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->cmd_type != REQ_TYPE_DRV_PRIV) {
 857                         req->errors = -EFAULT;
 858                         blk_mq_complete_request(req);
 859                         return BLK_MQ_RQ_QUEUE_OK;
 860                 }
 861         }
 862
 863         iod = nvme_alloc_iod(req, dev, GFP_ATOMIC);
 864         if (!iod)
 865                 return BLK_MQ_RQ_QUEUE_BUSY;
 866
 867         if (req->cmd_flags & REQ_DISCARD) {
 868                 void *range;
 869                 /*
 870                  * We reuse the small pool to allocate the 16-byte range here
 871                  * as it is not worth having a special pool for these or
 872                  * additional cases to handle freeing the iod.
 873                  */
 874                 range = dma_pool_alloc(dev->prp_small_pool, GFP_ATOMIC,
 875                                                 &iod->first_dma);
 876                 if (!range)
 877                         goto retry_cmd;
 878                 iod_list(iod)[0] = (__le64 *)range;
 879                 iod->npages = 0;
 880         } else if (req->nr_phys_segments) {
 881                 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
 882
 883                 sg_init_table(iod->sg, req->nr_phys_segments);
 884                 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 885                 if (!iod->nents)
 886                         goto error_cmd;
 887
 888                 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
 889                         goto retry_cmd;
 890
 891                 if (blk_rq_bytes(req) !=
 892                     nvme_setup_prps(dev, iod, blk_rq_bytes(req), GFP_ATOMIC)) {
 893                         dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
 894                         goto retry_cmd;
 895                 }
 896                 if (blk_integrity_rq(req)) {
 897                         if (blk_rq_count_integrity_sg(req->q, req->bio) != 1)
 898                                 goto error_cmd;
 899
 900                         sg_init_table(iod->meta_sg, 1);
 901                         if (blk_rq_map_integrity_sg(
 902                                         req->q, req->bio, iod->meta_sg) != 1)
 903                                 goto error_cmd;
 904
 905                         if (rq_data_dir(req))
 906                                 nvme_dif_remap(req, nvme_dif_prep);
 907
 908                         if (!dma_map_sg(nvmeq->q_dmadev, iod->meta_sg, 1, dma_dir))
 909                                 goto error_cmd;
 910                 }
 911         }
 912
 913         nvme_set_info(cmd, iod, req_completion);
 914         spin_lock_irq(&nvmeq->q_lock);
 915         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 916                 nvme_submit_priv(nvmeq, req, iod);
 917         else if (req->cmd_flags & REQ_DISCARD)
 918                 nvme_submit_discard(nvmeq, ns, req, iod);
 919         else if (req->cmd_flags & REQ_FLUSH)
 920                 nvme_submit_flush(nvmeq, ns, req->tag);
 921         else
 922                 nvme_submit_iod(nvmeq, iod, ns);
 923
 924         nvme_process_cq(nvmeq);
 925         spin_unlock_irq(&nvmeq->q_lock);
 926         return BLK_MQ_RQ_QUEUE_OK;
 927
 928  error_cmd:
 929         nvme_free_iod(dev, iod);
 930         return BLK_MQ_RQ_QUEUE_ERROR;
 931  retry_cmd:
 932         nvme_free_iod(dev, iod);
 933         return BLK_MQ_RQ_QUEUE_BUSY;
 934 }
 935
 936 static int nvme_process_cq(struct nvme_queue *nvmeq)
 937 {
 938         u16 head, phase;
 939
 940         head = nvmeq->cq_head;
 941         phase = nvmeq->cq_phase;
 942
 943         for (;;) {
 944                 void *ctx;
 945                 nvme_completion_fn fn;
 946                 struct nvme_completion cqe = nvmeq->cqes[head];
 947                 if ((le16_to_cpu(cqe.status) & 1) != phase)
 948                         break;
 949                 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
 950                 if (++head == nvmeq->q_depth) {
 951                         head = 0;
 952                         phase = !phase;
 953                 }
 954                 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
 955                 fn(nvmeq, ctx, &cqe);
 956         }
 957
 958         /* If the controller ignores the cq head doorbell and continuously
 959          * writes to the queue, it is theoretically possible to wrap around
 960          * the queue twice and mistakenly return IRQ_NONE.  Linux only
 961          * requires that 0.1% of your interrupts are handled, so this isn't
 962          * a big problem.
 963          */
 964         if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
 965                 return 0;
 966
 967         writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 968         nvmeq->cq_head = head;
 969         nvmeq->cq_phase = phase;
 970
 971         nvmeq->cqe_seen = 1;
 972         return 1;
 973 }
 974
 975 static irqreturn_t nvme_irq(int irq, void *data)
 976 {
 977         irqreturn_t result;
 978         struct nvme_queue *nvmeq = data;
 979         spin_lock(&nvmeq->q_lock);
 980         nvme_process_cq(nvmeq);
 981         result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
 982         nvmeq->cqe_seen = 0;
 983         spin_unlock(&nvmeq->q_lock);
 984         return result;
 985 }
 986
 987 static irqreturn_t nvme_irq_check(int irq, void *data)
 988 {
 989         struct nvme_queue *nvmeq = data;
 990         struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
 991         if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
 992                 return IRQ_NONE;
 993         return IRQ_WAKE_THREAD;
 994 }
 995
 996 /*
 997  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 998  * if the result is positive, it's an NVM Express status code
 999  */
1000 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1001                 void *buffer, void __user *ubuffer, unsigned bufflen,
1002                 u32 *result, unsigned timeout)
1003 {
1004         bool write = cmd->common.opcode & 1;
1005         struct bio *bio = NULL;
1006         struct request *req;
1007         int ret;
1008
1009         req = blk_mq_alloc_request(q, write, GFP_KERNEL, false);
1010         if (IS_ERR(req))
1011                 return PTR_ERR(req);
1012
1013         req->cmd_type = REQ_TYPE_DRV_PRIV;
1014         req->cmd_flags |= REQ_FAILFAST_DRIVER;
1015         req->__data_len = 0;
1016         req->__sector = (sector_t) -1;
1017         req->bio = req->biotail = NULL;
1018
1019         req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
1020
1021         req->cmd = (unsigned char *)cmd;
1022         req->cmd_len = sizeof(struct nvme_command);
1023         req->special = (void *)0;
1024
1025         if (buffer && bufflen) {
1026                 ret = blk_rq_map_kern(q, req, buffer, bufflen, __GFP_WAIT);
1027                 if (ret)
1028                         goto out;
1029         } else if (ubuffer && bufflen) {
1030                 ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen, __GFP_WAIT);
1031                 if (ret)
1032                         goto out;
1033                 bio = req->bio;
1034         }
1035
1036         blk_execute_rq(req->q, NULL, req, 0);
1037         if (bio)
1038                 blk_rq_unmap_user(bio);
1039         if (result)
1040                 *result = (u32)(uintptr_t)req->special;
1041         ret = req->errors;
1042  out:
1043         blk_mq_free_request(req);
1044         return ret;
1045 }
1046
1047 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1048                 void *buffer, unsigned bufflen)
1049 {
1050         return __nvme_submit_sync_cmd(q, cmd, buffer, NULL, bufflen, NULL, 0);
1051 }
1052
1053 static int nvme_submit_async_admin_req(struct nvme_dev *dev)
1054 {
1055         struct nvme_queue *nvmeq = dev->queues[0];
1056         struct nvme_command c;
1057         struct nvme_cmd_info *cmd_info;
1058         struct request *req;
1059
1060         req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, true);
1061         if (IS_ERR(req))
1062                 return PTR_ERR(req);
1063
1064         req->cmd_flags |= REQ_NO_TIMEOUT;
1065         cmd_info = blk_mq_rq_to_pdu(req);
1066         nvme_set_info(cmd_info, NULL, async_req_completion);
1067
1068         memset(&c, 0, sizeof(c));
1069         c.common.opcode = nvme_admin_async_event;
1070         c.common.command_id = req->tag;
1071
1072         blk_mq_free_request(req);
1073         return __nvme_submit_cmd(nvmeq, &c);
1074 }
1075
1076 static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
1077                         struct nvme_command *cmd,
1078                         struct async_cmd_info *cmdinfo, unsigned timeout)
1079 {
1080         struct nvme_queue *nvmeq = dev->queues[0];
1081         struct request *req;
1082         struct nvme_cmd_info *cmd_rq;
1083
1084         req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
1085         if (IS_ERR(req))
1086                 return PTR_ERR(req);
1087
1088         req->timeout = timeout;
1089         cmd_rq = blk_mq_rq_to_pdu(req);
1090         cmdinfo->req = req;
1091         nvme_set_info(cmd_rq, cmdinfo, async_completion);
1092         cmdinfo->status = -EINTR;
1093
1094         cmd->common.command_id = req->tag;
1095
1096         return nvme_submit_cmd(nvmeq, cmd);
1097 }
1098
1099 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1100 {
1101         struct nvme_command c;
1102
1103         memset(&c, 0, sizeof(c));
1104         c.delete_queue.opcode = opcode;
1105         c.delete_queue.qid = cpu_to_le16(id);
1106
1107         return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1108 }
1109
1110 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1111                                                 struct nvme_queue *nvmeq)
1112 {
1113         struct nvme_command c;
1114         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1115
1116         /*
1117          * Note: we (ab)use the fact the the prp fields survive if no data
1118          * is attached to the request.
1119          */
1120         memset(&c, 0, sizeof(c));
1121         c.create_cq.opcode = nvme_admin_create_cq;
1122         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1123         c.create_cq.cqid = cpu_to_le16(qid);
1124         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1125         c.create_cq.cq_flags = cpu_to_le16(flags);
1126         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1127
1128         return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1129 }
1130
1131 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1132                                                 struct nvme_queue *nvmeq)
1133 {
1134         struct nvme_command c;
1135         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
1136
1137         /*
1138          * Note: we (ab)use the fact the the prp fields survive if no data
1139          * is attached to the request.
