drivers/block/nvme-core.c

   1 /*
   2  * NVM Express device driver
   3  * Copyright (c) 2011-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14
  15 #include <linux/nvme.h>
  16 #include <linux/bitops.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/blk-mq.h>
  19 #include <linux/cpu.h>
  20 #include <linux/delay.h>
  21 #include <linux/errno.h>
  22 #include <linux/fs.h>
  23 #include <linux/genhd.h>
  24 #include <linux/hdreg.h>
  25 #include <linux/idr.h>
  26 #include <linux/init.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/io.h>
  29 #include <linux/kdev_t.h>
  30 #include <linux/kthread.h>
  31 #include <linux/kernel.h>
  32 #include <linux/mm.h>
  33 #include <linux/module.h>
  34 #include <linux/moduleparam.h>
  35 #include <linux/pci.h>
  36 #include <linux/poison.h>
  37 #include <linux/ptrace.h>
  38 #include <linux/sched.h>
  39 #include <linux/slab.h>
  40 #include <linux/types.h>
  41 #include <scsi/sg.h>
  42 #include <asm-generic/io-64-nonatomic-lo-hi.h>
  43
  44 #define NVME_Q_DEPTH            1024
  45 #define NVME_AQ_DEPTH           64
  46 #define SQ_SIZE(depth)          (depth * sizeof(struct nvme_command))
  47 #define CQ_SIZE(depth)          (depth * sizeof(struct nvme_completion))
  48 #define ADMIN_TIMEOUT           (admin_timeout * HZ)
  49 #define SHUTDOWN_TIMEOUT        (shutdown_timeout * HZ)
  50 #define IOD_TIMEOUT             (retry_time * HZ)
  51
  52 static unsigned char admin_timeout = 60;
  53 module_param(admin_timeout, byte, 0644);
  54 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
  55
  56 unsigned char nvme_io_timeout = 30;
  57 module_param_named(io_timeout, nvme_io_timeout, byte, 0644);
  58 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
  59
  60 static unsigned char retry_time = 30;
  61 module_param(retry_time, byte, 0644);
  62 MODULE_PARM_DESC(retry_time, "time in seconds to retry failed I/O");
  63
  64 static unsigned char shutdown_timeout = 5;
  65 module_param(shutdown_timeout, byte, 0644);
  66 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
  67
  68 static int nvme_major;
  69 module_param(nvme_major, int, 0);
  70
  71 static int use_threaded_interrupts;
  72 module_param(use_threaded_interrupts, int, 0);
  73
  74 static DEFINE_SPINLOCK(dev_list_lock);
  75 static LIST_HEAD(dev_list);
  76 static struct task_struct *nvme_thread;
  77 static struct workqueue_struct *nvme_workq;
  78 static wait_queue_head_t nvme_kthread_wait;
  79 static struct notifier_block nvme_nb;
  80
  81 static void nvme_reset_failed_dev(struct work_struct *ws);
  82 static int nvme_process_cq(struct nvme_queue *nvmeq);
  83
  84 struct async_cmd_info {
  85         struct kthread_work work;
  86         struct kthread_worker *worker;
  87         struct request *req;
  88         u32 result;
  89         int status;
  90         void *ctx;
  91 };
  92
  93 /*
  94  * An NVM Express queue.  Each device has at least two (one for admin
  95  * commands and one for I/O commands).
