not worth having a special pool for these or
882 * additional cases to handle freeing the iod.
883 */
884 range = dma_pool_alloc(dev->prp_small_pool, GFP_ATOMIC,
885 &iod->first_dma);
886 if (!range)
887 goto retry_cmd;
888 iod_list(iod)[0] = (__le64 *)range;
889 iod->npages = 0;
890 } else if (req->nr_phys_segments) {
891 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
892
893 sg_init_table(iod->sg, req->nr_phys_segments);
894 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
895 if (!iod->nents)
896 goto error_cmd;
897
898 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
899 goto retry_cmd;
900
901 if (blk_rq_bytes(req) !=
902 nvme_setup_prps(dev, iod, blk_rq_bytes(req), GFP_ATOMIC)) {
903 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
904 goto retry_cmd;
905 }
906 if (blk_integrity_rq(req)) {
907 if (blk_rq_count_integrity_sg(req->q, req->bio) != 1)
908 goto error_cmd;
909
910 sg_init_table(iod->meta_sg, 1);
911 if (blk_rq_map_integrity_sg(
912 req->q, req->bio, iod->meta_sg) != 1)
913 goto error_cmd;
914
915 if (rq_data_dir(req))
916 nvme_dif_remap(req, nvme_dif_prep);
917
918 if (!dma_map_sg(nvmeq->q_dmadev, iod->meta_sg, 1, dma_dir))
919 goto error_cmd;
920 }
921 }
922
923 nvme_set_info(cmd, iod, req_completion);
924 spin_lock_irq(&nvmeq->q_lock);
925 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
926 nvme_submit_priv(nvmeq, req, iod);
927 else if (req->cmd_flags & REQ_DISCARD)
928 nvme_submit_discard(nvmeq, ns, req, iod);
929 else if (req->cmd_flags & REQ_FLUSH)
930 nvme_submit_flush(nvmeq, ns, req->tag);
931 else
932 nvme_submit_iod(nvmeq, iod, ns);
933
934 nvme_process_cq(nvmeq);
935 spin_unlock_irq(&nvmeq->q_lock);
936 return BLK_MQ_RQ_QUEUE_OK;
937
938 error_cmd:
939 nvme_free_iod(dev, iod);
940 return BLK_MQ_RQ_QUEUE_ERROR;
941 retry_cmd:
942 nvme_free_iod(dev, iod);
943 return BLK_MQ_RQ_QUEUE_BUSY;
944}
945
946static int nvme_process_cq(struct nvme_queue *nvmeq)
947{
948 u16 head, phase;
949
950 head = nvmeq->cq_head;
951 phase = nvmeq->cq_phase;
952
953 for (;;) {
954 void *ctx;
955 nvme_completion_fn fn;
956 struct nvme_completion cqe = nvmeq->cqes[head];
957 if ((le16_to_cpu(cqe.status) & 1) != phase)
958 break;
959 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
960 if (++head == nvmeq->q_depth) {
961 head = 0;
962 phase = !phase;
963 }
964 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
965 fn(nvmeq, ctx, &cqe);
966 }
967
968 /* If the controller ignores the cq head doorbell and continuously
969 * writes to the queue, it is theoretically possible to wrap around
970 * the queue twice and mistakenly return IRQ_NONE. Linux only
971 * requires that 0.1% of your interrupts are handled, so this isn't
972 * a big problem.
973 */
974 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
975 return 0;
976
977 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
978 nvmeq->cq_head = head;
979 nvmeq->cq_phase = phase;
980
981 nvmeq->cqe_seen = 1;
982 return 1;
983}
984
985static irqreturn_t nvme_irq(int irq, void *data)
986{
987 irqreturn_t result;
988 struct nvme_queue *nvmeq = data;
989 spin_lock(&nvmeq->q_lock);
990 nvme_process_cq(nvmeq);
991 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
992 nvmeq->cqe_seen = 0;
993 spin_unlock(&nvmeq->q_lock);
994 return result;
995}
996
997static irqreturn_t nvme_irq_check(int irq, void *data)
998{
999 struct nvme_queue *nvmeq = data;
1000 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
1001 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
1002 return IRQ_NONE;
1003 return IRQ_WAKE_THREAD;
1004}
1005
1006/*
1007 * Returns 0 on success. If the result is negative, it's a Linux error code;
1008 * if the result is positive, it's an NVM Express status code
1009 */
1010int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1011 void *buffer, void __user *ubuffer, unsigned bufflen,
1012 u32 *result, unsigned timeout)
1013{
1014 bool write = cmd->common.opcode & 1;
1015 struct bio *bio = NULL;
1016 struct request *req;
1017 int ret;
1018
1019 req = blk_mq_alloc_request(q, write, GFP_KERNEL, false);
1020 if (IS_ERR(req))
1021 return PTR_ERR(req);
1022
1023 req->cmd_type = REQ_TYPE_DRV_PRIV;
1024 req->cmd_flags |= REQ_FAILFAST_DRIVER;
1025 req->__data_len = 0;
1026 req->__sector = (sector_t) -1;
1027 req->bio = req->biotail = NULL;
1028
1029 req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
1030
1031 req->cmd = (unsigned char *)cmd;
1032 req->cmd_len = sizeof(struct nvme_command);
1033 req->special = (void *)0;
1034
1035 if (buffer && bufflen) {
1036 ret = blk_rq_map_kern(q, req, buffer, bufflen, __GFP_WAIT);
1037 if (ret)
1038 goto out;
1039 } else if (ubuffer && bufflen) {
1040 ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen, __GFP_WAIT);
1041 if (ret)
1042 goto out;
1043 bio = req->bio;
1044 }
1045
1046 blk_execute_rq(req->q, NULL, req, 0);
1047 if (bio)
1048 blk_rq_unmap_user(bio);
1049 if (result)
1050 *result = (u32)(uintptr_t)req->special;
1051 ret = req->errors;
1052 out:
1053 blk_mq_free_request(req);
1054 return ret;
1055}
1056
1057int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1058 void *buffer, unsigned bufflen)
1059{
1060 return __nvme_submit_sync_cmd(q, cmd, buffer, NULL, bufflen, NULL, 0);
1061}
1062
1063static int nvme_submit_async_admin_req(struct nvme_dev *dev)
1064{
1065 struct nvme_queue *nvmeq = dev->queues[0];
1066 struct nvme_command c;
1067 struct nvme_cmd_info *cmd_info;
1068 struct request *req;
1069
1070 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, true);
1071 if (IS_ERR(req))
1072 return PTR_ERR(req);
1073
1074 req->cmd_flags |= REQ_NO_TIMEOUT;
1075 cmd_info = blk_mq_rq_to_pdu(req);
1076 nvme_set_info(cmd_info, NULL, async_req_completion);
1077
1078 memset(&c, 0, sizeof(c));
1079 c.common.opcode = nvme_admin_async_event;
1080 c.common.command_id = req->tag;
1081
1082 blk_mq_free_request(req);
1083 return __nvme_submit_cmd(nvmeq, &c);
1084}
1085
1086static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
1087 struct nvme_command *cmd,
1088 struct async_cmd_info *cmdinfo, unsigned timeout)
1089{
1090 struct nvme_queue *nvmeq = dev->queues[0];
1091 struct request *req;
1092 struct nvme_cmd_info *cmd_rq;
1093
1094 req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
1095 if (IS_ERR(req))
1096 return PTR_ERR(req);
1097
1098 req->timeout = timeout;
1099 cmd_rq = blk_mq_rq_to_pdu(req);
1100 cmdinfo->req = req;
1101 nvme_set_info(cmd_rq, cmdinfo, async_completion);
1102 cmdinfo->status = -EINTR;
1103
1104 cmd->common.command_id = req->tag;
1105
1106 return nvme_submit_cmd(nvmeq, cmd);
1107}
1108
1109static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1110{
1111 struct nvme_command c;
1112
1113 memset(&c, 0, sizeof(c));
1114 c.delete_queue.opcode = opcode;
1115 c.delete_queue.qid = cpu_to_le16(id);
1116
1117 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1118}
1119
1120static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1121 struct nvme_queue *nvmeq)
1122{
1123 struct nvme_command c;
1124 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1125
1126 /*
1127 * Note: we (ab)use the fact the the prp fields survive if no data
1128 * is attached to the request.
1129 */
1130 memset(&c, 0, sizeof(c));
1131 c.create_cq.opcode = nvme_admin_create_cq;
1132 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1133 c.create_cq.cqid = cpu_to_le16(qid);
1134 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1135 c.create_cq.cq_flags = cpu_to_le16(flags);
1136 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1137
1138 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1139}
1140
1141static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1142 struct nvme_queue *nvmeq)
1143{
1144 struct nvme_command c;
1145 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1146
1147 /*
1148 * Note: we (ab)use the fact the the prp fields survive if no data
1149 * is attached to the request.