1140          */
1141         memset(&c, 0, sizeof(c));
1142         c.create_sq.opcode = nvme_admin_create_sq;
1143         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1144         c.create_sq.sqid = cpu_to_le16(qid);
1145         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1146         c.create_sq.sq_flags = cpu_to_le16(flags);
1147         c.create_sq.cqid = cpu_to_le16(qid);
1148
1149         return nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1150 }
1151
1152 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1153 {
1154         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1155 }
1156
1157 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1158 {
1159         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1160 }
1161
1162 int nvme_identify_ctrl(struct nvme_dev *dev, struct nvme_id_ctrl **id)
1163 {
1164         struct nvme_command c = {
1165                 .identify.opcode = nvme_admin_identify,
1166                 .identify.cns = cpu_to_le32(1),
1167         };
1168         int error;
1169
1170         *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1171         if (!*id)
1172                 return -ENOMEM;
1173
1174         error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1175                         sizeof(struct nvme_id_ctrl));
1176         if (error)
1177                 kfree(*id);
1178         return error;
1179 }
1180
1181 int nvme_identify_ns(struct nvme_dev *dev, unsigned nsid,
1182                 struct nvme_id_ns **id)
1183 {
1184         struct nvme_command c = {
1185                 .identify.opcode = nvme_admin_identify,
1186                 .identify.nsid = cpu_to_le32(nsid),
1187         };
1188         int error;
1189
1190         *id = kmalloc(sizeof(struct nvme_id_ns), GFP_KERNEL);
1191         if (!*id)
1192                 return -ENOMEM;
1193
1194         error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1195                         sizeof(struct nvme_id_ns));
1196         if (error)
1197                 kfree(*id);
1198         return error;
1199 }
1200
1201 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1202                                         dma_addr_t dma_addr, u32 *result)
1203 {
1204         struct nvme_command c;
1205
1206         memset(&c, 0, sizeof(c));
1207         c.features.opcode = nvme_admin_get_features;
1208         c.features.nsid = cpu_to_le32(nsid);
1209         c.features.prp1 = cpu_to_le64(dma_addr);
1210         c.features.fid = cpu_to_le32(fid);
1211
1212         return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1213                         result, 0);
1214 }
1215
1216 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1217                                         dma_addr_t dma_addr, u32 *result)
1218 {
1219         struct nvme_command c;
1220
1221         memset(&c, 0, sizeof(c));
1222         c.features.opcode = nvme_admin_set_features;
1223         c.features.prp1 = cpu_to_le64(dma_addr);
1224         c.features.fid = cpu_to_le32(fid);
1225         c.features.dword11 = cpu_to_le32(dword11);
1226
1227         return __nvme_submit_sync_cmd(dev->admin_q, &c, NULL, NULL, 0,
1228                         result, 0);
1229 }
1230
1231 int nvme_get_log_page(struct nvme_dev *dev, struct nvme_smart_log **log)
1232 {
1233         struct nvme_command c = {
1234                 .common.opcode = nvme_admin_get_log_page,
1235                 .common.nsid = cpu_to_le32(0xFFFFFFFF),
1236                 .common.cdw10[0] = cpu_to_le32(
1237                         (((sizeof(struct nvme_smart_log) / 4) - 1) << 16) |
1238                          NVME_LOG_SMART),
1239         };
1240         int error;
1241
1242         *log = kmalloc(sizeof(struct nvme_smart_log), GFP_KERNEL);
1243         if (!*log)
1244                 return -ENOMEM;
1245
1246         error = nvme_submit_sync_cmd(dev->admin_q, &c, *log,
1247                         sizeof(struct nvme_smart_log));
1248         if (error)
1249                 kfree(*log);
1250         return error;
1251 }
1252
1253 /**
1254  * nvme_abort_req - Attempt aborting a request
1255  *
1256  * Schedule controller reset if the command was already aborted once before and
1257  * still hasn't been returned to the driver, or if this is the admin queue.
1258  */
1259 static void nvme_abort_req(struct request *req)
1260 {
1261         struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1262         struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1263         struct nvme_dev *dev = nvmeq->dev;
1264         struct request *abort_req;
1265         struct nvme_cmd_info *abort_cmd;
1266         struct nvme_command cmd;
1267
1268         if (!nvmeq->qid || cmd_rq->aborted) {
1269                 unsigned long flags;
1270
1271                 spin_lock_irqsave(&dev_list_lock, flags);
1272                 if (work_busy(&dev->reset_work))
1273                         goto out;
1274                 list_del_init(&dev->node);
1275                 dev_warn(dev->dev, "I/O %d QID %d timeout, reset controller\n",
1276                                                         req->tag, nvmeq->qid);
1277                 dev->reset_workfn = nvme_reset_failed_dev;
1278                 queue_work(nvme_workq, &dev->reset_work);
1279  out:
1280                 spin_unlock_irqrestore(&dev_list_lock, flags);
1281                 return;
1282         }
1283
1284         if (!dev->abort_limit)
1285                 return;
1286
1287         abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1288                                                                         false);
1289         if (IS_ERR(abort_req))
1290                 return;
1291
1292         abort_cmd = blk_mq_rq_to_pdu(abort_req);
1293         nvme_set_info(abort_cmd, abort_req, abort_completion);
1294
1295         memset(&cmd, 0, sizeof(cmd));
1296         cmd.abort.opcode = nvme_admin_abort_cmd;
1297         cmd.abort.cid = req->tag;
1298         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1299         cmd.abort.command_id = abort_req->tag;
1300
1301         --dev->abort_limit;
1302         cmd_rq->aborted = 1;
1303
1304         dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1305                                                         nvmeq->qid);
1306         if (nvme_submit_cmd(dev->queues[0], &cmd) < 0) {
1307                 dev_warn(nvmeq->q_dmadev,
1308                                 "Could not abort I/O %d QID %d",
1309                                 req->tag, nvmeq->qid);
1310                 blk_mq_free_request(abort_req);
1311         }
1312 }
1313
1314 static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved)
1315 {
1316         struct nvme_queue *nvmeq = data;
1317         void *ctx;
1318         nvme_completion_fn fn;
1319         struct nvme_cmd_info *cmd;
1320         struct nvme_completion cqe;
1321
1322         if (!blk_mq_request_started(req))
1323                 return;
1324
1325         cmd = blk_mq_rq_to_pdu(req);
1326
1327         if (cmd->ctx == CMD_CTX_CANCELLED)
1328                 return;
1329
1330         if (blk_queue_dying(req->q))
1331                 cqe.status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1332         else
1333                 cqe.status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1334
1335
1336         dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1337                                                 req->tag, nvmeq->qid);
1338         ctx = cancel_cmd_info(cmd, &fn);
1339         fn(nvmeq, ctx, &cqe);
1340 }
1341
1342 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1343 {
1344         struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1345         struct nvme_queue *nvmeq = cmd->nvmeq;
1346
1347         dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1348                                                         nvmeq->qid);
1349         spin_lock_irq(&nvmeq->q_lock);
1350         nvme_abort_req(req);
1351         spin_unlock_irq(&nvmeq->q_lock);
1352
1353         /*
1354          * The aborted req will be completed on receiving the abort req.
1355          * We enable the timer again. If hit twice, it'll cause a device reset,
1356          * as the device then is in a faulty state.
1357          */
1358         return BLK_EH_RESET_TIMER;
1359 }
1360
1361 static void nvme_free_queue(struct nvme_queue *nvmeq)
1362 {
1363         dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1364                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1365         dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1366                                         nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1367         kfree(nvmeq);
1368 }
1369
1370 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1371 {
1372         int i;
1373
1374         for (i = dev->queue_count - 1; i >= lowest; i--) {
1375                 struct nvme_queue *nvmeq = dev->queues[i];
1376                 dev->queue_count--;
1377                 dev->queues[i] = NULL;
1378                 nvme_free_queue(nvmeq);
1379         }
1380 }
1381
1382 /**
1383  * nvme_suspend_queue - put queue into suspended state
1384  * @nvmeq - queue to suspend
1385  */
1386 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1387 {
1388         int vector;
1389
1390         spin_lock_irq(&nvmeq->q_lock);
1391         if (nvmeq->cq_vector == -1) {
1392                 spin_unlock_irq(&nvmeq->q_lock);
1393                 return 1;
1394         }
1395         vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1396         nvmeq->dev->online_queues--;
1397         nvmeq->cq_vector = -1;
1398         spin_unlock_irq(&nvmeq->q_lock);
1399
1400         if (!nvmeq->qid && nvmeq->dev->admin_q)
1401                 blk_mq_freeze_queue_start(nvmeq->dev->admin_q);
1402
1403         irq_set_affinity_hint(vector, NULL);
1404         free_irq(vector, nvmeq);
1405
1406         return 0;
1407 }
1408
1409 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1410 {
1411         spin_lock_irq(&nvmeq->q_lock);
1412         if (nvmeq->tags && *nvmeq->tags)
1413                 blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq);
1414         spin_unlock_irq(&nvmeq->q_lock);
1415 }
1416
1417 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1418 {
1419         struct nvme_queue *nvmeq = dev->queues[qid];
1420
1421         if (!nvmeq)
1422                 return;
1423         if (nvme_suspend_queue(nvmeq))
1424                 return;
1425
1426         /* Don't tell the adapter to delete the admin queue.