  96  */
  97 struct nvme_queue {
  98         struct llist_node node;
  99         struct device *q_dmadev;
 100         struct nvme_dev *dev;
 101         char irqname[24];       /* nvme4294967295-65535\0 */
 102         spinlock_t q_lock;
 103         struct nvme_command *sq_cmds;
 104         volatile struct nvme_completion *cqes;
 105         dma_addr_t sq_dma_addr;
 106         dma_addr_t cq_dma_addr;
 107         u32 __iomem *q_db;
 108         u16 q_depth;
 109         u16 cq_vector;
 110         u16 sq_head;
 111         u16 sq_tail;
 112         u16 cq_head;
 113         u16 qid;
 114         u8 cq_phase;
 115         u8 cqe_seen;
 116         struct async_cmd_info cmdinfo;
 117         struct blk_mq_hw_ctx *hctx;
 118 };
 119
 120 /*
 121  * Check we didin't inadvertently grow the command struct
 122  */
 123 static inline void _nvme_check_size(void)
 124 {
 125         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
 126         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
 127         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
 128         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
 129         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
 130         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
 131         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
 132         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
 133         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
 134         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
 135         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
 136         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
 137 }
 138
 139 typedef void (*nvme_completion_fn)(struct nvme_queue *, void *,
 140                                                 struct nvme_completion *);
 141
 142 struct nvme_cmd_info {
 143         nvme_completion_fn fn;
 144         void *ctx;
 145         int aborted;
 146         struct nvme_queue *nvmeq;
 147 };
 148
 149 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 150                                 unsigned int hctx_idx)
 151 {
 152         struct nvme_dev *dev = data;
 153         struct nvme_queue *nvmeq = dev->queues[0];
 154
 155         WARN_ON(nvmeq->hctx);
 156         nvmeq->hctx = hctx;
 157         hctx->driver_data = nvmeq;
 158         return 0;
 159 }
 160
 161 static int nvme_admin_init_request(void *data, struct request *req,
 162                                 unsigned int hctx_idx, unsigned int rq_idx,
 163                                 unsigned int numa_node)
 164 {
 165         struct nvme_dev *dev = data;
 166         struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
 167         struct nvme_queue *nvmeq = dev->queues[0];
 168
 169         BUG_ON(!nvmeq);
 170         cmd->nvmeq = nvmeq;
 171         return 0;
 172 }
 173
 174 static void nvme_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 175 {
 176         struct nvme_queue *nvmeq = hctx->driver_data;
 177
 178         nvmeq->hctx = NULL;
 179 }
 180
 181 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 182                           unsigned int hctx_idx)
 183 {
 184         struct nvme_dev *dev = data;
 185         struct nvme_queue *nvmeq = dev->queues[
 186                                         (hctx_idx % dev->queue_count) + 1];
 187
 188         if (!nvmeq->hctx)
 189                 nvmeq->hctx = hctx;
 190
 191         /* nvmeq queues are shared between namespaces. We assume here that
 192          * blk-mq map the tags so they match up with the nvme queue tags. */
 193         WARN_ON(nvmeq->hctx->tags != hctx->tags);
 194
 195         hctx->driver_data = nvmeq;
 196         return 0;
 197 }
 198
 199 static int nvme_init_request(void *data, struct request *req,
 200                                 unsigned int hctx_idx, unsigned int rq_idx,
 201                                 unsigned int numa_node)
 202 {
 203         struct nvme_dev *dev = data;
 204         struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
 205         struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
 206
 207         BUG_ON(!nvmeq);
 208         cmd->nvmeq = nvmeq;
 209         return 0;
 210 }
 211
 212 static void nvme_set_info(struct nvme_cmd_info *cmd, void *ctx,
 213                                 nvme_completion_fn handler)
 214 {
 215         cmd->fn = handler;
 216         cmd->ctx = ctx;
 217         cmd->aborted = 0;
 218 }
 219
 220 /* Special values must be less than 0x1000 */
 221 #define CMD_CTX_BASE            ((void *)POISON_POINTER_DELTA)
 222 #define CMD_CTX_CANCELLED       (0x30C + CMD_CTX_BASE)
 223 #define CMD_CTX_COMPLETED       (0x310 + CMD_CTX_BASE)
 224 #define CMD_CTX_INVALID         (0x314 + CMD_CTX_BASE)
 225
 226 static void special_completion(struct nvme_queue *nvmeq, void *ctx,
 227                                                 struct nvme_completion *cqe)
 228 {