1150 */
1151 memset(&c, 0, sizeof(c));
1152 c.create_sq.opcode = nvme_admin_create_sq;
1153 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1154 c.create_sq.sqid = cpu_to_le16(qid);
1155 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1156 c.create_sq.sq_flags = cpu_to_le16(flags);
1157 c.create_sq.cqid = cpu_to_le16(qid);
1158
1159 return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1160}
1161
1162static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1163{
1164 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1165}
1166
1167static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1168{
1169 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1170}
1171
1172int nvme_identify_ctrl(struct nvme_dev *dev, struct nvme_id_ctrl **id)
1173{
1174 struct nvme_command c = { };
1175 int error;
1176
1177 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1178 c.identify.opcode = nvme_admin_identify;
1179 c.identify.cns = cpu_to_le32(1);
1180
1181 *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1182 if (!*id)
1183 return -ENOMEM;
1184
1185 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1186 sizeof(struct nvme_id_ctrl));
1187 if (error)
1188 kfree(*id);
1189 return error;
1190}
1191
1192int nvme_identify_ns(struct nvme_dev *dev, unsigned nsid,
1193 struct nvme_id_ns **id)
1194{
1195 struct nvme_command c = { };
1196 int error;
1197
1198 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1199 c.identify.opcode = nvme_admin_identify,
1200 c.identify.nsid = cpu_to_le32(nsid),
1201
1202 *id = kmalloc(sizeof(struct nvme_id_ns), GFP_KERNEL);
1203 if (!*id)
1204 return -ENOMEM;
1205
1206 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1207 sizeof(struct nvme_id_ns));
1208 if (error)
1209 kfree(*id);
1210 return error;
1211}
1212
1213int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1214 dma_addr_t dma_addr, u32 *result)
1215{
1216 struct nvme_command c;
1217
1218 memset(&c, 0, sizeof(c));
1219 c.features.opcode = nvme_admin_get_features;
1220 c.features.nsid = cpu_to_le32(nsid);
1221 c.features.prp1 = cpu_to_le64(dma_addr);
1222 c.features.fid = cpu_to_le32(fid);
1223
1224 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1225 result, 0);
1226}
1227
1228int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1229 dma_addr_t dma_addr, u32 *result)
1230{
1231 struct nvme_command c;
1232
1233 memset(&c, 0, sizeof(c));
1234 c.features.opcode = nvme_admin_set_features;
1235 c.features.prp1 = cpu_to_le64(dma_addr);
1236 c.features.fid = cpu_to_le32(fid);
1237 c.features.dword11 = cpu_to_le32(dword11);
1238
1239 return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1240 result, 0);
1241}
1242
1243int nvme_get_log_page(struct nvme_dev *dev, struct nvme_smart_log **log)
1244{
1245 struct nvme_command c = { };
1246 int error;
1247
1248 c.common.opcode = nvme_admin_get_log_page,
1249 c.common.nsid = cpu_to_le32(0xFFFFFFFF),
1250 c.common.cdw10[0] = cpu_to_le32(
1251 (((sizeof(struct nvme_smart_log) / 4) - 1) << 16) |
1252 NVME_LOG_SMART),
1253
1254 *log = kmalloc(sizeof(struct nvme_smart_log), GFP_KERNEL);
1255 if (!*log)
1256 return -ENOMEM;
1257
1258 error = nvme_submit_sync_cmd(dev->admin_q, &c, *log,
1259 sizeof(struct nvme_smart_log));
1260 if (error)
1261 kfree(*log);
1262 return error;
1263}
1264
1265/**
1266 * nvme_abort_req - Attempt aborting a request
1267 *
1268 * Schedule controller reset if the command was already aborted once before and
1269 * still hasn't been returned to the driver, or if this is the admin queue.
1270 */
1271static void nvme_abort_req(struct request *req)
1272{
1273 struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1274 struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1275 struct nvme_dev *dev = nvmeq->dev;
1276 struct request *abort_req;
1277 struct nvme_cmd_info *abort_cmd;
1278 struct nvme_command cmd;
1279
1280 if (!nvmeq->qid || cmd_rq->aborted) {
1281 unsigned long flags;
1282
1283 spin_lock_irqsave(&dev_list_lock, flags);
1284 if (work_busy(&dev->reset_work))
1285 goto out;
1286 list_del_init(&dev->node);
1287 dev_warn(dev->dev, "I/O %d QID %d timeout, reset controller\n",
1288 req->tag, nvmeq->qid);
1289 dev->reset_workfn = nvme_reset_failed_dev;
1290 queue_work(nvme_workq, &dev->reset_work);
1291 out:
1292 spin_unlock_irqrestore(&dev_list_lock, flags);
1293 return;
1294 }
1295
1296 if (!dev->abort_limit)
1297 return;
1298
1299 abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1300 false);
1301 if (IS_ERR(abort_req))
1302 return;
1303
1304 abort_cmd = blk_mq_rq_to_pdu(abort_req);
1305 nvme_set_info(abort_cmd, abort_req, abort_completion);
1306
1307 memset(&cmd, 0, sizeof(cmd));
1308 cmd.abort.opcode = nvme_admin_abort_cmd;
1309 cmd.abort.cid = req->tag;
1310 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1311 cmd.abort.command_id = abort_req->tag;
1312
1313 --dev->abort_limit;
1314 cmd_rq->aborted = 1;
1315
1316 dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1317 nvmeq->qid);
1318 if (nvme_submit_cmd(dev->queues[0], &cmd) < 0) {
1319 dev_warn(nvmeq->q_dmadev,
1320 "Could not abort I/O %d QID %d",
1321 req->tag, nvmeq->qid);
1322 blk_mq_free_request(abort_req);
1323 }
1324}
1325
1326static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved)
1327{
1328 struct nvme_queue *nvmeq = data;
1329 void *ctx;
1330 nvme_completion_fn fn;
1331 struct nvme_cmd_info *cmd;
1332 struct nvme_completion cqe;
1333
1334 if (!blk_mq_request_started(req))
1335 return;
1336
1337 cmd = blk_mq_rq_to_pdu(req);
1338
1339 if (cmd->ctx == CMD_CTX_CANCELLED)
1340 return;
1341
1342 if (blk_queue_dying(req->q))
1343 cqe.status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1344 else
1345 cqe.status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1346
1347
1348 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1349 req->tag, nvmeq->qid);
1350 ctx = cancel_cmd_info(cmd, &fn);
1351 fn(nvmeq, ctx, &cqe);
1352}
1353
1354static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1355{
1356 struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1357 struct nvme_queue *nvmeq = cmd->nvmeq;
1358
1359 dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1360 nvmeq->qid);
1361 spin_lock_irq(&nvmeq->q_lock);
1362 nvme_abort_req(req);
1363 spin_unlock_irq(&nvmeq->q_lock);
1364
1365 /*
1366 * The aborted req will be completed on receiving the abort req.
1367 * We enable the timer again. If hit twice, it'll cause a device reset,
1368 * as the device then is in a faulty state.
1369 */
1370 return BLK_EH_RESET_TIMER;
1371}
1372
1373static void nvme_free_queue(struct nvme_queue *nvmeq)
1374{
1375 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1376 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1377 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1378 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1379 kfree(nvmeq);
1380}
1381
1382static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1383{
1384 int i;
1385
1386 for (i = dev->queue_count - 1; i >= lowest; i--) {
1387 struct nvme_queue *nvmeq = dev->queues[i];
1388 dev->queue_count--;
1389 dev->queues[i] = NULL;
1390 nvme_free_queue(nvmeq);
1391 }
1392}
1393
1394/**
1395 * nvme_suspend_queue - put queue into suspended state
1396 * @nvmeq - queue to suspend
1397 */
1398static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1399{
1400 int vector;
1401
1402 spin_lock_irq(&nvmeq->q_lock);
1403 if (nvmeq->cq_vector == -1) {
1404 spin_unlock_irq(&nvmeq->q_lock);
1405 return 1;
1406 }
1407 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1408 nvmeq->dev->online_queues--;
1409 nvmeq->cq_vector = -1;
1410 spin_unlock_irq(&nvmeq->q_lock);
1411
1412 if (!nvmeq->qid && nvmeq->dev->admin_q)
1413 blk_mq_freeze_queue_start(nvmeq->dev->admin_q);
1414
1415 irq_set_affinity_hint(vector, NULL);
1416 free_irq(vector, nvmeq);
1417
1418 return 0;
1419}
1420
1421static void nvme_clear_queue(struct nvme_queue *nvmeq)
1422{
1423 spin_lock_irq(&nvmeq->q_lock);
1424 if (nvmeq->tags && *nvmeq->tags)
1425 blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq);
1426 spin_unlock_irq(&nvmeq->q_lock);
1427}
1428
1429static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1430{
1431 struct nvme_queue *nvmeq = dev->queues[qid];
1432
1433 if (!nvmeq)
1434 return;
1435 if (nvme_suspend_queue(nvmeq))
1436 return;
1437
1438 /* Don't tell the adapter to delete the admin queue.