1427          * Don't tell a removed adapter to delete IO queues. */
1428         if (qid && readl(&dev->bar->csts) != -1) {
1429                 adapter_delete_sq(dev, qid);
1430                 adapter_delete_cq(dev, qid);
1431         }
1432
1433         spin_lock_irq(&nvmeq->q_lock);
1434         nvme_process_cq(nvmeq);
1435         spin_unlock_irq(&nvmeq->q_lock);
1436 }
1437
1438 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1439                                                         int depth)
1440 {
1441         struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1442         if (!nvmeq)
1443                 return NULL;
1444
1445         nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1446                                           &nvmeq->cq_dma_addr, GFP_KERNEL);
1447         if (!nvmeq->cqes)
1448                 goto free_nvmeq;
1449
1450         nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1451                                         &nvmeq->sq_dma_addr, GFP_KERNEL);
1452         if (!nvmeq->sq_cmds)
1453                 goto free_cqdma;
1454
1455         nvmeq->q_dmadev = dev->dev;
1456         nvmeq->dev = dev;
1457         snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1458                         dev->instance, qid);
1459         spin_lock_init(&nvmeq->q_lock);
1460         nvmeq->cq_head = 0;
1461         nvmeq->cq_phase = 1;
1462         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1463         nvmeq->q_depth = depth;
1464         nvmeq->qid = qid;
1465         dev->queues[qid] = nvmeq;
1466
1467         /* make sure queue descriptor is set before queue count, for kthread */
1468         mb();
1469         dev->queue_count++;
1470
1471         return nvmeq;
1472
1473  free_cqdma:
1474         dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1475                                                         nvmeq->cq_dma_addr);
1476  free_nvmeq:
1477         kfree(nvmeq);
1478         return NULL;
1479 }
1480
1481 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1482                                                         const char *name)
1483 {
1484         if (use_threaded_interrupts)
1485                 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1486                                         nvme_irq_check, nvme_irq, IRQF_SHARED,
1487                                         name, nvmeq);
1488         return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1489                                 IRQF_SHARED, name, nvmeq);
1490 }
1491
1492 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1493 {
1494         struct nvme_dev *dev = nvmeq->dev;
1495
1496         spin_lock_irq(&nvmeq->q_lock);
1497         nvmeq->sq_tail = 0;
1498         nvmeq->cq_head = 0;
1499         nvmeq->cq_phase = 1;
1500         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1501         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1502         dev->online_queues++;
1503         spin_unlock_irq(&nvmeq->q_lock);
1504 }
1505
1506 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1507 {
1508         struct nvme_dev *dev = nvmeq->dev;
1509         int result;
1510
1511         nvmeq->cq_vector = qid - 1;
1512         result = adapter_alloc_cq(dev, qid, nvmeq);
1513         if (result < 0)
1514                 return result;
1515
1516         result = adapter_alloc_sq(dev, qid, nvmeq);
1517         if (result < 0)
1518                 goto release_cq;
1519
1520         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1521         if (result < 0)
1522                 goto release_sq;
1523
1524         nvme_init_queue(nvmeq, qid);
1525         return result;
1526
1527  release_sq:
1528         adapter_delete_sq(dev, qid);
1529  release_cq:
1530         adapter_delete_cq(dev, qid);
1531         return result;
1532 }
1533
1534 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1535 {
1536         unsigned long timeout;
1537         u32 bit = enabled ? NVME_CSTS_RDY : 0;
1538
1539         timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1540
1541         while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1542                 msleep(100);
1543                 if (fatal_signal_pending(current))
1544                         return -EINTR;
1545                 if (time_after(jiffies, timeout)) {
1546                         dev_err(dev->dev,
1547                                 "Device not ready; aborting %s\n", enabled ?
1548                                                 "initialisation" : "reset");
1549                         return -ENODEV;
1550                 }
1551         }
1552
1553         return 0;
1554 }
1555
1556 /*
1557  * If the device has been passed off to us in an enabled state, just clear
1558  * the enabled bit.  The spec says we should set the 'shutdown notification
1559  * bits', but doing so may cause the device to complete commands to the
1560  * admin queue ... and we don't know what memory that might be pointing at!
1561  */
1562 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1563 {
1564         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1565         dev->ctrl_config &= ~NVME_CC_ENABLE;
1566         writel(dev->ctrl_config, &dev->bar->cc);
1567
1568         return nvme_wait_ready(dev, cap, false);
1569 }
1570
1571 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1572 {
1573         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1574         dev->ctrl_config |= NVME_CC_ENABLE;
1575         writel(dev->ctrl_config, &dev->bar->cc);
1576
1577         return nvme_wait_ready(dev, cap, true);
1578 }
1579
1580 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1581 {
1582         unsigned long timeout;
1583
1584         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1585         dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1586
1587         writel(dev->ctrl_config, &dev->bar->cc);
1588
1589         timeout = SHUTDOWN_TIMEOUT + jiffies;
1590         while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1591                                                         NVME_CSTS_SHST_CMPLT) {
1592                 msleep(100);
1593                 if (fatal_signal_pending(current))
1594                         return -EINTR;
1595                 if (time_after(jiffies, timeout)) {
1596                         dev_err(dev->dev,
1597                                 "Device shutdown incomplete; abort shutdown\n");
1598                         return -ENODEV;
1599                 }
1600         }
1601
1602         return 0;
1603 }
1604
1605 static struct blk_mq_ops nvme_mq_admin_ops = {
1606         .queue_rq       = nvme_queue_rq,
1607         .map_queue      = blk_mq_map_queue,
1608         .init_hctx      = nvme_admin_init_hctx,
1609         .init_request   = nvme_admin_init_request,
1610         .timeout        = nvme_timeout,
1611 };
1612
1613 static struct blk_mq_ops nvme_mq_ops = {
1614         .queue_rq       = nvme_queue_rq,
1615         .map_queue      = blk_mq_map_queue,
1616         .init_hctx      = nvme_init_hctx,
1617         .init_request   = nvme_init_request,
1618         .timeout        = nvme_timeout,
1619 };
1620
1621 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1622 {
1623         if (dev->admin_q && !blk_queue_dying(dev->admin_q)) {
1624                 blk_cleanup_queue(dev->admin_q);
1625                 blk_mq_free_tag_set(&dev->admin_tagset);
1626         }
1627 }
1628
1629 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1630 {
1631         if (!dev->admin_q) {
1632                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1633                 dev->admin_tagset.nr_hw_queues = 1;
1634                 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1635                 dev->admin_tagset.reserved_tags = 1;
1636                 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1637                 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1638                 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1639                 dev->admin_tagset.driver_data = dev;
1640
1641                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1642                         return -ENOMEM;
1643
1644                 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1645                 if (IS_ERR(dev->admin_q)) {
1646                         blk_mq_free_tag_set(&dev->admin_tagset);
1647                         return -ENOMEM;
1648                 }
1649                 if (!blk_get_queue(dev->admin_q)) {
1650                         nvme_dev_remove_admin(dev);
1651                         return -ENODEV;
1652                 }
1653         } else
1654                 blk_mq_unfreeze_queue(dev->admin_q);
1655
1656         return 0;
1657 }
1658
1659 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1660 {
1661         int result;
1662         u32 aqa;
1663         u64 cap = readq(&dev->bar->cap);
1664         struct nvme_queue *nvmeq;
1665         unsigned page_shift = PAGE_SHIFT;
1666         unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1667         unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1668
1669         if (page_shift < dev_page_min) {
1670                 dev_err(dev->dev,
1671                                 "Minimum device page size (%u) too large for "
1672                                 "host (%u)\n", 1 << dev_page_min,
1673                                 1 << page_shift);
1674                 return -ENODEV;
1675         }
1676         if (page_shift > dev_page_max) {
1677                 dev_info(dev->dev,
1678                                 "Device maximum page size (%u) smaller than "
1679                                 "host (%u); enabling work-around\n",
1680                                 1 << dev_page_max, 1 << page_shift);
1681                 page_shift = dev_page_max;
1682         }
1683
1684         result = nvme_disable_ctrl(dev, cap);
1685         if (result < 0)
1686                 return result;
1687
1688         nvmeq = dev->queues[0];
1689         if (!nvmeq) {
1690                 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1691                 if (!nvmeq)
1692                         return -ENOMEM;
1693         }
1694
1695         aqa = nvmeq->q_depth - 1;
1696         aqa |= aqa << 16;
1697
1698         dev->page_size = 1 << page_shift;
1699
1700         dev->ctrl_config = NVME_CC_CSS_NVM;
1701         dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1702         dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1703         dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1704
1705         writel(aqa, &dev->bar->aqa);
1706         writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1707         writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1708
1709         result = nvme_enable_ctrl(dev, cap);
1710         if (result)
1711                 goto free_nvmeq;
1712
1713         nvmeq->cq_vector = 0;
1714         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1715         if (result)
1716                 goto free_nvmeq;
1717
1718         return result;
1719
1720  free_nvmeq:
1721         nvme_free_queues(dev, 0);
1722         return result;
1723 }
1724
1725 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1726 {
1727         struct nvme_dev *dev = ns->dev;
1728         struct nvme_user_io io;
1729         struct nvme_command c;
1730         unsigned length, meta_len;
1731         int status, write;
1732         dma_addr_t meta_dma = 0;
1733         void *meta = NULL;
1734         void __user *metadata;
1735
1736         if (copy_from_user(&io, uio, sizeof(io)))
1737                 return -EFAULT;
1738
1739         switch (io.opcode) {
1740         case nvme_cmd_write:
1741         case nvme_cmd_read:
1742         case nvme_cmd_compare:
1743                 break;
1744         default:
1745                 return -EINVAL;
1746         }
1747
1748         length = (io.nblocks + 1) << ns->lba_shift;
1749         meta_len = (io.nblocks + 1) * ns->ms;
1750         metadata = (void __user *)(unsigned long)io.metadata;
1751         write = io.opcode & 1;
1752
1753         if (ns->ext) {
1754                 length += meta_len;
1755                 meta_len = 0;
1756         }
1757         if (meta_len) {
1758                 if (((io.metadata & 3) || !io.metadata) && !ns->ext)
1759                         return -EINVAL;
1760
1761                 meta = dma_alloc_coherent(dev->dev, meta_len,
1762                                                 &meta_dma, GFP_KERNEL);
1763
1764                 if (!meta) {
1765                         status = -ENOMEM;
1766                         goto unmap;
1767                 }
1768                 if (write) {
1769                         if (copy_from_user(meta, metadata, meta_len)) {
1770                                 status = -EFAULT;