 229         if (ctx == CMD_CTX_CANCELLED)
 230                 return;
 231         if (ctx == CMD_CTX_COMPLETED) {
 232                 dev_warn(nvmeq->q_dmadev,
 233                                 "completed id %d twice on queue %d\n",
 234                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 235                 return;
 236         }
 237         if (ctx == CMD_CTX_INVALID) {
 238                 dev_warn(nvmeq->q_dmadev,
 239                                 "invalid id %d completed on queue %d\n",
 240                                 cqe->command_id, le16_to_cpup(&cqe->sq_id));
 241                 return;
 242         }
 243         dev_warn(nvmeq->q_dmadev, "Unknown special completion %p\n", ctx);
 244 }
 245
 246 static void *cancel_cmd_info(struct nvme_cmd_info *cmd, nvme_completion_fn *fn)
 247 {
 248         void *ctx;
 249
 250         if (fn)
 251                 *fn = cmd->fn;
 252         ctx = cmd->ctx;
 253         cmd->fn = special_completion;
 254         cmd->ctx = CMD_CTX_CANCELLED;
 255         return ctx;
 256 }
 257
 258 static void async_req_completion(struct nvme_queue *nvmeq, void *ctx,
 259                                                 struct nvme_completion *cqe)
 260 {
 261         struct request *req = ctx;
 262
 263         u32 result = le32_to_cpup(&cqe->result);
 264         u16 status = le16_to_cpup(&cqe->status) >> 1;
 265
 266         if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ)
 267                 ++nvmeq->dev->event_limit;
 268         if (status == NVME_SC_SUCCESS)
 269                 dev_warn(nvmeq->q_dmadev,
 270                         "async event result %08x\n", result);
 271
 272         blk_mq_free_hctx_request(nvmeq->hctx, req);
 273 }
 274
 275 static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
 276                                                 struct nvme_completion *cqe)
 277 {
 278         struct request *req = ctx;
 279
 280         u16 status = le16_to_cpup(&cqe->status) >> 1;
 281         u32 result = le32_to_cpup(&cqe->result);
 282
 283         blk_mq_free_hctx_request(nvmeq->hctx, req);
 284
 285         dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
 286         ++nvmeq->dev->abort_limit;
 287 }
 288
 289 static void async_completion(struct nvme_queue *nvmeq, void *ctx,
 290                                                 struct nvme_completion *cqe)
 291 {
 292         struct async_cmd_info *cmdinfo = ctx;
 293         cmdinfo->result = le32_to_cpup(&cqe->result);
 294         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 295         queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
 296         blk_mq_free_hctx_request(nvmeq->hctx, cmdinfo->req);
 297 }
 298
 299 static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
 300                                   unsigned int tag)
 301 {
 302         struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
 303         struct request *req = blk_mq_tag_to_rq(hctx->tags, tag);
 304
 305         return blk_mq_rq_to_pdu(req);
 306 }
 307
 308 /*
 309  * Called with local interrupts disabled and the q_lock held.  May not sleep.
 310  */
 311 static void *nvme_finish_cmd(struct nvme_queue *nvmeq, int tag,
 312                                                 nvme_completion_fn *fn)
 313 {
 314         struct nvme_cmd_info *cmd = get_cmd_from_tag(nvmeq, tag);
 315         void *ctx;
 316         if (tag >= nvmeq->q_depth) {
 317                 *fn = special_completion;
 318                 return CMD_CTX_INVALID;
 319         }
 320         if (fn)
 321                 *fn = cmd->fn;
 322         ctx = cmd->ctx;
 323         cmd->fn = special_completion;
 324         cmd->ctx = CMD_CTX_COMPLETED;
 325         return ctx;
 326 }
 327
 328 /**
 329  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 330  * @nvmeq: The queue to use
 331  * @cmd: The command to send
 332  *
 333  * Safe to use from interrupt context
 334  */
 335 static int __nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
 336 {
 337         u16 tail = nvmeq->sq_tail;
 338
 339         memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
 340         if (++tail == nvmeq->q_depth)
 341                 tail = 0;
 342         writel(tail, nvmeq->q_db);
 343         nvmeq->sq_tail = tail;
 344
 345         return 0;
 346 }
 347
 348 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
 349 {
 350         unsigned long flags;
 351         int ret;
 352         spin_lock_irqsave(&nvmeq->q_lock, flags);
 353         ret = __nvme_submit_cmd(nvmeq, cmd);
 354         spin_unlock_irqrestore(&nvmeq->q_lock, flags);
 355         return ret;
 356 }
 357
 358 static __le64 **iod_list(struct nvme_iod *iod)
 359 {
 360         return ((void *)iod) + iod->offset;
 361 }
 362
 363 /*
 364  * Will slightly overestimate the number of pages needed.  This is OK
 365  * as it only leads to a small amount of wasted memory for the lifetime of
 366  * the I/O.
