drivers/block/nvme-core.c

   1 /*
   2  * NVM Express device driver
   3  * Copyright (c) 2011-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * 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/cpu.h>
  24 #include <linux/delay.h>
  25 #include <linux/errno.h>
  26 #include <linux/fs.h>
  27 #include <linux/genhd.h>
  28 #include <linux/hdreg.h>
  29 #include <linux/idr.h>
  30 #include <linux/init.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/io.h>
  33 #include <linux/kdev_t.h>
  34 #include <linux/kthread.h>
  35 #include <linux/kernel.h>
  36 #include <linux/mm.h>
  37 #include <linux/module.h>
  38 #include <linux/moduleparam.h>
  39 #include <linux/pci.h>
  40 #include <linux/percpu.h>
  41 #include <linux/poison.h>
  42 #include <linux/ptrace.h>
  43 #include <linux/sched.h>
  44 #include <linux/slab.h>
  45 #include <linux/types.h>
  46 #include <scsi/sg.h>
  47 #include <asm-generic/io-64-nonatomic-lo-hi.h>
  48
  49 #define NVME_Q_DEPTH 1024
  50 #define SQ_SIZE(depth)          (depth * sizeof(struct nvme_command))
  51 #define CQ_SIZE(depth)          (depth * sizeof(struct nvme_completion))
  52 #define ADMIN_TIMEOUT   (60 * HZ)
  53 #define IOD_TIMEOUT     (4 * NVME_IO_TIMEOUT)
  54
  55 unsigned char io_timeout = 30;
  56 module_param(io_timeout, byte, 0644);
  57 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
  58
  59 static int nvme_major;
  60 module_param(nvme_major, int, 0);
  61
  62 static int use_threaded_interrupts;
  63 module_param(use_threaded_interrupts, int, 0);
  64
  65 static DEFINE_SPINLOCK(dev_list_lock);
  66 static LIST_HEAD(dev_list);
  67 static struct task_struct *nvme_thread;
  68 static struct workqueue_struct *nvme_workq;
  69 static wait_queue_head_t nvme_kthread_wait;
  70
  71 static void nvme_reset_failed_dev(struct work_struct *ws);
  72
  73 struct async_cmd_info {
  74         struct kthread_work work;
  75         struct kthread_worker *worker;
  76         u32 result;
  77         int status;
  78         void *ctx;
  79 };
  80
  81 /*
  82  * An NVM Express queue.  Each device has at least two (one for admin
  83  * commands and one for I/O commands).
  84  */
  85 struct nvme_queue {
  86         struct rcu_head r_head;
  87         struct device *q_dmadev;
  88         struct nvme_dev *dev;
  89         char irqname[24];       /* nvme4294967295-65535\0 */
  90         spinlock_t q_lock;
  91         struct nvme_command *sq_cmds;
  92         volatile struct nvme_completion *cqes;
  93         dma_addr_t sq_dma_addr;
  94         dma_addr_t cq_dma_addr;
  95         wait_queue_head_t sq_full;
  96         wait_queue_t sq_cong_wait;
  97         struct bio_list sq_cong;
  98         struct list_head iod_bio;
  99         u32 __iomem *q_db;
 100         u16 q_depth;
 101         u16 cq_vector;
 102         u16 sq_head;
 103         u16 sq_tail;
 104         u16 cq_head;
 105         u16 qid;
 106         u8 cq_phase;
 107         u8 cqe_seen;
 108         u8 q_suspended;
 109         cpumask_var_t cpu_mask;
 110         struct async_cmd_info cmdinfo;
 111         unsigned long cmdid_data[];
 112 };
 113
 114 /*
 115  * Check we didin't inadvertently grow the command struct
 116  */
 117 static inline void _nvme_check_size(void)
 118 {
 119         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
 120         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
 121         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
 122         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
 123         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
 124         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
 125         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
 126         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
 127         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
 128         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
 129         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
 130         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
 131 }
 132
 133 typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
 134                                                 struct nvme_completion *);
 135
 136 struct nvme_cmd_info {
 137         nvme_completion_fn fn;
 138         void *ctx;
 139         unsigned long timeout;
 140         int aborted;
 141 };
 142
 143 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
 144 {
 145         return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
 146 }
 147
 148 static unsigned nvme_queue_extra(int depth)
 149 {
 150         return DIV_ROUND_UP(depth, 8) + (depth * sizeof(struct nvme_cmd_info));
 151 }
 152
 153 /**
 154  * alloc_cmdid() - Allocate a Command ID
 155  * @nvmeq: The queue that will be used for this command
 156  * @ctx: A pointer that will be passed to the handler
 157  * @handler: The function to call on completion
 158  *
 159  * Allocate a Command ID for a queue.  The data passed in will
 160  * be passed to the completion handler.  This is implemented by using
 161  * the bottom two bits of the ctx pointer to store the handler ID.
 162  * Passing in a pointer that's not 4-byte aligned will cause a BUG.
 163  * We can change this if it becomes a problem.
 164  *
 165  * May be called with local interrupts disabled and the q_lock held,
 166  * or with interrupts enabled and no locks held.
 167  */
 168 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx,
 169                                 nvme_completion_fn handler, unsigned timeout)
 170 {
 171         int depth = nvmeq->q_depth - 1;
 172         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 173         int cmdid;
 174
 175         do {
 176                 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
 177                 if (cmdid >= depth)
 178                         return -EBUSY;
 179         } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
 180
 181         info[cmdid].fn = handler;
 182         info[cmdid].ctx = ctx;
 183         info[cmdid].timeout = jiffies + timeout;
 184         info[cmdid].aborted = 0;
 185         return cmdid;
 186 }
 187
 188 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
 189                                 nvme_completion_fn handler, unsigned timeout)
 190 {
 191         int cmdid;
 192         wait_event_killable(nvmeq->sq_full,
 193                 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
 194         return (cmdid < 0) ? -EINTR : cmdid;
 195 }
 196
 197 /* Special values must be less than 0x1000 */
 198 #define CMD_CTX_BASE            ((void *)POISON_POINTER_DELTA)
 199 #define CMD_CTX_CANCELLED       (0x30C + CMD_CTX_BASE)
 200 #define CMD_CTX_COMPLETED       (0x310 + CMD_CTX_BASE)
 201 #define CMD_CTX_INVALID         (0x314 + CMD_CTX_BASE)
 202 #define CMD_CTX_FLUSH           (0x318 + CMD_CTX_BASE)
 203 #define CMD_CTX_ABORT           (0x31C + CMD_CTX_BASE)
 204
 205 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
 206                                                 struct nvme_completion *cqe)
 207 {
 208         if (ctx == CMD_CTX_CANCELLED)
 209                 return;
 210         if (ctx == CMD_CTX_FLUSH)
 211                 return;
 212         if (ctx == CMD_CTX_ABORT) {
 213                 ++nvmeq->dev->abort_limit;
 214                 return;
 215         }
 216         if (ctx == CMD_CTX_COMPLETED) {
 217                 dev_warn(nvmeq->q_dmadev,
 218                                 "completed id %d twice on queue %d\n",
 219                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 220                 return;
 221         }
 222         if (ctx == CMD_CTX_INVALID) {
 223                 dev_warn(nvmeq->q_dmadev,
 224                                 "invalid id %d completed on queue %d\n",
 225                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 226                 return;
 227         }
 228
 229         dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
 230 }
 231
 232 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
 233                                                 struct nvme_completion *cqe)
 234 {
 235         struct async_cmd_info *cmdinfo = ctx;
 236         cmdinfo->result = le32_to_cpup(&cqe->result);
 237         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 238         queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
 239 }
 240
 241 /*
 242  * Called with local interrupts disabled and the q_lock held.  May not sleep.
 243  */
 244 static void *free_cmdid(struct nvme_queue *nvmeq, int cmdid,
 245                                                 nvme_completion_fn *fn)
 246 {
 247         void *ctx;
 248         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 249
 250         if (cmdid >= nvmeq->q_depth) {
 251                 *fn = special_completion;
 252                 return CMD_CTX_INVALID;
 253         }
 254         if (fn)
 255                 *fn = info[cmdid].fn;
 256         ctx = info[cmdid].ctx;
 257         info[cmdid].fn = special_completion;
 258         info[cmdid].ctx = CMD_CTX_COMPLETED;
 259         clear_bit(cmdid, nvmeq->cmdid_data);
 260         wake_up(&nvmeq->sq_full);
 261         return ctx;
 262 }
 263
 264 static void *cancel_cmdid(struct nvme_queue *nvmeq, int cmdid,
 265                                                 nvme_completion_fn *fn)
 266 {
 267         void *ctx;
 268         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
 269         if (fn)
 270                 *fn = info[cmdid].fn;
 271         ctx = info[cmdid].ctx;
 272         info[cmdid].fn = special_completion;
 273         info[cmdid].ctx = CMD_CTX_CANCELLED;
 274         return ctx;
 275 }
 276
 277 static struct nvme_queue *raw_nvmeq(struct nvme_dev *dev, int qid)
 278 {
 279         return rcu_dereference_raw(dev->queues[qid]);
 280 }
 281
 282 static struct nvme_queue *get_nvmeq(struct nvme_dev *dev) __acquires(RCU)
 283 {
 284         unsigned queue_id = get_cpu_var(*dev->io_queue);
 285         rcu_read_lock();
 286         return rcu_dereference(dev->queues[queue_id]);
 287 }
 288
 289 static void put_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
 290 {
 291         rcu_read_unlock();
 292         put_cpu_var(nvmeq->dev->io_queue);
 293 }
 294
 295 static struct nvme_queue *lock_nvmeq(struct nvme_dev *dev, int q_idx)
 296                                                         __acquires(RCU)
 297 {
 298         rcu_read_lock();
 299         return rcu_dereference(dev->queues[q_idx]);
 300 }
 301
 302 static void unlock_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
 303 {
 304         rcu_read_unlock();
 305 }
 306
 307 /**
 308  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 309  * @nvmeq: The queue to use
 310  * @cmd: The command to send
 311  *
 312  * Safe to use from interrupt context
 313  */
 314 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
 315 {
 316         unsigned long flags;
 317         u16 tail;
 318         spin_lock_irqsave(&nvmeq->q_lock, flags);
 319         if (nvmeq->q_suspended) {
 320                 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
 321                 return -EBUSY;
 322         }
 323         tail = nvmeq->sq_tail;
 324         memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
 325         if (++tail == nvmeq->q_depth)
 326                 tail = 0;
 327         writel(tail, nvmeq->q_db);
 328         nvmeq->sq_tail = tail;
 329         spin_unlock_irqrestore(&nvmeq->q_lock, flags);
 330
 331         return 0;
 332 }
 333
 334 static __le64 **iod_list(struct nvme_iod *iod)
 335 {
 336         return ((void *)iod) + iod->offset;
 337 }
 338
 339 /*
 340  * Will slightly overestimate the number of pages needed.  This is OK
 341  * as it only leads to a small amount of wasted memory for the lifetime of
 342  * the I/O.