1439 * Don't tell a removed adapter to delete IO queues. */
1440 if (qid && readl(&dev->bar->csts) != -1) {
1441 adapter_delete_sq(dev, qid);
1442 adapter_delete_cq(dev, qid);
1443 }
1444
1445 spin_lock_irq(&nvmeq->q_lock);
1446 nvme_process_cq(nvmeq);
1447 spin_unlock_irq(&nvmeq->q_lock);
1448}
1449
1450static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1451 int depth)
1452{
1453 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1454 if (!nvmeq)
1455 return NULL;
1456
1457 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1458 &nvmeq->cq_dma_addr, GFP_KERNEL);
1459 if (!nvmeq->cqes)
1460 goto free_nvmeq;
1461
1462 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1463 &nvmeq->sq_dma_addr, GFP_KERNEL);
1464 if (!nvmeq->sq_cmds)
1465 goto free_cqdma;
1466
1467 nvmeq->q_dmadev = dev->dev;
1468 nvmeq->dev = dev;
1469 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1470 dev->instance, qid);
1471 spin_lock_init(&nvmeq->q_lock);
1472 nvmeq->cq_head = 0;
1473 nvmeq->cq_phase = 1;
1474 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1475 nvmeq->q_depth = depth;
1476 nvmeq->qid = qid;
1477 dev->queues[qid] = nvmeq;
1478
1479 /* make sure queue descriptor is set before queue count, for kthread */
1480 mb();
1481 dev->queue_count++;
1482
1483 return nvmeq;
1484
1485 free_cqdma:
1486 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1487 nvmeq->cq_dma_addr);
1488 free_nvmeq:
1489 kfree(nvmeq);
1490 return NULL;
1491}
1492
1493static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1494 const char *name)
1495{
1496 if (use_threaded_interrupts)
1497 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1498 nvme_irq_check, nvme_irq, IRQF_SHARED,
1499 name, nvmeq);
1500 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1501 IRQF_SHARED, name, nvmeq);
1502}
1503
1504static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1505{
1506 struct nvme_dev *dev = nvmeq->dev;
1507
1508 spin_lock_irq(&nvmeq->q_lock);
1509 nvmeq->sq_tail = 0;
1510 nvmeq->cq_head = 0;
1511 nvmeq->cq_phase = 1;
1512 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1513 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1514 dev->online_queues++;
1515 spin_unlock_irq(&nvmeq->q_lock);
1516}
1517
1518static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1519{
1520 struct nvme_dev *dev = nvmeq->dev;
1521 int result;
1522
1523 nvmeq->cq_vector = qid - 1;
1524 result = adapter_alloc_cq(dev, qid, nvmeq);
1525 if (result < 0)
1526 return result;
1527
1528 result = adapter_alloc_sq(dev, qid, nvmeq);
1529 if (result < 0)
1530 goto release_cq;
1531
1532 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1533 if (result < 0)
1534 goto release_sq;
1535
1536 nvme_init_queue(nvmeq, qid);
1537 return result;
1538
1539 release_sq:
1540 adapter_delete_sq(dev, qid);
1541 release_cq:
1542 adapter_delete_cq(dev, qid);
1543 return result;
1544}
1545
1546static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1547{
1548 unsigned long timeout;
1549 u32 bit = enabled ? NVME_CSTS_RDY : 0;
1550
1551 timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1552
1553 while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1554 msleep(100);
1555 if (fatal_signal_pending(current))
1556 return -EINTR;
1557 if (time_after(jiffies, timeout)) {
1558 dev_err(dev->dev,
1559 "Device not ready; aborting %s\n", enabled ?
1560 "initialisation" : "reset");
1561 return -ENODEV;
1562 }
1563 }
1564
1565 return 0;
1566}
1567
1568/*
1569 * If the device has been passed off to us in an enabled state, just clear
1570 * the enabled bit. The spec says we should set the 'shutdown notification
1571 * bits', but doing so may cause the device to complete commands to the
1572 * admin queue ... and we don't know what memory that might be pointing at!
1573 */
1574static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1575{
1576 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1577 dev->ctrl_config &= ~NVME_CC_ENABLE;
1578 writel(dev->ctrl_config, &dev->bar->cc);
1579
1580 return nvme_wait_ready(dev, cap, false);
1581}
1582
1583static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1584{
1585 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1586 dev->ctrl_config |= NVME_CC_ENABLE;
1587 writel(dev->ctrl_config, &dev->bar->cc);
1588
1589 return nvme_wait_ready(dev, cap, true);
1590}
1591
1592static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1593{
1594 unsigned long timeout;
1595
1596 dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1597 dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1598
1599 writel(dev->ctrl_config, &dev->bar->cc);
1600
1601 timeout = SHUTDOWN_TIMEOUT + jiffies;
1602 while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1603 NVME_CSTS_SHST_CMPLT) {
1604 msleep(100);
1605 if (fatal_signal_pending(current))
1606 return -EINTR;
1607 if (time_after(jiffies, timeout)) {
1608 dev_err(dev->dev,
1609 "Device shutdown incomplete; abort shutdown\n");
1610 return -ENODEV;
1611 }
1612 }
1613
1614 return 0;
1615}
1616
1617static struct blk_mq_ops nvme_mq_admin_ops = {
1618 .queue_rq = nvme_queue_rq,
1619 .map_queue = blk_mq_map_queue,
1620 .init_hctx = nvme_admin_init_hctx,
1621 .exit_hctx = nvme_admin_exit_hctx,
1622 .init_request = nvme_admin_init_request,
1623 .timeout = nvme_timeout,
1624};
1625
1626static struct blk_mq_ops nvme_mq_ops = {
1627 .queue_rq = nvme_queue_rq,
1628 .map_queue = blk_mq_map_queue,
1629 .init_hctx = nvme_init_hctx,
1630 .init_request = nvme_init_request,
1631 .timeout = nvme_timeout,
1632};
1633
1634static void nvme_dev_remove_admin(struct nvme_dev *dev)
1635{
1636 if (dev->admin_q && !blk_queue_dying(dev->admin_q)) {
1637 blk_cleanup_queue(dev->admin_q);
1638 blk_mq_free_tag_set(&dev->admin_tagset);
1639 }
1640}
1641
1642static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1643{
1644 if (!dev->admin_q) {
1645 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1646 dev->admin_tagset.nr_hw_queues = 1;
1647 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1648 dev->admin_tagset.reserved_tags = 1;
1649 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1650 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1651 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1652 dev->admin_tagset.driver_data = dev;
1653
1654 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1655 return -ENOMEM;
1656
1657 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1658 if (IS_ERR(dev->admin_q)) {
1659 blk_mq_free_tag_set(&dev->admin_tagset);
1660 return -ENOMEM;
1661 }
1662 if (!blk_get_queue(dev->admin_q)) {
1663 nvme_dev_remove_admin(dev);
1664 dev->admin_q = NULL;
1665 return -ENODEV;
1666 }
1667 } else
1668 blk_mq_unfreeze_queue(dev->admin_q);
1669
1670 return 0;
1671}
1672
1673static int nvme_configure_admin_queue(struct nvme_dev *dev)
1674{
1675 int result;
1676 u32 aqa;
1677 u64 cap = readq(&dev->bar->cap);
1678 struct nvme_queue *nvmeq;
1679 unsigned page_shift = PAGE_SHIFT;
1680 unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1681 unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1682
1683 if (page_shift < dev_page_min) {
1684 dev_err(dev->dev,
1685 "Minimum device page size (%u) too large for "
1686 "host (%u)\n", 1 << dev_page_min,
1687 1 << page_shift);
1688 return -ENODEV;
1689 }
1690 if (page_shift > dev_page_max) {
1691 dev_info(dev->dev,
1692 "Device maximum page size (%u) smaller than "
1693 "host (%u); enabling work-around\n",
1694 1 << dev_page_max, 1 << page_shift);
1695 page_shift = dev_page_max;
1696 }
1697
1698 result = nvme_disable_ctrl(dev, cap);
1699 if (result < 0)
1700 return result;
1701
1702 nvmeq = dev->queues[0];
1703 if (!nvmeq) {
1704 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1705 if (!nvmeq)
1706 return -ENOMEM;
1707 }
1708
1709 aqa = nvmeq->q_depth - 1;
1710 aqa |= aqa << 16;
1711
1712 dev->page_size = 1 << page_shift;
1713
1714 dev->ctrl_config = NVME_CC_CSS_NVM;
1715 dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1716 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1717 dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1718
1719 writel(aqa, &dev->bar->aqa);
1720 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1721 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1722
1723 result = nvme_enable_ctrl(dev, cap);
1724 if (result)
1725 goto free_nvmeq;
1726
1727 nvmeq->cq_vector = 0;
1728 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1729 if (result)
1730 goto free_nvmeq;
1731
1732 return result;
1733
1734 free_nvmeq:
1735 nvme_free_queues(dev, 0);
1736 return result;
1737}
1738
1739static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1740{
1741 struct nvme_dev *dev = ns->dev;
1742 struct nvme_user_io io;
1743 struct nvme_command c;
1744 unsigned length, meta_len;
1745 int status, write;
1746 dma_addr_t meta_dma = 0;
1747 void *meta = NULL;
1748 void __user *metadata;
1749
1750 if (copy_from_user(&io, uio, sizeof(io)))
1751 return -EFAULT;
1752
1753 switch (io.opcode) {
1754 case nvme_cmd_write:
1755 case nvme_cmd_read:
1756 case nvme_cmd_compare:
1757 break;
1758 default:
1759 return -EINVAL;
1760 }
1761
1762 length = (io.nblocks + 1) << ns->lba_shift;
1763 meta_len = (io.nblocks + 1) * ns->ms;
1764 metadata = (void __user *)(unsigned long)io.metadata;
1765 write = io.opcode & 1;
1766
1767 if (ns->ext) {
1768 length += meta_len;
1769 meta_len = 0;
1770 }
1771 if (meta_len) {
1772 if (((io.metadata & 3) || !io.metadata) && !ns->ext)
1773 return -EINVAL;
1774
1775 meta = dma_alloc_coherent(dev->dev, meta_len,
1776 &meta_dma, GFP_KERNEL);
1777
1778 if (!meta) {
1779 status = -ENOMEM;
1780 goto unmap;
1781 }
1782 if (write) {
1783 if (copy_from_user(meta, metadata, meta_len)) {
