drivers/block/nvme-core.c

   1 /*
   2  * NVM Express device driver
   3  * Copyright (c) 2011, 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  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  17  */
  18
  19 #include <linux/nvme.h>
  20 #include <linux/bio.h>
  21 #include <linux/bitops.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/delay.h>
  24 #include <linux/errno.h>
  25 #include <linux/fs.h>
  26 #include <linux/genhd.h>
  27 #include <linux/idr.h>
  28 #include <linux/init.h>
  29 #include <linux/interrupt.h>
  30 #include <linux/io.h>
  31 #include <linux/kdev_t.h>
  32 #include <linux/kthread.h>
  33 #include <linux/kernel.h>
  34 #include <linux/mm.h>
  35 #include <linux/module.h>
  36 #include <linux/moduleparam.h>
  37 #include <linux/pci.h>
  38 #include <linux/poison.h>
  39 #include <linux/ptrace.h>
  40 #include <linux/sched.h>
  41 #include <linux/slab.h>
  42 #include <linux/types.h>
  43 #include <scsi/sg.h>
  44 #include <asm-generic/io-64-nonatomic-lo-hi.h>
  45
  46 #define NVME_Q_DEPTH 1024
  47 #define SQ_SIZE(depth)          (depth * sizeof(struct nvme_command))
  48 #define CQ_SIZE(depth)          (depth * sizeof(struct nvme_completion))
  49 #define ADMIN_TIMEOUT   (60 * HZ)
  50
  51 static int nvme_major;
  52 module_param(nvme_major, int, 0);
  53
  54 static int use_threaded_interrupts;
  55 module_param(use_threaded_interrupts, int, 0);
  56
  57 static DEFINE_SPINLOCK(dev_list_lock);
  58 static LIST_HEAD(dev_list);
  59 static struct task_struct *nvme_thread;
  60 static struct workqueue_struct *nvme_workq;
  61
  62 static void nvme_reset_failed_dev(struct work_struct *ws);
  63
  64 struct async_cmd_info {
  65         struct kthread_work work;
  66         struct kthread_worker *worker;
  67         u32 result;
  68         int status;
  69         void *ctx;
  70 };
  71
  72 /*
  73  * An NVM Express queue.  Each device has at least two (one for admin
  74  * commands and one for I/O commands).
  75  */
  76 struct nvme_queue {
  77         struct device *q_dmadev;
  78         struct nvme_dev *dev;
  79         char irqname[24];       /* nvme4294967295-65535\0 */
  80         spinlock_t q_lock;
  81         struct nvme_command *sq_cmds;
  82         volatile struct nvme_completion *cqes;
  83         dma_addr_t sq_dma_addr;
  84         dma_addr_t cq_dma_addr;
  85         wait_queue_head_t sq_full;
  86         wait_queue_t sq_cong_wait;
  87         struct bio_list sq_cong;
  88         u32 __iomem *q_db;
  89         u16 q_depth;
  90         u16 cq_vector;
  91         u16 sq_head;
  92         u16 sq_tail;
  93         u16 cq_head;
  94         u16 qid;
  95         u8 cq_phase;
  96         u8 cqe_seen;
  97         u8 q_suspended;
  98         struct async_cmd_info cmdinfo;
  99         unsigned long cmdid_data[];
 100 };
 101
 102 /*
 103  * Check we didin't inadvertently grow the command struct
 104  */
 105 static inline void _nvme_check_size(void)
 106 {
 107         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
 108         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
 109         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
 110         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
 111         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
 112         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
 113         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
 114         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
 115         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
 116         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
 117         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
 118         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
 119 }
 120
 121 typedef void (*nvme_completion_fn)(struct nvme_dev *, void *,
 122                                                 struct nvme_completion *);
 123
 124 struct nvme_cmd_info {
 125         nvme_completion_fn fn;
 126         void *ctx;
 127         unsigned long timeout;
 128         int aborted;
 129 };
 130
 131 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
 132 {
 133         return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
 134 }
 135
 136 static unsigned nvme_queue_extra(int depth)
 137 {
 138         return DIV_ROUND_UP(depth, 8) + (depth * sizeof(struct nvme_cmd_info));
 139 }
 140
 141 /**
 142  * alloc_cmdid() - Allocate a Command ID
 143  * @nvmeq: The queue that will be used for this command
 144  * @ctx: A pointer that will be passed to the handler
 145  * @handler: The function to call on completion
 146  *
 147  * Allocate a Command ID for a queue.  The data passed in will
 148  * be passed to the completion handler.  This is implemented by using
 149  * the bottom two bits of the ctx pointer to store the handler ID.
 150  * Passing in a pointer that's not 4-byte aligned will cause a BUG.
 151  * We can change this if it becomes a problem.
 152  *
 153  * May be called with local interrupts disabled and the q_lock held,
 154  * or with interrupts enabled and no locks held.
 155  */
 156 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx,
 157                                 nvme_completion_fn handler, unsigned timeout)
 158 {
 159         int depth = nvmeq->q_depth - 1;
 160         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 161         int cmdid;
 162
 163         do {
 164                 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
 165                 if (cmdid >= depth)
 166                         return -EBUSY;
 167         } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
 168
 169         info[cmdid].fn = handler;
 170         info[cmdid].ctx = ctx;
 171         info[cmdid].timeout = jiffies + timeout;
 172         info[cmdid].aborted = 0;
 173         return cmdid;
 174 }
 175
 176 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
 177                                 nvme_completion_fn handler, unsigned timeout)
 178 {
 179         int cmdid;
 180         wait_event_killable(nvmeq->sq_full,
 181                 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
 182         return (cmdid < 0) ? -EINTR : cmdid;
 183 }
 184
 185 /* Special values must be less than 0x1000 */
 186 #define CMD_CTX_BASE            ((void *)POISON_POINTER_DELTA)
 187 #define CMD_CTX_CANCELLED       (0x30C + CMD_CTX_BASE)
 188 #define CMD_CTX_COMPLETED       (0x310 + CMD_CTX_BASE)
 189 #define CMD_CTX_INVALID         (0x314 + CMD_CTX_BASE)
 190 #define CMD_CTX_FLUSH           (0x318 + CMD_CTX_BASE)
 191 #define CMD_CTX_ABORT           (0x31C + CMD_CTX_BASE)
 192
 193 static void special_completion(struct nvme_dev *dev, void *ctx,
 194                                                 struct nvme_completion *cqe)
 195 {
 196         if (ctx == CMD_CTX_CANCELLED)
 197                 return;
 198         if (ctx == CMD_CTX_FLUSH)
 199                 return;
 200         if (ctx == CMD_CTX_ABORT) {
 201                 ++dev->abort_limit;
 202                 return;
 203         }
 204         if (ctx == CMD_CTX_COMPLETED) {
 205                 dev_warn(&dev->pci_dev->dev,
 206                                 "completed id %d twice on queue %d\n",
 207                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 208                 return;
 209         }
 210         if (ctx == CMD_CTX_INVALID) {
 211                 dev_warn(&dev->pci_dev->dev,
 212                                 "invalid id %d completed on queue %d\n",
 213                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 214                 return;
 215         }
 216
 217         dev_warn(&dev->pci_dev->dev, "Unknown special completion %p\n", ctx);
 218 }
 219
 220 static void async_completion(struct nvme_dev *dev, void *ctx,
 221                                                 struct nvme_completion *cqe)
 222 {
 223         struct async_cmd_info *cmdinfo = ctx;
 224         cmdinfo->result = le32_to_cpup(&cqe->result);
 225         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 226         queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
 227 }
 228
 229 /*
 230  * Called with local interrupts disabled and the q_lock held.  May not sleep.
 231  */
 232 static void *free_cmdid(struct nvme_queue *nvmeq, int cmdid,
 233                                                 nvme_completion_fn *fn)
 234 {
 235         void *ctx;
 236         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 237
 238         if (cmdid >= nvmeq->q_depth) {
 239                 *fn = special_completion;
 240                 return CMD_CTX_INVALID;
 241         }
 242         if (fn)
 243                 *fn = info[cmdid].fn;
 244         ctx = info[cmdid].ctx;
 245         info[cmdid].fn = special_completion;
 246         info[cmdid].ctx = CMD_CTX_COMPLETED;
 247         clear_bit(cmdid, nvmeq->cmdid_data);
 248         wake_up(&nvmeq->sq_full);
 249         return ctx;
 250 }
 251
 252 static void *cancel_cmdid(struct nvme_queue *nvmeq, int cmdid,
 253                                                 nvme_completion_fn *fn)
 254 {
 255         void *ctx;
 256         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 257         if (fn)
 258                 *fn = info[cmdid].fn;
 259         ctx = info[cmdid].ctx;
 260         info[cmdid].fn = special_completion;
 261         info[cmdid].ctx = CMD_CTX_CANCELLED;
 262         return ctx;
 263 }
 264
 265 struct nvme_queue *get_nvmeq(struct nvme_dev *dev)
 266 {
 267         return dev->queues[get_cpu() + 1];
 268 }
 269
 270 void put_nvmeq(struct nvme_queue *nvmeq)
 271 {
 272         put_cpu();
 273 }
 274
 275 /**
 276  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 277  * @nvmeq: The queue to use
 278  * @cmd: The command to send
 279  *
 280  * Safe to use from interrupt context
 281  */
 282 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
 283 {
 284         unsigned long flags;
 285         u16 tail;
 286         spin_lock_irqsave(&nvmeq->q_lock, flags);
 287         tail = nvmeq->sq_tail;
 288         memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
 289         if (++tail == nvmeq->q_depth)
 290                 tail = 0;
 291         writel(tail, nvmeq->q_db);
 292         nvmeq->sq_tail = tail;
 293         spin_unlock_irqrestore(&nvmeq->q_lock, flags);
 294
 295         return 0;
 296 }
 297
 298 static __le64 **iod_list(struct nvme_iod *iod)
 299 {
 300         return ((void *)iod) + iod->offset;
 301 }
 302
 303 /*
 304  * Will slightly overestimate the number of pages needed.  This is OK
 305  * as it only leads to a small amount of wasted memory for the lifetime of
 306  * the I/O.