 343  */
 344 static int nvme_npages(unsigned size)
 345 {
 346         unsigned nprps = DIV_ROUND_UP(size + PAGE_SIZE, PAGE_SIZE);
 347         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 348 }
 349
 350 static struct nvme_iod *
 351 nvme_alloc_iod(unsigned nseg, unsigned nbytes, gfp_t gfp)
 352 {
 353         struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
 354                                 sizeof(__le64 *) * nvme_npages(nbytes) +
 355                                 sizeof(struct scatterlist) * nseg, gfp);
 356
 357         if (iod) {
 358                 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
 359                 iod->npages = -1;
 360                 iod->length = nbytes;
 361                 iod->nents = 0;
 362                 iod->first_dma = 0ULL;
 363                 iod->start_time = jiffies;
 364         }
 365
 366         return iod;
 367 }
 368
 369 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
 370 {
 371         const int last_prp = PAGE_SIZE / 8 - 1;
 372         int i;
 373         __le64 **list = iod_list(iod);
 374         dma_addr_t prp_dma = iod->first_dma;
 375
 376         if (iod->npages == 0)
 377                 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
 378         for (i = 0; i < iod->npages; i++) {
 379                 __le64 *prp_list = list[i];
 380                 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
 381                 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
 382                 prp_dma = next_prp_dma;
 383         }
 384         kfree(iod);
 385 }
 386
 387 static void nvme_start_io_acct(struct bio *bio)
 388 {
 389         struct gendisk *disk = bio->bi_bdev->bd_disk;
 390         const int rw = bio_data_dir(bio);
 391         int cpu = part_stat_lock();
 392         part_round_stats(cpu, &disk->part0);
 393         part_stat_inc(cpu, &disk->part0, ios[rw]);
 394         part_stat_add(cpu, &disk->part0, sectors[rw], bio_sectors(bio));
 395         part_inc_in_flight(&disk->part0, rw);
 396         part_stat_unlock();
 397 }
 398
 399 static void nvme_end_io_acct(struct bio *bio, unsigned long start_time)
 400 {
 401         struct gendisk *disk = bio->bi_bdev->bd_disk;
 402         const int rw = bio_data_dir(bio);
 403         unsigned long duration = jiffies - start_time;
 404         int cpu = part_stat_lock();
 405         part_stat_add(cpu, &disk->part0, ticks[rw], duration);
 406         part_round_stats(cpu, &disk->part0);
 407         part_dec_in_flight(&disk->part0, rw);
 408         part_stat_unlock();
 409 }
 410
 411 static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
 412                                                 struct nvme_completion *cqe)
 413 {
 414         struct nvme_iod *iod = ctx;
 415         struct bio *bio = iod->private;
 416         u16 status = le16_to_cpup(&cqe->status) >> 1;
 417
 418         if (unlikely(status)) {
 419                 if (!(status & NVME_SC_DNR ||
 420                                 bio->bi_rw & REQ_FAILFAST_MASK) &&
 421                                 (jiffies - iod->start_time) < IOD_TIMEOUT) {
 422                         if (!waitqueue_active(&nvmeq->sq_full))
 423                                 add_wait_queue(&nvmeq->sq_full,
 424                                                         &nvmeq->sq_cong_wait);
 425                         list_add_tail(&iod->node, &nvmeq->iod_bio);
 426                         wake_up(&nvmeq->sq_full);
 427                         return;
 428                 }
 429         }
 430         if (iod->nents) {
 431                 dma_unmap_sg(nvmeq->q_dmadev, iod->sg, iod->nents,
 432                         bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 433                 nvme_end_io_acct(bio, iod->start_time);
 434         }
 435         nvme_free_iod(nvmeq->dev, iod);
 436         if (status)
 437                 bio_endio(bio, -EIO);
 438         else
 439                 bio_endio(bio, 0);
 440 }
 441
 442 /* length is in bytes.  gfp flags indicates whether we may sleep. */
 443 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod, int total_len,
 444                                                                 gfp_t gfp)
 445 {
 446         struct dma_pool *pool;
 447         int length = total_len;
 448         struct scatterlist *sg = iod->sg;
 449         int dma_len = sg_dma_len(sg);
 450         u64 dma_addr = sg_dma_address(sg);
 451         int offset = offset_in_page(dma_addr);
 452         __le64 *prp_list;
 453         __le64 **list = iod_list(iod);
 454         dma_addr_t prp_dma;
 455         int nprps, i;
 456
 457         length -= (PAGE_SIZE - offset);
 458         if (length <= 0)
 459                 return total_len;
 460
 461         dma_len -= (PAGE_SIZE - offset);
 462         if (dma_len) {
 463                 dma_addr += (PAGE_SIZE - offset);
 464         } else {
 465                 sg = sg_next(sg);
 466                 dma_addr = sg_dma_address(sg);
 467                 dma_len = sg_dma_len(sg);
 468         }
 469
 470         if (length <= PAGE_SIZE) {
 471                 iod->first_dma = dma_addr;
 472                 return total_len;
 473         }
 474
 475         nprps = DIV_ROUND_UP(length, PAGE_SIZE);
 476         if (nprps <= (256 / 8)) {
 477                 pool = dev->prp_small_pool;
 478                 iod->npages = 0;
 479         } else {
 480                 pool = dev->prp_page_pool;
 481                 iod->npages = 1;
 482         }
 483
 484         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 485         if (!prp_list) {
 486                 iod->first_dma = dma_addr;
 487                 iod->npages = -1;
 488                 return (total_len - length) + PAGE_SIZE;
 489         }
 490         list[0] = prp_list;
 491         iod->first_dma = prp_dma;
 492         i = 0;
 493         for (;;) {
 494                 if (i == PAGE_SIZE / 8) {
 495                         __le64 *old_prp_list = prp_list;
 496                         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 497                         if (!prp_list)
 498                                 return total_len - length;
 499                         list[iod->npages++] = prp_list;
 500                         prp_list[0] = old_prp_list[i - 1];
 501                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 502                         i = 1;
 503                 }
 504                 prp_list[i++] = cpu_to_le64(dma_addr);
 505                 dma_len -= PAGE_SIZE;
 506                 dma_addr += PAGE_SIZE;
 507                 length -= PAGE_SIZE;
 508                 if (length <= 0)
 509                         break;
 510                 if (dma_len > 0)
 511                         continue;
 512                 BUG_ON(dma_len < 0);
 513                 sg = sg_next(sg);
 514                 dma_addr = sg_dma_address(sg);
 515                 dma_len = sg_dma_len(sg);
 516         }
 517
 518         return total_len;
 519 }
 520
 521 static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
 522                                  int len)
 523 {
 524         struct bio *split = bio_split(bio, len >> 9, GFP_ATOMIC, NULL);
 525         if (!split)
 526                 return -ENOMEM;
 527
 528         bio_chain(split, bio);
 529
 530         if (!waitqueue_active(&nvmeq->sq_full))
 531                 add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
 532         bio_list_add(&nvmeq->sq_cong, split);
 533         bio_list_add(&nvmeq->sq_cong, bio);
 534         wake_up(&nvmeq->sq_full);
 535
 536         return 0;
 537 }
 538
 539 /* NVMe scatterlists require no holes in the virtual address */
 540 #define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2)   ((vec2)->bv_offset || \
 541                         (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
 542
 543 static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
 544                 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
 545 {
 546         struct bio_vec bvec, bvprv;
 547         struct bvec_iter iter;
 548         struct scatterlist *sg = NULL;
 549         int length = 0, nsegs = 0, split_len = bio->bi_iter.bi_size;
 550         int first = 1;
 551
 552         if (nvmeq->dev->stripe_size)
 553                 split_len = nvmeq->dev->stripe_size -
 554                         ((bio->bi_iter.bi_sector << 9) &
 555                          (nvmeq->dev->stripe_size - 1));
 556
 557         sg_init_table(iod->sg, psegs);
 558         bio_for_each_segment(bvec, bio, iter) {
 559                 if (!first && BIOVEC_PHYS_MERGEABLE(&bvprv, &bvec)) {
 560                         sg->length += bvec.bv_len;
 561                 } else {
 562                         if (!first && BIOVEC_NOT_VIRT_MERGEABLE(&bvprv, &bvec))
 563                                 return nvme_split_and_submit(bio, nvmeq,
 564                                                              length);
 565
 566                         sg = sg ? sg + 1 : iod->sg;
 567                         sg_set_page(sg, bvec.bv_page,
 568                                     bvec.bv_len, bvec.bv_offset);
 569                         nsegs++;
 570                 }
 571
 572                 if (split_len - length < bvec.bv_len)
 573                         return nvme_split_and_submit(bio, nvmeq, split_len);
 574                 length += bvec.bv_len;
 575                 bvprv = bvec;
 576                 first = 0;
 577         }
 578         iod->nents = nsegs;
 579         sg_mark_end(sg);
 580         if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
 581                 return -ENOMEM;
 582
 583         BUG_ON(length != bio->bi_iter.bi_size);
 584         return length;
 585 }
 586
 587 static int nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 588                 struct bio *bio, struct nvme_iod *iod, int cmdid)
 589 {
 590         struct nvme_dsm_range *range =
 591                                 (struct nvme_dsm_range *)iod_list(iod)[0];
 592         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 593
 594         range->cattr = cpu_to_le32(0);
 595         range->nlb = cpu_to_le32(bio->bi_iter.bi_size >> ns->lba_shift);
 596         range->slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
 597
 598         memset(cmnd, 0, sizeof(*cmnd));
 599         cmnd->dsm.opcode = nvme_cmd_dsm;
 600         cmnd->dsm.command_id = cmdid;
 601         cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
 602         cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
 603         cmnd->dsm.nr = 0;
 604         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 605
 606         if (++nvmeq->sq_tail == nvmeq->q_depth)
 607                 nvmeq->sq_tail = 0;
 608         writel(nvmeq->sq_tail, nvmeq->q_db);
 609
 610         return 0;
 611 }
 612
 613 static int nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 614                                                                 int cmdid)
 615 {
 616         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 617
 618         memset(cmnd, 0, sizeof(*cmnd));
 619         cmnd->common.opcode = nvme_cmd_flush;
 620         cmnd->common.command_id = cmdid;
 621         cmnd->common.nsid = cpu_to_le32(ns->ns_id);
 622
 623         if (++nvmeq->sq_tail == nvmeq->q_depth)
 624                 nvmeq->sq_tail = 0;
 625         writel(nvmeq->sq_tail, nvmeq->q_db);
 626
 627         return 0;
 628 }
 629
 630 int nvme_submit_flush_data(struct nvme_queue *nvmeq, struct nvme_ns *ns)
 631 {
 632         int cmdid = alloc_cmdid(nvmeq, (void *)CMD_CTX_FLUSH,
 633                                         special_completion, NVME_IO_TIMEOUT);
 634         if (unlikely(cmdid < 0))
 635                 return cmdid;
 636
 637         return nvme_submit_flush(nvmeq, ns, cmdid);
 638 }
 639
 640 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod)
 641 {
 642         struct bio *bio = iod->private;
 643         struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
 644         struct nvme_command *cmnd;
 645         int cmdid;
 646         u16 control;
 647         u32 dsmgmt;
 648
 649         cmdid = alloc_cmdid(nvmeq, iod, bio_completion, NVME_IO_TIMEOUT);
 650         if (unlikely(cmdid < 0))
 651                 return cmdid;
 652
 653         if (bio->bi_rw & REQ_DISCARD)
 654                 return nvme_submit_discard(nvmeq, ns, bio, iod, cmdid);
 655         if ((bio->bi_rw & REQ_FLUSH) && !iod->nents)
 656                 return nvme_submit_flush(nvmeq, ns, cmdid);
 657
 658         control = 0;
 659         if (bio->bi_rw & REQ_FUA)
 660                 control |= NVME_RW_FUA;
 661         if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 662                 control |= NVME_RW_LR;
 663
 664         dsmgmt = 0;
 665         if (bio->bi_rw & REQ_RAHEAD)
 666                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 667
 668         cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 669         memset(cmnd, 0, sizeof(*cmnd));
 670
 671         cmnd->rw.opcode = bio_data_dir(bio) ? nvme_cmd_write : nvme_cmd_read;
 672         cmnd->rw.command_id = cmdid;
 673         cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
 674         cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 675         cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
 676         cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, bio->bi_iter.bi_sector));
 677         cmnd->rw.length =
 678                 cpu_to_le16((bio->bi_iter.bi_size >> ns->lba_shift) - 1);
 679         cmnd->rw.control = cpu_to_le16(control);
 680         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 681
 682         if (++nvmeq->sq_tail == nvmeq->q_depth)
 683                 nvmeq->sq_tail = 0;
 684         writel(nvmeq->sq_tail, nvmeq->q_db);
 685
 686         return 0;
 687 }
 688
 689 /*
 690  * Called with local interrupts disabled and the q_lock held.  May not sleep.
 691  */
 692 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 693                                                                 struct bio *bio)
 694 {
 695         struct nvme_iod *iod;
 696         int psegs = bio_phys_segments(ns->queue, bio);
 697         int result;
 698
 699         if ((bio->bi_rw & REQ_FLUSH) && psegs) {
 700                 result = nvme_submit_flush_data(nvmeq, ns);
 701                 if (result)
 702                         return result;
 703         }
 704
 705         iod = nvme_alloc_iod(psegs, bio->bi_iter.bi_size, GFP_ATOMIC);
 706         if (!iod)
 707                 return -ENOMEM;
 708
 709         iod->private = bio;
 710         if (bio->bi_rw & REQ_DISCARD) {
 711                 void *range;
 712                 /*
 713                  * We reuse the small pool to allocate the 16-byte range here
 714                  * as it is not worth having a special pool for these or
 715                  * additional cases to handle freeing the iod.