1784 status = -EFAULT;
1785 goto unmap;
1786 }
1787 }
1788 }
1789
1790 memset(&c, 0, sizeof(c));
1791 c.rw.opcode = io.opcode;
1792 c.rw.flags = io.flags;
1793 c.rw.nsid = cpu_to_le32(ns->ns_id);
1794 c.rw.slba = cpu_to_le64(io.slba);
1795 c.rw.length = cpu_to_le16(io.nblocks);
1796 c.rw.control = cpu_to_le16(io.control);
1797 c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1798 c.rw.reftag = cpu_to_le32(io.reftag);
1799 c.rw.apptag = cpu_to_le16(io.apptag);
1800 c.rw.appmask = cpu_to_le16(io.appmask);
1801 c.rw.metadata = cpu_to_le64(meta_dma);
1802
1803 status = __nvme_submit_sync_cmd(ns->queue, &c, NULL,
1804 (void __user *)io.addr, length, NULL, 0);
1805 unmap:
1806 if (meta) {
1807 if (status == NVME_SC_SUCCESS && !write) {
1808 if (copy_to_user(metadata, meta, meta_len))
1809 status = -EFAULT;
1810 }
1811 dma_free_coherent(dev->dev, meta_len, meta, meta_dma);
1812 }
1813 return status;
1814}
1815
1816static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1817 struct nvme_passthru_cmd __user *ucmd)
1818{
1819 struct nvme_passthru_cmd cmd;
1820 struct nvme_command c;
1821 unsigned timeout = 0;
1822 int status;
1823
1824 if (!capable(CAP_SYS_ADMIN))
1825 return -EACCES;
1826 if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1827 return -EFAULT;
1828
1829 memset(&c, 0, sizeof(c));
1830 c.common.opcode = cmd.opcode;
1831 c.common.flags = cmd.flags;
1832 c.common.nsid = cpu_to_le32(cmd.nsid);
1833 c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1834 c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1835 c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1836 c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1837 c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1838 c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1839 c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1840 c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1841
1842 if (cmd.timeout_ms)
1843 timeout = msecs_to_jiffies(cmd.timeout_ms);
1844
1845 status = __nvme_submit_sync_cmd(ns ? ns->queue : dev->admin_q, &c,
1846 NULL, (void __user *)cmd.addr, cmd.data_len,
1847 &cmd.result, timeout);
1848 if (status >= 0) {
1849 if (put_user(cmd.result, &ucmd->result))
1850 return -EFAULT;
1851 }
1852
1853 return status;
1854}
1855
1856static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1857 unsigned long arg)
1858{
1859 struct nvme_ns *ns = bdev->bd_disk->private_data;
1860
1861 switch (cmd) {
1862 case NVME_IOCTL_ID:
1863 force_successful_syscall_return();
1864 return ns->ns_id;
1865 case NVME_IOCTL_ADMIN_CMD:
1866 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1867 case NVME_IOCTL_IO_CMD:
1868 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1869 case NVME_IOCTL_SUBMIT_IO:
1870 return nvme_submit_io(ns, (void __user *)arg);
1871 case SG_GET_VERSION_NUM:
1872 return nvme_sg_get_version_num((void __user *)arg);
1873 case SG_IO:
1874 return nvme_sg_io(ns, (void __user *)arg);
1875 default:
1876 return -ENOTTY;
1877 }
1878}
1879
1880#ifdef CONFIG_COMPAT
1881static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1882 unsigned int cmd, unsigned long arg)
1883{
1884 switch (cmd) {
1885 case SG_IO:
1886 return -ENOIOCTLCMD;
1887 }
1888 return nvme_ioctl(bdev, mode, cmd, arg);
1889}
1890#else
1891#define nvme_compat_ioctl NULL
1892#endif
1893
1894static int nvme_open(struct block_device *bdev, fmode_t mode)
1895{
1896 int ret = 0;
1897 struct nvme_ns *ns;
1898
1899 spin_lock(&dev_list_lock);
1900 ns = bdev->bd_disk->private_data;
1901 if (!ns)
1902 ret = -ENXIO;
1903 else if (!kref_get_unless_zero(&ns->dev->kref))
1904 ret = -ENXIO;
1905 spin_unlock(&dev_list_lock);
1906
1907 return ret;
1908}
1909
1910static void nvme_free_dev(struct kref *kref);
1911
1912static void nvme_release(struct gendisk *disk, fmode_t mode)
1913{
1914 struct nvme_ns *ns = disk->private_data;
1915 struct nvme_dev *dev = ns->dev;
1916
1917 kref_put(&dev->kref, nvme_free_dev);
1918}
1919
1920static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1921{
1922 /* some standard values */
1923 geo->heads = 1 << 6;
1924 geo->sectors = 1 << 5;
1925 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1926 return 0;
1927}
1928
1929static void nvme_config_discard(struct nvme_ns *ns)
1930{
1931 u32 logical_block_size = queue_logical_block_size(ns->queue);
1932 ns->queue->limits.discard_zeroes_data = 0;
1933 ns->queue->limits.discard_alignment = logical_block_size;
1934 ns->queue->limits.discard_granularity = logical_block_size;
1935 ns->queue->limits.max_discard_sectors = 0xffffffff;
1936 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1937}
1938
1939static int nvme_revalidate_disk(struct gendisk *disk)
1940{
1941 struct nvme_ns *ns = disk->private_data;
1942 struct nvme_dev *dev = ns->dev;
1943 struct nvme_id_ns *id;
1944 u8 lbaf, pi_type;
1945 u16 old_ms;
1946 unsigned short bs;
1947
1948 if (nvme_identify_ns(dev, ns->ns_id, &id)) {
1949 dev_warn(dev->dev, "%s: Identify failure nvme%dn%d\n", __func__,
1950 dev->instance, ns->ns_id);
1951 return -ENODEV;
1952 }
1953 if (id->ncap == 0) {
1954 kfree(id);
1955 return -ENODEV;
1956 }
1957
1958 old_ms = ns->ms;
1959 lbaf = id->flbas & NVME_NS_FLBAS_LBA_MASK;
1960 ns->lba_shift = id->lbaf[lbaf].ds;
1961 ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1962 ns->ext = ns->ms && (id->flbas & NVME_NS_FLBAS_META_EXT);
1963
1964 /*
1965 * If identify namespace failed, use default 512 byte block size so
1966 * block layer can use before failing read/write for 0 capacity.
1967 */
1968 if (ns->lba_shift == 0)
1969 ns->lba_shift = 9;
1970 bs = 1 << ns->lba_shift;
1971
1972 /* XXX: PI implementation requires metadata equal t10 pi tuple size */
1973 pi_type = ns->ms == sizeof(struct t10_pi_tuple) ?
1974 id->dps & NVME_NS_DPS_PI_MASK : 0;
1975
1976 if (blk_get_integrity(disk) && (ns->pi_type != pi_type ||
1977 ns->ms != old_ms ||
1978 bs != queue_logical_block_size(disk->queue) ||
1979 (ns->ms && ns->ext)))
1980 blk_integrity_unregister(disk);
1981
1982 ns->pi_type = pi_type;
1983 blk_queue_logical_block_size(ns->queue, bs);
1984
1985 if (ns->ms && !blk_get_integrity(disk) && (disk->flags & GENHD_FL_UP) &&
1986 !ns->ext)
1987 nvme_init_integrity(ns);
1988
1989 if (ns->ms && !blk_get_integrity(disk))
1990 set_capacity(disk, 0);
1991 else
1992 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1993
1994 if (dev->oncs & NVME_CTRL_ONCS_DSM)
1995 nvme_config_discard(ns);
1996
1997 kfree(id);
1998 return 0;
1999}
2000
2001static const struct block_device_operations nvme_fops = {
2002 .owner = THIS_MODULE,
2003 .ioctl = nvme_ioctl,
2004 .compat_ioctl = nvme_compat_ioctl,
2005 .open = nvme_open,
2006 .release = nvme_release,
2007 .getgeo = nvme_getgeo,
2008 .revalidate_disk= nvme_revalidate_disk,
2009};
2010
2011static int nvme_kthread(void *data)
2012{
2013 struct nvme_dev *dev, *next;
2014
2015 while (!kthread_should_stop()) {
2016 set_current_state(TASK_INTERRUPTIBLE);
2017 spin_lock(&dev_list_lock);
2018 list_for_each_entry_safe(dev, next, &dev_list, node) {
2019 int i;
2020 if (readl(&dev->bar->csts) & NVME_CSTS_CFS) {
2021 if (work_busy(&dev->reset_work))
2022 continue;
2023 list_del_init(&dev->node);
2024 dev_warn(dev->dev,
2025 "Failed status: %x, reset controller\n",
2026 readl(&dev->bar->csts));
2027 dev->reset_workfn = nvme_reset_failed_dev;
2028 queue_work(nvme_workq, &dev->reset_work);
2029 continue;
2030 }
2031 for (i = 0; i < dev->queue_count; i++) {
2032 struct nvme_queue *nvmeq = dev->queues[i];
2033 if (!nvmeq)
2034 continue;
2035 spin_lock_irq(&nvmeq->q_lock);
2036 nvme_process_cq(nvmeq);
2037
2038 while ((i == 0) && (dev->event_limit > 0)) {
2039 if (nvme_submit_async_admin_req(dev))
2040 break;
2041 dev->event_limit--;
2042 }
2043 spin_unlock_irq(&nvmeq->q_lock);
2044 }
2045 }
2046 spin_unlock(&dev_list_lock);
2047 schedule_timeout(round_jiffies_relative(HZ));
2048 }
2049 return 0;
2050}
2051
2052static void nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid)
2053{
2054 struct nvme_ns *ns;
2055 struct gendisk *disk;
2056 int node = dev_to_node(dev->dev);
2057
2058 ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
2059 if (!ns)
2060 return;
2061
2062 ns->queue = blk_mq_init_queue(&dev->tagset);
2063 if (IS_ERR(ns->queue))
2064 goto out_free_ns;
2065 queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
2066 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
2067 queue_flag_set_unlocked(QUEUE_FLAG_SG_GAPS, ns->queue);
2068 ns->dev = dev;
2069 ns->queue->queuedata = ns;
2070
2071 disk = alloc_disk_node(0, node);
2072 if (!disk)
2073 goto out_free_queue;
2074
2075 ns->ns_id = nsid;
2076 ns->disk = disk;
2077 ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
2078 list_add_tail(&ns->list, &dev->namespaces);
2079
2080 blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2081 if (dev->max_hw_sectors)
2082 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2083 if (dev->stripe_size)
2084 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
2085 if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2086 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2087
2088 disk->major = nvme_major;
2089 disk->first_minor = 0;
2090 disk->fops = &nvme_fops;
2091 disk->private_data = ns;
2092 disk->queue = ns->queue;
2093 disk->driverfs_dev = dev->device;
2094 disk->flags = GENHD_FL_EXT_DEVT;
2095 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2096
2097 /*
2098 * Initialize capacity to 0 until we establish the namespace format and
2099 * setup integrity extentions if necessary. The revalidate_disk after
2100 * add_disk allows the driver to register with integrity if the format
2101 * requires it.