 307  */
 308 static int nvme_npages(unsigned size)
 309 {
 310         unsigned nprps = DIV_ROUND_UP(size + PAGE_SIZE, PAGE_SIZE);
 311         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 312 }
 313
 314 static struct nvme_iod *
 315 nvme_alloc_iod(unsigned nseg, unsigned nbytes, gfp_t gfp)
 316 {
 317         struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
 318                                 sizeof(__le64 *) * nvme_npages(nbytes) +
 319                                 sizeof(struct scatterlist) * nseg, gfp);
 320
 321         if (iod) {
 322                 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
 323                 iod->npages = -1;
 324                 iod->length = nbytes;
 325                 iod->nents = 0;
 326                 iod->start_time = jiffies;
 327         }
 328
 329         return iod;
 330 }
 331
 332 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
 333 {
 334         const int last_prp = PAGE_SIZE / 8 - 1;
 335         int i;
 336         __le64 **list = iod_list(iod);
 337         dma_addr_t prp_dma = iod->first_dma;
 338
 339         if (iod->npages == 0)
 340                 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
 341         for (i = 0; i < iod->npages; i++) {
 342                 __le64 *prp_list = list[i];
 343                 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
 344                 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
 345                 prp_dma = next_prp_dma;
 346         }
 347         kfree(iod);
 348 }
 349
 350 static void nvme_start_io_acct(struct bio *bio)
 351 {
 352         struct gendisk *disk = bio->bi_bdev->bd_disk;
 353         const int rw = bio_data_dir(bio);
 354         int cpu = part_stat_lock();
 355         part_round_stats(cpu, &disk->part0);
 356         part_stat_inc(cpu, &disk->part0, ios[rw]);
 357         part_stat_add(cpu, &disk->part0, sectors[rw], bio_sectors(bio));
 358         part_inc_in_flight(&disk->part0, rw);
 359         part_stat_unlock();
 360 }
 361
 362 static void nvme_end_io_acct(struct bio *bio, unsigned long start_time)
 363 {
 364         struct gendisk *disk = bio->bi_bdev->bd_disk;
 365         const int rw = bio_data_dir(bio);
 366         unsigned long duration = jiffies - start_time;
 367         int cpu = part_stat_lock();
 368         part_stat_add(cpu, &disk->part0, ticks[rw], duration);
 369         part_round_stats(cpu, &disk->part0);
 370         part_dec_in_flight(&disk->part0, rw);
 371         part_stat_unlock();
 372 }
 373
 374 static void bio_completion(struct nvme_dev *dev, void *ctx,
 375                                                 struct nvme_completion *cqe)
 376 {
 377         struct nvme_iod *iod = ctx;
 378         struct bio *bio = iod->private;
 379         u16 status = le16_to_cpup(&cqe->status) >> 1;
 380
 381         if (iod->nents) {
 382                 dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
 383                         bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 384                 nvme_end_io_acct(bio, iod->start_time);
 385         }
 386         nvme_free_iod(dev, iod);
 387         if (status)
 388                 bio_endio(bio, -EIO);
 389         else
 390                 bio_endio(bio, 0);
 391 }
 392
 393 /* length is in bytes.  gfp flags indicates whether we may sleep. */
 394 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_common_command *cmd,
 395                         struct nvme_iod *iod, int total_len, gfp_t gfp)
 396 {
 397         struct dma_pool *pool;
 398         int length = total_len;
 399         struct scatterlist *sg = iod->sg;
 400         int dma_len = sg_dma_len(sg);
 401         u64 dma_addr = sg_dma_address(sg);
 402         int offset = offset_in_page(dma_addr);
 403         __le64 *prp_list;
 404         __le64 **list = iod_list(iod);
 405         dma_addr_t prp_dma;
 406         int nprps, i;
 407
 408         cmd->prp1 = cpu_to_le64(dma_addr);
 409         length -= (PAGE_SIZE - offset);
 410         if (length <= 0)
 411                 return total_len;
 412
 413         dma_len -= (PAGE_SIZE - offset);
 414         if (dma_len) {
 415                 dma_addr += (PAGE_SIZE - offset);
 416         } else {
 417                 sg = sg_next(sg);
 418                 dma_addr = sg_dma_address(sg);
 419                 dma_len = sg_dma_len(sg);
 420         }
 421
 422         if (length <= PAGE_SIZE) {
 423                 cmd->prp2 = cpu_to_le64(dma_addr);
 424                 return total_len;
 425         }
 426
 427         nprps = DIV_ROUND_UP(length, PAGE_SIZE);
 428         if (nprps <= (256 / 8)) {
 429                 pool = dev->prp_small_pool;
 430                 iod->npages = 0;
 431         } else {
 432                 pool = dev->prp_page_pool;
 433                 iod->npages = 1;
 434         }
 435
 436         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 437         if (!prp_list) {
 438                 cmd->prp2 = cpu_to_le64(dma_addr);
 439                 iod->npages = -1;
 440                 return (total_len - length) + PAGE_SIZE;
 441         }
 442         list[0] = prp_list;
 443         iod->first_dma = prp_dma;
 444         cmd->prp2 = cpu_to_le64(prp_dma);
 445         i = 0;
 446         for (;;) {
 447                 if (i == PAGE_SIZE / 8) {
 448                         __le64 *old_prp_list = prp_list;
 449                         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 450                         if (!prp_list)
 451                                 return total_len - length;
 452                         list[iod->npages++] = prp_list;
 453                         prp_list[0] = old_prp_list[i - 1];
 454                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 455                         i = 1;
 456                 }
 457                 prp_list[i++] = cpu_to_le64(dma_addr);
 458                 dma_len -= PAGE_SIZE;
 459                 dma_addr += PAGE_SIZE;
 460                 length -= PAGE_SIZE;
 461                 if (length <= 0)
 462                         break;
 463                 if (dma_len > 0)
 464                         continue;
 465                 BUG_ON(dma_len < 0);
 466                 sg = sg_next(sg);
 467                 dma_addr = sg_dma_address(sg);
 468                 dma_len = sg_dma_len(sg);
 469         }
 470
 471         return total_len;
 472 }
 473
 474 static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
 475                                  int len)
 476 {
 477         struct bio *split = bio_split(bio, len >> 9, GFP_ATOMIC, NULL);
 478         if (!split)
 479                 return -ENOMEM;
 480
 481         bio_chain(split, bio);
 482
 483         if (bio_list_empty(&nvmeq->sq_cong))
 484                 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
 485         bio_list_add(&nvmeq->sq_cong, split);
 486         bio_list_add(&nvmeq->sq_cong, bio);
 487
 488         return 0;
 489 }
 490
 491 /* NVMe scatterlists require no holes in the virtual address */
 492 #define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2)   ((vec2)->bv_offset || \
 493                         (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
 494
 495 static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
 496                 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
 497 {
 498         struct bio_vec bvec, bvprv;
 499         struct bvec_iter iter;
 500         struct scatterlist *sg = NULL;
 501         int length = 0, nsegs = 0, split_len = bio->bi_iter.bi_size;
 502         int first = 1;
 503
 504         if (nvmeq->dev->stripe_size)
 505                 split_len = nvmeq->dev->stripe_size -
 506                         ((bio->bi_iter.bi_sector << 9) &
 507                          (nvmeq->dev->stripe_size - 1));
 508
 509         sg_init_table(iod->sg, psegs);
 510         bio_for_each_segment(bvec, bio, iter) {
 511                 if (!first && BIOVEC_PHYS_MERGEABLE(&bvprv, &bvec)) {
 512                         sg->length += bvec.bv_len;
 513                 } else {
 514                         if (!first && BIOVEC_NOT_VIRT_MERGEABLE(&bvprv, &bvec))
 515                                 return nvme_split_and_submit(bio, nvmeq,
 516                                                              length);
 517
 518                         sg = sg ? sg + 1 : iod->sg;
 519                         sg_set_page(sg, bvec.bv_page,
 520                                     bvec.bv_len, bvec.bv_offset);
 521                         nsegs++;
 522                 }
 523
 524                 if (split_len - length < bvec.bv_len)
 525                         return nvme_split_and_submit(bio, nvmeq, split_len);
 526                 length += bvec.bv_len;
 527                 bvprv = bvec;
 528                 first = 0;
 529         }
 530         iod->nents = nsegs;
 531         sg_mark_end(sg);
 532         if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
 533                 return -ENOMEM;
 534
 535         BUG_ON(length != bio->bi_iter.bi_size);
 536         return length;
 537 }
 538
 539 /*
 540  * We reuse the small pool to allocate the 16-byte range here as it is not
 541  * worth having a special pool for these or additional cases to handle freeing
 542  * the iod.
 543  */
 544 static int nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 545                 struct bio *bio, struct nvme_iod *iod, int cmdid)
 546 {
 547         struct nvme_dsm_range *range;
 548         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 549
 550         range = dma_pool_alloc(nvmeq->dev->prp_small_pool, GFP_ATOMIC,
 551                                                         &iod->first_dma);
 552         if (!range)
 553                 return -ENOMEM;
 554
 555         iod_list(iod)[0] = (__le64 *)range;
 556         iod->npages = 0;
 557
 558         range->cattr = cpu_to_le32(0);
 559         range->nlb = cpu_to_le32(bio->bi_iter.bi_size >> ns->lba_shift);
 560         range->slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
 561
 562         memset(cmnd, 0, sizeof(*cmnd));
 563         cmnd->dsm.opcode = nvme_cmd_dsm;
 564         cmnd->dsm.command_id = cmdid;
 565         cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
 566         cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
 567         cmnd->dsm.nr = 0;
 568         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 569
 570         if (++nvmeq->sq_tail == nvmeq->q_depth)
 571                 nvmeq->sq_tail = 0;
 572         writel(nvmeq->sq_tail, nvmeq->q_db);
 573
 574         return 0;
 575 }
 576
 577 static int nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 578                                                                 int cmdid)
 579 {
 580         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 581
 582         memset(cmnd, 0, sizeof(*cmnd));
 583         cmnd->common.opcode = nvme_cmd_flush;
 584         cmnd->common.command_id = cmdid;
 585         cmnd->common.nsid = cpu_to_le32(ns->ns_id);
 586
 587         if (++nvmeq->sq_tail == nvmeq->q_depth)
 588                 nvmeq->sq_tail = 0;
 589         writel(nvmeq->sq_tail, nvmeq->q_db);
 590
 591         return 0;
 592 }
 593
 594 int nvme_submit_flush_data(struct nvme_queue *nvmeq, struct nvme_ns *ns)
 595 {
 596         int cmdid = alloc_cmdid(nvmeq, (void *)CMD_CTX_FLUSH,
 597                                         special_completion, NVME_IO_TIMEOUT);
 598         if (unlikely(cmdid < 0))
 599                 return cmdid;
 600
 601         return nvme_submit_flush(nvmeq, ns, cmdid);
 602 }
 603
 604 /*
 605  * Called with local interrupts disabled and the q_lock held.  May not sleep.