 716                  */
 717                 range = dma_pool_alloc(nvmeq->dev->prp_small_pool,
 718                                                 GFP_ATOMIC,
 719                                                 &iod->first_dma);
 720                 if (!range) {
 721                         result = -ENOMEM;
 722                         goto free_iod;
 723                 }
 724                 iod_list(iod)[0] = (__le64 *)range;
 725                 iod->npages = 0;
 726         } else if (psegs) {
 727                 result = nvme_map_bio(nvmeq, iod, bio,
 728                         bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE,
 729                         psegs);
 730                 if (result <= 0)
 731                         goto free_iod;
 732                 if (nvme_setup_prps(nvmeq->dev, iod, result, GFP_ATOMIC) !=
 733                                                                 result) {
 734                         result = -ENOMEM;
 735                         goto free_iod;
 736                 }
 737                 nvme_start_io_acct(bio);
 738         }
 739         if (unlikely(nvme_submit_iod(nvmeq, iod))) {
 740                 if (!waitqueue_active(&nvmeq->sq_full))
 741                         add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
 742                 list_add_tail(&iod->node, &nvmeq->iod_bio);
 743         }
 744         return 0;
 745
 746  free_iod:
 747         nvme_free_iod(nvmeq->dev, iod);
 748         return result;
 749 }
 750
 751 static int nvme_process_cq(struct nvme_queue *nvmeq)
 752 {
 753         u16 head, phase;
 754
 755         head = nvmeq->cq_head;
 756         phase = nvmeq->cq_phase;
 757
 758         for (;;) {
 759                 void *ctx;
 760                 nvme_completion_fn fn;
 761                 struct nvme_completion cqe = nvmeq->cqes[head];
 762                 if ((le16_to_cpu(cqe.status) & 1) != phase)
 763                         break;
 764                 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
 765                 if (++head == nvmeq->q_depth) {
 766                         head = 0;
 767                         phase = !phase;
 768                 }
 769
 770                 ctx = free_cmdid(nvmeq, cqe.command_id, &fn);
 771                 fn(nvmeq, ctx, &cqe);
 772         }
 773
 774         /* If the controller ignores the cq head doorbell and continuously
 775          * writes to the queue, it is theoretically possible to wrap around
 776          * the queue twice and mistakenly return IRQ_NONE.  Linux only
 777          * requires that 0.1% of your interrupts are handled, so this isn't
 778          * a big problem.
 779          */
 780         if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
 781                 return 0;
 782
 783         writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 784         nvmeq->cq_head = head;
 785         nvmeq->cq_phase = phase;
 786
 787         nvmeq->cqe_seen = 1;
 788         return 1;
 789 }
 790
 791 static void nvme_make_request(struct request_queue *q, struct bio *bio)
 792 {
 793         struct nvme_ns *ns = q->queuedata;
 794         struct nvme_queue *nvmeq = get_nvmeq(ns->dev);
 795         int result = -EBUSY;
 796
 797         if (!nvmeq) {
 798                 put_nvmeq(NULL);
 799                 bio_endio(bio, -EIO);
 800                 return;
 801         }
 802
 803         spin_lock_irq(&nvmeq->q_lock);
 804         if (!nvmeq->q_suspended && bio_list_empty(&nvmeq->sq_cong))
 805                 result = nvme_submit_bio_queue(nvmeq, ns, bio);
 806         if (unlikely(result)) {
 807                 if (!waitqueue_active(&nvmeq->sq_full))
 808                         add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
 809                 bio_list_add(&nvmeq->sq_cong, bio);
 810         }
 811
 812         nvme_process_cq(nvmeq);
 813         spin_unlock_irq(&nvmeq->q_lock);
 814         put_nvmeq(nvmeq);
 815 }
 816
 817 static irqreturn_t nvme_irq(int irq, void *data)
 818 {
 819         irqreturn_t result;
 820         struct nvme_queue *nvmeq = data;
 821         spin_lock(&nvmeq->q_lock);
 822         nvme_process_cq(nvmeq);
 823         result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
 824         nvmeq->cqe_seen = 0;
 825         spin_unlock(&nvmeq->q_lock);
 826         return result;
 827 }
 828
 829 static irqreturn_t nvme_irq_check(int irq, void *data)
 830 {
 831         struct nvme_queue *nvmeq = data;
 832         struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
 833         if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
 834                 return IRQ_NONE;
 835         return IRQ_WAKE_THREAD;
 836 }
 837
 838 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
 839 {
 840         spin_lock_irq(&nvmeq->q_lock);
 841         cancel_cmdid(nvmeq, cmdid, NULL);
 842         spin_unlock_irq(&nvmeq->q_lock);
 843 }
 844
 845 struct sync_cmd_info {
 846         struct task_struct *task;
 847         u32 result;
 848         int status;
 849 };
 850
 851 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
 852                                                 struct nvme_completion *cqe)
 853 {
 854         struct sync_cmd_info *cmdinfo = ctx;
 855         cmdinfo->result = le32_to_cpup(&cqe->result);
 856         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 857         wake_up_process(cmdinfo->task);
 858 }
 859
 860 /*
 861  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 862  * if the result is positive, it's an NVM Express status code
 863  */
 864 static int nvme_submit_sync_cmd(struct nvme_dev *dev, int q_idx,
 865                                                 struct nvme_command *cmd,
 866                                                 u32 *result, unsigned timeout)
 867 {
 868         int cmdid, ret;
 869         struct sync_cmd_info cmdinfo;
 870         struct nvme_queue *nvmeq;
 871
 872         nvmeq = lock_nvmeq(dev, q_idx);
 873         if (!nvmeq) {
 874                 unlock_nvmeq(nvmeq);
 875                 return -ENODEV;
 876         }
 877
 878         cmdinfo.task = current;
 879         cmdinfo.status = -EINTR;
 880
 881         cmdid = alloc_cmdid(nvmeq, &cmdinfo, sync_completion, timeout);
 882         if (cmdid < 0) {
 883                 unlock_nvmeq(nvmeq);
 884                 return cmdid;
 885         }
 886         cmd->common.command_id = cmdid;
 887
 888         set_current_state(TASK_KILLABLE);
 889         ret = nvme_submit_cmd(nvmeq, cmd);
 890         if (ret) {
 891                 free_cmdid(nvmeq, cmdid, NULL);
 892                 unlock_nvmeq(nvmeq);
 893                 set_current_state(TASK_RUNNING);
 894                 return ret;
 895         }
 896         unlock_nvmeq(nvmeq);
 897         schedule_timeout(timeout);
 898
 899         if (cmdinfo.status == -EINTR) {
 900                 nvmeq = lock_nvmeq(dev, q_idx);
 901                 if (nvmeq)
 902                         nvme_abort_command(nvmeq, cmdid);
 903                 unlock_nvmeq(nvmeq);
 904                 return -EINTR;
 905         }
 906
 907         if (result)
 908                 *result = cmdinfo.result;
 909
 910         return cmdinfo.status;
 911 }
 912
 913 static int nvme_submit_async_cmd(struct nvme_queue *nvmeq,
 914                         struct nvme_command *cmd,
 915                         struct async_cmd_info *cmdinfo, unsigned timeout)
 916 {
 917         int cmdid;
 918
 919         cmdid = alloc_cmdid_killable(nvmeq, cmdinfo, async_completion, timeout);
 920         if (cmdid < 0)
 921                 return cmdid;
 922         cmdinfo->status = -EINTR;
 923         cmd->common.command_id = cmdid;
 924         return nvme_submit_cmd(nvmeq, cmd);
 925 }
 926
 927 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 928                                                                 u32 *result)
 929 {
 930         return nvme_submit_sync_cmd(dev, 0, cmd, result, ADMIN_TIMEOUT);
 931 }
 932
 933 int nvme_submit_io_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 934                                                                 u32 *result)
 935 {
 936         return nvme_submit_sync_cmd(dev, smp_processor_id() + 1, cmd, result,
 937                                                         NVME_IO_TIMEOUT);
 938 }
 939
 940 static int nvme_submit_admin_cmd_async(struct nvme_dev *dev,
 941                 struct nvme_command *cmd, struct async_cmd_info *cmdinfo)
 942 {
 943         return nvme_submit_async_cmd(raw_nvmeq(dev, 0), cmd, cmdinfo,
 944                                                                 ADMIN_TIMEOUT);
 945 }
 946
 947 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
 948 {
 949         int status;
 950         struct nvme_command c;
 951
 952         memset(&c, 0, sizeof(c));
 953         c.delete_queue.opcode = opcode;
 954         c.delete_queue.qid = cpu_to_le16(id);
 955
 956         status = nvme_submit_admin_cmd(dev, &c, NULL);
 957         if (status)
 958                 return -EIO;
 959         return 0;
 960 }
 961
 962 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
 963                                                 struct nvme_queue *nvmeq)
 964 {
 965         int status;
 966         struct nvme_command c;
 967         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
 968
 969         memset(&c, 0, sizeof(c));
 970         c.create_cq.opcode = nvme_admin_create_cq;
 971         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
 972         c.create_cq.cqid = cpu_to_le16(qid);
 973         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 974         c.create_cq.cq_flags = cpu_to_le16(flags);
 975         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
 976
 977         status = nvme_submit_admin_cmd(dev, &c, NULL);
 978         if (status)
 979                 return -EIO;
 980         return 0;
 981 }
 982
 983 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
 984                                                 struct nvme_queue *nvmeq)
 985 {
 986         int status;
 987         struct nvme_command c;
 988         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
 989
 990         memset(&c, 0, sizeof(c));
 991         c.create_sq.opcode = nvme_admin_create_sq;
 992         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
 993         c.create_sq.sqid = cpu_to_le16(qid);
 994         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 995         c.create_sq.sq_flags = cpu_to_le16(flags);
 996         c.create_sq.cqid = cpu_to_le16(qid);
 997
 998         status = nvme_submit_admin_cmd(dev, &c, NULL);
 999         if (status)
1000                 return -EIO;
1001         return 0;
1002 }
1003
1004 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1005 {
1006         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1007 }
1008
1009 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1010 {
1011         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1012 }
1013
1014 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
1015                                                         dma_addr_t dma_addr)
1016 {
1017         struct nvme_command c;
1018
1019         memset(&c, 0, sizeof(c));
1020         c.identify.opcode = nvme_admin_identify;
1021         c.identify.nsid = cpu_to_le32(nsid);
1022         c.identify.prp1 = cpu_to_le64(dma_addr);
1023         c.identify.cns = cpu_to_le32(cns);
1024
1025         return nvme_submit_admin_cmd(dev, &c, NULL);
1026 }
1027
1028 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
1029                                         dma_addr_t dma_addr, u32 *result)
1030 {
1031         struct nvme_command c;
1032
1033         memset(&c, 0, sizeof(c));
1034         c.features.opcode = nvme_admin_get_features;
1035         c.features.nsid = cpu_to_le32(nsid);
1036         c.features.prp1 = cpu_to_le64(dma_addr);
1037         c.features.fid = cpu_to_le32(fid);
1038
1039         return nvme_submit_admin_cmd(dev, &c, result);
1040 }
1041
1042 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
1043                                         dma_addr_t dma_addr, u32 *result)
1044 {
1045         struct nvme_command c;
1046
1047         memset(&c, 0, sizeof(c));
1048         c.features.opcode = nvme_admin_set_features;
1049         c.features.prp1 = cpu_to_le64(dma_addr);
1050         c.features.fid = cpu_to_le32(fid);
1051         c.features.dword11 = cpu_to_le32(dword11);
1052
1053         return nvme_submit_admin_cmd(dev, &c, result);
1054 }
1055
1056 /**
1057  * nvme_abort_cmd - Attempt aborting a command
1058  * @cmdid: Command id of a timed out IO
1059  * @queue: The queue with timed out IO
1060  *
1061  * Schedule controller reset if the command was already aborted once before and
1062  * still hasn't been returned to the driver, or if this is the admin queue.