2102 */
2103 set_capacity(disk, 0);
2104 if (nvme_revalidate_disk(ns->disk))
2105 goto out_free_disk;
2106
2107 add_disk(ns->disk);
2108 if (ns->ms)
2109 revalidate_disk(ns->disk);
2110 return;
2111 out_free_disk:
2112 kfree(disk);
2113 list_del(&ns->list);
2114 out_free_queue:
2115 blk_cleanup_queue(ns->queue);
2116 out_free_ns:
2117 kfree(ns);
2118}
2119
2120static void nvme_create_io_queues(struct nvme_dev *dev)
2121{
2122 unsigned i;
2123
2124 for (i = dev->queue_count; i <= dev->max_qid; i++)
2125 if (!nvme_alloc_queue(dev, i, dev->q_depth))
2126 break;
2127
2128 for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
2129 if (nvme_create_queue(dev->queues[i], i))
2130 break;
2131}
2132
2133static int set_queue_count(struct nvme_dev *dev, int count)
2134{
2135 int status;
2136 u32 result;
2137 u32 q_count = (count - 1) | ((count - 1) << 16);
2138
2139 status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2140 &result);
2141 if (status < 0)
2142 return status;
2143 if (status > 0) {
2144 dev_err(dev->dev, "Could not set queue count (%d)\n", status);
2145 return 0;
2146 }
2147 return min(result & 0xffff, result >> 16) + 1;
2148}
2149
2150static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2151{
2152 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2153}
2154
2155static int nvme_setup_io_queues(struct nvme_dev *dev)
2156{
2157 struct nvme_queue *adminq = dev->queues[0];
2158 struct pci_dev *pdev = to_pci_dev(dev->dev);
2159 int result, i, vecs, nr_io_queues, size;
2160
2161 nr_io_queues = num_possible_cpus();
2162 result = set_queue_count(dev, nr_io_queues);
2163 if (result <= 0)
2164 return result;
2165 if (result < nr_io_queues)
2166 nr_io_queues = result;
2167
2168 size = db_bar_size(dev, nr_io_queues);
2169 if (size > 8192) {
2170 iounmap(dev->bar);
2171 do {
2172 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2173 if (dev->bar)
2174 break;
2175 if (!--nr_io_queues)
2176 return -ENOMEM;
2177 size = db_bar_size(dev, nr_io_queues);
2178 } while (1);
2179 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2180 adminq->q_db = dev->dbs;
2181 }
2182
2183 /* Deregister the admin queue's interrupt */
2184 free_irq(dev->entry[0].vector, adminq);
2185
2186 /*
2187 * If we enable msix early due to not intx, disable it again before
2188 * setting up the full range we need.
2189 */
2190 if (!pdev->irq)
2191 pci_disable_msix(pdev);
2192
2193 for (i = 0; i < nr_io_queues; i++)
2194 dev->entry[i].entry = i;
2195 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2196 if (vecs < 0) {
2197 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2198 if (vecs < 0) {
2199 vecs = 1;
2200 } else {
2201 for (i = 0; i < vecs; i++)
2202 dev->entry[i].vector = i + pdev->irq;
2203 }
2204 }
2205
2206 /*
2207 * Should investigate if there's a performance win from allocating
2208 * more queues than interrupt vectors; it might allow the submission
2209 * path to scale better, even if the receive path is limited by the
2210 * number of interrupts.
2211 */
2212 nr_io_queues = vecs;
2213 dev->max_qid = nr_io_queues;
2214
2215 result = queue_request_irq(dev, adminq, adminq->irqname);
2216 if (result)
2217 goto free_queues;
2218
2219 /* Free previously allocated queues that are no longer usable */
2220 nvme_free_queues(dev, nr_io_queues + 1);
2221 nvme_create_io_queues(dev);
2222
2223 return 0;
2224
2225 free_queues:
2226 nvme_free_queues(dev, 1);
2227 return result;
2228}
2229
2230static void nvme_free_namespace(struct nvme_ns *ns)
2231{
2232 list_del(&ns->list);
2233
2234 spin_lock(&dev_list_lock);
2235 ns->disk->private_data = NULL;
2236 spin_unlock(&dev_list_lock);
2237
2238 put_disk(ns->disk);
2239 kfree(ns);
2240}
2241
2242static int ns_cmp(void *priv, struct list_head *a, struct list_head *b)
2243{
2244 struct nvme_ns *nsa = container_of(a, struct nvme_ns, list);
2245 struct nvme_ns *nsb = container_of(b, struct nvme_ns, list);
2246
2247 return nsa->ns_id - nsb->ns_id;
2248}
2249
2250static struct nvme_ns *nvme_find_ns(struct nvme_dev *dev, unsigned nsid)
2251{
2252 struct nvme_ns *ns;
2253
2254 list_for_each_entry(ns, &dev->namespaces, list) {
2255 if (ns->ns_id == nsid)
2256 return ns;
2257 if (ns->ns_id > nsid)
2258 break;
2259 }
2260 return NULL;
2261}
2262
2263static inline bool nvme_io_incapable(struct nvme_dev *dev)
2264{
2265 return (!dev->bar || readl(&dev->bar->csts) & NVME_CSTS_CFS ||
2266 dev->online_queues < 2);
2267}
2268
2269static void nvme_ns_remove(struct nvme_ns *ns)
2270{
2271 bool kill = nvme_io_incapable(ns->dev) && !blk_queue_dying(ns->queue);
2272
2273 if (kill)
2274 blk_set_queue_dying(ns->queue);
2275 if (ns->disk->flags & GENHD_FL_UP) {
2276 if (blk_get_integrity(ns->disk))
2277 blk_integrity_unregister(ns->disk);
2278 del_gendisk(ns->disk);
2279 }
2280 if (kill || !blk_queue_dying(ns->queue)) {
2281 blk_mq_abort_requeue_list(ns->queue);
2282 blk_cleanup_queue(ns->queue);
2283 }
2284}
2285
2286static void nvme_scan_namespaces(struct nvme_dev *dev, unsigned nn)
2287{
2288 struct nvme_ns *ns, *next;
2289 unsigned i;
2290
2291 for (i = 1; i <= nn; i++) {
2292 ns = nvme_find_ns(dev, i);
2293 if (ns) {
2294 if (revalidate_disk(ns->disk)) {
2295 nvme_ns_remove(ns);
2296 nvme_free_namespace(ns);
2297 }
2298 } else
2299 nvme_alloc_ns(dev, i);
2300 }
2301 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2302 if (ns->ns_id > nn) {
2303 nvme_ns_remove(ns);
2304 nvme_free_namespace(ns);
2305 }
2306 }
2307 list_sort(NULL, &dev->namespaces, ns_cmp);
2308}
2309
2310static void nvme_dev_scan(struct work_struct *work)
2311{
2312 struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work);
2313 struct nvme_id_ctrl *ctrl;
2314
2315 if (!dev->tagset.tags)
2316 return;
2317 if (nvme_identify_ctrl(dev, &ctrl))
2318 return;
2319 nvme_scan_namespaces(dev, le32_to_cpup(&ctrl->nn));
2320 kfree(ctrl);
2321}
2322
2323/*
2324 * Return: error value if an error occurred setting up the queues or calling
2325 * Identify Device. 0 if these succeeded, even if adding some of the
2326 * namespaces failed. At the moment, these failures are silent. TBD which
2327 * failures should be reported.