 606  */
 607 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 608                                                                 struct bio *bio)
 609 {
 610         struct nvme_command *cmnd;
 611         struct nvme_iod *iod;
 612         enum dma_data_direction dma_dir;
 613         int cmdid, length, result;
 614         u16 control;
 615         u32 dsmgmt;
 616         int psegs = bio_phys_segments(ns->queue, bio);
 617
 618         if ((bio->bi_rw & REQ_FLUSH) && psegs) {
 619                 result = nvme_submit_flush_data(nvmeq, ns);
 620                 if (result)
 621                         return result;
 622         }
 623
 624         result = -ENOMEM;
 625         iod = nvme_alloc_iod(psegs, bio->bi_iter.bi_size, GFP_ATOMIC);
 626         if (!iod)
 627                 goto nomem;
 628         iod->private = bio;
 629
 630         result = -EBUSY;
 631         cmdid = alloc_cmdid(nvmeq, iod, bio_completion, NVME_IO_TIMEOUT);
 632         if (unlikely(cmdid < 0))
 633                 goto free_iod;
 634
 635         if (bio->bi_rw & REQ_DISCARD) {
 636                 result = nvme_submit_discard(nvmeq, ns, bio, iod, cmdid);
 637                 if (result)
 638                         goto free_cmdid;
 639                 return result;
 640         }
 641         if ((bio->bi_rw & REQ_FLUSH) && !psegs)
 642                 return nvme_submit_flush(nvmeq, ns, cmdid);
 643
 644         control = 0;
 645         if (bio->bi_rw & REQ_FUA)
 646                 control |= NVME_RW_FUA;
 647         if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 648                 control |= NVME_RW_LR;
 649
 650         dsmgmt = 0;
 651         if (bio->bi_rw & REQ_RAHEAD)
 652                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 653
 654         cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 655
 656         memset(cmnd, 0, sizeof(*cmnd));
 657         if (bio_data_dir(bio)) {
 658                 cmnd->rw.opcode = nvme_cmd_write;
 659                 dma_dir = DMA_TO_DEVICE;
 660         } else {
 661                 cmnd->rw.opcode = nvme_cmd_read;
 662                 dma_dir = DMA_FROM_DEVICE;
 663         }
 664
 665         result = nvme_map_bio(nvmeq, iod, bio, dma_dir, psegs);
 666         if (result <= 0)
 667                 goto free_cmdid;
 668         length = result;
 669
 670         cmnd->rw.command_id = cmdid;
 671         cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
 672         length = nvme_setup_prps(nvmeq->dev, &cmnd->common, iod, length,
 673                                                                 GFP_ATOMIC);
 674         cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
 675         cmnd->rw.length = cpu_to_le16((length >> ns->lba_shift) - 1);
 676         cmnd->rw.control = cpu_to_le16(control);
 677         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 678
 679         nvme_start_io_acct(bio);
 680         if (++nvmeq->sq_tail == nvmeq->q_depth)
 681                 nvmeq->sq_tail = 0;
 682         writel(nvmeq->sq_tail, nvmeq->q_db);
 683
 684         return 0;
 685
 686  free_cmdid:
 687         free_cmdid(nvmeq, cmdid, NULL);
 688  free_iod:
 689         nvme_free_iod(nvmeq->dev, iod);
 690  nomem:
 691         return result;
 692 }
 693
 694 static int nvme_process_cq(struct nvme_queue *nvmeq)
 695 {
 696         u16 head, phase;
 697
 698         head = nvmeq->cq_head;
 699         phase = nvmeq->cq_phase;
 700
 701         for (;;) {
 702                 void *ctx;
 703                 nvme_completion_fn fn;
 704                 struct nvme_completion cqe = nvmeq->cqes[head];
 705                 if ((le16_to_cpu(cqe.status) & 1) != phase)
 706                         break;
 707                 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
 708                 if (++head == nvmeq->q_depth) {
 709                         head = 0;
 710                         phase = !phase;
 711                 }
 712
 713                 ctx = free_cmdid(nvmeq, cqe.command_id, &fn);
 714                 fn(nvmeq->dev, ctx, &cqe);
 715         }
 716
 717         /* If the controller ignores the cq head doorbell and continuously
 718          * writes to the queue, it is theoretically possible to wrap around
 719          * the queue twice and mistakenly return IRQ_NONE.  Linux only
 720          * requires that 0.1% of your interrupts are handled, so this isn't
 721          * a big problem.
 722          */
 723         if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
 724                 return 0;
 725
 726         writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 727         nvmeq->cq_head = head;
 728         nvmeq->cq_phase = phase;
 729
 730         nvmeq->cqe_seen = 1;
 731         return 1;
 732 }
 733
 734 static void nvme_make_request(struct request_queue *q, struct bio *bio)
 735 {
 736         struct nvme_ns *ns = q->queuedata;
 737         struct nvme_queue *nvmeq = get_nvmeq(ns->dev);
 738         int result = -EBUSY;
 739
 740         if (!nvmeq) {
 741                 put_nvmeq(NULL);
 742                 bio_endio(bio, -EIO);
 743                 return;
 744         }
 745
 746         spin_lock_irq(&nvmeq->q_lock);
 747         if (!nvmeq->q_suspended && bio_list_empty(&nvmeq->sq_cong))
 748                 result = nvme_submit_bio_queue(nvmeq, ns, bio);
 749         if (unlikely(result)) {
 750                 if (bio_list_empty(&nvmeq->sq_cong))
 751                         add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
 752                 bio_list_add(&nvmeq->sq_cong, bio);
 753         }
 754
 755         nvme_process_cq(nvmeq);
 756         spin_unlock_irq(&nvmeq->q_lock);
 757         put_nvmeq(nvmeq);
 758 }
 759
 760 static irqreturn_t nvme_irq(int irq, void *data)
 761 {
 762         irqreturn_t result;
 763         struct nvme_queue *nvmeq = data;
 764         spin_lock(&nvmeq->q_lock);
 765         nvme_process_cq(nvmeq);
 766         result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
 767         nvmeq->cqe_seen = 0;
 768         spin_unlock(&nvmeq->q_lock);
 769         return result;
 770 }
 771
 772 static irqreturn_t nvme_irq_check(int irq, void *data)
 773 {
 774         struct nvme_queue *nvmeq = data;
 775         struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
 776         if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
 777                 return IRQ_NONE;
 778         return IRQ_WAKE_THREAD;
 779 }
 780
 781 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
 782 {
 783         spin_lock_irq(&nvmeq->q_lock);
 784         cancel_cmdid(nvmeq, cmdid, NULL);
 785         spin_unlock_irq(&nvmeq->q_lock);
 786 }
 787
 788 struct sync_cmd_info {
 789         struct task_struct *task;
 790         u32 result;
 791         int status;
 792 };
 793
 794 static void sync_completion(struct nvme_dev *dev, void *ctx,
 795                                                 struct nvme_completion *cqe)
 796 {
 797         struct sync_cmd_info *cmdinfo = ctx;
 798         cmdinfo->result = le32_to_cpup(&cqe->result);
 799         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 800         wake_up_process(cmdinfo->task);
 801 }
 802
 803 /*
 804  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 805  * if the result is positive, it's an NVM Express status code
 806  */
 807 int nvme_submit_sync_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
 808                                                 u32 *result, unsigned timeout)
 809 {
 810         int cmdid;
 811         struct sync_cmd_info cmdinfo;
 812
 813         cmdinfo.task = current;
 814         cmdinfo.status = -EINTR;
 815
 816         cmdid = alloc_cmdid_killable(nvmeq, &cmdinfo, sync_completion,
 817                                                                 timeout);
 818         if (cmdid < 0)
 819                 return cmdid;
 820         cmd->common.command_id = cmdid;
 821
 822         set_current_state(TASK_KILLABLE);
 823         nvme_submit_cmd(nvmeq, cmd);
 824         schedule_timeout(timeout);
 825
 826         if (cmdinfo.status == -EINTR) {
 827                 nvme_abort_command(nvmeq, cmdid);
 828                 return -EINTR;
 829         }
 830
 831         if (result)
 832                 *result = cmdinfo.result;
 833
 834         return cmdinfo.status;
 835 }
 836
 837 static int nvme_submit_async_cmd(struct nvme_queue *nvmeq,
 838                         struct nvme_command *cmd,
 839                         struct async_cmd_info *cmdinfo, unsigned timeout)
 840 {
 841         int cmdid;
 842
 843         cmdid = alloc_cmdid_killable(nvmeq, cmdinfo, async_completion, timeout);
 844         if (cmdid < 0)
 845                 return cmdid;
 846         cmdinfo->status = -EINTR;
 847         cmd->common.command_id = cmdid;
 848         nvme_submit_cmd(nvmeq, cmd);
 849         return 0;
 850 }
 851
 852 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 853                                                                 u32 *result)
 854 {
 855         return nvme_submit_sync_cmd(dev->queues[0], cmd, result, ADMIN_TIMEOUT);
 856 }
 857
 858 static int nvme_submit_admin_cmd_async(struct nvme_dev *dev,
 859                 struct nvme_command *cmd, struct async_cmd_info *cmdinfo)
 860 {
 861         return nvme_submit_async_cmd(dev->queues[0], cmd, cmdinfo,
 862                                                                 ADMIN_TIMEOUT);
 863 }
 864
 865 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
 866 {
 867         int status;
 868         struct nvme_command c;
 869
 870         memset(&c, 0, sizeof(c));
 871         c.delete_queue.opcode = opcode;
 872         c.delete_queue.qid = cpu_to_le16(id);
 873
 874         status = nvme_submit_admin_cmd(dev, &c, NULL);
 875         if (status)
 876                 return -EIO;
 877         return 0;
 878 }
 879
 880 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
 881                                                 struct nvme_queue *nvmeq)
 882 {
 883         int status;
 884         struct nvme_command c;
 885         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
 886
 887         memset(&c, 0, sizeof(c));
 888         c.create_cq.opcode = nvme_admin_create_cq;
 889         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
 890         c.create_cq.cqid = cpu_to_le16(qid);
 891         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 892         c.create_cq.cq_flags = cpu_to_le16(flags);
 893         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
 894
 895         status = nvme_submit_admin_cmd(dev, &c, NULL);
 896         if (status)
 897                 return -EIO;
 898         return 0;
 899 }
 900
 901 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
 902                                                 struct nvme_queue *nvmeq)
 903 {
 904         int status;
 905         struct nvme_command c;
 906         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
 907
 908         memset(&c, 0, sizeof(c));
 909         c.create_sq.opcode = nvme_admin_create_sq;
 910         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
 911         c.create_sq.sqid = cpu_to_le16(qid);
 912         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 913         c.create_sq.sq_flags = cpu_to_le16(flags);
 914         c.create_sq.cqid = cpu_to_le16(qid);
 915
 916         status = nvme_submit_admin_cmd(dev, &c, NULL);
 917         if (status)
 918                 return -EIO;
 919         return 0;
 920 }
 921
 922 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
 923 {
 924         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
 925 }
 926
 927 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
 928 {
 929         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
 930 }
 931
 932 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
 933                                                         dma_addr_t dma_addr)
 934 {
 935         struct nvme_command c;
 936
 937         memset(&c, 0, sizeof(c));
 938         c.identify.opcode = nvme_admin_identify;
 939         c.identify.nsid = cpu_to_le32(nsid);
 940         c.identify.prp1 = cpu_to_le64(dma_addr);
 941         c.identify.cns = cpu_to_le32(cns);
 942
 943         return nvme_submit_admin_cmd(dev, &c, NULL);
 944 }
 945
 946 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
 947                                         dma_addr_t dma_addr, u32 *result)
 948 {
 949         struct nvme_command c;
 950
 951         memset(&c, 0, sizeof(c));
 952         c.features.opcode = nvme_admin_get_features;
 953         c.features.nsid = cpu_to_le32(nsid);
 954         c.features.prp1 = cpu_to_le64(dma_addr);
 955         c.features.fid = cpu_to_le32(fid);
 956
 957         return nvme_submit_admin_cmd(dev, &c, result);
 958 }
 959
 960 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
 961                                         dma_addr_t dma_addr, u32 *result)
 962 {
 963         struct nvme_command c;
 964
 965         memset(&c, 0, sizeof(c));
 966         c.features.opcode = nvme_admin_set_features;
 967         c.features.prp1 = cpu_to_le64(dma_addr);
 968         c.features.fid = cpu_to_le32(fid);
 969         c.features.dword11 = cpu_to_le32(dword11);
 970
 971         return nvme_submit_admin_cmd(dev, &c, result);
 972 }
 973
 974 /**
 975  * nvme_abort_cmd - Attempt aborting a command
 976  * @cmdid: Command id of a timed out IO
 977  * @queue: The queue with timed out IO
 978  *
 979  * Schedule controller reset if the command was already aborted once before and
 980  * still hasn't been returned to the driver, or if this is the admin queue.