1063  */
1064 static void nvme_abort_cmd(int cmdid, struct nvme_queue *nvmeq)
1065 {
1066         int a_cmdid;
1067         struct nvme_command cmd;
1068         struct nvme_dev *dev = nvmeq->dev;
1069         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1070         struct nvme_queue *adminq;
1071
1072         if (!nvmeq->qid || info[cmdid].aborted) {
1073                 if (work_busy(&dev->reset_work))
1074                         return;
1075                 list_del_init(&dev->node);
1076                 dev_warn(&dev->pci_dev->dev,
1077                         "I/O %d QID %d timeout, reset controller\n", cmdid,
1078                                                                 nvmeq->qid);
1079                 PREPARE_WORK(&dev->reset_work, nvme_reset_failed_dev);
1080                 queue_work(nvme_workq, &dev->reset_work);
1081                 return;
1082         }
1083
1084         if (!dev->abort_limit)
1085                 return;
1086
1087         adminq = rcu_dereference(dev->queues[0]);
1088         a_cmdid = alloc_cmdid(adminq, CMD_CTX_ABORT, special_completion,
1089                                                                 ADMIN_TIMEOUT);
1090         if (a_cmdid < 0)
1091                 return;
1092
1093         memset(&cmd, 0, sizeof(cmd));
1094         cmd.abort.opcode = nvme_admin_abort_cmd;
1095         cmd.abort.cid = cmdid;
1096         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1097         cmd.abort.command_id = a_cmdid;
1098
1099         --dev->abort_limit;
1100         info[cmdid].aborted = 1;
1101         info[cmdid].timeout = jiffies + ADMIN_TIMEOUT;
1102
1103         dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", cmdid,
1104                                                         nvmeq->qid);
1105         nvme_submit_cmd(adminq, &cmd);
1106 }
1107
1108 /**
1109  * nvme_cancel_ios - Cancel outstanding I/Os
1110  * @queue: The queue to cancel I/Os on
1111  * @timeout: True to only cancel I/Os which have timed out
1112  */
1113 static void nvme_cancel_ios(struct nvme_queue *nvmeq, bool timeout)
1114 {
1115         int depth = nvmeq->q_depth - 1;
1116         struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
1117         unsigned long now = jiffies;
1118         int cmdid;
1119
1120         for_each_set_bit(cmdid, nvmeq->cmdid_data, depth) {
1121                 void *ctx;
1122                 nvme_completion_fn fn;
1123                 static struct nvme_completion cqe = {
1124                         .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1125                 };
1126
1127                 if (timeout && !time_after(now, info[cmdid].timeout))
1128                         continue;
1129                 if (info[cmdid].ctx == CMD_CTX_CANCELLED)
1130                         continue;
1131                 if (timeout && nvmeq->dev->initialized) {
1132                         nvme_abort_cmd(cmdid, nvmeq);
1133                         continue;
1134                 }
1135                 dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n", cmdid,
1136                                                                 nvmeq->qid);
1137                 ctx = cancel_cmdid(nvmeq, cmdid, &fn);
1138                 fn(nvmeq, ctx, &cqe);
1139         }
1140 }
1141
1142 static void nvme_free_queue(struct rcu_head *r)
1143 {
1144         struct nvme_queue *nvmeq = container_of(r, struct nvme_queue, r_head);
1145
1146         spin_lock_irq(&nvmeq->q_lock);
1147         while (bio_list_peek(&nvmeq->sq_cong)) {
1148                 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1149                 bio_endio(bio, -EIO);
1150         }
1151         while (!list_empty(&nvmeq->iod_bio)) {
1152                 static struct nvme_completion cqe = {
1153                         .status = cpu_to_le16(
1154                                 (NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1),
1155                 };
1156                 struct nvme_iod *iod = list_first_entry(&nvmeq->iod_bio,
1157                                                         struct nvme_iod,
1158                                                         node);
1159                 list_del(&iod->node);
1160                 bio_completion(nvmeq, iod, &cqe);
1161         }
1162         spin_unlock_irq(&nvmeq->q_lock);
1163
1164         dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1165                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1166         dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1167                                         nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1168         if (nvmeq->qid)
1169                 free_cpumask_var(nvmeq->cpu_mask);
1170         kfree(nvmeq);
1171 }
1172
1173 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1174 {
1175         int i;
1176
1177         for (i = dev->queue_count - 1; i >= lowest; i--) {
1178                 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
1179                 rcu_assign_pointer(dev->queues[i], NULL);
1180                 call_rcu(&nvmeq->r_head, nvme_free_queue);
1181                 dev->queue_count--;
1182         }
1183 }
1184
1185 /**
1186  * nvme_suspend_queue - put queue into suspended state
1187  * @nvmeq - queue to suspend
1188  *
1189  * Returns 1 if already suspended, 0 otherwise.
1190  */
1191 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1192 {
1193         int vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1194
1195         spin_lock_irq(&nvmeq->q_lock);
1196         if (nvmeq->q_suspended) {
1197                 spin_unlock_irq(&nvmeq->q_lock);
1198                 return 1;
1199         }
1200         nvmeq->q_suspended = 1;
1201         nvmeq->dev->online_queues--;
1202         spin_unlock_irq(&nvmeq->q_lock);
1203
1204         irq_set_affinity_hint(vector, NULL);
1205         free_irq(vector, nvmeq);
1206
1207         return 0;
1208 }
1209
1210 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1211 {
1212         spin_lock_irq(&nvmeq->q_lock);
1213         nvme_process_cq(nvmeq);
1214         nvme_cancel_ios(nvmeq, false);
1215         spin_unlock_irq(&nvmeq->q_lock);
1216 }
1217
1218 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1219 {
1220         struct nvme_queue *nvmeq = raw_nvmeq(dev, qid);
1221
1222         if (!nvmeq)
1223                 return;
1224         if (nvme_suspend_queue(nvmeq))
1225                 return;
1226
1227         /* Don't tell the adapter to delete the admin queue.
1228          * Don't tell a removed adapter to delete IO queues. */
1229         if (qid && readl(&dev->bar->csts) != -1) {
1230                 adapter_delete_sq(dev, qid);
1231                 adapter_delete_cq(dev, qid);
1232         }
1233         nvme_clear_queue(nvmeq);
1234 }
1235
1236 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1237                                                         int depth, int vector)
1238 {
1239         struct device *dmadev = &dev->pci_dev->dev;
1240         unsigned extra = nvme_queue_extra(depth);
1241         struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
1242         if (!nvmeq)
1243                 return NULL;
1244
1245         nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
1246                                         &nvmeq->cq_dma_addr, GFP_KERNEL);
1247         if (!nvmeq->cqes)
1248                 goto free_nvmeq;
1249         memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
1250
1251         nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1252                                         &nvmeq->sq_dma_addr, GFP_KERNEL);
1253         if (!nvmeq->sq_cmds)
1254                 goto free_cqdma;
1255
1256         if (qid && !zalloc_cpumask_var(&nvmeq->cpu_mask, GFP_KERNEL))
1257                 goto free_sqdma;
1258
1259         nvmeq->q_dmadev = dmadev;
1260         nvmeq->dev = dev;
1261         snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1262                         dev->instance, qid);
1263         spin_lock_init(&nvmeq->q_lock);
1264         nvmeq->cq_head = 0;
1265         nvmeq->cq_phase = 1;
1266         init_waitqueue_head(&nvmeq->sq_full);
1267         init_waitqueue_entry(&nvmeq->sq_cong_wait, nvme_thread);
1268         bio_list_init(&nvmeq->sq_cong);
1269         INIT_LIST_HEAD(&nvmeq->iod_bio);
1270         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1271         nvmeq->q_depth = depth;
1272         nvmeq->cq_vector = vector;
1273         nvmeq->qid = qid;
1274         nvmeq->q_suspended = 1;
1275         dev->queue_count++;
1276         rcu_assign_pointer(dev->queues[qid], nvmeq);
1277
1278         return nvmeq;
1279
1280  free_sqdma:
1281         dma_free_coherent(dmadev, SQ_SIZE(depth), (void *)nvmeq->sq_cmds,
1282                                                         nvmeq->sq_dma_addr);
1283  free_cqdma:
1284         dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1285                                                         nvmeq->cq_dma_addr);
1286  free_nvmeq:
1287         kfree(nvmeq);
1288         return NULL;
1289 }
1290
1291 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1292                                                         const char *name)
1293 {
1294         if (use_threaded_interrupts)
1295                 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1296                                         nvme_irq_check, nvme_irq, IRQF_SHARED,
1297                                         name, nvmeq);
1298         return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1299                                 IRQF_SHARED, name, nvmeq);
1300 }
1301
1302 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1303 {
1304         struct nvme_dev *dev = nvmeq->dev;
1305         unsigned extra = nvme_queue_extra(nvmeq->q_depth);
1306
1307         nvmeq->sq_tail = 0;
1308         nvmeq->cq_head = 0;
1309         nvmeq->cq_phase = 1;
1310         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1311         memset(nvmeq->cmdid_data, 0, extra);
1312         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1313         nvme_cancel_ios(nvmeq, false);
1314         nvmeq->q_suspended = 0;
1315         dev->online_queues++;
1316 }
1317
1318 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1319 {
1320         struct nvme_dev *dev = nvmeq->dev;
1321         int result;
1322
1323         result = adapter_alloc_cq(dev, qid, nvmeq);
1324         if (result < 0)
1325                 return result;
1326
1327         result = adapter_alloc_sq(dev, qid, nvmeq);
1328         if (result < 0)
1329                 goto release_cq;
1330
1331         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1332         if (result < 0)
1333                 goto release_sq;
1334
1335         spin_lock_irq(&nvmeq->q_lock);
1336         nvme_init_queue(nvmeq, qid);
1337         spin_unlock_irq(&nvmeq->q_lock);
1338
1339         return result;
1340
1341  release_sq:
1342         adapter_delete_sq(dev, qid);
1343  release_cq:
1344         adapter_delete_cq(dev, qid);
1345         return result;
1346 }
1347
1348 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1349 {
1350         unsigned long timeout;
1351         u32 bit = enabled ? NVME_CSTS_RDY : 0;
1352
1353         timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1354
1355         while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1356                 msleep(100);
1357                 if (fatal_signal_pending(current))
1358                         return -EINTR;
1359                 if (time_after(jiffies, timeout)) {
1360                         dev_err(&dev->pci_dev->dev,
1361                                 "Device not ready; aborting initialisation\n");
1362                         return -ENODEV;
1363                 }
1364         }
1365
1366         return 0;
1367 }
1368
1369 /*
1370  * If the device has been passed off to us in an enabled state, just clear
1371  * the enabled bit.  The spec says we should set the 'shutdown notification
1372  * bits', but doing so may cause the device to complete commands to the
1373  * admin queue ... and we don't know what memory that might be pointing at!