1771                                 goto unmap;
1772                         }
1773                 }
1774         }
1775
1776         memset(&c, 0, sizeof(c));
1777         c.rw.opcode = io.opcode;
1778         c.rw.flags = io.flags;
1779         c.rw.nsid = cpu_to_le32(ns->ns_id);
1780         c.rw.slba = cpu_to_le64(io.slba);
1781         c.rw.length = cpu_to_le16(io.nblocks);
1782         c.rw.control = cpu_to_le16(io.control);
1783         c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1784         c.rw.reftag = cpu_to_le32(io.reftag);
1785         c.rw.apptag = cpu_to_le16(io.apptag);
1786         c.rw.appmask = cpu_to_le16(io.appmask);
1787         c.rw.metadata = cpu_to_le64(meta_dma);
1788
1789         status = __nvme_submit_sync_cmd(ns->queue, &c, NULL,
1790                         (void __user *)io.addr, length, NULL, 0);
1791  unmap:
1792         if (meta) {
1793                 if (status == NVME_SC_SUCCESS && !write) {
1794                         if (copy_to_user(metadata, meta, meta_len))
1795                                 status = -EFAULT;
1796                 }
1797                 dma_free_coherent(dev->dev, meta_len, meta, meta_dma);
1798         }
1799         return status;
1800 }
1801
1802 static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1803                         struct nvme_passthru_cmd __user *ucmd)
1804 {
1805         struct nvme_passthru_cmd cmd;
1806         struct nvme_command c;
1807         unsigned timeout = 0;
1808         int status;
1809
1810         if (!capable(CAP_SYS_ADMIN))
1811                 return -EACCES;
1812         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1813                 return -EFAULT;
1814
1815         memset(&c, 0, sizeof(c));
1816         c.common.opcode = cmd.opcode;
1817         c.common.flags = cmd.flags;
1818         c.common.nsid = cpu_to_le32(cmd.nsid);
1819         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1820         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1821         c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1822         c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1823         c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1824         c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1825         c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1826         c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1827
1828         if (cmd.timeout_ms)
1829                 timeout = msecs_to_jiffies(cmd.timeout_ms);
1830
1831         status = __nvme_submit_sync_cmd(ns ? ns->queue : dev->admin_q, &c,
1832                         NULL, (void __user *)cmd.addr, cmd.data_len,
1833                         &cmd.result, timeout);
1834         if (status >= 0) {
1835                 if (put_user(cmd.result, &ucmd->result))
1836                         return -EFAULT;
1837         }
1838
1839         return status;
1840 }
1841
1842 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1843                                                         unsigned long arg)
1844 {
1845         struct nvme_ns *ns = bdev->bd_disk->private_data;
1846
1847         switch (cmd) {
1848         case NVME_IOCTL_ID:
1849                 force_successful_syscall_return();
1850                 return ns->ns_id;
1851         case NVME_IOCTL_ADMIN_CMD:
1852                 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1853         case NVME_IOCTL_IO_CMD:
1854                 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1855         case NVME_IOCTL_SUBMIT_IO:
1856                 return nvme_submit_io(ns, (void __user *)arg);
1857         case SG_GET_VERSION_NUM:
1858                 return nvme_sg_get_version_num((void __user *)arg);
1859         case SG_IO:
1860                 return nvme_sg_io(ns, (void __user *)arg);
1861         default:
1862                 return -ENOTTY;
1863         }
1864 }
1865
1866 #ifdef CONFIG_COMPAT
1867 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1868                                         unsigned int cmd, unsigned long arg)
1869 {
1870         switch (cmd) {
1871         case SG_IO:
1872                 return -ENOIOCTLCMD;
1873         }
1874         return nvme_ioctl(bdev, mode, cmd, arg);
1875 }
1876 #else
1877 #define nvme_compat_ioctl       NULL
1878 #endif
1879
1880 static int nvme_open(struct block_device *bdev, fmode_t mode)
1881 {
1882         int ret = 0;
1883         struct nvme_ns *ns;
1884
1885         spin_lock(&dev_list_lock);
1886         ns = bdev->bd_disk->private_data;
1887         if (!ns)
1888                 ret = -ENXIO;
1889         else if (!kref_get_unless_zero(&ns->dev->kref))
1890                 ret = -ENXIO;
1891         spin_unlock(&dev_list_lock);
1892
1893         return ret;
1894 }
1895
1896 static void nvme_free_dev(struct kref *kref);
1897
1898 static void nvme_release(struct gendisk *disk, fmode_t mode)
1899 {
1900         struct nvme_ns *ns = disk->private_data;
1901         struct nvme_dev *dev = ns->dev;
1902
1903         kref_put(&dev->kref, nvme_free_dev);
1904 }
1905
1906 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1907 {
1908         /* some standard values */
1909         geo->heads = 1 << 6;
1910         geo->sectors = 1 << 5;
1911         geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1912         return 0;
1913 }
1914
1915 static void nvme_config_discard(struct nvme_ns *ns)
1916 {
1917         u32 logical_block_size = queue_logical_block_size(ns->queue);
1918         ns->queue->limits.discard_zeroes_data = 0;
1919         ns->queue->limits.discard_alignment = logical_block_size;
1920         ns->queue->limits.discard_granularity = logical_block_size;
1921         ns->queue->limits.max_discard_sectors = 0xffffffff;
1922         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1923 }
1924
1925 static int nvme_revalidate_disk(struct gendisk *disk)
1926 {
1927         struct nvme_ns *ns = disk->private_data;
1928         struct nvme_dev *dev = ns->dev;
1929         struct nvme_id_ns *id;
1930         u8 lbaf, pi_type;
1931         u16 old_ms;
1932         unsigned short bs;
1933
1934         if (nvme_identify_ns(dev, ns->ns_id, &id)) {
1935                 dev_warn(dev->dev, "%s: Identify failure nvme%dn%d\n", __func__,
1936                                                 dev->instance, ns->ns_id);
1937                 return -ENODEV;
1938         }
1939         if (id->ncap == 0) {
1940                 kfree(id);
1941                 return -ENODEV;
1942         }
1943
1944         old_ms = ns->ms;
1945         lbaf = id->flbas & NVME_NS_FLBAS_LBA_MASK;
1946         ns->lba_shift = id->lbaf[lbaf].ds;
1947         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1948         ns->ext = ns->ms && (id->flbas & NVME_NS_FLBAS_META_EXT);
1949
1950         /*
1951          * If identify namespace failed, use default 512 byte block size so
1952          * block layer can use before failing read/write for 0 capacity.
1953          */
1954         if (ns->lba_shift == 0)
1955                 ns->lba_shift = 9;
1956         bs = 1 << ns->lba_shift;
1957
1958         /* XXX: PI implementation requires metadata equal t10 pi tuple size */
1959         pi_type = ns->ms == sizeof(struct t10_pi_tuple) ?
1960                                         id->dps & NVME_NS_DPS_PI_MASK : 0;
1961
1962         if (blk_get_integrity(disk) && (ns->pi_type != pi_type ||
1963                                 ns->ms != old_ms ||
1964                                 bs != queue_logical_block_size(disk->queue) ||
1965                                 (ns->ms && ns->ext)))
1966                 blk_integrity_unregister(disk);
1967
1968         ns->pi_type = pi_type;
1969         blk_queue_logical_block_size(ns->queue, bs);
1970
1971         if (ns->ms && !blk_get_integrity(disk) && (disk->flags & GENHD_FL_UP) &&
1972                                                                 !ns->ext)
1973                 nvme_init_integrity(ns);
1974
1975         if (ns->ms && !blk_get_integrity(disk))
1976                 set_capacity(disk, 0);
1977         else
1978                 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1979
1980         if (dev->oncs & NVME_CTRL_ONCS_DSM)
1981                 nvme_config_discard(ns);
1982
1983         kfree(id);
1984         return 0;
1985 }
1986
1987 static const struct block_device_operations nvme_fops = {
1988         .owner          = THIS_MODULE,
1989         .ioctl          = nvme_ioctl,
1990         .compat_ioctl   = nvme_compat_ioctl,
1991         .open           = nvme_open,
1992         .release        = nvme_release,
1993         .getgeo         = nvme_getgeo,
1994         .revalidate_disk= nvme_revalidate_disk,
1995 };
1996
1997 static int nvme_kthread(void *data)
1998 {
1999         struct nvme_dev *dev, *next;
2000
2001         while (!kthread_should_stop()) {
2002                 set_current_state(TASK_INTERRUPTIBLE);
2003                 spin_lock(&dev_list_lock);
2004                 list_for_each_entry_safe(dev, next, &dev_list, node) {
2005                         int i;
2006                         if (readl(&dev->bar->csts) & NVME_CSTS_CFS) {
2007                                 if (work_busy(&dev->reset_work))
2008                                         continue;
2009                                 list_del_init(&dev->node);
2010                                 dev_warn(dev->dev,
2011                                         "Failed status: %x, reset controller\n",
2012                                         readl(&dev->bar->csts));
2013                                 dev->reset_workfn = nvme_reset_failed_dev;
2014                                 queue_work(nvme_workq, &dev->reset_work);
2015                                 continue;
2016                         }
2017                         for (i = 0; i < dev->queue_count; i++) {
2018                                 struct nvme_queue *nvmeq = dev->queues[i];
2019                                 if (!nvmeq)
2020                                         continue;
2021                                 spin_lock_irq(&nvmeq->q_lock);
2022                                 nvme_process_cq(nvmeq);
2023
2024                                 while ((i == 0) && (dev->event_limit > 0)) {
2025                                         if (nvme_submit_async_admin_req(dev))
2026                                                 break;
2027                                         dev->event_limit--;
2028                                 }
2029                                 spin_unlock_irq(&nvmeq->q_lock);
2030                         }
2031                 }
2032                 spin_unlock(&dev_list_lock);
2033                 schedule_timeout(round_jiffies_relative(HZ));
2034         }
2035         return 0;
2036 }
2037
2038 static void nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid)
2039 {
2040         struct nvme_ns *ns;
2041         struct gendisk *disk;
2042         int node = dev_to_node(dev->dev);
2043
2044         ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
2045         if (!ns)
2046                 return;
2047
2048         ns->queue = blk_mq_init_queue(&dev->tagset);
2049         if (IS_ERR(ns->queue))
2050                 goto out_free_ns;
2051         queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
2052         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
2053         queue_flag_set_unlocked(QUEUE_FLAG_SG_GAPS, ns->queue);
2054         ns->dev = dev;
2055         ns->queue->queuedata = ns;
2056
2057         disk = alloc_disk_node(0, node);
2058         if (!disk)
2059                 goto out_free_queue;
2060
2061         ns->ns_id = nsid;
2062         ns->disk = disk;
2063         ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
2064         list_add_tail(&ns->list, &dev->namespaces);
2065
2066         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
2067         if (dev->max_hw_sectors)
2068                 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
2069         if (dev->stripe_size)
2070                 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
2071         if (dev->vwc & NVME_CTRL_VWC_PRESENT)
2072                 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
2073
2074         disk->major = nvme_major;
2075         disk->first_minor = 0;
2076         disk->fops = &nvme_fops;
2077         disk->private_data = ns;
2078         disk->queue = ns->queue;
2079         disk->driverfs_dev = dev->device;
2080         disk->flags = GENHD_FL_EXT_DEVT;
2081         sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
2082
2083         /*
2084          * Initialize capacity to 0 until we establish the namespace format and
2085          * setup integrity extentions if necessary. The revalidate_disk after
2086          * add_disk allows the driver to register with integrity if the format
2087          * requires it.