 981  */
 982 static void nvme_abort_cmd(int cmdid, struct nvme_queue *nvmeq)
 983 {
 984         int a_cmdid;
 985         struct nvme_command cmd;
 986         struct nvme_dev *dev = nvmeq->dev;
 987         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 988
 989         if (!nvmeq->qid || info[cmdid].aborted) {
 990                 if (work_busy(&dev->reset_work))
 991                         return;
 992                 list_del_init(&dev->node);
 993                 dev_warn(&dev->pci_dev->dev,
 994                         "I/O %d QID %d timeout, reset controller\n", cmdid,
 995                                                                 nvmeq->qid);
 996                 PREPARE_WORK(&dev->reset_work, nvme_reset_failed_dev);
 997                 queue_work(nvme_workq, &dev->reset_work);
 998                 return;
 999         }
1000
1001         if (!dev->abort_limit)
1002                 return;
1003
1004         a_cmdid = alloc_cmdid(dev->queues[0], CMD_CTX_ABORT, special_completion,
1005                                                                 ADMIN_TIMEOUT);
1006         if (a_cmdid < 0)
1007                 return;
1008
1009         memset(&cmd, 0, sizeof(cmd));
1010         cmd.abort.opcode = nvme_admin_abort_cmd;
1011         cmd.abort.cid = cmdid;
1012         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1013         cmd.abort.command_id = a_cmdid;
1014
1015         --dev->abort_limit;
1016         info[cmdid].aborted = 1;
1017         info[cmdid].timeout = jiffies + ADMIN_TIMEOUT;
1018
1019         dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", cmdid,
1020                                                         nvmeq->qid);
1021         nvme_submit_cmd(dev->queues[0], &cmd);
1022 }
1023
1024 /**
1025  * nvme_cancel_ios - Cancel outstanding I/Os
1026  * @queue: The queue to cancel I/Os on
1027  * @timeout: True to only cancel I/Os which have timed out
1028  */
1029 static void nvme_cancel_ios(struct nvme_queue *nvmeq, bool timeout)
1030 {
1031         int depth = nvmeq->q_depth - 1;
1032         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1033         unsigned long now = jiffies;
1034         int cmdid;
1035
1036         for_each_set_bit(cmdid, nvmeq->cmdid_data, depth) {
1037                 void *ctx;
1038                 nvme_completion_fn fn;
1039                 static struct nvme_completion cqe = {
1040                         .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1041                 };
1042
1043                 if (timeout && !time_after(now, info[cmdid].timeout))
1044                         continue;
1045                 if (info[cmdid].ctx == CMD_CTX_CANCELLED)
1046                         continue;
1047                 if (timeout && nvmeq->dev->initialized) {
1048                         nvme_abort_cmd(cmdid, nvmeq);
1049                         continue;
1050                 }
1051                 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n", cmdid,
1052                                                                 nvmeq->qid);
1053                 ctx = cancel_cmdid(nvmeq, cmdid, &fn);
1054                 fn(nvmeq->dev, ctx, &cqe);
1055         }
1056 }
1057
1058 static void nvme_free_queue(struct nvme_queue *nvmeq)
1059 {
1060         spin_lock_irq(&nvmeq->q_lock);
1061         while (bio_list_peek(&nvmeq->sq_cong)) {
1062                 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1063                 bio_endio(bio, -EIO);
1064         }
1065         spin_unlock_irq(&nvmeq->q_lock);
1066
1067         dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1068                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1069         dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1070                                         nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1071         kfree(nvmeq);
1072 }
1073
1074 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1075 {
1076         int i;
1077
1078         for (i = dev->queue_count - 1; i >= lowest; i--) {
1079                 nvme_free_queue(dev->queues[i]);
1080                 dev->queue_count--;
1081                 dev->queues[i] = NULL;
1082         }
1083 }
1084
1085 /**
1086  * nvme_suspend_queue - put queue into suspended state
1087  * @nvmeq - queue to suspend
1088  *
1089  * Returns 1 if already suspended, 0 otherwise.
1090  */
1091 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1092 {
1093         int vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1094
1095         spin_lock_irq(&nvmeq->q_lock);
1096         if (nvmeq->q_suspended) {
1097                 spin_unlock_irq(&nvmeq->q_lock);
1098                 return 1;
1099         }
1100         nvmeq->q_suspended = 1;
1101         spin_unlock_irq(&nvmeq->q_lock);
1102
1103         irq_set_affinity_hint(vector, NULL);
1104         free_irq(vector, nvmeq);
1105
1106         return 0;
1107 }
1108
1109 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1110 {
1111         spin_lock_irq(&nvmeq->q_lock);
1112         nvme_process_cq(nvmeq);
1113         nvme_cancel_ios(nvmeq, false);
1114         spin_unlock_irq(&nvmeq->q_lock);
1115 }
1116
1117 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1118 {
1119         struct nvme_queue *nvmeq = dev->queues[qid];
1120
1121         if (!nvmeq)
1122                 return;
1123         if (nvme_suspend_queue(nvmeq))
1124                 return;
1125
1126         /* Don't tell the adapter to delete the admin queue.
1127          * Don't tell a removed adapter to delete IO queues. */
1128         if (qid && readl(&dev->bar->csts) != -1) {
1129                 adapter_delete_sq(dev, qid);
1130                 adapter_delete_cq(dev, qid);
1131         }
1132         nvme_clear_queue(nvmeq);
1133 }
1134
1135 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1136                                                         int depth, int vector)
1137 {
1138         struct device *dmadev = &dev->pci_dev->dev;
1139         unsigned extra = nvme_queue_extra(depth);
1140         struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
1141         if (!nvmeq)
1142                 return NULL;
1143
1144         nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
1145                                         &nvmeq->cq_dma_addr, GFP_KERNEL);
1146         if (!nvmeq->cqes)
1147                 goto free_nvmeq;
1148         memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
1149
1150         nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1151                                         &nvmeq->sq_dma_addr, GFP_KERNEL);
1152         if (!nvmeq->sq_cmds)
1153                 goto free_cqdma;
1154
1155         nvmeq->q_dmadev = dmadev;
1156         nvmeq->dev = dev;
1157         snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1158                         dev->instance, qid);
1159         spin_lock_init(&nvmeq->q_lock);
1160         nvmeq->cq_head = 0;
1161         nvmeq->cq_phase = 1;
1162         init_waitqueue_head(&nvmeq->sq_full);
1163         init_waitqueue_entry(&nvmeq->sq_cong_wait, nvme_thread);
1164         bio_list_init(&nvmeq->sq_cong);
1165         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1166         nvmeq->q_depth = depth;
1167         nvmeq->cq_vector = vector;
1168         nvmeq->qid = qid;
1169         nvmeq->q_suspended = 1;
1170         dev->queue_count++;
1171
1172         return nvmeq;
1173
1174  free_cqdma:
1175         dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1176                                                         nvmeq->cq_dma_addr);
1177  free_nvmeq:
1178         kfree(nvmeq);
1179         return NULL;
1180 }
1181
1182 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1183                                                         const char *name)
1184 {
1185         if (use_threaded_interrupts)
1186                 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1187                                         nvme_irq_check, nvme_irq, IRQF_SHARED,
1188                                         name, nvmeq);
1189         return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1190                                 IRQF_SHARED, name, nvmeq);
1191 }
1192
1193 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1194 {
1195         struct nvme_dev *dev = nvmeq->dev;
1196         unsigned extra = nvme_queue_extra(nvmeq->q_depth);
1197
1198         nvmeq->sq_tail = 0;
1199         nvmeq->cq_head = 0;
1200         nvmeq->cq_phase = 1;
1201         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1202         memset(nvmeq->cmdid_data, 0, extra);
1203         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1204         nvme_cancel_ios(nvmeq, false);
1205         nvmeq->q_suspended = 0;
1206 }
1207
1208 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1209 {
1210         struct nvme_dev *dev = nvmeq->dev;
1211         int result;
1212
1213         result = adapter_alloc_cq(dev, qid, nvmeq);
1214         if (result < 0)
1215                 return result;
1216
1217         result = adapter_alloc_sq(dev, qid, nvmeq);
1218         if (result < 0)
1219                 goto release_cq;
1220
1221         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1222         if (result < 0)
1223                 goto release_sq;
1224
1225         spin_lock_irq(&nvmeq->q_lock);
1226         nvme_init_queue(nvmeq, qid);
1227         spin_unlock_irq(&nvmeq->q_lock);
1228
1229         return result;
1230
1231  release_sq:
1232         adapter_delete_sq(dev, qid);
1233  release_cq:
1234         adapter_delete_cq(dev, qid);
1235         return result;
1236 }
1237
1238 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1239 {
1240         unsigned long timeout;
1241         u32 bit = enabled ? NVME_CSTS_RDY : 0;
1242
1243         timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1244
1245         while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1246                 msleep(100);
1247                 if (fatal_signal_pending(current))
1248                         return -EINTR;
1249                 if (time_after(jiffies, timeout)) {
1250                         dev_err(&dev->pci_dev->dev,
1251                                 "Device not ready; aborting initialisation\n");
1252                         return -ENODEV;
1253                 }
1254         }
1255
1256         return 0;
1257 }
1258
1259 /*
1260  * If the device has been passed off to us in an enabled state, just clear
1261  * the enabled bit.  The spec says we should set the 'shutdown notification
1262  * bits', but doing so may cause the device to complete commands to the
1263  * admin queue ... and we don't know what memory that might be pointing at!