1374  */
1375 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1376 {
1377         u32 cc = readl(&dev->bar->cc);
1378
1379         if (cc & NVME_CC_ENABLE)
1380                 writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
1381         return nvme_wait_ready(dev, cap, false);
1382 }
1383
1384 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1385 {
1386         return nvme_wait_ready(dev, cap, true);
1387 }
1388
1389 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1390 {
1391         unsigned long timeout;
1392         u32 cc;
1393
1394         cc = (readl(&dev->bar->cc) & ~NVME_CC_SHN_MASK) | NVME_CC_SHN_NORMAL;
1395         writel(cc, &dev->bar->cc);
1396
1397         timeout = 2 * HZ + jiffies;
1398         while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1399                                                         NVME_CSTS_SHST_CMPLT) {
1400                 msleep(100);
1401                 if (fatal_signal_pending(current))
1402                         return -EINTR;
1403                 if (time_after(jiffies, timeout)) {
1404                         dev_err(&dev->pci_dev->dev,
1405                                 "Device shutdown incomplete; abort shutdown\n");
1406                         return -ENODEV;
1407                 }
1408         }
1409
1410         return 0;
1411 }
1412
1413 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1414 {
1415         int result;
1416         u32 aqa;
1417         u64 cap = readq(&dev->bar->cap);
1418         struct nvme_queue *nvmeq;
1419
1420         result = nvme_disable_ctrl(dev, cap);
1421         if (result < 0)
1422                 return result;
1423
1424         nvmeq = raw_nvmeq(dev, 0);
1425         if (!nvmeq) {
1426                 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
1427                 if (!nvmeq)
1428                         return -ENOMEM;
1429         }
1430
1431         aqa = nvmeq->q_depth - 1;
1432         aqa |= aqa << 16;
1433
1434         dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
1435         dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
1436         dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1437         dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1438
1439         writel(aqa, &dev->bar->aqa);
1440         writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1441         writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1442         writel(dev->ctrl_config, &dev->bar->cc);
1443
1444         result = nvme_enable_ctrl(dev, cap);
1445         if (result)
1446                 return result;
1447
1448         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1449         if (result)
1450                 return result;
1451
1452         spin_lock_irq(&nvmeq->q_lock);
1453         nvme_init_queue(nvmeq, 0);
1454         spin_unlock_irq(&nvmeq->q_lock);
1455         return result;
1456 }
1457
1458 struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1459                                 unsigned long addr, unsigned length)
1460 {
1461         int i, err, count, nents, offset;
1462         struct scatterlist *sg;
1463         struct page **pages;
1464         struct nvme_iod *iod;
1465
1466         if (addr & 3)
1467                 return ERR_PTR(-EINVAL);
1468         if (!length || length > INT_MAX - PAGE_SIZE)
1469                 return ERR_PTR(-EINVAL);
1470
1471         offset = offset_in_page(addr);
1472         count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1473         pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1474         if (!pages)
1475                 return ERR_PTR(-ENOMEM);
1476
1477         err = get_user_pages_fast(addr, count, 1, pages);
1478         if (err < count) {
1479                 count = err;
1480                 err = -EFAULT;
1481                 goto put_pages;
1482         }
1483
1484         iod = nvme_alloc_iod(count, length, GFP_KERNEL);
1485         sg = iod->sg;
1486         sg_init_table(sg, count);
1487         for (i = 0; i < count; i++) {
1488                 sg_set_page(&sg[i], pages[i],
1489                             min_t(unsigned, length, PAGE_SIZE - offset),
1490                             offset);
1491                 length -= (PAGE_SIZE - offset);
1492                 offset = 0;
1493         }
1494         sg_mark_end(&sg[i - 1]);
1495         iod->nents = count;
1496
1497         err = -ENOMEM;
1498         nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1499                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1500         if (!nents)
1501                 goto free_iod;
1502
1503         kfree(pages);
1504         return iod;
1505
1506  free_iod:
1507         kfree(iod);
1508  put_pages:
1509         for (i = 0; i < count; i++)
1510                 put_page(pages[i]);
1511         kfree(pages);
1512         return ERR_PTR(err);
1513 }
1514
1515 void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1516                         struct nvme_iod *iod)
1517 {
1518         int i;
1519
1520         dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1521                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1522
1523         for (i = 0; i < iod->nents; i++)
1524                 put_page(sg_page(&iod->sg[i]));
1525 }
1526
1527 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1528 {
1529         struct nvme_dev *dev = ns->dev;
1530         struct nvme_user_io io;
1531         struct nvme_command c;
1532         unsigned length, meta_len;
1533         int status, i;
1534         struct nvme_iod *iod, *meta_iod = NULL;
1535         dma_addr_t meta_dma_addr;
1536         void *meta, *uninitialized_var(meta_mem);
1537
1538         if (copy_from_user(&io, uio, sizeof(io)))
1539                 return -EFAULT;
1540         length = (io.nblocks + 1) << ns->lba_shift;
1541         meta_len = (io.nblocks + 1) * ns->ms;
1542
1543         if (meta_len && ((io.metadata & 3) || !io.metadata))
1544                 return -EINVAL;
1545
1546         switch (io.opcode) {
1547         case nvme_cmd_write:
1548         case nvme_cmd_read:
1549         case nvme_cmd_compare:
1550                 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1551                 break;
1552         default:
1553                 return -EINVAL;
1554         }
1555
1556         if (IS_ERR(iod))
1557                 return PTR_ERR(iod);
1558
1559         memset(&c, 0, sizeof(c));
1560         c.rw.opcode = io.opcode;
1561         c.rw.flags = io.flags;
1562         c.rw.nsid = cpu_to_le32(ns->ns_id);
1563         c.rw.slba = cpu_to_le64(io.slba);
1564         c.rw.length = cpu_to_le16(io.nblocks);
1565         c.rw.control = cpu_to_le16(io.control);
1566         c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1567         c.rw.reftag = cpu_to_le32(io.reftag);
1568         c.rw.apptag = cpu_to_le16(io.apptag);
1569         c.rw.appmask = cpu_to_le16(io.appmask);
1570
1571         if (meta_len) {
1572                 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1573                                                                 meta_len);
1574                 if (IS_ERR(meta_iod)) {
1575                         status = PTR_ERR(meta_iod);
1576                         meta_iod = NULL;
1577                         goto unmap;
1578                 }
1579
1580                 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1581                                                 &meta_dma_addr, GFP_KERNEL);
1582                 if (!meta_mem) {
1583                         status = -ENOMEM;
1584                         goto unmap;
1585                 }
1586
1587                 if (io.opcode & 1) {
1588                         int meta_offset = 0;
1589
1590                         for (i = 0; i < meta_iod->nents; i++) {
1591                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1592                                                 meta_iod->sg[i].offset;
1593                                 memcpy(meta_mem + meta_offset, meta,
1594                                                 meta_iod->sg[i].length);
1595                                 kunmap_atomic(meta);
1596                                 meta_offset += meta_iod->sg[i].length;
1597                         }
1598                 }
1599
1600                 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1601         }
1602
1603         length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1604         c.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1605         c.rw.prp2 = cpu_to_le64(iod->first_dma);
1606
1607         if (length != (io.nblocks + 1) << ns->lba_shift)
1608                 status = -ENOMEM;
1609         else
1610                 status = nvme_submit_io_cmd(dev, &c, NULL);
1611
1612         if (meta_len) {
1613                 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1614                         int meta_offset = 0;
1615
1616                         for (i = 0; i < meta_iod->nents; i++) {
1617                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1618                                                 meta_iod->sg[i].offset;
1619                                 memcpy(meta, meta_mem + meta_offset,
1620                                                 meta_iod->sg[i].length);
1621                                 kunmap_atomic(meta);
1622                                 meta_offset += meta_iod->sg[i].length;
1623                         }
1624                 }
1625
1626                 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1627                                                                 meta_dma_addr);
1628         }
1629
1630  unmap:
1631         nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1632         nvme_free_iod(dev, iod);
1633
1634         if (meta_iod) {
1635                 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1636                 nvme_free_iod(dev, meta_iod);
1637         }
1638
1639         return status;
1640 }
1641
1642 static int nvme_user_admin_cmd(struct nvme_dev *dev,
1643                                         struct nvme_admin_cmd __user *ucmd)
1644 {
1645         struct nvme_admin_cmd cmd;
1646         struct nvme_command c;
1647         int status, length;
1648         struct nvme_iod *uninitialized_var(iod);
1649         unsigned timeout;
1650
1651         if (!capable(CAP_SYS_ADMIN))
1652                 return -EACCES;
1653         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1654                 return -EFAULT;
1655
1656         memset(&c, 0, sizeof(c));
1657         c.common.opcode = cmd.opcode;
1658         c.common.flags = cmd.flags;
1659         c.common.nsid = cpu_to_le32(cmd.nsid);
1660         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1661         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1662         c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1663         c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1664         c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1665         c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1666         c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1667         c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1668
1669         length = cmd.data_len;
1670         if (cmd.data_len) {
1671                 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1672                                                                 length);
1673                 if (IS_ERR(iod))
1674                         return PTR_ERR(iod);
1675                 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1676                 c.common.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1677                 c.common.prp2 = cpu_to_le64(iod->first_dma);
1678         }
1679
1680         timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1681                                                                 ADMIN_TIMEOUT;
1682         if (length != cmd.data_len)
1683                 status = -ENOMEM;
1684         else
1685                 status = nvme_submit_sync_cmd(dev, 0, &c, &cmd.result, timeout);
1686
1687         if (cmd.data_len) {
1688                 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1689                 nvme_free_iod(dev, iod);
1690         }
1691
1692         if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1693                                                         sizeof(cmd.result)))
1694                 status = -EFAULT;
1695
1696         return status;
1697 }
1698
1699 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1700                                                         unsigned long arg)
1701 {
1702         struct nvme_ns *ns = bdev->bd_disk->private_data;
1703
1704         switch (cmd) {
1705         case NVME_IOCTL_ID:
1706                 force_successful_syscall_return();
1707                 return ns->ns_id;
1708         case NVME_IOCTL_ADMIN_CMD:
1709                 return nvme_user_admin_cmd(ns->dev, (void __user *)arg);
1710         case NVME_IOCTL_SUBMIT_IO:
1711                 return nvme_submit_io(ns, (void __user *)arg);
1712         case SG_GET_VERSION_NUM:
1713                 return nvme_sg_get_version_num((void __user *)arg);
1714         case SG_IO:
1715                 return nvme_sg_io(ns, (void __user *)arg);
1716         default:
1717                 return -ENOTTY;
1718         }
1719 }
1720
1721 #ifdef CONFIG_COMPAT
1722 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1723                                         unsigned int cmd, unsigned long arg)
1724 {
1725         struct nvme_ns *ns = bdev->bd_disk->private_data;
1726
1727         switch (cmd) {
1728         case SG_IO:
1729                 return nvme_sg_io32(ns, arg);
1730         }
1731         return nvme_ioctl(bdev, mode, cmd, arg);
1732 }
1733 #else
1734 #define nvme_compat_ioctl       NULL
1735 #endif
1736
1737 static int nvme_open(struct block_device *bdev, fmode_t mode)
1738 {
1739         struct nvme_ns *ns = bdev->bd_disk->private_data;
1740         struct nvme_dev *dev = ns->dev;
1741
1742         kref_get(&dev->kref);
1743         return 0;
1744 }
1745
1746 static void nvme_free_dev(struct kref *kref);
1747
1748 static void nvme_release(struct gendisk *disk, fmode_t mode)
1749 {
1750         struct nvme_ns *ns = disk->private_data;
1751         struct nvme_dev *dev = ns->dev;
1752
1753         kref_put(&dev->kref, nvme_free_dev);
1754 }
1755
1756 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1757 {
1758         /* some standard values */
1759         geo->heads = 1 << 6;
1760         geo->sectors = 1 << 5;
1761         geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1762         return 0;
1763 }
1764
1765 static const struct block_device_operations nvme_fops = {
1766         .owner          = THIS_MODULE,
1767         .ioctl          = nvme_ioctl,
1768         .compat_ioctl   = nvme_compat_ioctl,
1769         .open           = nvme_open,
1770         .release        = nvme_release,
1771         .getgeo         = nvme_getgeo,
1772 };
1773
1774 static void nvme_resubmit_iods(struct nvme_queue *nvmeq)
1775 {
1776         struct nvme_iod *iod, *next;
1777
1778         list_for_each_entry_safe(iod, next, &nvmeq->iod_bio, node) {
1779                 if (unlikely(nvme_submit_iod(nvmeq, iod)))
1780                         break;
1781                 list_del(&iod->node);
1782                 if (bio_list_empty(&nvmeq->sq_cong) &&
1783                                                 list_empty(&nvmeq->iod_bio))
1784                         remove_wait_queue(&nvmeq->sq_full,
1785                                                 &nvmeq->sq_cong_wait);
1786         }
1787 }
1788
1789 static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
1790 {