2088          */
2089         set_capacity(disk, 0);
2090         if (nvme_revalidate_disk(ns->disk))
2091                 goto out_free_disk;
2092
2093         add_disk(ns->disk);
2094         if (ns->ms)
2095                 revalidate_disk(ns->disk);
2096         return;
2097  out_free_disk:
2098         kfree(disk);
2099         list_del(&ns->list);
2100  out_free_queue:
2101         blk_cleanup_queue(ns->queue);
2102  out_free_ns:
2103         kfree(ns);
2104 }
2105
2106 static void nvme_create_io_queues(struct nvme_dev *dev)
2107 {
2108         unsigned i;
2109
2110         for (i = dev->queue_count; i <= dev->max_qid; i++)
2111                 if (!nvme_alloc_queue(dev, i, dev->q_depth))
2112                         break;
2113
2114         for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
2115                 if (nvme_create_queue(dev->queues[i], i))
2116                         break;
2117 }
2118
2119 static int set_queue_count(struct nvme_dev *dev, int count)
2120 {
2121         int status;
2122         u32 result;
2123         u32 q_count = (count - 1) | ((count - 1) << 16);
2124
2125         status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2126                                                                 &result);
2127         if (status < 0)
2128                 return status;
2129         if (status > 0) {
2130                 dev_err(dev->dev, "Could not set queue count (%d)\n", status);
2131                 return 0;
2132         }
2133         return min(result & 0xffff, result >> 16) + 1;
2134 }
2135
2136 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2137 {
2138         return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2139 }
2140
2141 static int nvme_setup_io_queues(struct nvme_dev *dev)
2142 {
2143         struct nvme_queue *adminq = dev->queues[0];
2144         struct pci_dev *pdev = to_pci_dev(dev->dev);
2145         int result, i, vecs, nr_io_queues, size;
2146
2147         nr_io_queues = num_possible_cpus();
2148         result = set_queue_count(dev, nr_io_queues);
2149         if (result <= 0)
2150                 return result;
2151         if (result < nr_io_queues)
2152                 nr_io_queues = result;
2153
2154         size = db_bar_size(dev, nr_io_queues);
2155         if (size > 8192) {
2156                 iounmap(dev->bar);
2157                 do {
2158                         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2159                         if (dev->bar)
2160                                 break;
2161                         if (!--nr_io_queues)
2162                                 return -ENOMEM;
2163                         size = db_bar_size(dev, nr_io_queues);
2164                 } while (1);
2165                 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2166                 adminq->q_db = dev->dbs;
2167         }
2168
2169         /* Deregister the admin queue's interrupt */
2170         free_irq(dev->entry[0].vector, adminq);
2171
2172         /*
2173          * If we enable msix early due to not intx, disable it again before
2174          * setting up the full range we need.
2175          */
2176         if (!pdev->irq)
2177                 pci_disable_msix(pdev);
2178
2179         for (i = 0; i < nr_io_queues; i++)
2180                 dev->entry[i].entry = i;
2181         vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2182         if (vecs < 0) {
2183                 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2184                 if (vecs < 0) {
2185                         vecs = 1;
2186                 } else {
2187                         for (i = 0; i < vecs; i++)
2188                                 dev->entry[i].vector = i + pdev->irq;
2189                 }
2190         }
2191
2192         /*
2193          * Should investigate if there's a performance win from allocating
2194          * more queues than interrupt vectors; it might allow the submission
2195          * path to scale better, even if the receive path is limited by the
2196          * number of interrupts.
2197          */
2198         nr_io_queues = vecs;
2199         dev->max_qid = nr_io_queues;
2200
2201         result = queue_request_irq(dev, adminq, adminq->irqname);
2202         if (result)
2203                 goto free_queues;
2204
2205         /* Free previously allocated queues that are no longer usable */
2206         nvme_free_queues(dev, nr_io_queues + 1);
2207         nvme_create_io_queues(dev);
2208
2209         return 0;
2210
2211  free_queues:
2212         nvme_free_queues(dev, 1);
2213         return result;
2214 }
2215
2216 static void nvme_free_namespace(struct nvme_ns *ns)
2217 {
2218         list_del(&ns->list);
2219
2220         spin_lock(&dev_list_lock);
2221         ns->disk->private_data = NULL;
2222         spin_unlock(&dev_list_lock);
2223
2224         put_disk(ns->disk);
2225         kfree(ns);
2226 }
2227
2228 static int ns_cmp(void *priv, struct list_head *a, struct list_head *b)
2229 {
2230         struct nvme_ns *nsa = container_of(a, struct nvme_ns, list);
2231         struct nvme_ns *nsb = container_of(b, struct nvme_ns, list);
2232
2233         return nsa->ns_id - nsb->ns_id;
2234 }
2235
2236 static struct nvme_ns *nvme_find_ns(struct nvme_dev *dev, unsigned nsid)
2237 {
2238         struct nvme_ns *ns;
2239
2240         list_for_each_entry(ns, &dev->namespaces, list) {
2241                 if (ns->ns_id == nsid)
2242                         return ns;
2243                 if (ns->ns_id > nsid)
2244                         break;
2245         }
2246         return NULL;
2247 }
2248
2249 static inline bool nvme_io_incapable(struct nvme_dev *dev)
2250 {
2251         return (!dev->bar || readl(&dev->bar->csts) & NVME_CSTS_CFS ||
2252                                                         dev->online_queues < 2);
2253 }
2254
2255 static void nvme_ns_remove(struct nvme_ns *ns)
2256 {
2257         bool kill = nvme_io_incapable(ns->dev) && !blk_queue_dying(ns->queue);
2258
2259         if (kill)
2260                 blk_set_queue_dying(ns->queue);
2261         if (ns->disk->flags & GENHD_FL_UP) {
2262                 if (blk_get_integrity(ns->disk))
2263                         blk_integrity_unregister(ns->disk);
2264                 del_gendisk(ns->disk);
2265         }
2266         if (kill || !blk_queue_dying(ns->queue)) {
2267                 blk_mq_abort_requeue_list(ns->queue);
2268                 blk_cleanup_queue(ns->queue);
2269         }
2270 }
2271
2272 static void nvme_scan_namespaces(struct nvme_dev *dev, unsigned nn)
2273 {
2274         struct nvme_ns *ns, *next;
2275         unsigned i;
2276
2277         for (i = 1; i <= nn; i++) {
2278                 ns = nvme_find_ns(dev, i);
2279                 if (ns) {
2280                         if (revalidate_disk(ns->disk)) {
2281                                 nvme_ns_remove(ns);
2282                                 nvme_free_namespace(ns);
2283                         }
2284                 } else
2285                         nvme_alloc_ns(dev, i);
2286         }
2287         list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2288                 if (ns->ns_id > nn) {
2289                         nvme_ns_remove(ns);
2290                         nvme_free_namespace(ns);
2291                 }
2292         }
2293         list_sort(NULL, &dev->namespaces, ns_cmp);
2294 }
2295
2296 static void nvme_dev_scan(struct work_struct *work)
2297 {
2298         struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work);
2299         struct nvme_id_ctrl *ctrl;
2300
2301         if (!dev->tagset.tags)
2302                 return;
2303         if (nvme_identify_ctrl(dev, &ctrl))
2304                 return;
2305         nvme_scan_namespaces(dev, le32_to_cpup(&ctrl->nn));
2306         kfree(ctrl);
2307 }
2308
2309 /*
2310  * Return: error value if an error occurred setting up the queues or calling
2311  * Identify Device.  0 if these succeeded, even if adding some of the
2312  * namespaces failed.  At the moment, these failures are silent.  TBD which
2313  * failures should be reported.