1264  */
1265 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1266 {
1267         u32 cc = readl(&dev->bar->cc);
1268
1269         if (cc & NVME_CC_ENABLE)
1270                 writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
1271         return nvme_wait_ready(dev, cap, false);
1272 }
1273
1274 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1275 {
1276         return nvme_wait_ready(dev, cap, true);
1277 }
1278
1279 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1280 {
1281         unsigned long timeout;
1282         u32 cc;
1283
1284         cc = (readl(&dev->bar->cc) & ~NVME_CC_SHN_MASK) | NVME_CC_SHN_NORMAL;
1285         writel(cc, &dev->bar->cc);
1286
1287         timeout = 2 * HZ + jiffies;
1288         while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1289                                                         NVME_CSTS_SHST_CMPLT) {
1290                 msleep(100);
1291                 if (fatal_signal_pending(current))
1292                         return -EINTR;
1293                 if (time_after(jiffies, timeout)) {
1294                         dev_err(&dev->pci_dev->dev,
1295                                 "Device shutdown incomplete; abort shutdown\n");
1296                         return -ENODEV;
1297                 }
1298         }
1299
1300         return 0;
1301 }
1302
1303 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1304 {
1305         int result;
1306         u32 aqa;
1307         u64 cap = readq(&dev->bar->cap);
1308         struct nvme_queue *nvmeq;
1309
1310         result = nvme_disable_ctrl(dev, cap);
1311         if (result < 0)
1312                 return result;
1313
1314         nvmeq = dev->queues[0];
1315         if (!nvmeq) {
1316                 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
1317                 if (!nvmeq)
1318                         return -ENOMEM;
1319                 dev->queues[0] = nvmeq;
1320         }
1321
1322         aqa = nvmeq->q_depth - 1;
1323         aqa |= aqa << 16;
1324
1325         dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
1326         dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
1327         dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1328         dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1329
1330         writel(aqa, &dev->bar->aqa);
1331         writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1332         writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1333         writel(dev->ctrl_config, &dev->bar->cc);
1334
1335         result = nvme_enable_ctrl(dev, cap);
1336         if (result)
1337                 return result;
1338
1339         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1340         if (result)
1341                 return result;
1342
1343         spin_lock_irq(&nvmeq->q_lock);
1344         nvme_init_queue(nvmeq, 0);
1345         spin_unlock_irq(&nvmeq->q_lock);
1346         return result;
1347 }
1348
1349 struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1350                                 unsigned long addr, unsigned length)
1351 {
1352         int i, err, count, nents, offset;
1353         struct scatterlist *sg;
1354         struct page **pages;
1355         struct nvme_iod *iod;
1356
1357         if (addr & 3)
1358                 return ERR_PTR(-EINVAL);
1359         if (!length || length > INT_MAX - PAGE_SIZE)
1360                 return ERR_PTR(-EINVAL);
1361
1362         offset = offset_in_page(addr);
1363         count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1364         pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1365         if (!pages)
1366                 return ERR_PTR(-ENOMEM);
1367
1368         err = get_user_pages_fast(addr, count, 1, pages);
1369         if (err < count) {
1370                 count = err;
1371                 err = -EFAULT;
1372                 goto put_pages;
1373         }
1374
1375         iod = nvme_alloc_iod(count, length, GFP_KERNEL);
1376         sg = iod->sg;
1377         sg_init_table(sg, count);
1378         for (i = 0; i < count; i++) {
1379                 sg_set_page(&sg[i], pages[i],
1380                             min_t(unsigned, length, PAGE_SIZE - offset),
1381                             offset);
1382                 length -= (PAGE_SIZE - offset);
1383                 offset = 0;
1384         }
1385         sg_mark_end(&sg[i - 1]);
1386         iod->nents = count;
1387
1388         err = -ENOMEM;
1389         nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1390                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1391         if (!nents)
1392                 goto free_iod;
1393
1394         kfree(pages);
1395         return iod;
1396
1397  free_iod:
1398         kfree(iod);
1399  put_pages:
1400         for (i = 0; i < count; i++)
1401                 put_page(pages[i]);
1402         kfree(pages);
1403         return ERR_PTR(err);
1404 }
1405
1406 void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1407                         struct nvme_iod *iod)
1408 {
1409         int i;
1410
1411         dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1412                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1413
1414         for (i = 0; i < iod->nents; i++)
1415                 put_page(sg_page(&iod->sg[i]));
1416 }
1417
1418 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1419 {
1420         struct nvme_dev *dev = ns->dev;
1421         struct nvme_queue *nvmeq;
1422         struct nvme_user_io io;
1423         struct nvme_command c;
1424         unsigned length, meta_len;
1425         int status, i;
1426         struct nvme_iod *iod, *meta_iod = NULL;
1427         dma_addr_t meta_dma_addr;
1428         void *meta, *uninitialized_var(meta_mem);
1429
1430         if (copy_from_user(&io, uio, sizeof(io)))
1431                 return -EFAULT;
1432         length = (io.nblocks + 1) << ns->lba_shift;
1433         meta_len = (io.nblocks + 1) * ns->ms;
1434
1435         if (meta_len && ((io.metadata & 3) || !io.metadata))
1436                 return -EINVAL;
1437
1438         switch (io.opcode) {
1439         case nvme_cmd_write:
1440         case nvme_cmd_read:
1441         case nvme_cmd_compare:
1442                 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1443                 break;
1444         default:
1445                 return -EINVAL;
1446         }
1447
1448         if (IS_ERR(iod))
1449                 return PTR_ERR(iod);
1450
1451         memset(&c, 0, sizeof(c));
1452         c.rw.opcode = io.opcode;
1453         c.rw.flags = io.flags;
1454         c.rw.nsid = cpu_to_le32(ns->ns_id);
1455         c.rw.slba = cpu_to_le64(io.slba);
1456         c.rw.length = cpu_to_le16(io.nblocks);
1457         c.rw.control = cpu_to_le16(io.control);
1458         c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1459         c.rw.reftag = cpu_to_le32(io.reftag);
1460         c.rw.apptag = cpu_to_le16(io.apptag);
1461         c.rw.appmask = cpu_to_le16(io.appmask);
1462
1463         if (meta_len) {
1464                 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1465                                                                 meta_len);
1466                 if (IS_ERR(meta_iod)) {
1467                         status = PTR_ERR(meta_iod);
1468                         meta_iod = NULL;
1469                         goto unmap;
1470                 }
1471
1472                 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1473                                                 &meta_dma_addr, GFP_KERNEL);
1474                 if (!meta_mem) {
1475                         status = -ENOMEM;
1476                         goto unmap;
1477                 }
1478
1479                 if (io.opcode & 1) {
1480                         int meta_offset = 0;
1481
1482                         for (i = 0; i < meta_iod->nents; i++) {
1483                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1484                                                 meta_iod->sg[i].offset;
1485                                 memcpy(meta_mem + meta_offset, meta,
1486                                                 meta_iod->sg[i].length);
1487                                 kunmap_atomic(meta);
1488                                 meta_offset += meta_iod->sg[i].length;
1489                         }
1490                 }
1491
1492                 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1493         }
1494
1495         length = nvme_setup_prps(dev, &c.common, iod, length, GFP_KERNEL);
1496
1497         nvmeq = get_nvmeq(dev);
1498         /*
1499          * Since nvme_submit_sync_cmd sleeps, we can't keep preemption
1500          * disabled.  We may be preempted at any point, and be rescheduled
1501          * to a different CPU.  That will cause cacheline bouncing, but no
1502          * additional races since q_lock already protects against other CPUs.
1503          */
1504         put_nvmeq(nvmeq);
1505         if (length != (io.nblocks + 1) << ns->lba_shift)
1506                 status = -ENOMEM;
1507         else if (!nvmeq || nvmeq->q_suspended)
1508                 status = -EBUSY;
1509         else
1510                 status = nvme_submit_sync_cmd(nvmeq, &c, NULL, NVME_IO_TIMEOUT);
1511
1512         if (meta_len) {
1513                 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1514                         int meta_offset = 0;
1515
1516                         for (i = 0; i < meta_iod->nents; i++) {
1517                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1518                                                 meta_iod->sg[i].offset;
1519                                 memcpy(meta, meta_mem + meta_offset,
1520                                                 meta_iod->sg[i].length);
1521                                 kunmap_atomic(meta);
1522                                 meta_offset += meta_iod->sg[i].length;
1523                         }
1524                 }
1525
1526                 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1527                                                                 meta_dma_addr);
1528         }
1529
1530  unmap:
1531         nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1532         nvme_free_iod(dev, iod);
1533
1534         if (meta_iod) {
1535                 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1536                 nvme_free_iod(dev, meta_iod);
1537         }
1538
1539         return status;
1540 }
1541
1542 static int nvme_user_admin_cmd(struct nvme_dev *dev,
1543                                         struct nvme_admin_cmd __user *ucmd)
1544 {
1545         struct nvme_admin_cmd cmd;
1546         struct nvme_command c;
1547         int status, length;
1548         struct nvme_iod *uninitialized_var(iod);
1549         unsigned timeout;
1550
1551         if (!capable(CAP_SYS_ADMIN))
1552                 return -EACCES;
1553         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1554                 return -EFAULT;
1555
1556         memset(&c, 0, sizeof(c));
1557         c.common.opcode = cmd.opcode;
1558         c.common.flags = cmd.flags;
1559         c.common.nsid = cpu_to_le32(cmd.nsid);
1560         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1561         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1562         c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1563         c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1564         c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1565         c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1566         c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1567         c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1568
1569         length = cmd.data_len;
1570         if (cmd.data_len) {
1571                 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1572                                                                 length);
1573                 if (IS_ERR(iod))
1574                         return PTR_ERR(iod);
1575                 length = nvme_setup_prps(dev, &c.common, iod, length,
1576                                                                 GFP_KERNEL);
1577         }
1578
1579         timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1580                                                                 ADMIN_TIMEOUT;
1581         if (length != cmd.data_len)
1582                 status = -ENOMEM;
1583         else
1584                 status = nvme_submit_sync_cmd(dev->queues[0], &c, &cmd.result,
1585                                                                 timeout);
1586
1587         if (cmd.data_len) {
1588                 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1589                 nvme_free_iod(dev, iod);
1590         }
1591
1592         if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1593                                                         sizeof(cmd.result)))
1594                 status = -EFAULT;
1595
1596         return status;
1597 }
1598
1599 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1600                                                         unsigned long arg)
1601 {
1602         struct nvme_ns *ns = bdev->bd_disk->private_data;
1603
1604         switch (cmd) {
1605         case NVME_IOCTL_ID:
1606                 force_successful_syscall_return();
1607                 return ns->ns_id;