1791         while (bio_list_peek(&nvmeq->sq_cong)) {
1792                 struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
1793                 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
1794
1795                 if (bio_list_empty(&nvmeq->sq_cong) &&
1796                                                 list_empty(&nvmeq->iod_bio))
1797                         remove_wait_queue(&nvmeq->sq_full,
1798                                                         &nvmeq->sq_cong_wait);
1799                 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
1800                         if (!waitqueue_active(&nvmeq->sq_full))
1801                                 add_wait_queue(&nvmeq->sq_full,
1802                                                         &nvmeq->sq_cong_wait);
1803                         bio_list_add_head(&nvmeq->sq_cong, bio);
1804                         break;
1805                 }
1806         }
1807 }
1808
1809 static int nvme_kthread(void *data)
1810 {
1811         struct nvme_dev *dev, *next;
1812
1813         while (!kthread_should_stop()) {
1814                 set_current_state(TASK_INTERRUPTIBLE);
1815                 spin_lock(&dev_list_lock);
1816                 list_for_each_entry_safe(dev, next, &dev_list, node) {
1817                         int i;
1818                         if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1819                                                         dev->initialized) {
1820                                 if (work_busy(&dev->reset_work))
1821                                         continue;
1822                                 list_del_init(&dev->node);
1823                                 dev_warn(&dev->pci_dev->dev,
1824                                         "Failed status, reset controller\n");
1825                                 PREPARE_WORK(&dev->reset_work,
1826                                                         nvme_reset_failed_dev);
1827                                 queue_work(nvme_workq, &dev->reset_work);
1828                                 continue;
1829                         }
1830                         rcu_read_lock();
1831                         for (i = 0; i < dev->queue_count; i++) {
1832                                 struct nvme_queue *nvmeq =
1833                                                 rcu_dereference(dev->queues[i]);
1834                                 if (!nvmeq)
1835                                         continue;
1836                                 spin_lock_irq(&nvmeq->q_lock);
1837                                 if (nvmeq->q_suspended)
1838                                         goto unlock;
1839                                 nvme_process_cq(nvmeq);
1840                                 nvme_cancel_ios(nvmeq, true);
1841                                 nvme_resubmit_bios(nvmeq);
1842                                 nvme_resubmit_iods(nvmeq);
1843  unlock:
1844                                 spin_unlock_irq(&nvmeq->q_lock);
1845                         }
1846                         rcu_read_unlock();
1847                 }
1848                 spin_unlock(&dev_list_lock);
1849                 schedule_timeout(round_jiffies_relative(HZ));
1850         }
1851         return 0;
1852 }
1853
1854 static void nvme_config_discard(struct nvme_ns *ns)
1855 {
1856         u32 logical_block_size = queue_logical_block_size(ns->queue);
1857         ns->queue->limits.discard_zeroes_data = 0;
1858         ns->queue->limits.discard_alignment = logical_block_size;
1859         ns->queue->limits.discard_granularity = logical_block_size;
1860         ns->queue->limits.max_discard_sectors = 0xffffffff;
1861         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1862 }
1863
1864 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1865                         struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1866 {
1867         struct nvme_ns *ns;
1868         struct gendisk *disk;
1869         int lbaf;
1870
1871         if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1872                 return NULL;
1873
1874         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1875         if (!ns)
1876                 return NULL;
1877         ns->queue = blk_alloc_queue(GFP_KERNEL);
1878         if (!ns->queue)
1879                 goto out_free_ns;
1880         ns->queue->queue_flags = QUEUE_FLAG_DEFAULT;
1881         queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1882         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1883         blk_queue_make_request(ns->queue, nvme_make_request);
1884         ns->dev = dev;
1885         ns->queue->queuedata = ns;
1886
1887         disk = alloc_disk(0);
1888         if (!disk)
1889                 goto out_free_queue;
1890         ns->ns_id = nsid;
1891         ns->disk = disk;
1892         lbaf = id->flbas & 0xf;
1893         ns->lba_shift = id->lbaf[lbaf].ds;
1894         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1895         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1896         if (dev->max_hw_sectors)
1897                 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
1898
1899         disk->major = nvme_major;
1900         disk->first_minor = 0;
1901         disk->fops = &nvme_fops;
1902         disk->private_data = ns;
1903         disk->queue = ns->queue;
1904         disk->driverfs_dev = &dev->pci_dev->dev;
1905         disk->flags = GENHD_FL_EXT_DEVT;
1906         sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
1907         set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1908
1909         if (dev->oncs & NVME_CTRL_ONCS_DSM)
1910                 nvme_config_discard(ns);
1911
1912         return ns;
1913
1914  out_free_queue:
1915         blk_cleanup_queue(ns->queue);
1916  out_free_ns:
1917         kfree(ns);
1918         return NULL;
1919 }
1920
1921 static int nvme_find_closest_node(int node)
1922 {
1923         int n, val, min_val = INT_MAX, best_node = node;
1924
1925         for_each_online_node(n) {
1926                 if (n == node)
1927                         continue;
1928                 val = node_distance(node, n);
1929                 if (val < min_val) {
1930                         min_val = val;
1931                         best_node = n;
1932                 }
1933         }
1934         return best_node;
1935 }
1936
1937 static void nvme_set_queue_cpus(cpumask_t *qmask, struct nvme_queue *nvmeq,
1938                                                                 int count)
1939 {
1940         int cpu;
1941         for_each_cpu(cpu, qmask) {
1942                 if (cpumask_weight(nvmeq->cpu_mask) >= count)
1943                         break;
1944                 if (!cpumask_test_and_set_cpu(cpu, nvmeq->cpu_mask))
1945                         *per_cpu_ptr(nvmeq->dev->io_queue, cpu) = nvmeq->qid;
1946         }
1947 }
1948
1949 static void nvme_add_cpus(cpumask_t *mask, const cpumask_t *unassigned_cpus,
1950         const cpumask_t *new_mask, struct nvme_queue *nvmeq, int cpus_per_queue)
1951 {
1952         int next_cpu;
1953         for_each_cpu(next_cpu, new_mask) {
1954                 cpumask_or(mask, mask, get_cpu_mask(next_cpu));
1955                 cpumask_or(mask, mask, topology_thread_cpumask(next_cpu));
1956                 cpumask_and(mask, mask, unassigned_cpus);
1957                 nvme_set_queue_cpus(mask, nvmeq, cpus_per_queue);
1958         }
1959 }
1960
1961 static void nvme_create_io_queues(struct nvme_dev *dev)
1962 {
1963         unsigned i, max;
1964
1965         max = min(dev->max_qid, num_online_cpus());
1966         for (i = dev->queue_count; i <= max; i++)
1967                 if (!nvme_alloc_queue(dev, i, dev->q_depth, i - 1))
1968                         break;
1969
1970         max = min(dev->queue_count - 1, num_online_cpus());
1971         for (i = dev->online_queues; i <= max; i++)
1972                 if (nvme_create_queue(raw_nvmeq(dev, i), i))
1973                         break;
1974 }
1975
1976 /*
1977  * If there are fewer queues than online cpus, this will try to optimally
1978  * assign a queue to multiple cpus by grouping cpus that are "close" together:
1979  * thread siblings, core, socket, closest node, then whatever else is
1980  * available.
1981  */
1982 static void nvme_assign_io_queues(struct nvme_dev *dev)
1983 {
1984         unsigned cpu, cpus_per_queue, queues, remainder, i;
1985         cpumask_var_t unassigned_cpus;
1986
1987         nvme_create_io_queues(dev);
1988
1989         queues = min(dev->online_queues - 1, num_online_cpus());
1990         if (!queues)
1991                 return;
1992
1993         cpus_per_queue = num_online_cpus() / queues;
1994         remainder = queues - (num_online_cpus() - queues * cpus_per_queue);
1995
1996         if (!alloc_cpumask_var(&unassigned_cpus, GFP_KERNEL))
1997                 return;
1998
1999         cpumask_copy(unassigned_cpus, cpu_online_mask);
2000         cpu = cpumask_first(unassigned_cpus);
2001         for (i = 1; i <= queues; i++) {
2002                 struct nvme_queue *nvmeq = lock_nvmeq(dev, i);
2003                 cpumask_t mask;
2004
2005                 cpumask_clear(nvmeq->cpu_mask);
2006                 if (!cpumask_weight(unassigned_cpus)) {
2007                         unlock_nvmeq(nvmeq);
2008                         break;
2009                 }
2010
2011                 mask = *get_cpu_mask(cpu);
2012                 nvme_set_queue_cpus(&mask, nvmeq, cpus_per_queue);
2013                 if (cpus_weight(mask) < cpus_per_queue)
2014                         nvme_add_cpus(&mask, unassigned_cpus,
2015                                 topology_thread_cpumask(cpu),
2016                                 nvmeq, cpus_per_queue);
2017                 if (cpus_weight(mask) < cpus_per_queue)
2018                         nvme_add_cpus(&mask, unassigned_cpus,
2019                                 topology_core_cpumask(cpu),
2020                                 nvmeq, cpus_per_queue);
2021                 if (cpus_weight(mask) < cpus_per_queue)
2022                         nvme_add_cpus(&mask, unassigned_cpus,
2023                                 cpumask_of_node(cpu_to_node(cpu)),
2024                                 nvmeq, cpus_per_queue);
2025                 if (cpus_weight(mask) < cpus_per_queue)
2026                         nvme_add_cpus(&mask, unassigned_cpus,
2027                                 cpumask_of_node(
2028                                         nvme_find_closest_node(
2029                                                 cpu_to_node(cpu))),
2030                                 nvmeq, cpus_per_queue);
2031                 if (cpus_weight(mask) < cpus_per_queue)
2032                         nvme_add_cpus(&mask, unassigned_cpus,
2033                                 unassigned_cpus,
2034                                 nvmeq, cpus_per_queue);
2035
2036                 WARN(cpumask_weight(nvmeq->cpu_mask) != cpus_per_queue,
2037                         "nvme%d qid:%d mis-matched queue-to-cpu assignment\n",
2038                         dev->instance, i);
2039
2040                 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2041                                                         nvmeq->cpu_mask);
2042                 cpumask_andnot(unassigned_cpus, unassigned_cpus,
2043                                                 nvmeq->cpu_mask);
2044                 cpu = cpumask_next(cpu, unassigned_cpus);
2045                 if (remainder && !--remainder)
2046                         cpus_per_queue++;
2047                 unlock_nvmeq(nvmeq);
2048         }
2049         WARN(cpumask_weight(unassigned_cpus), "nvme%d unassigned online cpus\n",
2050                                                                 dev->instance);
2051         i = 0;
2052         cpumask_andnot(unassigned_cpus, cpu_possible_mask, cpu_online_mask);
2053         for_each_cpu(cpu, unassigned_cpus)
2054                 *per_cpu_ptr(dev->io_queue, cpu) = (i++ % queues) + 1;
2055         free_cpumask_var(unassigned_cpus);
2056 }
2057
2058 static int set_queue_count(struct nvme_dev *dev, int count)
2059 {
2060         int status;
2061         u32 result;
2062         u32 q_count = (count - 1) | ((count - 1) << 16);
2063
2064         status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
2065                                                                 &result);
2066         if (status)
2067                 return status < 0 ? -EIO : -EBUSY;
2068         return min(result & 0xffff, result >> 16) + 1;
2069 }
2070
2071 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
2072 {
2073         return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
2074 }
2075
2076 static int nvme_cpu_notify(struct notifier_block *self,
2077                                 unsigned long action, void *hcpu)
2078 {
2079         struct nvme_dev *dev = container_of(self, struct nvme_dev, nb);
2080         switch (action) {
2081         case CPU_ONLINE:
2082         case CPU_DEAD:
2083                 nvme_assign_io_queues(dev);
2084                 break;
2085         }
2086         return NOTIFY_OK;
2087 }
2088
2089 static int nvme_setup_io_queues(struct nvme_dev *dev)
2090 {
2091         struct nvme_queue *adminq = raw_nvmeq(dev, 0);
2092         struct pci_dev *pdev = dev->pci_dev;
2093         int result, i, vecs, nr_io_queues, size;
2094
2095         nr_io_queues = num_possible_cpus();
2096         result = set_queue_count(dev, nr_io_queues);
2097         if (result < 0)
2098                 return result;
2099         if (result < nr_io_queues)
2100                 nr_io_queues = result;
2101
2102         size = db_bar_size(dev, nr_io_queues);
2103         if (size > 8192) {
2104                 iounmap(dev->bar);
2105                 do {
2106                         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
2107                         if (dev->bar)
2108                                 break;
2109                         if (!--nr_io_queues)
2110                                 return -ENOMEM;
2111                         size = db_bar_size(dev, nr_io_queues);
2112                 } while (1);
2113                 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2114                 adminq->q_db = dev->dbs;
2115         }
2116
2117         /* Deregister the admin queue's interrupt */
2118         free_irq(dev->entry[0].vector, adminq);
2119
2120         vecs = nr_io_queues;
2121         for (i = 0; i < vecs; i++)
2122                 dev->entry[i].entry = i;
2123         for (;;) {
2124                 result = pci_enable_msix(pdev, dev->entry, vecs);
2125                 if (result <= 0)
2126                         break;
2127                 vecs = result;
2128         }
2129
2130         if (result < 0) {
2131                 vecs = nr_io_queues;
2132                 if (vecs > 32)
2133                         vecs = 32;
2134                 for (;;) {
2135                         result = pci_enable_msi_block(pdev, vecs);
2136                         if (result == 0) {
2137                                 for (i = 0; i < vecs; i++)
2138                                         dev->entry[i].vector = i + pdev->irq;
2139                                 break;
2140                         } else if (result < 0) {
2141                                 vecs = 1;
2142                                 break;
2143                         }
2144                         vecs = result;
2145                 }
2146         }
2147
2148         /*
2149          * Should investigate if there's a performance win from allocating
2150          * more queues than interrupt vectors; it might allow the submission
2151          * path to scale better, even if the receive path is limited by the
2152          * number of interrupts.
