drivers/block/nvme-core.c

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