2314  */
2315 static int nvme_dev_add(struct nvme_dev *dev)
2316 {
2317         struct pci_dev *pdev = to_pci_dev(dev->dev);
2318         int res;
2319         unsigned nn;
2320         struct nvme_id_ctrl *ctrl;
2321         int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2322
2323         res = nvme_identify_ctrl(dev, &ctrl);
2324         if (res) {
2325                 dev_err(dev->dev, "Identify Controller failed (%d)\n", res);
2326                 return -EIO;
2327         }
2328
2329         nn = le32_to_cpup(&ctrl->nn);
2330         dev->oncs = le16_to_cpup(&ctrl->oncs);
2331         dev->abort_limit = ctrl->acl + 1;
2332         dev->vwc = ctrl->vwc;
2333         memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2334         memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2335         memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2336         if (ctrl->mdts)
2337                 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2338         if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2339                         (pdev->device == 0x0953) && ctrl->vs[3]) {
2340                 unsigned int max_hw_sectors;
2341
2342                 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2343                 max_hw_sectors = dev->stripe_size >> (shift - 9);
2344                 if (dev->max_hw_sectors) {
2345                         dev->max_hw_sectors = min(max_hw_sectors,
2346                                                         dev->max_hw_sectors);
2347                 } else
2348                         dev->max_hw_sectors = max_hw_sectors;
2349         }
2350         kfree(ctrl);
2351
2352         dev->tagset.ops = &nvme_mq_ops;
2353         dev->tagset.nr_hw_queues = dev->online_queues - 1;
2354         dev->tagset.timeout = NVME_IO_TIMEOUT;
2355         dev->tagset.numa_node = dev_to_node(dev->dev);
2356         dev->tagset.queue_depth =
2357                                 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2358         dev->tagset.cmd_size = nvme_cmd_size(dev);
2359         dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2360         dev->tagset.driver_data = dev;
2361
2362         if (blk_mq_alloc_tag_set(&dev->tagset))
2363                 return 0;
2364
2365         schedule_work(&dev->scan_work);
2366         return 0;
2367 }
2368
2369 static int nvme_dev_map(struct nvme_dev *dev)
2370 {
2371         u64 cap;
2372         int bars, result = -ENOMEM;
2373         struct pci_dev *pdev = to_pci_dev(dev->dev);
2374
2375         if (pci_enable_device_mem(pdev))
2376                 return result;
2377
2378         dev->entry[0].vector = pdev->irq;
2379         pci_set_master(pdev);
2380         bars = pci_select_bars(pdev, IORESOURCE_MEM);
2381         if (!bars)
2382                 goto disable_pci;
2383
2384         if (pci_request_selected_regions(pdev, bars, "nvme"))
2385                 goto disable_pci;
2386
2387         if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
2388             dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
2389                 goto disable;
2390
2391         dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2392         if (!dev->bar)
2393                 goto disable;
2394
2395         if (readl(&dev->bar->csts) == -1) {
2396                 result = -ENODEV;
2397                 goto unmap;
2398         }
2399
2400         /*
2401          * Some devices don't advertse INTx interrupts, pre-enable a single
2402          * MSIX vec for setup. We'll adjust this later.
2403          */
2404         if (!pdev->irq) {
2405                 result = pci_enable_msix(pdev, dev->entry, 1);
2406                 if (result < 0)
2407                         goto unmap;
2408         }
2409
2410         cap = readq(&dev->bar->cap);
2411         dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2412         dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2413         dev->dbs = ((void __iomem *)dev->bar) + 4096;
2414
2415         return 0;
2416
2417  unmap:
2418         iounmap(dev->bar);
2419         dev->bar = NULL;
2420  disable:
2421         pci_release_regions(pdev);
2422  disable_pci:
2423         pci_disable_device(pdev);
2424         return result;
2425 }
2426
2427 static void nvme_dev_unmap(struct nvme_dev *dev)
2428 {
2429         struct pci_dev *pdev = to_pci_dev(dev->dev);
2430
2431         if (pdev->msi_enabled)
2432                 pci_disable_msi(pdev);
2433         else if (pdev->msix_enabled)
2434                 pci_disable_msix(pdev);
2435
2436         if (dev->bar) {
2437                 iounmap(dev->bar);
2438                 dev->bar = NULL;
2439                 pci_release_regions(pdev);
2440         }
2441
2442         if (pci_is_enabled(pdev))
2443                 pci_disable_device(pdev);
2444 }
2445
2446 struct nvme_delq_ctx {
2447         struct task_struct *waiter;
2448         struct kthread_worker *worker;
2449         atomic_t refcount;
2450 };
2451
2452 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2453 {
2454         dq->waiter = current;
2455         mb();
2456
2457         for (;;) {
2458                 set_current_state(TASK_KILLABLE);
2459                 if (!atomic_read(&dq->refcount))
2460                         break;
2461                 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2462                                         fatal_signal_pending(current)) {
2463                         /*
2464                          * Disable the controller first since we can't trust it
2465                          * at this point, but leave the admin queue enabled
2466                          * until all queue deletion requests are flushed.
2467                          * FIXME: This may take a while if there are more h/w
2468                          * queues than admin tags.
2469                          */
2470                         set_current_state(TASK_RUNNING);
2471                         nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2472                         nvme_clear_queue(dev->queues[0]);
2473                         flush_kthread_worker(dq->worker);
2474                         nvme_disable_queue(dev, 0);
2475                         return;
2476                 }
2477         }
2478         set_current_state(TASK_RUNNING);
2479 }
2480
2481 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2482 {
2483         atomic_dec(&dq->refcount);
2484         if (dq->waiter)
2485                 wake_up_process(dq->waiter);
2486 }
2487
2488 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2489 {
2490         atomic_inc(&dq->refcount);
2491         return dq;
2492 }
2493
2494 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2495 {
2496         struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2497         nvme_put_dq(dq);
2498 }
2499
2500 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2501                                                 kthread_work_func_t fn)
2502 {
2503         struct nvme_command c;
2504
2505         memset(&c, 0, sizeof(c));
2506         c.delete_queue.opcode = opcode;
2507         c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2508
2509         init_kthread_work(&nvmeq->cmdinfo.work, fn);
2510         return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2511                                                                 ADMIN_TIMEOUT);
2512 }
2513
2514 static void nvme_del_cq_work_handler(struct kthread_work *work)
2515 {
2516         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2517                                                         cmdinfo.work);
2518         nvme_del_queue_end(nvmeq);
2519 }
2520
2521 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2522 {
2523         return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2524                                                 nvme_del_cq_work_handler);
2525 }
2526
2527 static void nvme_del_sq_work_handler(struct kthread_work *work)
2528 {
2529         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2530                                                         cmdinfo.work);
2531         int status = nvmeq->cmdinfo.status;
2532
2533         if (!status)
2534                 status = nvme_delete_cq(nvmeq);
2535         if (status)
2536                 nvme_del_queue_end(nvmeq);
2537 }
2538
2539 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2540 {
2541         return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2542                                                 nvme_del_sq_work_handler);
2543 }
2544
2545 static void nvme_del_queue_start(struct kthread_work *work)
2546 {
2547         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2548                                                         cmdinfo.work);
2549         if (nvme_delete_sq(nvmeq))
2550                 nvme_del_queue_end(nvmeq);
2551 }
2552
2553 static void nvme_disable_io_queues(struct nvme_dev *dev)
2554 {
2555         int i;
2556         DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2557         struct nvme_delq_ctx dq;
2558         struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2559                                         &worker, "nvme%d", dev->instance);
2560
2561         if (IS_ERR(kworker_task)) {
2562                 dev_err(dev->dev,
2563                         "Failed to create queue del task\n");
2564                 for (i = dev->queue_count - 1; i > 0; i--)
2565                         nvme_disable_queue(dev, i);
2566                 return;
2567         }
2568
2569         dq.waiter = NULL;
2570         atomic_set(&dq.refcount, 0);
2571         dq.worker = &worker;
2572         for (i = dev->queue_count - 1; i > 0; i--) {
2573                 struct nvme_queue *nvmeq = dev->queues[i];
2574
2575                 if (nvme_suspend_queue(nvmeq))
2576                         continue;
2577                 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2578                 nvmeq->cmdinfo.worker = dq.worker;
2579                 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2580                 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2581         }
2582         nvme_wait_dq(&dq, dev);
2583         kthread_stop(kworker_task);
2584 }
2585
2586 /*
2587 * Remove the node from the device list and check
2588 * for whether or not we need to stop the nvme_thread.