2328 */
2329static int nvme_dev_add(struct nvme_dev *dev)
2330{
2331 struct pci_dev *pdev = to_pci_dev(dev->dev);
2332 int res;
2333 unsigned nn;
2334 struct nvme_id_ctrl *ctrl;
2335 int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2336
2337 res = nvme_identify_ctrl(dev, &ctrl);
2338 if (res) {
2339 dev_err(dev->dev, "Identify Controller failed (%d)\n", res);
2340 return -EIO;
2341 }
2342
2343 nn = le32_to_cpup(&ctrl->nn);
2344 dev->oncs = le16_to_cpup(&ctrl->oncs);
2345 dev->abort_limit = ctrl->acl + 1;
2346 dev->vwc = ctrl->vwc;
2347 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2348 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2349 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2350 if (ctrl->mdts)
2351 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2352 if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2353 (pdev->device == 0x0953) && ctrl->vs[3]) {
2354 unsigned int max_hw_sectors;
2355
2356 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2357 max_hw_sectors = dev->stripe_size >> (shift - 9);
2358 if (dev->max_hw_sectors) {
2359 dev->max_hw_sectors = min(max_hw_sectors,
2360 dev->max_hw_sectors);
2361 } else
2362 dev->max_hw_sectors = max_hw_sectors;
2363 }
2364 kfree(ctrl);
2365
2366 if (!dev->tagset.tags) {
2367 dev->tagset.ops = &nvme_mq_ops;
2368 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2369 dev->tagset.timeout = NVME_IO_TIMEOUT;
2370 dev->tagset.numa_node = dev_to_node(dev->dev);
2371 dev->tagset.queue_depth =
2372 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2373 dev->tagset.cmd_size = nvme_cmd_size(dev);
2374 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2375 dev->tagset.driver_data = dev;
2376
2377 if (blk_mq_alloc_tag_set(&dev->tagset))
2378 return 0;
2379 }
2380 schedule_work(&dev->scan_work);
2381 return 0;
2382}
2383
2384static int nvme_dev_map(struct nvme_dev *dev)
2385{
2386 u64 cap;
2387 int bars, result = -ENOMEM;
2388 struct pci_dev *pdev = to_pci_dev(dev->dev);
2389
2390 if (pci_enable_device_mem(pdev))
2391 return result;
2392
2393 dev->entry[0].vector = pdev->irq;
2394 pci_set_master(pdev);
2395 bars = pci_select_bars(pdev, IORESOURCE_MEM);
2396 if (!bars)
2397 goto disable_pci;
2398
2399 if (pci_request_selected_regions(pdev, bars, "nvme"))
2400 goto disable_pci;
2401
2402 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
2403 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
2404 goto disable;
2405
2406 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2407 if (!dev->bar)
2408 goto disable;
2409
2410 if (readl(&dev->bar->csts) == -1) {
2411 result = -ENODEV;
2412 goto unmap;
2413 }
2414
2415 /*
2416 * Some devices don't advertse INTx interrupts, pre-enable a single
2417 * MSIX vec for setup. We'll adjust this later.
2418 */
2419 if (!pdev->irq) {
2420 result = pci_enable_msix(pdev, dev->entry, 1);
2421 if (result < 0)
2422 goto unmap;
2423 }
2424
2425 cap = readq(&dev->bar->cap);
2426 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2427 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2428 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2429
2430 return 0;
2431
2432 unmap:
2433 iounmap(dev->bar);
2434 dev->bar = NULL;
2435 disable:
2436 pci_release_regions(pdev);
2437 disable_pci:
2438 pci_disable_device(pdev);
2439 return result;
2440}
2441
2442static void nvme_dev_unmap(struct nvme_dev *dev)
2443{
2444 struct pci_dev *pdev = to_pci_dev(dev->dev);
2445
2446 if (pdev->msi_enabled)
2447 pci_disable_msi(pdev);
2448 else if (pdev->msix_enabled)
2449 pci_disable_msix(pdev);
2450
2451 if (dev->bar) {
2452 iounmap(dev->bar);
2453 dev->bar = NULL;
2454 pci_release_regions(pdev);
2455 }
2456
2457 if (pci_is_enabled(pdev))
2458 pci_disable_device(pdev);
2459}
2460
2461struct nvme_delq_ctx {
2462 struct task_struct *waiter;
2463 struct kthread_worker *worker;
2464 atomic_t refcount;
2465};
2466
2467static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2468{
2469 dq->waiter = current;
2470 mb();
2471
2472 for (;;) {
2473 set_current_state(TASK_KILLABLE);
2474 if (!atomic_read(&dq->refcount))
2475 break;
2476 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2477 fatal_signal_pending(current)) {
2478 /*
2479 * Disable the controller first since we can't trust it
2480 * at this point, but leave the admin queue enabled
2481 * until all queue deletion requests are flushed.
2482 * FIXME: This may take a while if there are more h/w
2483 * queues than admin tags.
2484 */
2485 set_current_state(TASK_RUNNING);
2486 nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2487 nvme_clear_queue(dev->queues[0]);
2488 flush_kthread_worker(dq->worker);
2489 nvme_disable_queue(dev, 0);
2490 return;
2491 }
2492 }
2493 set_current_state(TASK_RUNNING);
2494}
2495
2496static void nvme_put_dq(struct nvme_delq_ctx *dq)
2497{
2498 atomic_dec(&dq->refcount);
2499 if (dq->waiter)
2500 wake_up_process(dq->waiter);
2501}
2502
2503static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2504{
2505 atomic_inc(&dq->refcount);
2506 return dq;
2507}
2508
2509static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2510{
2511 struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2512 nvme_put_dq(dq);
2513}
2514
2515static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2516 kthread_work_func_t fn)
2517{
2518 struct nvme_command c;
2519
2520 memset(&c, 0, sizeof(c));
2521 c.delete_queue.opcode = opcode;
2522 c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2523
2524 init_kthread_work(&nvmeq->cmdinfo.work, fn);
2525 return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2526 ADMIN_TIMEOUT);
2527}
2528
2529static void nvme_del_cq_work_handler(struct kthread_work *work)
2530{
2531 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2532 cmdinfo.work);
2533 nvme_del_queue_end(nvmeq);
2534}
2535
2536static int nvme_delete_cq(struct nvme_queue *nvmeq)
2537{
2538 return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2539 nvme_del_cq_work_handler);
2540}
2541
2542static void nvme_del_sq_work_handler(struct kthread_work *work)
2543{
2544 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2545 cmdinfo.work);
2546 int status = nvmeq->cmdinfo.status;
2547
2548 if (!status)
2549 status = nvme_delete_cq(nvmeq);
2550 if (status)
2551 nvme_del_queue_end(nvmeq);
2552}
2553
2554static int nvme_delete_sq(struct nvme_queue *nvmeq)
2555{
2556 return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2557 nvme_del_sq_work_handler);
2558}
2559
2560static void nvme_del_queue_start(struct kthread_work *work)
2561{
2562 struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2563 cmdinfo.work);
2564 if (nvme_delete_sq(nvmeq))
2565 nvme_del_queue_end(nvmeq);
2566}
2567
2568static void nvme_disable_io_queues(struct nvme_dev *dev)
2569{
2570 int i;
2571 DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2572 struct nvme_delq_ctx dq;
2573 struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2574 &worker, "nvme%d", dev->instance);
2575
2576 if (IS_ERR(kworker_task)) {
2577 dev_err(dev->dev,
2578 "Failed to create queue del task\n");
2579 for (i = dev->queue_count - 1; i > 0; i--)
2580 nvme_disable_queue(dev, i);
2581 return;
2582 }
2583
2584 dq.waiter = NULL;
2585 atomic_set(&dq.refcount, 0);
2586 dq.worker = &worker;
2587 for (i = dev->queue_count - 1; i > 0; i--) {
2588 struct nvme_queue *nvmeq = dev->queues[i];
2589
2590 if (nvme_suspend_queue(nvmeq))
2591 continue;
2592 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2593 nvmeq->cmdinfo.worker = dq.worker;
2594 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2595 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2596 }
2597 nvme_wait_dq(&dq, dev);
2598 kthread_stop(kworker_task);
2599}
2600
2601/*
2602* Remove the node from the device list and check
2603* for whether or not we need to stop the nvme_thread.