1608         case NVME_IOCTL_ADMIN_CMD:
1609                 return nvme_user_admin_cmd(ns->dev, (void __user *)arg);
1610         case NVME_IOCTL_SUBMIT_IO:
1611                 return nvme_submit_io(ns, (void __user *)arg);
1612         case SG_GET_VERSION_NUM:
1613                 return nvme_sg_get_version_num((void __user *)arg);
1614         case SG_IO:
1615                 return nvme_sg_io(ns, (void __user *)arg);
1616         default:
1617                 return -ENOTTY;
1618         }
1619 }
1620
1621 #ifdef CONFIG_COMPAT
1622 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1623                                         unsigned int cmd, unsigned long arg)
1624 {
1625         struct nvme_ns *ns = bdev->bd_disk->private_data;
1626
1627         switch (cmd) {
1628         case SG_IO:
1629                 return nvme_sg_io32(ns, arg);
1630         }
1631         return nvme_ioctl(bdev, mode, cmd, arg);
1632 }
1633 #else
1634 #define nvme_compat_ioctl       NULL
1635 #endif
1636
1637 static int nvme_open(struct block_device *bdev, fmode_t mode)
1638 {
1639         struct nvme_ns *ns = bdev->bd_disk->private_data;
1640         struct nvme_dev *dev = ns->dev;
1641
1642         kref_get(&dev->kref);
1643         return 0;
1644 }
1645
1646 static void nvme_free_dev(struct kref *kref);
1647
1648 static void nvme_release(struct gendisk *disk, fmode_t mode)
1649 {
1650         struct nvme_ns *ns = disk->private_data;
1651         struct nvme_dev *dev = ns->dev;
1652
1653         kref_put(&dev->kref, nvme_free_dev);
1654 }
1655
1656 static const struct block_device_operations nvme_fops = {
1657         .owner          = THIS_MODULE,
1658         .ioctl          = nvme_ioctl,
1659         .compat_ioctl   = nvme_compat_ioctl,
1660         .open           = nvme_open,
1661         .release        = nvme_release,
1662 };
1663
1664 static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
1665 {
1666         while (bio_list_peek(&nvmeq->sq_cong)) {
1667                 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1668                 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
1669
1670                 if (bio_list_empty(&nvmeq->sq_cong))
1671                         remove_wait_queue(&nvmeq->sq_full,
1672                                                         &nvmeq->sq_cong_wait);
1673                 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
1674                         if (bio_list_empty(&nvmeq->sq_cong))
1675                                 add_wait_queue(&nvmeq->sq_full,
1676                                                         &nvmeq->sq_cong_wait);
1677                         bio_list_add_head(&nvmeq->sq_cong, bio);
1678                         break;
1679                 }
1680         }
1681 }
1682
1683 static int nvme_kthread(void *data)
1684 {
1685         struct nvme_dev *dev, *next;
1686
1687         while (!kthread_should_stop()) {
1688                 set_current_state(TASK_INTERRUPTIBLE);
1689                 spin_lock(&dev_list_lock);
1690                 list_for_each_entry_safe(dev, next, &dev_list, node) {
1691                         int i;
1692                         if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1693                                                         dev->initialized) {
1694                                 if (work_busy(&dev->reset_work))
1695                                         continue;
1696                                 list_del_init(&dev->node);
1697                                 dev_warn(&dev->pci_dev->dev,
1698                                         "Failed status, reset controller\n");
1699                                 PREPARE_WORK(&dev->reset_work,
1700                                                         nvme_reset_failed_dev);
1701                                 queue_work(nvme_workq, &dev->reset_work);
1702                                 continue;
1703                         }
1704                         for (i = 0; i < dev->queue_count; i++) {
1705                                 struct nvme_queue *nvmeq = dev->queues[i];
1706                                 if (!nvmeq)
1707                                         continue;
1708                                 spin_lock_irq(&nvmeq->q_lock);
1709                                 if (nvmeq->q_suspended)
1710                                         goto unlock;
1711                                 nvme_process_cq(nvmeq);
1712                                 nvme_cancel_ios(nvmeq, true);
1713                                 nvme_resubmit_bios(nvmeq);
1714  unlock:
1715                                 spin_unlock_irq(&nvmeq->q_lock);
1716                         }
1717                 }
1718                 spin_unlock(&dev_list_lock);
1719                 schedule_timeout(round_jiffies_relative(HZ));
1720         }
1721         return 0;
1722 }
1723
1724 static void nvme_config_discard(struct nvme_ns *ns)
1725 {
1726         u32 logical_block_size = queue_logical_block_size(ns->queue);
1727         ns->queue->limits.discard_zeroes_data = 0;
1728         ns->queue->limits.discard_alignment = logical_block_size;
1729         ns->queue->limits.discard_granularity = logical_block_size;
1730         ns->queue->limits.max_discard_sectors = 0xffffffff;
1731         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1732 }
1733
1734 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1735                         struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1736 {
1737         struct nvme_ns *ns;
1738         struct gendisk *disk;
1739         int lbaf;
1740
1741         if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1742                 return NULL;
1743
1744         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1745         if (!ns)
1746                 return NULL;
1747         ns->queue = blk_alloc_queue(GFP_KERNEL);
1748         if (!ns->queue)
1749                 goto out_free_ns;
1750         ns->queue->queue_flags = QUEUE_FLAG_DEFAULT;
1751         queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1752         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1753         blk_queue_make_request(ns->queue, nvme_make_request);
1754         ns->dev = dev;
1755         ns->queue->queuedata = ns;
1756
1757         disk = alloc_disk(0);
1758         if (!disk)
1759                 goto out_free_queue;
1760         ns->ns_id = nsid;
1761         ns->disk = disk;
1762         lbaf = id->flbas & 0xf;
1763         ns->lba_shift = id->lbaf[lbaf].ds;
1764         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1765         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1766         if (dev->max_hw_sectors)
1767                 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
1768
1769         disk->major = nvme_major;
1770         disk->first_minor = 0;
1771         disk->fops = &nvme_fops;
1772         disk->private_data = ns;
1773         disk->queue = ns->queue;
1774         disk->driverfs_dev = &dev->pci_dev->dev;
1775         disk->flags = GENHD_FL_EXT_DEVT;
1776         sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
1777         set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1778
1779         if (dev->oncs & NVME_CTRL_ONCS_DSM)
1780                 nvme_config_discard(ns);
1781
1782         return ns;
1783
1784  out_free_queue:
1785         blk_cleanup_queue(ns->queue);
1786  out_free_ns:
1787         kfree(ns);
1788         return NULL;
1789 }
1790
1791 static int set_queue_count(struct nvme_dev *dev, int count)
1792 {
1793         int status;
1794         u32 result;
1795         u32 q_count = (count - 1) | ((count - 1) << 16);
1796
1797         status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
1798                                                                 &result);
1799         if (status)
1800                 return status < 0 ? -EIO : -EBUSY;
1801         return min(result & 0xffff, result >> 16) + 1;
1802 }
1803
1804 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1805 {
1806         return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1807 }
1808
1809 static int nvme_setup_io_queues(struct nvme_dev *dev)
1810 {
1811         struct nvme_queue *adminq = dev->queues[0];
1812         struct pci_dev *pdev = dev->pci_dev;
1813         int result, cpu, i, vecs, nr_io_queues, size, q_depth;
1814
1815         nr_io_queues = num_online_cpus();
1816         result = set_queue_count(dev, nr_io_queues);
1817         if (result < 0)
1818                 return result;
1819         if (result < nr_io_queues)
1820                 nr_io_queues = result;
1821
1822         size = db_bar_size(dev, nr_io_queues);
1823         if (size > 8192) {
1824                 iounmap(dev->bar);
1825                 do {
1826                         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1827                         if (dev->bar)
1828                                 break;
1829                         if (!--nr_io_queues)
1830                                 return -ENOMEM;
1831                         size = db_bar_size(dev, nr_io_queues);
1832                 } while (1);
1833                 dev->dbs = ((void __iomem *)dev->bar) + 4096;
1834                 dev->queues[0]->q_db = dev->dbs;
1835         }
1836
1837         /* Deregister the admin queue's interrupt */
1838         free_irq(dev->entry[0].vector, adminq);
1839
1840         vecs = nr_io_queues;
1841         for (i = 0; i < vecs; i++)
1842                 dev->entry[i].entry = i;
1843         for (;;) {
1844                 result = pci_enable_msix(pdev, dev->entry, vecs);
1845                 if (result <= 0)
1846                         break;
1847                 vecs = result;
1848         }
1849
1850         if (result < 0) {
1851                 vecs = nr_io_queues;
1852                 if (vecs > 32)
1853                         vecs = 32;
1854                 for (;;) {
1855                         result = pci_enable_msi_block(pdev, vecs);
1856                         if (result == 0) {
1857                                 for (i = 0; i < vecs; i++)
1858                                         dev->entry[i].vector = i + pdev->irq;
1859                                 break;
1860                         } else if (result < 0) {
1861                                 vecs = 1;
1862                                 break;
1863                         }
1864                         vecs = result;
1865                 }
1866         }
1867
1868         /*
1869          * Should investigate if there's a performance win from allocating
1870          * more queues than interrupt vectors; it might allow the submission
1871          * path to scale better, even if the receive path is limited by the
1872          * number of interrupts.
1873          */
1874         nr_io_queues = vecs;
1875
1876         result = queue_request_irq(dev, adminq, adminq->irqname);
1877         if (result) {
1878                 adminq->q_suspended = 1;
1879                 goto free_queues;
1880         }
1881
1882         /* Free previously allocated queues that are no longer usable */
1883         spin_lock(&dev_list_lock);
1884         for (i = dev->queue_count - 1; i > nr_io_queues; i--) {
1885                 struct nvme_queue *nvmeq = dev->queues[i];
1886
1887                 spin_lock_irq(&nvmeq->q_lock);
1888                 nvme_cancel_ios(nvmeq, false);
1889                 spin_unlock_irq(&nvmeq->q_lock);
1890
1891                 nvme_free_queue(nvmeq);
1892                 dev->queue_count--;
1893                 dev->queues[i] = NULL;
1894         }
1895         spin_unlock(&dev_list_lock);
1896
1897         cpu = cpumask_first(cpu_online_mask);
1898         for (i = 0; i < nr_io_queues; i++) {
1899                 irq_set_affinity_hint(dev->entry[i].vector, get_cpu_mask(cpu));
1900                 cpu = cpumask_next(cpu, cpu_online_mask);
1901         }
1902
1903         q_depth = min_t(int, NVME_CAP_MQES(readq(&dev->bar->cap)) + 1,
1904                                                                 NVME_Q_DEPTH);
1905         for (i = dev->queue_count - 1; i < nr_io_queues; i++) {
1906                 dev->queues[i + 1] = nvme_alloc_queue(dev, i + 1, q_depth, i);
1907                 if (!dev->queues[i + 1]) {
1908                         result = -ENOMEM;
1909                         goto free_queues;
1910                 }
1911         }
1912
1913         for (; i < num_possible_cpus(); i++) {
1914                 int target = i % rounddown_pow_of_two(dev->queue_count - 1);
1915                 dev->queues[i + 1] = dev->queues[target + 1];
1916         }
1917
1918         for (i = 1; i < dev->queue_count; i++) {
1919                 result = nvme_create_queue(dev->queues[i], i);
1920                 if (result) {
1921                         for (--i; i > 0; i--)
1922                                 nvme_disable_queue(dev, i);
1923                         goto free_queues;
1924                 }
1925         }
1926
1927         return 0;
1928
1929  free_queues:
1930         nvme_free_queues(dev, 1);
1931         return result;
1932 }
1933
1934 /*
1935  * Return: error value if an error occurred setting up the queues or calling
1936  * Identify Device.  0 if these succeeded, even if adding some of the
1937  * namespaces failed.  At the moment, these failures are silent.  TBD which
1938  * failures should be reported.