2589 */
2590 static void nvme_dev_list_remove(struct nvme_dev *dev)
2591 {
2592         struct task_struct *tmp = NULL;
2593
2594         spin_lock(&dev_list_lock);
2595         list_del_init(&dev->node);
2596         if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2597                 tmp = nvme_thread;
2598                 nvme_thread = NULL;
2599         }
2600         spin_unlock(&dev_list_lock);
2601
2602         if (tmp)
2603                 kthread_stop(tmp);
2604 }
2605
2606 static void nvme_freeze_queues(struct nvme_dev *dev)
2607 {
2608         struct nvme_ns *ns;
2609
2610         list_for_each_entry(ns, &dev->namespaces, list) {
2611                 blk_mq_freeze_queue_start(ns->queue);
2612
2613                 spin_lock_irq(ns->queue->queue_lock);
2614                 queue_flag_set(QUEUE_FLAG_STOPPED, ns->queue);
2615                 spin_unlock_irq(ns->queue->queue_lock);
2616
2617                 blk_mq_cancel_requeue_work(ns->queue);
2618                 blk_mq_stop_hw_queues(ns->queue);
2619         }
2620 }
2621
2622 static void nvme_unfreeze_queues(struct nvme_dev *dev)
2623 {
2624         struct nvme_ns *ns;
2625
2626         list_for_each_entry(ns, &dev->namespaces, list) {
2627                 queue_flag_clear_unlocked(QUEUE_FLAG_STOPPED, ns->queue);
2628                 blk_mq_unfreeze_queue(ns->queue);
2629                 blk_mq_start_stopped_hw_queues(ns->queue, true);
2630                 blk_mq_kick_requeue_list(ns->queue);
2631         }
2632 }
2633
2634 static void nvme_dev_shutdown(struct nvme_dev *dev)
2635 {
2636         int i;
2637         u32 csts = -1;
2638
2639         nvme_dev_list_remove(dev);
2640
2641         if (dev->bar) {
2642                 nvme_freeze_queues(dev);
2643                 csts = readl(&dev->bar->csts);
2644         }
2645         if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2646                 for (i = dev->queue_count - 1; i >= 0; i--) {
2647                         struct nvme_queue *nvmeq = dev->queues[i];
2648                         nvme_suspend_queue(nvmeq);
2649                 }
2650         } else {
2651                 nvme_disable_io_queues(dev);
2652                 nvme_shutdown_ctrl(dev);
2653                 nvme_disable_queue(dev, 0);
2654         }
2655         nvme_dev_unmap(dev);
2656
2657         for (i = dev->queue_count - 1; i >= 0; i--)
2658                 nvme_clear_queue(dev->queues[i]);
2659 }
2660
2661 static void nvme_dev_remove(struct nvme_dev *dev)
2662 {
2663         struct nvme_ns *ns;
2664
2665         list_for_each_entry(ns, &dev->namespaces, list)
2666                 nvme_ns_remove(ns);
2667 }
2668
2669 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2670 {
2671         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2672                                                 PAGE_SIZE, PAGE_SIZE, 0);
2673         if (!dev->prp_page_pool)
2674                 return -ENOMEM;
2675
2676         /* Optimisation for I/Os between 4k and 128k */
2677         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2678                                                 256, 256, 0);
2679         if (!dev->prp_small_pool) {
2680                 dma_pool_destroy(dev->prp_page_pool);
2681                 return -ENOMEM;
2682         }
2683         return 0;
2684 }
2685
2686 static void nvme_release_prp_pools(struct nvme_dev *dev)
2687 {
2688         dma_pool_destroy(dev->prp_page_pool);
2689         dma_pool_destroy(dev->prp_small_pool);
2690 }
2691
2692 static DEFINE_IDA(nvme_instance_ida);
2693
2694 static int nvme_set_instance(struct nvme_dev *dev)
2695 {
2696         int instance, error;
2697
2698         do {
2699                 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2700                         return -ENODEV;
2701
2702                 spin_lock(&dev_list_lock);
2703                 error = ida_get_new(&nvme_instance_ida, &instance);
2704                 spin_unlock(&dev_list_lock);
2705         } while (error == -EAGAIN);
2706
2707         if (error)
2708                 return -ENODEV;
2709
2710         dev->instance = instance;
2711         return 0;
2712 }
2713
2714 static void nvme_release_instance(struct nvme_dev *dev)
2715 {
2716         spin_lock(&dev_list_lock);
2717         ida_remove(&nvme_instance_ida, dev->instance);
2718         spin_unlock(&dev_list_lock);
2719 }
2720
2721 static void nvme_free_namespaces(struct nvme_dev *dev)
2722 {
2723         struct nvme_ns *ns, *next;
2724
2725         list_for_each_entry_safe(ns, next, &dev->namespaces, list)
2726                 nvme_free_namespace(ns);
2727 }
2728
2729 static void nvme_free_dev(struct kref *kref)
2730 {
2731         struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2732
2733         put_device(dev->dev);
2734         put_device(dev->device);
2735         nvme_free_namespaces(dev);
2736         nvme_release_instance(dev);
2737         blk_mq_free_tag_set(&dev->tagset);
2738         blk_put_queue(dev->admin_q);
2739         kfree(dev->queues);
2740         kfree(dev->entry);
2741         kfree(dev);
2742 }
2743
2744 static int nvme_dev_open(struct inode *inode, struct file *f)
2745 {
2746         struct nvme_dev *dev;
2747         int instance = iminor(inode);
2748         int ret = -ENODEV;
2749
2750         spin_lock(&dev_list_lock);
2751         list_for_each_entry(dev, &dev_list, node) {
2752                 if (dev->instance == instance) {
2753                         if (!dev->admin_q) {
2754                                 ret = -EWOULDBLOCK;
2755                                 break;
2756                         }
2757                         if (!kref_get_unless_zero(&dev->kref))
2758                                 break;
2759                         f->private_data = dev;
2760                         ret = 0;
2761                         break;
2762                 }
2763         }
2764         spin_unlock(&dev_list_lock);
2765
2766         return ret;
2767 }
2768
2769 static int nvme_dev_release(struct inode *inode, struct file *f)
2770 {
2771         struct nvme_dev *dev = f->private_data;
2772         kref_put(&dev->kref, nvme_free_dev);
2773         return 0;
2774 }
2775
2776 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2777 {
2778         struct nvme_dev *dev = f->private_data;
2779         struct nvme_ns *ns;
2780
2781         switch (cmd) {
2782         case NVME_IOCTL_ADMIN_CMD:
2783                 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2784         case NVME_IOCTL_IO_CMD:
2785                 if (list_empty(&dev->namespaces))
2786                         return -ENOTTY;
2787                 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2788                 return nvme_user_cmd(dev, ns, (void __user *)arg);
2789         case NVME_IOCTL_RESET:
2790                 dev_warn(dev->dev, "resetting controller\n");
2791                 return nvme_reset(dev);
2792         default:
2793                 return -ENOTTY;
2794         }
2795 }
2796
2797 static const struct file_operations nvme_dev_fops = {
2798         .owner          = THIS_MODULE,
2799         .open           = nvme_dev_open,
2800         .release        = nvme_dev_release,
2801         .unlocked_ioctl = nvme_dev_ioctl,
2802         .compat_ioctl   = nvme_dev_ioctl,
2803 };
2804
2805 static void nvme_set_irq_hints(struct nvme_dev *dev)
2806 {
2807         struct nvme_queue *nvmeq;
2808         int i;
2809
2810         for (i = 0; i < dev->online_queues; i++) {
2811                 nvmeq = dev->queues[i];
2812
2813                 if (!nvmeq->tags || !(*nvmeq->tags))
2814                         continue;
2815
2816                 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2817                                         blk_mq_tags_cpumask(*nvmeq->tags));
2818         }
2819 }
2820
2821 static int nvme_dev_start(struct nvme_dev *dev)
2822 {
2823         int result;
2824         bool start_thread = false;
2825
2826         result = nvme_dev_map(dev);
2827         if (result)
2828                 return result;
2829
2830         result = nvme_configure_admin_queue(dev);
2831         if (result)
2832                 goto unmap;
2833
2834         spin_lock(&dev_list_lock);
2835         if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2836                 start_thread = true;
2837                 nvme_thread = NULL;
2838         }
2839         list_add(&dev->node, &dev_list);
2840         spin_unlock(&dev_list_lock);
2841
2842         if (start_thread) {
2843                 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2844                 wake_up_all(&nvme_kthread_wait);
2845         } else
2846                 wait_event_killable(nvme_kthread_wait, nvme_thread);
2847
2848         if (IS_ERR_OR_NULL(nvme_thread)) {
2849                 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2850                 goto disable;
2851         }
2852
2853         nvme_init_queue(dev->queues[0], 0);
2854         result = nvme_alloc_admin_tags(dev);
2855         if (result)
2856                 goto disable;
2857
2858         result = nvme_setup_io_queues(dev);
2859         if (result)
2860                 goto free_tags;
2861
2862         nvme_set_irq_hints(dev);
2863
2864         dev->event_limit = 1;
2865         return result;
2866
2867  free_tags:
2868         nvme_dev_remove_admin(dev);
2869  disable:
2870         nvme_disable_queue(dev, 0);
2871         nvme_dev_list_remove(dev);
2872  unmap:
2873         nvme_dev_unmap(dev);
2874         return result;
2875 }
2876
2877 static int nvme_remove_dead_ctrl(void *arg)
2878 {
2879         struct nvme_dev *dev = (struct nvme_dev *)arg;
2880         struct pci_dev *pdev = to_pci_dev(dev->dev);
2881
2882         if (pci_get_drvdata(pdev))
2883                 pci_stop_and_remove_bus_device_locked(pdev);
2884         kref_put(&dev->kref, nvme_free_dev);
2885         return 0;
2886 }
2887
2888 static void nvme_remove_disks(struct work_struct *ws)
2889 {
2890         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2891
2892         nvme_free_queues(dev, 1);
2893         nvme_dev_remove(dev);
2894 }
2895
2896 static int nvme_dev_resume(struct nvme_dev *dev)
2897 {
2898         int ret;
2899
2900         ret = nvme_dev_start(dev);
2901         if (ret)
2902                 return ret;
2903         if (dev->online_queues < 2) {
2904                 spin_lock(&dev_list_lock);
2905                 dev->reset_workfn = nvme_remove_disks;
2906                 queue_work(nvme_workq, &dev->reset_work);
2907                 spin_unlock(&dev_list_lock);
2908         } else {
2909                 nvme_unfreeze_queues(dev);
2910                 schedule_work(&dev->scan_work);
2911                 nvme_set_irq_hints(dev);
2912         }
2913         return 0;
2914 }
2915
2916 static void nvme_dev_reset(struct nvme_dev *dev)
2917 {
2918         nvme_dev_shutdown(dev);
2919         if (nvme_dev_resume(dev)) {