2604*/
2605static void nvme_dev_list_remove(struct nvme_dev *dev)
2606{
2607 struct task_struct *tmp = NULL;
2608
2609 spin_lock(&dev_list_lock);
2610 list_del_init(&dev->node);
2611 if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2612 tmp = nvme_thread;
2613 nvme_thread = NULL;
2614 }
2615 spin_unlock(&dev_list_lock);
2616
2617 if (tmp)
2618 kthread_stop(tmp);
2619}
2620
2621static void nvme_freeze_queues(struct nvme_dev *dev)
2622{
2623 struct nvme_ns *ns;
2624
2625 list_for_each_entry(ns, &dev->namespaces, list) {
2626 blk_mq_freeze_queue_start(ns->queue);
2627
2628 spin_lock_irq(ns->queue->queue_lock);
2629 queue_flag_set(QUEUE_FLAG_STOPPED, ns->queue);
2630 spin_unlock_irq(ns->queue->queue_lock);
2631
2632 blk_mq_cancel_requeue_work(ns->queue);
2633 blk_mq_stop_hw_queues(ns->queue);
2634 }
2635}
2636
2637static void nvme_unfreeze_queues(struct nvme_dev *dev)
2638{
2639 struct nvme_ns *ns;
2640
2641 list_for_each_entry(ns, &dev->namespaces, list) {
2642 queue_flag_clear_unlocked(QUEUE_FLAG_STOPPED, ns->queue);
2643 blk_mq_unfreeze_queue(ns->queue);
2644 blk_mq_start_stopped_hw_queues(ns->queue, true);
2645 blk_mq_kick_requeue_list(ns->queue);
2646 }
2647}
2648
2649static void nvme_dev_shutdown(struct nvme_dev *dev)
2650{
2651 int i;
2652 u32 csts = -1;
2653
2654 nvme_dev_list_remove(dev);
2655
2656 if (dev->bar) {
2657 nvme_freeze_queues(dev);
2658 csts = readl(&dev->bar->csts);
2659 }
2660 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2661 for (i = dev->queue_count - 1; i >= 0; i--) {
2662 struct nvme_queue *nvmeq = dev->queues[i];
2663 nvme_suspend_queue(nvmeq);
2664 }
2665 } else {
2666 nvme_disable_io_queues(dev);
2667 nvme_shutdown_ctrl(dev);
2668 nvme_disable_queue(dev, 0);
2669 }
2670 nvme_dev_unmap(dev);
2671
2672 for (i = dev->queue_count - 1; i >= 0; i--)
2673 nvme_clear_queue(dev->queues[i]);
2674}
2675
2676static void nvme_dev_remove(struct nvme_dev *dev)
2677{
2678 struct nvme_ns *ns;
2679
2680 list_for_each_entry(ns, &dev->namespaces, list)
2681 nvme_ns_remove(ns);
2682}
2683
2684static int nvme_setup_prp_pools(struct nvme_dev *dev)
2685{
2686 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2687 PAGE_SIZE, PAGE_SIZE, 0);
2688 if (!dev->prp_page_pool)
2689 return -ENOMEM;
2690
2691 /* Optimisation for I/Os between 4k and 128k */
2692 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2693 256, 256, 0);
2694 if (!dev->prp_small_pool) {
2695 dma_pool_destroy(dev->prp_page_pool);
2696 return -ENOMEM;
2697 }
2698 return 0;
2699}
2700
2701static void nvme_release_prp_pools(struct nvme_dev *dev)
2702{
2703 dma_pool_destroy(dev->prp_page_pool);
2704 dma_pool_destroy(dev->prp_small_pool);
2705}
2706
2707static DEFINE_IDA(nvme_instance_ida);
2708
2709static int nvme_set_instance(struct nvme_dev *dev)
2710{
2711 int instance, error;
2712
2713 do {
2714 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2715 return -ENODEV;
2716
2717 spin_lock(&dev_list_lock);
2718 error = ida_get_new(&nvme_instance_ida, &instance);
2719 spin_unlock(&dev_list_lock);
2720 } while (error == -EAGAIN);
2721
2722 if (error)
2723 return -ENODEV;
2724
2725 dev->instance = instance;
2726 return 0;
2727}
2728
2729static void nvme_release_instance(struct nvme_dev *dev)
2730{
2731 spin_lock(&dev_list_lock);
2732 ida_remove(&nvme_instance_ida, dev->instance);
2733 spin_unlock(&dev_list_lock);
2734}
2735
2736static void nvme_free_namespaces(struct nvme_dev *dev)
2737{
2738 struct nvme_ns *ns, *next;
2739
2740 list_for_each_entry_safe(ns, next, &dev->namespaces, list)
2741 nvme_free_namespace(ns);
2742}
2743
2744static void nvme_free_dev(struct kref *kref)
2745{
2746 struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2747
2748 put_device(dev->dev);
2749 put_device(dev->device);
2750 nvme_free_namespaces(dev);
2751 nvme_release_instance(dev);
2752 if (dev->tagset.tags)
2753 blk_mq_free_tag_set(&dev->tagset);
2754 if (dev->admin_q)
2755 blk_put_queue(dev->admin_q);
2756 kfree(dev->queues);
2757 kfree(dev->entry);
2758 kfree(dev);
2759}
2760
2761static int nvme_dev_open(struct inode *inode, struct file *f)
2762{
2763 struct nvme_dev *dev;
2764 int instance = iminor(inode);
2765 int ret = -ENODEV;
2766
2767 spin_lock(&dev_list_lock);
2768 list_for_each_entry(dev, &dev_list, node) {
2769 if (dev->instance == instance) {
2770 if (!dev->admin_q) {
2771 ret = -EWOULDBLOCK;
2772 break;
2773 }
2774 if (!kref_get_unless_zero(&dev->kref))
2775 break;
2776 f->private_data = dev;
2777 ret = 0;
2778 break;
2779 }
2780 }
2781 spin_unlock(&dev_list_lock);
2782
2783 return ret;
2784}
2785
2786static int nvme_dev_release(struct inode *inode, struct file *f)
2787{
2788 struct nvme_dev *dev = f->private_data;
2789 kref_put(&dev->kref, nvme_free_dev);
2790 return 0;
2791}
2792
2793static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2794{
2795 struct nvme_dev *dev = f->private_data;
2796 struct nvme_ns *ns;
2797
2798 switch (cmd) {
2799 case NVME_IOCTL_ADMIN_CMD:
2800 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2801 case NVME_IOCTL_IO_CMD:
2802 if (list_empty(&dev->namespaces))
2803 return -ENOTTY;
2804 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2805 return nvme_user_cmd(dev, ns, (void __user *)arg);
2806 case NVME_IOCTL_RESET:
2807 dev_warn(dev->dev, "resetting controller\n");
2808 return nvme_reset(dev);
2809 default:
2810 return -ENOTTY;
2811 }
2812}
2813
2814static const struct file_operations nvme_dev_fops = {
2815 .owner = THIS_MODULE,
2816 .open = nvme_dev_open,
2817 .release = nvme_dev_release,
2818 .unlocked_ioctl = nvme_dev_ioctl,
2819 .compat_ioctl = nvme_dev_ioctl,
2820};
2821
2822static void nvme_set_irq_hints(struct nvme_dev *dev)
2823{
2824 struct nvme_queue *nvmeq;
2825 int i;
2826
2827 for (i = 0; i < dev->online_queues; i++) {
2828 nvmeq = dev->queues[i];
2829
2830 if (!nvmeq->tags || !(*nvmeq->tags))
2831 continue;
2832
2833 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2834 blk_mq_tags_cpumask(*nvmeq->tags));
2835 }
2836}
2837
2838static int nvme_dev_start(struct nvme_dev *dev)
2839{
2840 int result;
2841 bool start_thread = false;
2842
2843 result = nvme_dev_map(dev);
2844 if (result)
2845 return result;
2846
2847 result = nvme_configure_admin_queue(dev);
2848 if (result)
2849 goto unmap;
2850
2851 spin_lock(&dev_list_lock);
2852 if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2853 start_thread = true;
2854 nvme_thread = NULL;
2855 }
2856 list_add(&dev->node, &dev_list);
2857 spin_unlock(&dev_list_lock);
2858
2859 if (start_thread) {
2860 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2861 wake_up_all(&nvme_kthread_wait);
2862 } else
2863 wait_event_killable(nvme_kthread_wait, nvme_thread);
2864
2865 if (IS_ERR_OR_NULL(nvme_thread)) {
2866 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2867 goto disable;
2868 }
2869
2870 nvme_init_queue(dev->queues[0], 0);
2871 result = nvme_alloc_admin_tags(dev);
2872 if (result)
2873 goto disable;
2874
2875 result = nvme_setup_io_queues(dev);
2876 if (result)
2877 goto free_tags;
2878
2879 nvme_set_irq_hints(dev);
2880
2881 dev->event_limit = 1;
2882 return result;
2883
2884 free_tags:
2885 nvme_dev_remove_admin(dev);
2886 blk_put_queue(dev->admin_q);
2887 dev->admin_q = NULL;
2888 dev->queues[0]->tags = NULL;
2889 disable:
2890 nvme_disable_queue(dev, 0);
2891 nvme_dev_list_remove(dev);
2892 unmap:
2893 nvme_dev_unmap(dev);
2894 return result;
2895}
2896
2897static int nvme_remove_dead_ctrl(void *arg)
2898{
2899 struct nvme_dev *dev = (struct nvme_dev *)arg;
2900 struct pci_dev *pdev = to_pci_dev(dev->dev);
2901
2902 if (pci_get_drvdata(pdev))
2903 pci_stop_and_remove_bus_device_locked(pdev);
2904 kref_put(&dev->kref, nvme_free_dev);
2905 return 0;
2906}
2907
2908static void nvme_remove_disks(struct work_struct *ws)
2909{
2910 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2911
2912 nvme_free_queues(dev, 1);
2913 nvme_dev_remove(dev);
2914}
2915
2916static int nvme_dev_resume(struct nvme_dev *dev)
2917{
2918 int ret;
2919
2920 ret = nvme_dev_start(dev);
2921 if (ret)
2922 return ret;
2923 if (dev->online_queues < 2) {
2924 spin_lock(&dev_list_lock);
2925 dev->reset_workfn = nvme_remove_disks;
2926 queue_work(nvme_workq, &dev->reset_work);
2927 spin_unlock(&dev_list_lock);
2928 } else {
2929 nvme_unfreeze_queues(dev);
2930 nvme_dev_add(dev);
2931 nvme_set_irq_hints(dev);
2932 }
2933 return 0;
2934}
2935
2936static void nvme_dead_ctrl(struct nvme_dev *dev)
2937{
2938 dev_warn(dev->dev, "Device failed to resume\n");
2939 kref_get(&dev->kref);
2940 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2941 dev->instance))) {
2942 dev_err(dev->dev,
2943 "Failed to start controller remove task\n");
2944 kref_put(&dev->kref, nvme_free_dev);
2945 }
2946}
2947