1939  */
1940 static int nvme_dev_add(struct nvme_dev *dev)
1941 {
1942         struct pci_dev *pdev = dev->pci_dev;
1943         int res;
1944         unsigned nn, i;
1945         struct nvme_ns *ns;
1946         struct nvme_id_ctrl *ctrl;
1947         struct nvme_id_ns *id_ns;
1948         void *mem;
1949         dma_addr_t dma_addr;
1950         int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
1951
1952         mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
1953         if (!mem)
1954                 return -ENOMEM;
1955
1956         res = nvme_identify(dev, 0, 1, dma_addr);
1957         if (res) {
1958                 res = -EIO;
1959                 goto out;
1960         }
1961
1962         ctrl = mem;
1963         nn = le32_to_cpup(&ctrl->nn);
1964         dev->oncs = le16_to_cpup(&ctrl->oncs);
1965         dev->abort_limit = ctrl->acl + 1;
1966         memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
1967         memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
1968         memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
1969         if (ctrl->mdts)
1970                 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
1971         if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
1972                         (pdev->device == 0x0953) && ctrl->vs[3])
1973                 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
1974
1975         id_ns = mem;
1976         for (i = 1; i <= nn; i++) {
1977                 res = nvme_identify(dev, i, 0, dma_addr);
1978                 if (res)
1979                         continue;
1980
1981                 if (id_ns->ncap == 0)
1982                         continue;
1983
1984                 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
1985                                                         dma_addr + 4096, NULL);
1986                 if (res)
1987                         memset(mem + 4096, 0, 4096);
1988
1989                 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
1990                 if (ns)
1991                         list_add_tail(&ns->list, &dev->namespaces);
1992         }
1993         list_for_each_entry(ns, &dev->namespaces, list)
1994                 add_disk(ns->disk);
1995         res = 0;
1996
1997  out:
1998         dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
1999         return res;
2000 }
2001
2002 static int nvme_dev_map(struct nvme_dev *dev)
2003 {
2004         int bars, result = -ENOMEM;
2005         struct pci_dev *pdev = dev->pci_dev;
2006
2007         if (pci_enable_device_mem(pdev))
2008                 return result;
2009
2010         dev->entry[0].vector = pdev->irq;
2011         pci_set_master(pdev);
2012         bars = pci_select_bars(pdev, IORESOURCE_MEM);
2013         if (pci_request_selected_regions(pdev, bars, "nvme"))
2014                 goto disable_pci;
2015
2016         if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2017             dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2018                 goto disable;
2019
2020         dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2021         if (!dev->bar)
2022                 goto disable;
2023         if (readl(&dev->bar->csts) == -1) {
2024                 result = -ENODEV;
2025                 goto unmap;
2026         }
2027         dev->db_stride = 1 << NVME_CAP_STRIDE(readq(&dev->bar->cap));
2028         dev->dbs = ((void __iomem *)dev->bar) + 4096;
2029
2030         return 0;
2031
2032  unmap:
2033         iounmap(dev->bar);
2034         dev->bar = NULL;
2035  disable:
2036         pci_release_regions(pdev);
2037  disable_pci:
2038         pci_disable_device(pdev);
2039         return result;
2040 }
2041
2042 static void nvme_dev_unmap(struct nvme_dev *dev)
2043 {
2044         if (dev->pci_dev->msi_enabled)
2045                 pci_disable_msi(dev->pci_dev);
2046         else if (dev->pci_dev->msix_enabled)
2047                 pci_disable_msix(dev->pci_dev);
2048
2049         if (dev->bar) {
2050                 iounmap(dev->bar);
2051                 dev->bar = NULL;
2052                 pci_release_regions(dev->pci_dev);
2053         }
2054
2055         if (pci_is_enabled(dev->pci_dev))
2056                 pci_disable_device(dev->pci_dev);
2057 }
2058
2059 struct nvme_delq_ctx {
2060         struct task_struct *waiter;
2061         struct kthread_worker *worker;
2062         atomic_t refcount;
2063 };
2064
2065 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2066 {
2067         dq->waiter = current;
2068         mb();
2069
2070         for (;;) {
2071                 set_current_state(TASK_KILLABLE);
2072                 if (!atomic_read(&dq->refcount))
2073                         break;
2074                 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2075                                         fatal_signal_pending(current)) {
2076                         set_current_state(TASK_RUNNING);
2077
2078                         nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2079                         nvme_disable_queue(dev, 0);
2080
2081                         send_sig(SIGKILL, dq->worker->task, 1);
2082                         flush_kthread_worker(dq->worker);
2083                         return;
2084                 }
2085         }
2086         set_current_state(TASK_RUNNING);
2087 }
2088
2089 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2090 {
2091         atomic_dec(&dq->refcount);
2092         if (dq->waiter)
2093                 wake_up_process(dq->waiter);
2094 }
2095
2096 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2097 {
2098         atomic_inc(&dq->refcount);
2099         return dq;
2100 }
2101
2102 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2103 {
2104         struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2105
2106         nvme_clear_queue(nvmeq);
2107         nvme_put_dq(dq);
2108 }
2109
2110 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2111                                                 kthread_work_func_t fn)
2112 {
2113         struct nvme_command c;
2114
2115         memset(&c, 0, sizeof(c));
2116         c.delete_queue.opcode = opcode;
2117         c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2118
2119         init_kthread_work(&nvmeq->cmdinfo.work, fn);
2120         return nvme_submit_admin_cmd_async(nvmeq->dev, &c, &nvmeq->cmdinfo);
2121 }
2122
2123 static void nvme_del_cq_work_handler(struct kthread_work *work)
2124 {
2125         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2126                                                         cmdinfo.work);
2127         nvme_del_queue_end(nvmeq);
2128 }
2129
2130 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2131 {
2132         return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2133                                                 nvme_del_cq_work_handler);
2134 }
2135
2136 static void nvme_del_sq_work_handler(struct kthread_work *work)
2137 {
2138         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2139                                                         cmdinfo.work);
2140         int status = nvmeq->cmdinfo.status;
2141
2142         if (!status)
2143                 status = nvme_delete_cq(nvmeq);
2144         if (status)
2145                 nvme_del_queue_end(nvmeq);
2146 }
2147
2148 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2149 {
2150         return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2151                                                 nvme_del_sq_work_handler);
2152 }
2153
2154 static void nvme_del_queue_start(struct kthread_work *work)
2155 {
2156         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2157                                                         cmdinfo.work);
2158         allow_signal(SIGKILL);
2159         if (nvme_delete_sq(nvmeq))
2160                 nvme_del_queue_end(nvmeq);
2161 }
2162
2163 static void nvme_disable_io_queues(struct nvme_dev *dev)
2164 {
2165         int i;
2166         DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2167         struct nvme_delq_ctx dq;
2168         struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2169                                         &worker, "nvme%d", dev->instance);
2170
2171         if (IS_ERR(kworker_task)) {
2172                 dev_err(&dev->pci_dev->dev,
2173                         "Failed to create queue del task\n");
2174                 for (i = dev->queue_count - 1; i > 0; i--)
2175                         nvme_disable_queue(dev, i);
2176                 return;
2177         }
2178
2179         dq.waiter = NULL;
2180         atomic_set(&dq.refcount, 0);
2181         dq.worker = &worker;
2182         for (i = dev->queue_count - 1; i > 0; i--) {
2183                 struct nvme_queue *nvmeq = dev->queues[i];
2184
2185                 if (nvme_suspend_queue(nvmeq))
2186                         continue;
2187                 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2188                 nvmeq->cmdinfo.worker = dq.worker;
2189                 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2190                 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2191         }
2192         nvme_wait_dq(&dq, dev);
2193         kthread_stop(kworker_task);
2194 }
2195
2196 static void nvme_dev_shutdown(struct nvme_dev *dev)
2197 {
2198         int i;
2199
2200         dev->initialized = 0;
2201
2202         spin_lock(&dev_list_lock);
2203         list_del_init(&dev->node);
2204         spin_unlock(&dev_list_lock);
2205
2206         if (!dev->bar || (dev->bar && readl(&dev->bar->csts) == -1)) {
2207                 for (i = dev->queue_count - 1; i >= 0; i--) {
2208                         struct nvme_queue *nvmeq = dev->queues[i];
2209                         nvme_suspend_queue(nvmeq);
2210                         nvme_clear_queue(nvmeq);
2211                 }
2212         } else {
2213                 nvme_disable_io_queues(dev);
2214                 nvme_shutdown_ctrl(dev);
2215                 nvme_disable_queue(dev, 0);
2216         }
2217         nvme_dev_unmap(dev);
2218 }
2219
2220 static void nvme_dev_remove(struct nvme_dev *dev)
2221 {
2222         struct nvme_ns *ns;
2223
2224         list_for_each_entry(ns, &dev->namespaces, list) {
2225                 if (ns->disk->flags & GENHD_FL_UP)
2226                         del_gendisk(ns->disk);
2227                 if (!blk_queue_dying(ns->queue))
2228                         blk_cleanup_queue(ns->queue);
2229         }
2230 }
2231
2232 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2233 {
2234         struct device *dmadev = &dev->pci_dev->dev;
2235         dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2236                                                 PAGE_SIZE, PAGE_SIZE, 0);
2237         if (!dev->prp_page_pool)
2238                 return -ENOMEM;
2239
2240         /* Optimisation for I/Os between 4k and 128k */
2241         dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2242                                                 256, 256, 0);
2243         if (!dev->prp_small_pool) {
2244                 dma_pool_destroy(dev->prp_page_pool);
2245                 return -ENOMEM;
2246         }
2247         return 0;
2248 }
2249
2250 static void nvme_release_prp_pools(struct nvme_dev *dev)
2251 {
2252         dma_pool_destroy(dev->prp_page_pool);
2253         dma_pool_destroy(dev->prp_small_pool);
2254 }
2255
2256 static DEFINE_IDA(nvme_instance_ida);
2257
2258 static int nvme_set_instance(struct nvme_dev *dev)
2259 {
2260         int instance, error;
2261
2262         do {
2263                 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2264                         return -ENODEV;
2265
2266                 spin_lock(&dev_list_lock);
2267                 error = ida_get_new(&nvme_instance_ida, &instance);
2268                 spin_unlock(&dev_list_lock);
2269         } while (error == -EAGAIN);
2270
2271         if (error)