2920                 dev_warn(dev->dev, "Device failed to resume\n");
2921                 kref_get(&dev->kref);
2922                 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2923                                                         dev->instance))) {
2924                         dev_err(dev->dev,
2925                                 "Failed to start controller remove task\n");
2926                         kref_put(&dev->kref, nvme_free_dev);
2927                 }
2928         }
2929 }
2930
2931 static void nvme_reset_failed_dev(struct work_struct *ws)
2932 {
2933         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2934         nvme_dev_reset(dev);
2935 }
2936
2937 static void nvme_reset_workfn(struct work_struct *work)
2938 {
2939         struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2940         dev->reset_workfn(work);
2941 }
2942
2943 static int nvme_reset(struct nvme_dev *dev)
2944 {
2945         int ret = -EBUSY;
2946
2947         if (!dev->admin_q || blk_queue_dying(dev->admin_q))
2948                 return -ENODEV;
2949
2950         spin_lock(&dev_list_lock);
2951         if (!work_pending(&dev->reset_work)) {
2952                 dev->reset_workfn = nvme_reset_failed_dev;
2953                 queue_work(nvme_workq, &dev->reset_work);
2954                 ret = 0;
2955         }
2956         spin_unlock(&dev_list_lock);
2957
2958         if (!ret) {
2959                 flush_work(&dev->reset_work);
2960                 return 0;
2961         }
2962
2963         return ret;
2964 }
2965
2966 static ssize_t nvme_sysfs_reset(struct device *dev,
2967                                 struct device_attribute *attr, const char *buf,
2968                                 size_t count)
2969 {
2970         struct nvme_dev *ndev = dev_get_drvdata(dev);
2971         int ret;
2972
2973         ret = nvme_reset(ndev);
2974         if (ret < 0)
2975                 return ret;
2976
2977         return count;
2978 }
2979 static DEVICE_ATTR(reset_controller, S_IWUSR, NULL, nvme_sysfs_reset);
2980
2981 static void nvme_async_probe(struct work_struct *work);
2982 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2983 {
2984         int node, result = -ENOMEM;
2985         struct nvme_dev *dev;
2986
2987         node = dev_to_node(&pdev->dev);
2988         if (node == NUMA_NO_NODE)
2989                 set_dev_node(&pdev->dev, 0);
2990
2991         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2992         if (!dev)
2993                 return -ENOMEM;
2994         dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
2995                                                         GFP_KERNEL, node);
2996         if (!dev->entry)
2997                 goto free;
2998         dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2999                                                         GFP_KERNEL, node);
3000         if (!dev->queues)
3001                 goto free;
3002
3003         INIT_LIST_HEAD(&dev->namespaces);
3004         dev->reset_workfn = nvme_reset_failed_dev;
3005         INIT_WORK(&dev->reset_work, nvme_reset_workfn);
3006         dev->dev = get_device(&pdev->dev);
3007         pci_set_drvdata(pdev, dev);
3008         result = nvme_set_instance(dev);
3009         if (result)
3010                 goto put_pci;
3011
3012         result = nvme_setup_prp_pools(dev);
3013         if (result)
3014                 goto release;
3015
3016         kref_init(&dev->kref);
3017         dev->device = device_create(nvme_class, &pdev->dev,
3018                                 MKDEV(nvme_char_major, dev->instance),
3019                                 dev, "nvme%d", dev->instance);
3020         if (IS_ERR(dev->device)) {
3021                 result = PTR_ERR(dev->device);
3022                 goto release_pools;
3023         }
3024         get_device(dev->device);
3025         dev_set_drvdata(dev->device, dev);
3026
3027         result = device_create_file(dev->device, &dev_attr_reset_controller);
3028         if (result)
3029                 goto put_dev;
3030
3031         INIT_LIST_HEAD(&dev->node);
3032         INIT_WORK(&dev->scan_work, nvme_dev_scan);
3033         INIT_WORK(&dev->probe_work, nvme_async_probe);
3034         schedule_work(&dev->probe_work);
3035         return 0;
3036
3037  put_dev:
3038         device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3039         put_device(dev->device);
3040  release_pools:
3041         nvme_release_prp_pools(dev);
3042  release:
3043         nvme_release_instance(dev);
3044  put_pci:
3045         put_device(dev->dev);
3046  free:
3047         kfree(dev->queues);
3048         kfree(dev->entry);
3049         kfree(dev);
3050         return result;
3051 }
3052
3053 static void nvme_async_probe(struct work_struct *work)
3054 {
3055         struct nvme_dev *dev = container_of(work, struct nvme_dev, probe_work);
3056         int result;
3057
3058         result = nvme_dev_start(dev);
3059         if (result)
3060                 goto reset;
3061
3062         if (dev->online_queues > 1)
3063                 result = nvme_dev_add(dev);
3064         if (result)
3065                 goto reset;
3066
3067         nvme_set_irq_hints(dev);
3068         return;
3069  reset:
3070         spin_lock(&dev_list_lock);
3071         if (!work_busy(&dev->reset_work)) {
3072                 dev->reset_workfn = nvme_reset_failed_dev;
3073                 queue_work(nvme_workq, &dev->reset_work);
3074         }
3075         spin_unlock(&dev_list_lock);
3076 }
3077
3078 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
3079 {
3080         struct nvme_dev *dev = pci_get_drvdata(pdev);
3081
3082         if (prepare)
3083                 nvme_dev_shutdown(dev);
3084         else
3085                 nvme_dev_resume(dev);
3086 }
3087
3088 static void nvme_shutdown(struct pci_dev *pdev)
3089 {
3090         struct nvme_dev *dev = pci_get_drvdata(pdev);
3091         nvme_dev_shutdown(dev);
3092 }
3093
3094 static void nvme_remove(struct pci_dev *pdev)
3095 {
3096         struct nvme_dev *dev = pci_get_drvdata(pdev);
3097
3098         spin_lock(&dev_list_lock);
3099         list_del_init(&dev->node);
3100         spin_unlock(&dev_list_lock);
3101
3102         pci_set_drvdata(pdev, NULL);
3103         flush_work(&dev->probe_work);
3104         flush_work(&dev->reset_work);
3105         flush_work(&dev->scan_work);
3106         device_remove_file(dev->device, &dev_attr_reset_controller);
3107         nvme_dev_shutdown(dev);
3108         nvme_dev_remove(dev);
3109         nvme_dev_remove_admin(dev);
3110         device_destroy(nvme_class, MKDEV(nvme_char_major, dev->instance));
3111         nvme_free_queues(dev, 0);
3112         nvme_release_prp_pools(dev);
3113         kref_put(&dev->kref, nvme_free_dev);
3114 }
3115
3116 /* These functions are yet to be implemented */
3117 #define nvme_error_detected NULL
3118 #define nvme_dump_registers NULL
3119 #define nvme_link_reset NULL
3120 #define nvme_slot_reset NULL
3121 #define nvme_error_resume NULL
3122
3123 #ifdef CONFIG_PM_SLEEP
3124 static int nvme_suspend(struct device *dev)
3125 {
3126         struct pci_dev *pdev = to_pci_dev(dev);
3127         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3128
3129         nvme_dev_shutdown(ndev);
3130         return 0;
3131 }
3132
3133 static int nvme_resume(struct device *dev)
3134 {
3135         struct pci_dev *pdev = to_pci_dev(dev);
3136         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3137
3138         if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
3139                 ndev->reset_workfn = nvme_reset_failed_dev;
3140                 queue_work(nvme_workq, &ndev->reset_work);
3141         }
3142         return 0;
3143 }
3144 #endif
3145
3146 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
3147
3148 static const struct pci_error_handlers nvme_err_handler = {
3149         .error_detected = nvme_error_detected,
3150         .mmio_enabled   = nvme_dump_registers,
3151         .link_reset     = nvme_link_reset,
3152         .slot_reset     = nvme_slot_reset,
3153         .resume         = nvme_error_resume,
3154         .reset_notify   = nvme_reset_notify,
3155 };
3156
3157 /* Move to pci_ids.h later */
3158 #define PCI_CLASS_STORAGE_EXPRESS       0x010802
3159
3160 static const struct pci_device_id nvme_id_table[] = {
3161         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3162         { 0, }
3163 };
3164 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3165
3166 static struct pci_driver nvme_driver = {
3167         .name           = "nvme",
3168         .id_table       = nvme_id_table,
3169         .probe          = nvme_probe,
3170         .remove         = nvme_remove,
3171         .shutdown       = nvme_shutdown,
3172         .driver         = {
3173                 .pm     = &nvme_dev_pm_ops,
3174         },
3175         .err_handler    = &nvme_err_handler,
3176 };
3177
3178 static int __init nvme_init(void)
3179 {
3180         int result;
3181
3182         init_waitqueue_head(&nvme_kthread_wait);
3183
3184         nvme_workq = create_singlethread_workqueue("nvme");
3185         if (!nvme_workq)
3186                 return -ENOMEM;
3187
3188         result = register_blkdev(nvme_major, "nvme");
3189         if (result < 0)
3190                 goto kill_workq;
3191         else if (result > 0)
3192                 nvme_major = result;
3193
3194         result = __register_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme",
3195                                                         &nvme_dev_fops);
3196         if (result < 0)
3197                 goto unregister_blkdev;
3198         else if (result > 0)
3199                 nvme_char_major = result;
3200
3201         nvme_class = class_create(THIS_MODULE, "nvme");
3202         if (IS_ERR(nvme_class)) {
3203                 result = PTR_ERR(nvme_class);
3204                 goto unregister_chrdev;
3205         }
3206
3207         result = pci_register_driver(&nvme_driver);
3208         if (result)
3209                 goto destroy_class;
3210         return 0;
3211
3212  destroy_class:
3213         class_destroy(nvme_class);
3214  unregister_chrdev:
3215         __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3216  unregister_blkdev:
3217         unregister_blkdev(nvme_major, "nvme");
3218  kill_workq:
3219         destroy_workqueue(nvme_workq);
3220         return result;
3221 }
3222
3223 static void __exit nvme_exit(void)
3224 {
3225         pci_unregister_driver(&nvme_driver);
3226         unregister_blkdev(nvme_major, "nvme");
3227         destroy_workqueue(nvme_workq);
3228         class_destroy(nvme_class);
3229         __unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
3230         BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
3231         _nvme_check_size();
3232 }
3233
3234 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3235 MODULE_LICENSE("GPL");
3236 MODULE_VERSION("1.0");
3237 module_init(nvme_init);
3238 module_exit(nvme_exit);