2948static void nvme_dev_reset(struct nvme_dev *dev)
2949{
2950 bool in_probe = work_busy(&dev->probe_work);
2951
2952 nvme_dev_shutdown(dev);
2953
2954 /* Synchronize with device probe so that work will see failure status
2955 * and exit gracefully without trying to schedule another reset */
2956 flush_work(&dev->probe_work);
2957
2958 /* Fail this device if reset occured during probe to avoid
2959 * infinite initialization loops. */
2960 if (in_probe) {
2961 nvme_dead_ctrl(dev);
2962 return;
2963 }
2964 /* Schedule device resume asynchronously so the reset work is available
2965 * to cleanup errors that may occur during reinitialization */
2966 schedule_work(&dev->probe_work);
2967}
2968
2969static void nvme_reset_failed_dev(struct work_struct *ws)
2970{
2971 struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2972 nvme_dev_reset(dev);
2973}
2974
2975static void nvme_reset_workfn(struct work_struct *work)
2976{
2977 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2978 dev->reset_workfn(work);
2979}
2980
2981static int nvme_reset(struct nvme_dev *dev)
2982{
2983 int ret = -EBUSY;
2984
2985 if (!dev->admin_q || blk_queue_dying(dev->admin_q))
2986 return -ENODEV;
2987
2988 spin_lock(&dev_list_lock);
2989 if (!work_pending(&dev->reset_work)) {
2990 dev->reset_workfn = nvme_reset_failed_dev;
2991 queue_work(nvme_workq, &dev->reset_work);
2992 ret = 0;
2993 }
2994 spin_unlock(&dev_list_lock);
2995
2996 if (!ret) {
2997 flush_work(&dev->reset_work);
2998 flush_work(&dev->probe_work);
2999 return 0;
3000 }
3001
3002 return ret;
3003}
3004
3005static ssize_t nvme_sysfs_reset(struct device *dev,
3006 struct device_attribute *attr, const char *buf,
3007 size_t count)
3008{
3009 struct nvme_dev *ndev = dev_get_drvdata(dev);
3010 int ret;
3011
3012 ret = nvme_reset(ndev);
3013 if (ret < 0)
3014 return ret;
3015
3016 return count;
3017}
3018static DEVICE_ATTR(reset_controller, S_IWUSR, NULL, nvme_sysfs_reset);
3019
3020static void nvme_async_probe(struct work_struct *work);
3021static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3022{
3023 int node, result = -ENOMEM;
3024 struct nvme_dev *dev;
3025
3026 node = dev_to_node(&pdev->dev);
3027 if (node == NUMA_NO_NODE)
3028 set_dev_node(&pdev->dev, 0);
3029
3030 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
3031 if (!dev)
3032 return -ENOMEM;
3033 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
3034 GFP_KERNEL, node);
3035 if (!dev->entry)
3036 goto free;
3037 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
3038 GFP_KERNEL, node);
3039 if (!dev->queues)
3040 goto free;
3041
3042 INIT_LIST_HEAD(&dev->namespaces);
3043 dev->reset_workfn = nvme_reset_failed_dev;
3044 INIT_WORK(&dev->reset_work, nvme_reset_workfn);
3045 dev->dev = get_device(&pdev->dev);
3046 pci_set_drvdata(pdev, dev);
3047 result = nvme_set_instance(dev);
3048 if (result)
3049 goto put_pci;
3050
3051 result = nvme_setup_prp_pools(dev);
3052 if (result)
3053 goto release;
3054
3055 kref_init(&dev->kref);
3056 dev->device = device_create(nvme_class, &pdev->dev,
3057 MKDEV(nvme_char_major, dev->instance),
3058 dev, "nvme%d", dev->instance);
3059 if (IS_ERR(dev->device)) {
3060 result = PTR_ERR(dev->device);
3061 goto release_pools;
3062 }
3063 get_device(dev->device);
3064 dev_set_drvdata(dev->device, dev);
3065
3066 result = device_create_file(dev->device, &dev_attr_reset_controller);
3067 if (result)
3068 goto put_dev;
3069
3070 INIT_LIST_HEAD(&dev->node);
3071 INIT_WORK(&dev->scan_work, nvme_dev_scan);
3072 INIT_WORK(&dev->probe_work, nvme_async_probe);
3073 schedule_work(&dev->probe_work);
3074 return 0;
3075
3076 put_dev:
3077 device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3078 put_device(dev->device);
3079 release_pools:
3080 nvme_release_prp_pools(dev);
3081 release:
3082 nvme_release_instance(dev);
3083 put_pci:
3084 put_device(dev->dev);
3085 free:
3086 kfree(dev->queues);
3087 kfree(dev->entry);
3088 kfree(dev);
3089 return result;
3090}
3091
3092static void nvme_async_probe(struct work_struct *work)
3093{
3094 struct nvme_dev *dev = container_of(work, struct nvme_dev, probe_work);
3095
3096 if (nvme_dev_resume(dev) && !work_busy(&dev->reset_work))
3097 nvme_dead_ctrl(dev);
3098}
3099
3100static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
3101{
3102 struct nvme_dev *dev = pci_get_drvdata(pdev);
3103
3104 if (prepare)
3105 nvme_dev_shutdown(dev);
3106 else
3107 nvme_dev_resume(dev);
3108}
3109
3110static void nvme_shutdown(struct pci_dev *pdev)
3111{
3112 struct nvme_dev *dev = pci_get_drvdata(pdev);
3113 nvme_dev_shutdown(dev);
3114}
3115
3116static void nvme_remove(struct pci_dev *pdev)
3117{
3118 struct nvme_dev *dev = pci_get_drvdata(pdev);
3119
3120 spin_lock(&dev_list_lock);
3121 list_del_init(&dev->node);
3122 spin_unlock(&dev_list_lock);
3123
3124 pci_set_drvdata(pdev, NULL);
3125 flush_work(&dev->probe_work);
3126 flush_work(&dev->reset_work);
3127 flush_work(&dev->scan_work);
3128 device_remove_file(dev->device, &dev_attr_reset_controller);
3129 nvme_dev_remove(dev);
3130 nvme_dev_shutdown(dev);
3131 nvme_dev_remove_admin(dev);
3132 device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3133 nvme_free_queues(dev, 0);
3134 nvme_release_prp_pools(dev);
3135 kref_put(&dev->kref, nvme_free_dev);
3136}
3137
3138/* These functions are yet to be implemented */
3139#define nvme_error_detected NULL
3140#define nvme_dump_registers NULL
3141#define nvme_link_reset NULL
3142#define nvme_slot_reset NULL
3143#define nvme_error_resume NULL
3144
3145#ifdef CONFIG_PM_SLEEP
3146static int nvme_suspend(struct device *dev)
3147{
3148 struct pci_dev *pdev = to_pci_dev(dev);
3149 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3150
3151 nvme_dev_shutdown(ndev);
3152 return 0;
3153}
3154
3155static int nvme_resume(struct device *dev)
3156{
3157 struct pci_dev *pdev = to_pci_dev(dev);
3158 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3159
3160 if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
3161 ndev->reset_workfn = nvme_reset_failed_dev;
3162 queue_work(nvme_workq, &ndev->reset_work);
3163 }
3164 return 0;
3165}
3166#endif
3167
3168static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
3169
3170static const struct pci_error_handlers nvme_err_handler = {
3171 .error_detected = nvme_error_detected,
3172 .mmio_enabled = nvme_dump_registers,
3173 .link_reset = nvme_link_reset,
3174 .slot_reset = nvme_slot_reset,
3175 .resume = nvme_error_resume,
3176 .reset_notify = nvme_reset_notify,
3177};
3178
3179/* Move to pci_ids.h later */
3180#define PCI_CLASS_STORAGE_EXPRESS 0x010802
3181
3182static const struct pci_device_id nvme_id_table[] = {
3183 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3184 { 0, }
3185};
3186MODULE_DEVICE_TABLE(pci, nvme_id_table);
3187
3188static struct pci_driver nvme_driver = {
3189 .name = "nvme",
3190 .id_table = nvme_id_table,
3191 .probe = nvme_probe,
3192 .remove = nvme_remove,
3193 .shutdown = nvme_shutdown,
3194 .driver = {
3195 .pm = &nvme_dev_pm_ops,
3196 },
3197 .err_handler = &nvme_err_handler,
3198};
3199
3200static int __init nvme_init(void)
3201{
3202 int result;
3203
3204 init_waitqueue_head(&nvme_kthread_wait);
3205
3206 nvme_workq = create_singlethread_workqueue("nvme");
3207 if (!nvme_workq)
3208 return -ENOMEM;
3209
3210 result = register_blkdev(nvme_major, "nvme");
3211 if (result < 0)
3212 goto kill_workq;
3213 else if (result > 0)
3214 nvme_major = result;
3215
3216 result = __register_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme",
3217 &nvme_dev_fops);
3218 if (result < 0)
3219 goto unregister_blkdev;
3220 else if (result > 0)
3221 nvme_char_major = result;
3222
3223 nvme_class = class_create(THIS_MODULE, "nvme");
3224 if (IS_ERR(nvme_class)) {
3225 result = PTR_ERR(nvme_class);
3226 goto unregister_chrdev;
3227 }
3228
3229 result = pci_register_driver(&nvme_driver);
3230 if (result)
3231 goto destroy_class;
3232 return 0;
3233
3234 destroy_class:
3235 class_destroy(nvme_class);
3236 unregister_chrdev:
3237 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3238 unregister_blkdev:
3239 unregister_blkdev(nvme_major, "nvme");
3240 kill_workq:
3241 destroy_workqueue(nvme_workq);
3242 return result;
3243}
3244
3245static void __exit nvme_exit(void)
3246{
3247 pci_unregister_driver(&nvme_driver);
3248 unregister_blkdev(nvme_major, "nvme");
3249 destroy_workqueue(nvme_workq);
3250 class_destroy(nvme_class);
3251 __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3252 BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3253 _nvme_check_size();
3254}
3255
3256MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3257MODULE_LICENSE("GPL");
3258MODULE_VERSION("1.0");
3259module_init(nvme_init);
3260module_exit(nvme_exit);
Linux 6.x block layer and io_uring tree(s)