2153          */
2154         nr_io_queues = vecs;
2155         dev->max_qid = nr_io_queues;
2156
2157         result = queue_request_irq(dev, adminq, adminq->irqname);
2158         if (result) {
2159                 adminq->q_suspended = 1;
2160                 goto free_queues;
2161         }
2162
2163         /* Free previously allocated queues that are no longer usable */
2164         nvme_free_queues(dev, nr_io_queues + 1);
2165         nvme_assign_io_queues(dev);
2166
2167         dev->nb.notifier_call = &nvme_cpu_notify;
2168         result = register_hotcpu_notifier(&dev->nb);
2169         if (result)
2170                 goto free_queues;
2171
2172         return 0;
2173
2174  free_queues:
2175         nvme_free_queues(dev, 1);
2176         return result;
2177 }
2178
2179 /*
2180  * Return: error value if an error occurred setting up the queues or calling
2181  * Identify Device.  0 if these succeeded, even if adding some of the
2182  * namespaces failed.  At the moment, these failures are silent.  TBD which
2183  * failures should be reported.
2184  */
2185 static int nvme_dev_add(struct nvme_dev *dev)
2186 {
2187         struct pci_dev *pdev = dev->pci_dev;
2188         int res;
2189         unsigned nn, i;
2190         struct nvme_ns *ns;
2191         struct nvme_id_ctrl *ctrl;
2192         struct nvme_id_ns *id_ns;
2193         void *mem;
2194         dma_addr_t dma_addr;
2195         int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2196
2197         mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
2198         if (!mem)
2199                 return -ENOMEM;
2200
2201         res = nvme_identify(dev, 0, 1, dma_addr);
2202         if (res) {
2203                 res = -EIO;
2204                 goto out;
2205         }
2206
2207         ctrl = mem;
2208         nn = le32_to_cpup(&ctrl->nn);
2209         dev->oncs = le16_to_cpup(&ctrl->oncs);
2210         dev->abort_limit = ctrl->acl + 1;
2211         memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2212         memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2213         memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2214         if (ctrl->mdts)
2215                 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2216         if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2217                         (pdev->device == 0x0953) && ctrl->vs[3])
2218                 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2219
2220         id_ns = mem;
2221         for (i = 1; i <= nn; i++) {
2222                 res = nvme_identify(dev, i, 0, dma_addr);
2223                 if (res)
2224                         continue;
2225
2226                 if (id_ns->ncap == 0)
2227                         continue;
2228
2229                 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
2230                                                         dma_addr + 4096, NULL);
2231                 if (res)
2232                         memset(mem + 4096, 0, 4096);
2233
2234                 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
2235                 if (ns)
2236                         list_add_tail(&ns->list, &dev->namespaces);
2237         }
2238         list_for_each_entry(ns, &dev->namespaces, list)
2239                 add_disk(ns->disk);
2240         res = 0;
2241
2242  out:
2243         dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
2244         return res;
2245 }
2246
2247 static int nvme_dev_map(struct nvme_dev *dev)
2248 {
2249         u64 cap;
2250         int bars, result = -ENOMEM;
2251         struct pci_dev *pdev = dev->pci_dev;
2252
2253         if (pci_enable_device_mem(pdev))
2254                 return result;
2255
2256         dev->entry[0].vector = pdev->irq;
2257         pci_set_master(pdev);
2258         bars = pci_select_bars(pdev, IORESOURCE_MEM);
2259         if (pci_request_selected_regions(pdev, bars, "nvme"))
2260                 goto disable_pci;
2261
2262         if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2263             dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2264                 goto disable;
2265
2266         dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2267         if (!dev->bar)
2268                 goto disable;
2269         if (readl(&dev->bar->csts) == -1) {
2270                 result = -ENODEV;
2271                 goto unmap;
2272         }
2273         cap = readq(&dev->bar->cap);
2274         dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2275         dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2276         dev->dbs = ((void __iomem *)dev->bar) + 4096;
2277
2278         return 0;
2279
2280  unmap:
2281         iounmap(dev->bar);
2282         dev->bar = NULL;
2283  disable:
2284         pci_release_regions(pdev);
2285  disable_pci:
2286         pci_disable_device(pdev);
2287         return result;
2288 }
2289
2290 static void nvme_dev_unmap(struct nvme_dev *dev)
2291 {
2292         if (dev->pci_dev->msi_enabled)
2293                 pci_disable_msi(dev->pci_dev);
2294         else if (dev->pci_dev->msix_enabled)
2295                 pci_disable_msix(dev->pci_dev);
2296
2297         if (dev->bar) {
2298                 iounmap(dev->bar);
2299                 dev->bar = NULL;
2300                 pci_release_regions(dev->pci_dev);
2301         }
2302
2303         if (pci_is_enabled(dev->pci_dev))
2304                 pci_disable_device(dev->pci_dev);
2305 }
2306
2307 struct nvme_delq_ctx {
2308         struct task_struct *waiter;
2309         struct kthread_worker *worker;
2310         atomic_t refcount;
2311 };
2312
2313 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2314 {
2315         dq->waiter = current;
2316         mb();
2317
2318         for (;;) {
2319                 set_current_state(TASK_KILLABLE);
2320                 if (!atomic_read(&dq->refcount))
2321                         break;
2322                 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2323                                         fatal_signal_pending(current)) {
2324                         set_current_state(TASK_RUNNING);
2325
2326                         nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2327                         nvme_disable_queue(dev, 0);
2328
2329                         send_sig(SIGKILL, dq->worker->task, 1);
2330                         flush_kthread_worker(dq->worker);
2331                         return;
2332                 }
2333         }
2334         set_current_state(TASK_RUNNING);
2335 }
2336
2337 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2338 {
2339         atomic_dec(&dq->refcount);
2340         if (dq->waiter)
2341                 wake_up_process(dq->waiter);
2342 }
2343
2344 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2345 {
2346         atomic_inc(&dq->refcount);
2347         return dq;
2348 }
2349
2350 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2351 {
2352         struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2353
2354         nvme_clear_queue(nvmeq);
2355         nvme_put_dq(dq);
2356 }
2357
2358 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2359                                                 kthread_work_func_t fn)
2360 {
2361         struct nvme_command c;
2362
2363         memset(&c, 0, sizeof(c));
2364         c.delete_queue.opcode = opcode;
2365         c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2366
2367         init_kthread_work(&nvmeq->cmdinfo.work, fn);
2368         return nvme_submit_admin_cmd_async(nvmeq->dev, &c, &nvmeq->cmdinfo);
2369 }
2370
2371 static void nvme_del_cq_work_handler(struct kthread_work *work)
2372 {
2373         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2374                                                         cmdinfo.work);
2375         nvme_del_queue_end(nvmeq);
2376 }
2377
2378 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2379 {
2380         return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2381                                                 nvme_del_cq_work_handler);
2382 }
2383
2384 static void nvme_del_sq_work_handler(struct kthread_work *work)
2385 {
2386         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2387                                                         cmdinfo.work);
2388         int status = nvmeq->cmdinfo.status;
2389
2390         if (!status)
2391                 status = nvme_delete_cq(nvmeq);
2392         if (status)
2393                 nvme_del_queue_end(nvmeq);
2394 }
2395
2396 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2397 {
2398         return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2399                                                 nvme_del_sq_work_handler);
2400 }
2401
2402 static void nvme_del_queue_start(struct kthread_work *work)
2403 {
2404         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2405                                                         cmdinfo.work);
2406         allow_signal(SIGKILL);
2407         if (nvme_delete_sq(nvmeq))
2408                 nvme_del_queue_end(nvmeq);
2409 }
2410
2411 static void nvme_disable_io_queues(struct nvme_dev *dev)
2412 {
2413         int i;
2414         DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2415         struct nvme_delq_ctx dq;
2416         struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2417                                         &worker, "nvme%d", dev->instance);
2418
2419         if (IS_ERR(kworker_task)) {
2420                 dev_err(&dev->pci_dev->dev,
2421                         "Failed to create queue del task\n");
2422                 for (i = dev->queue_count - 1; i > 0; i--)
2423                         nvme_disable_queue(dev, i);
2424                 return;
2425         }
2426
2427         dq.waiter = NULL;
2428         atomic_set(&dq.refcount, 0);
2429         dq.worker = &worker;
2430         for (i = dev->queue_count - 1; i > 0; i--) {
2431                 struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2432
2433                 if (nvme_suspend_queue(nvmeq))
2434                         continue;
2435                 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2436                 nvmeq->cmdinfo.worker = dq.worker;
2437                 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2438                 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2439         }
2440         nvme_wait_dq(&dq, dev);
2441         kthread_stop(kworker_task);
2442 }
2443
2444 /*
2445 * Remove the node from the device list and check
2446 * for whether or not we need to stop the nvme_thread.