2272                 return -ENODEV;
2273
2274         dev->instance = instance;
2275         return 0;
2276 }
2277
2278 static void nvme_release_instance(struct nvme_dev *dev)
2279 {
2280         spin_lock(&dev_list_lock);
2281         ida_remove(&nvme_instance_ida, dev->instance);
2282         spin_unlock(&dev_list_lock);
2283 }
2284
2285 static void nvme_free_namespaces(struct nvme_dev *dev)
2286 {
2287         struct nvme_ns *ns, *next;
2288
2289         list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2290                 list_del(&ns->list);
2291                 put_disk(ns->disk);
2292                 kfree(ns);
2293         }
2294 }
2295
2296 static void nvme_free_dev(struct kref *kref)
2297 {
2298         struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2299
2300         nvme_free_namespaces(dev);
2301         kfree(dev->queues);
2302         kfree(dev->entry);
2303         kfree(dev);
2304 }
2305
2306 static int nvme_dev_open(struct inode *inode, struct file *f)
2307 {
2308         struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2309                                                                 miscdev);
2310         kref_get(&dev->kref);
2311         f->private_data = dev;
2312         return 0;
2313 }
2314
2315 static int nvme_dev_release(struct inode *inode, struct file *f)
2316 {
2317         struct nvme_dev *dev = f->private_data;
2318         kref_put(&dev->kref, nvme_free_dev);
2319         return 0;
2320 }
2321
2322 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2323 {
2324         struct nvme_dev *dev = f->private_data;
2325         switch (cmd) {
2326         case NVME_IOCTL_ADMIN_CMD:
2327                 return nvme_user_admin_cmd(dev, (void __user *)arg);
2328         default:
2329                 return -ENOTTY;
2330         }
2331 }
2332
2333 static const struct file_operations nvme_dev_fops = {
2334         .owner          = THIS_MODULE,
2335         .open           = nvme_dev_open,
2336         .release        = nvme_dev_release,
2337         .unlocked_ioctl = nvme_dev_ioctl,
2338         .compat_ioctl   = nvme_dev_ioctl,
2339 };
2340
2341 static int nvme_dev_start(struct nvme_dev *dev)
2342 {
2343         int result;
2344
2345         result = nvme_dev_map(dev);
2346         if (result)
2347                 return result;
2348
2349         result = nvme_configure_admin_queue(dev);
2350         if (result)
2351                 goto unmap;
2352
2353         spin_lock(&dev_list_lock);
2354         list_add(&dev->node, &dev_list);
2355         spin_unlock(&dev_list_lock);
2356
2357         result = nvme_setup_io_queues(dev);
2358         if (result && result != -EBUSY)
2359                 goto disable;
2360
2361         return result;
2362
2363  disable:
2364         nvme_disable_queue(dev, 0);
2365         spin_lock(&dev_list_lock);
2366         list_del_init(&dev->node);
2367         spin_unlock(&dev_list_lock);
2368  unmap:
2369         nvme_dev_unmap(dev);
2370         return result;
2371 }
2372
2373 static int nvme_remove_dead_ctrl(void *arg)
2374 {
2375         struct nvme_dev *dev = (struct nvme_dev *)arg;
2376         struct pci_dev *pdev = dev->pci_dev;
2377
2378         if (pci_get_drvdata(pdev))
2379                 pci_stop_and_remove_bus_device(pdev);
2380         kref_put(&dev->kref, nvme_free_dev);
2381         return 0;
2382 }
2383
2384 static void nvme_remove_disks(struct work_struct *ws)
2385 {
2386         int i;
2387         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2388
2389         nvme_dev_remove(dev);
2390         spin_lock(&dev_list_lock);
2391         for (i = dev->queue_count - 1; i > 0; i--) {
2392                 BUG_ON(!dev->queues[i] || !dev->queues[i]->q_suspended);
2393                 nvme_free_queue(dev->queues[i]);
2394                 dev->queue_count--;
2395                 dev->queues[i] = NULL;
2396         }
2397         spin_unlock(&dev_list_lock);
2398 }
2399
2400 static int nvme_dev_resume(struct nvme_dev *dev)
2401 {
2402         int ret;
2403
2404         ret = nvme_dev_start(dev);
2405         if (ret && ret != -EBUSY)
2406                 return ret;
2407         if (ret == -EBUSY) {
2408                 spin_lock(&dev_list_lock);
2409                 PREPARE_WORK(&dev->reset_work, nvme_remove_disks);
2410                 queue_work(nvme_workq, &dev->reset_work);
2411                 spin_unlock(&dev_list_lock);
2412         }
2413         dev->initialized = 1;
2414         return 0;
2415 }
2416
2417 static void nvme_dev_reset(struct nvme_dev *dev)
2418 {
2419         nvme_dev_shutdown(dev);
2420         if (nvme_dev_resume(dev)) {
2421                 dev_err(&dev->pci_dev->dev, "Device failed to resume\n");
2422                 kref_get(&dev->kref);
2423                 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2424                                                         dev->instance))) {
2425                         dev_err(&dev->pci_dev->dev,
2426                                 "Failed to start controller remove task\n");
2427                         kref_put(&dev->kref, nvme_free_dev);
2428                 }
2429         }
2430 }
2431
2432 static void nvme_reset_failed_dev(struct work_struct *ws)
2433 {
2434         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2435         nvme_dev_reset(dev);
2436 }
2437
2438 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2439 {
2440         int result = -ENOMEM;
2441         struct nvme_dev *dev;
2442
2443         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2444         if (!dev)
2445                 return -ENOMEM;
2446         dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
2447                                                                 GFP_KERNEL);
2448         if (!dev->entry)
2449                 goto free;
2450         dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
2451                                                                 GFP_KERNEL);
2452         if (!dev->queues)
2453                 goto free;
2454
2455         INIT_LIST_HEAD(&dev->namespaces);
2456         INIT_WORK(&dev->reset_work, nvme_reset_failed_dev);
2457         dev->pci_dev = pdev;
2458         pci_set_drvdata(pdev, dev);
2459         result = nvme_set_instance(dev);
2460         if (result)
2461                 goto free;
2462
2463         result = nvme_setup_prp_pools(dev);
2464         if (result)
2465                 goto release;
2466
2467         result = nvme_dev_start(dev);
2468         if (result) {
2469                 if (result == -EBUSY)
2470                         goto create_cdev;
2471                 goto release_pools;
2472         }
2473
2474         kref_init(&dev->kref);
2475         result = nvme_dev_add(dev);
2476         if (result)
2477                 goto shutdown;
2478
2479  create_cdev:
2480         scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2481         dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2482         dev->miscdev.parent = &pdev->dev;
2483         dev->miscdev.name = dev->name;
2484         dev->miscdev.fops = &nvme_dev_fops;
2485         result = misc_register(&dev->miscdev);
2486         if (result)
2487                 goto remove;
2488
2489         dev->initialized = 1;
2490         return 0;
2491
2492  remove:
2493         nvme_dev_remove(dev);
2494         nvme_free_namespaces(dev);
2495  shutdown:
2496         nvme_dev_shutdown(dev);
2497  release_pools:
2498         nvme_free_queues(dev, 0);
2499         nvme_release_prp_pools(dev);
2500  release:
2501         nvme_release_instance(dev);
2502  free:
2503         kfree(dev->queues);
2504         kfree(dev->entry);
2505         kfree(dev);
2506         return result;
2507 }
2508
2509 static void nvme_shutdown(struct pci_dev *pdev)
2510 {
2511         struct nvme_dev *dev = pci_get_drvdata(pdev);
2512         nvme_dev_shutdown(dev);
2513 }
2514
2515 static void nvme_remove(struct pci_dev *pdev)
2516 {
2517         struct nvme_dev *dev = pci_get_drvdata(pdev);
2518
2519         spin_lock(&dev_list_lock);
2520         list_del_init(&dev->node);
2521         spin_unlock(&dev_list_lock);
2522
2523         pci_set_drvdata(pdev, NULL);
2524         flush_work(&dev->reset_work);
2525         misc_deregister(&dev->miscdev);
2526         nvme_dev_remove(dev);
2527         nvme_dev_shutdown(dev);
2528         nvme_free_queues(dev, 0);
2529         nvme_release_instance(dev);
2530         nvme_release_prp_pools(dev);
2531         kref_put(&dev->kref, nvme_free_dev);
2532 }
2533
2534 /* These functions are yet to be implemented */
2535 #define nvme_error_detected NULL
2536 #define nvme_dump_registers NULL
2537 #define nvme_link_reset NULL
2538 #define nvme_slot_reset NULL
2539 #define nvme_error_resume NULL
2540
2541 static int nvme_suspend(struct device *dev)
2542 {
2543         struct pci_dev *pdev = to_pci_dev(dev);
2544         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2545
2546         nvme_dev_shutdown(ndev);
2547         return 0;
2548 }
2549
2550 static int nvme_resume(struct device *dev)
2551 {
2552         struct pci_dev *pdev = to_pci_dev(dev);
2553         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2554
2555         if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2556                 PREPARE_WORK(&ndev->reset_work, nvme_reset_failed_dev);
2557                 queue_work(nvme_workq, &ndev->reset_work);
2558         }
2559         return 0;
2560 }
2561
2562 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2563
2564 static const struct pci_error_handlers nvme_err_handler = {
2565         .error_detected = nvme_error_detected,
2566         .mmio_enabled   = nvme_dump_registers,
2567         .link_reset     = nvme_link_reset,
2568         .slot_reset     = nvme_slot_reset,
2569         .resume         = nvme_error_resume,
2570 };
2571
2572 /* Move to pci_ids.h later */
2573 #define PCI_CLASS_STORAGE_EXPRESS       0x010802
2574
2575 static DEFINE_PCI_DEVICE_TABLE(nvme_id_table) = {
2576         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2577         { 0, }
2578 };
2579 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2580
2581 static struct pci_driver nvme_driver = {
2582         .name           = "nvme",
2583         .id_table       = nvme_id_table,
2584         .probe          = nvme_probe,
2585         .remove         = nvme_remove,
2586         .shutdown       = nvme_shutdown,
2587         .driver         = {
2588                 .pm     = &nvme_dev_pm_ops,
2589         },
2590         .err_handler    = &nvme_err_handler,
2591 };
2592
2593 static int __init nvme_init(void)
2594 {
2595         int result;
2596
2597         nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2598         if (IS_ERR(nvme_thread))
2599                 return PTR_ERR(nvme_thread);
2600
2601         result = -ENOMEM;
2602         nvme_workq = create_singlethread_workqueue("nvme");
2603         if (!nvme_workq)
2604                 goto kill_kthread;
2605
2606         result = register_blkdev(nvme_major, "nvme");
2607         if (result < 0)
2608                 goto kill_workq;
2609         else if (result > 0)
2610                 nvme_major = result;
2611
2612         result = pci_register_driver(&nvme_driver);
2613         if (result)
2614                 goto unregister_blkdev;
2615         return 0;
2616
2617  unregister_blkdev:
2618         unregister_blkdev(nvme_major, "nvme");
2619  kill_workq:
2620         destroy_workqueue(nvme_workq);
2621  kill_kthread:
2622         kthread_stop(nvme_thread);
2623         return result;
2624 }
2625
2626 static void __exit nvme_exit(void)
2627 {
2628         pci_unregister_driver(&nvme_driver);
2629         unregister_blkdev(nvme_major, "nvme");
2630         destroy_workqueue(nvme_workq);
2631         kthread_stop(nvme_thread);
2632 }
2633
2634 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2635 MODULE_LICENSE("GPL");
2636 MODULE_VERSION("0.8");
2637 module_init(nvme_init);
2638 module_exit(nvme_exit);