2447 */
2448 static void nvme_dev_list_remove(struct nvme_dev *dev)
2449 {
2450         struct task_struct *tmp = NULL;
2451
2452         spin_lock(&dev_list_lock);
2453         list_del_init(&dev->node);
2454         if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2455                 tmp = nvme_thread;
2456                 nvme_thread = NULL;
2457         }
2458         spin_unlock(&dev_list_lock);
2459
2460         if (tmp)
2461                 kthread_stop(tmp);
2462 }
2463
2464 static void nvme_dev_shutdown(struct nvme_dev *dev)
2465 {
2466         int i;
2467
2468         dev->initialized = 0;
2469         unregister_hotcpu_notifier(&dev->nb);
2470
2471         nvme_dev_list_remove(dev);
2472
2473         if (!dev->bar || (dev->bar && readl(&dev->bar->csts) == -1)) {
2474                 for (i = dev->queue_count - 1; i >= 0; i--) {
2475                         struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
2476                         nvme_suspend_queue(nvmeq);
2477                         nvme_clear_queue(nvmeq);
2478                 }
2479         } else {
2480                 nvme_disable_io_queues(dev);
2481                 nvme_shutdown_ctrl(dev);
2482                 nvme_disable_queue(dev, 0);
2483         }
2484         nvme_dev_unmap(dev);
2485 }
2486
2487 static void nvme_dev_remove(struct nvme_dev *dev)
2488 {
2489         struct nvme_ns *ns;
2490
2491         list_for_each_entry(ns, &dev->namespaces, list) {
2492                 if (ns->disk->flags & GENHD_FL_UP)
2493                         del_gendisk(ns->disk);
2494                 if (!blk_queue_dying(ns->queue))
2495                         blk_cleanup_queue(ns->queue);
2496         }
2497 }
2498
2499 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2500 {
2501         struct device *dmadev = &dev->pci_dev->dev;
2502         dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2503                                                 PAGE_SIZE, PAGE_SIZE, 0);
2504         if (!dev->prp_page_pool)
2505                 return -ENOMEM;
2506
2507         /* Optimisation for I/Os between 4k and 128k */
2508         dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2509                                                 256, 256, 0);
2510         if (!dev->prp_small_pool) {
2511                 dma_pool_destroy(dev->prp_page_pool);
2512                 return -ENOMEM;
2513         }
2514         return 0;
2515 }
2516
2517 static void nvme_release_prp_pools(struct nvme_dev *dev)
2518 {
2519         dma_pool_destroy(dev->prp_page_pool);
2520         dma_pool_destroy(dev->prp_small_pool);
2521 }
2522
2523 static DEFINE_IDA(nvme_instance_ida);
2524
2525 static int nvme_set_instance(struct nvme_dev *dev)
2526 {
2527         int instance, error;
2528
2529         do {
2530                 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2531                         return -ENODEV;
2532
2533                 spin_lock(&dev_list_lock);
2534                 error = ida_get_new(&nvme_instance_ida, &instance);
2535                 spin_unlock(&dev_list_lock);
2536         } while (error == -EAGAIN);
2537
2538         if (error)
2539                 return -ENODEV;
2540
2541         dev->instance = instance;
2542         return 0;
2543 }
2544
2545 static void nvme_release_instance(struct nvme_dev *dev)
2546 {
2547         spin_lock(&dev_list_lock);
2548         ida_remove(&nvme_instance_ida, dev->instance);
2549         spin_unlock(&dev_list_lock);
2550 }
2551
2552 static void nvme_free_namespaces(struct nvme_dev *dev)
2553 {
2554         struct nvme_ns *ns, *next;
2555
2556         list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2557                 list_del(&ns->list);
2558                 put_disk(ns->disk);
2559                 kfree(ns);
2560         }
2561 }
2562
2563 static void nvme_free_dev(struct kref *kref)
2564 {
2565         struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2566
2567         nvme_free_namespaces(dev);
2568         free_percpu(dev->io_queue);
2569         kfree(dev->queues);
2570         kfree(dev->entry);
2571         kfree(dev);
2572 }
2573
2574 static int nvme_dev_open(struct inode *inode, struct file *f)
2575 {
2576         struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2577                                                                 miscdev);
2578         kref_get(&dev->kref);
2579         f->private_data = dev;
2580         return 0;
2581 }
2582
2583 static int nvme_dev_release(struct inode *inode, struct file *f)
2584 {
2585         struct nvme_dev *dev = f->private_data;
2586         kref_put(&dev->kref, nvme_free_dev);
2587         return 0;
2588 }
2589
2590 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2591 {
2592         struct nvme_dev *dev = f->private_data;
2593         switch (cmd) {
2594         case NVME_IOCTL_ADMIN_CMD:
2595                 return nvme_user_admin_cmd(dev, (void __user *)arg);
2596         default:
2597                 return -ENOTTY;
2598         }
2599 }
2600
2601 static const struct file_operations nvme_dev_fops = {
2602         .owner          = THIS_MODULE,
2603         .open           = nvme_dev_open,
2604         .release        = nvme_dev_release,
2605         .unlocked_ioctl = nvme_dev_ioctl,
2606         .compat_ioctl   = nvme_dev_ioctl,
2607 };
2608
2609 static int nvme_dev_start(struct nvme_dev *dev)
2610 {
2611         int result;
2612         bool start_thread = false;
2613
2614         result = nvme_dev_map(dev);
2615         if (result)
2616                 return result;
2617
2618         result = nvme_configure_admin_queue(dev);
2619         if (result)
2620                 goto unmap;
2621
2622         spin_lock(&dev_list_lock);
2623         if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2624                 start_thread = true;
2625                 nvme_thread = NULL;
2626         }
2627         list_add(&dev->node, &dev_list);
2628         spin_unlock(&dev_list_lock);
2629
2630         if (start_thread) {
2631                 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2632                 wake_up(&nvme_kthread_wait);
2633         } else
2634                 wait_event_killable(nvme_kthread_wait, nvme_thread);
2635
2636         if (IS_ERR_OR_NULL(nvme_thread)) {
2637                 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2638                 goto disable;
2639         }
2640
2641         result = nvme_setup_io_queues(dev);
2642         if (result && result != -EBUSY)
2643                 goto disable;
2644
2645         return result;
2646
2647  disable:
2648         nvme_disable_queue(dev, 0);
2649         nvme_dev_list_remove(dev);
2650  unmap:
2651         nvme_dev_unmap(dev);
2652         return result;
2653 }
2654
2655 static int nvme_remove_dead_ctrl(void *arg)
2656 {
2657         struct nvme_dev *dev = (struct nvme_dev *)arg;
2658         struct pci_dev *pdev = dev->pci_dev;
2659
2660         if (pci_get_drvdata(pdev))
2661                 pci_stop_and_remove_bus_device(pdev);
2662         kref_put(&dev->kref, nvme_free_dev);
2663         return 0;
2664 }
2665
2666 static void nvme_remove_disks(struct work_struct *ws)
2667 {
2668         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2669
2670         nvme_dev_remove(dev);
2671         nvme_free_queues(dev, 1);
2672 }
2673
2674 static int nvme_dev_resume(struct nvme_dev *dev)
2675 {
2676         int ret;
2677
2678         ret = nvme_dev_start(dev);
2679         if (ret && ret != -EBUSY)
2680                 return ret;
2681         if (ret == -EBUSY) {
2682                 spin_lock(&dev_list_lock);
2683                 PREPARE_WORK(&dev->reset_work, nvme_remove_disks);
2684                 queue_work(nvme_workq, &dev->reset_work);
2685                 spin_unlock(&dev_list_lock);
2686         }
2687         dev->initialized = 1;
2688         return 0;
2689 }
2690
2691 static void nvme_dev_reset(struct nvme_dev *dev)
2692 {
2693         nvme_dev_shutdown(dev);
2694         if (nvme_dev_resume(dev)) {
2695                 dev_err(&dev->pci_dev->dev, "Device failed to resume\n");
2696                 kref_get(&dev->kref);
2697                 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2698                                                         dev->instance))) {
2699                         dev_err(&dev->pci_dev->dev,
2700                                 "Failed to start controller remove task\n");
2701                         kref_put(&dev->kref, nvme_free_dev);
2702                 }
2703         }
2704 }
2705
2706 static void nvme_reset_failed_dev(struct work_struct *ws)
2707 {
2708         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2709         nvme_dev_reset(dev);
2710 }
2711
2712 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2713 {
2714         int result = -ENOMEM;
2715         struct nvme_dev *dev;
2716
2717         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2718         if (!dev)
2719                 return -ENOMEM;
2720         dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
2721                                                                 GFP_KERNEL);
2722         if (!dev->entry)
2723                 goto free;
2724         dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
2725                                                                 GFP_KERNEL);
2726         if (!dev->queues)
2727                 goto free;
2728         dev->io_queue = alloc_percpu(unsigned short);
2729         if (!dev->io_queue)
2730                 goto free;
2731
2732         INIT_LIST_HEAD(&dev->namespaces);
2733         INIT_WORK(&dev->reset_work, nvme_reset_failed_dev);
2734         dev->pci_dev = pdev;
2735         pci_set_drvdata(pdev, dev);
2736         result = nvme_set_instance(dev);
2737         if (result)
2738                 goto free;
2739
2740         result = nvme_setup_prp_pools(dev);
2741         if (result)
2742                 goto release;
2743
2744         kref_init(&dev->kref);
2745         result = nvme_dev_start(dev);
2746         if (result) {
2747                 if (result == -EBUSY)
2748                         goto create_cdev;
2749                 goto release_pools;
2750         }
2751
2752         result = nvme_dev_add(dev);
2753         if (result)
2754                 goto shutdown;
2755
2756  create_cdev:
2757         scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2758         dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2759         dev->miscdev.parent = &pdev->dev;
2760         dev->miscdev.name = dev->name;
2761         dev->miscdev.fops = &nvme_dev_fops;
2762         result = misc_register(&dev->miscdev);
2763         if (result)
2764                 goto remove;
2765
2766         dev->initialized = 1;
2767         return 0;
2768
2769  remove:
2770         nvme_dev_remove(dev);
2771         nvme_free_namespaces(dev);
2772  shutdown:
2773         nvme_dev_shutdown(dev);
2774  release_pools:
2775         nvme_free_queues(dev, 0);
2776         nvme_release_prp_pools(dev);
2777  release:
2778         nvme_release_instance(dev);
2779  free:
2780         free_percpu(dev->io_queue);
2781         kfree(dev->queues);
2782         kfree(dev->entry);
2783         kfree(dev);
2784         return result;
2785 }
2786
2787 static void nvme_shutdown(struct pci_dev *pdev)
2788 {
2789         struct nvme_dev *dev = pci_get_drvdata(pdev);
2790         nvme_dev_shutdown(dev);
2791 }
2792
2793 static void nvme_remove(struct pci_dev *pdev)
2794 {
2795         struct nvme_dev *dev = pci_get_drvdata(pdev);
2796
2797         spin_lock(&dev_list_lock);
2798         list_del_init(&dev->node);
2799         spin_unlock(&dev_list_lock);
2800
2801         pci_set_drvdata(pdev, NULL);
2802         flush_work(&dev->reset_work);
2803         misc_deregister(&dev->miscdev);
2804         nvme_dev_remove(dev);
2805         nvme_dev_shutdown(dev);
2806         nvme_free_queues(dev, 0);
2807         rcu_barrier();
2808         nvme_release_instance(dev);
2809         nvme_release_prp_pools(dev);
2810         kref_put(&dev->kref, nvme_free_dev);
2811 }
2812
2813 /* These functions are yet to be implemented */
2814 #define nvme_error_detected NULL
2815 #define nvme_dump_registers NULL
2816 #define nvme_link_reset NULL
2817 #define nvme_slot_reset NULL
2818 #define nvme_error_resume NULL
2819
2820 #ifdef CONFIG_PM_SLEEP
2821 static int nvme_suspend(struct device *dev)
2822 {
2823         struct pci_dev *pdev = to_pci_dev(dev);
2824         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2825
2826         nvme_dev_shutdown(ndev);
2827         return 0;
2828 }
2829
2830 static int nvme_resume(struct device *dev)
2831 {
2832         struct pci_dev *pdev = to_pci_dev(dev);
2833         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2834
2835         if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2836                 PREPARE_WORK(&ndev->reset_work, nvme_reset_failed_dev);
2837                 queue_work(nvme_workq, &ndev->reset_work);
2838         }
2839         return 0;
2840 }
2841 #endif
2842
2843 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2844
2845 static const struct pci_error_handlers nvme_err_handler = {
2846         .error_detected = nvme_error_detected,
2847         .mmio_enabled   = nvme_dump_registers,
2848         .link_reset     = nvme_link_reset,
2849         .slot_reset     = nvme_slot_reset,
2850         .resume         = nvme_error_resume,
2851 };
2852
2853 /* Move to pci_ids.h later */
2854 #define PCI_CLASS_STORAGE_EXPRESS       0x010802
2855
2856 static const struct pci_device_id nvme_id_table[] = {
2857         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2858         { 0, }
2859 };
2860 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2861
2862 static struct pci_driver nvme_driver = {
2863         .name           = "nvme",
2864         .id_table       = nvme_id_table,
2865         .probe          = nvme_probe,
2866         .remove         = nvme_remove,
2867         .shutdown       = nvme_shutdown,
2868         .driver         = {
2869                 .pm     = &nvme_dev_pm_ops,
2870         },
2871         .err_handler    = &nvme_err_handler,
2872 };
2873
2874 static int __init nvme_init(void)
2875 {
2876         int result;
2877
2878         init_waitqueue_head(&nvme_kthread_wait);
2879
2880         nvme_workq = create_singlethread_workqueue("nvme");
2881         if (!nvme_workq)
2882                 return -ENOMEM;
2883
2884         result = register_blkdev(nvme_major, "nvme");
2885         if (result < 0)
2886                 goto kill_workq;
2887         else if (result > 0)
2888                 nvme_major = result;
2889
2890         result = pci_register_driver(&nvme_driver);
2891         if (result)
2892                 goto unregister_blkdev;
2893         return 0;
2894
2895  unregister_blkdev:
2896         unregister_blkdev(nvme_major, "nvme");
2897  kill_workq:
2898         destroy_workqueue(nvme_workq);
2899         return result;
2900 }
2901
2902 static void __exit nvme_exit(void)
2903 {
2904         pci_unregister_driver(&nvme_driver);
2905         unregister_blkdev(nvme_major, "nvme");
2906         destroy_workqueue(nvme_workq);
2907         BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
2908 }
2909
2910 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2911 MODULE_LICENSE("GPL");
2912 MODULE_VERSION("0.9");
2913 module_init(nvme_init);
2914 module_exit(nvme_exit);