 367  */
 368 static int nvme_npages(unsigned size, struct nvme_dev *dev)
 369 {
 370         unsigned nprps = DIV_ROUND_UP(size + dev->page_size, dev->page_size);
 371         return DIV_ROUND_UP(8 * nprps, dev->page_size - 8);
 372 }
 373
 374 static struct nvme_iod *
 375 nvme_alloc_iod(unsigned nseg, unsigned nbytes, struct nvme_dev *dev, gfp_t gfp)
 376 {
 377         struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
 378                                 sizeof(__le64 *) * nvme_npages(nbytes, dev) +
 379                                 sizeof(struct scatterlist) * nseg, gfp);
 380
 381         if (iod) {
 382                 iod->offset = offsetof(struct nvme_iod, sg[nseg]);
 383                 iod->npages = -1;
 384                 iod->length = nbytes;
 385                 iod->nents = 0;
 386                 iod->first_dma = 0ULL;
 387         }
 388
 389         return iod;
 390 }
 391
 392 void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
 393 {
 394         const int last_prp = dev->page_size / 8 - 1;
 395         int i;
 396         __le64 **list = iod_list(iod);
 397         dma_addr_t prp_dma = iod->first_dma;
 398
 399         if (iod->npages == 0)
 400                 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
 401         for (i = 0; i < iod->npages; i++) {
 402                 __le64 *prp_list = list[i];
 403                 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
 404                 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
 405                 prp_dma = next_prp_dma;
 406         }
 407         kfree(iod);
 408 }
 409
 410 static int nvme_error_status(u16 status)
 411 {
 412         switch (status & 0x7ff) {
 413         case NVME_SC_SUCCESS:
 414                 return 0;
 415         case NVME_SC_CAP_EXCEEDED:
 416                 return -ENOSPC;
 417         default:
 418                 return -EIO;
 419         }
 420 }
 421
 422 static void req_completion(struct nvme_queue *nvmeq, void *ctx,
 423                                                 struct nvme_completion *cqe)
 424 {
 425         struct nvme_iod *iod = ctx;
 426         struct request *req = iod->private;
 427         struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
 428
 429         u16 status = le16_to_cpup(&cqe->status) >> 1;
 430
 431         if (unlikely(status)) {
 432                 if (!(status & NVME_SC_DNR || blk_noretry_request(req))
 433                     && (jiffies - req->start_time) < req->timeout) {
 434                         blk_mq_requeue_request(req);
 435                         blk_mq_kick_requeue_list(req->q);
 436                         return;
 437                 }
 438                 req->errors = nvme_error_status(status);
 439         } else
 440                 req->errors = 0;
 441
 442         if (cmd_rq->aborted)
 443                 dev_warn(&nvmeq->dev->pci_dev->dev,
 444                         "completing aborted command with status:%04x\n",
 445                         status);
 446
 447         if (iod->nents)
 448                 dma_unmap_sg(&nvmeq->dev->pci_dev->dev, iod->sg, iod->nents,
 449                         rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 450         nvme_free_iod(nvmeq->dev, iod);
 451
 452         blk_mq_complete_request(req);
 453 }
 454
 455 /* length is in bytes.  gfp flags indicates whether we may sleep. */
 456 int nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod, int total_len,
 457                                                                 gfp_t gfp)
 458 {
 459         struct dma_pool *pool;
 460         int length = total_len;
 461         struct scatterlist *sg = iod->sg;
 462         int dma_len = sg_dma_len(sg);
 463         u64 dma_addr = sg_dma_address(sg);
 464         int offset = offset_in_page(dma_addr);
 465         __le64 *prp_list;
 466         __le64 **list = iod_list(iod);
 467         dma_addr_t prp_dma;
 468         int nprps, i;
 469         u32 page_size = dev->page_size;
 470
 471         length -= (page_size - offset);
 472         if (length <= 0)
 473                 return total_len;
 474
 475         dma_len -= (page_size - offset);
 476         if (dma_len) {
 477                 dma_addr += (page_size - offset);
 478         } else {
 479                 sg = sg_next(sg);
 480                 dma_addr = sg_dma_address(sg);
 481                 dma_len = sg_dma_len(sg);
 482         }
 483
 484         if (length <= page_size) {
 485                 iod->first_dma = dma_addr;
 486                 return total_len;
 487         }
 488
 489         nprps = DIV_ROUND_UP(length, page_size);
 490         if (nprps <= (256 / 8)) {
 491                 pool = dev->prp_small_pool;
 492                 iod->npages = 0;
 493         } else {
 494                 pool = dev->prp_page_pool;
 495                 iod->npages = 1;
 496         }
 497
 498         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 499         if (!prp_list) {
 500                 iod->first_dma = dma_addr;
 501                 iod->npages = -1;
 502                 return (total_len - length) + page_size;
 503         }
 504         list[0] = prp_list;
 505         iod->first_dma = prp_dma;
 506         i = 0;
 507         for (;;) {
 508                 if (i == page_size >> 3) {
 509                         __le64 *old_prp_list = prp_list;
 510                         prp_list = dma_pool_alloc(pool, gfp, &prp_dma);
 511                         if (!prp_list)
 512                                 return total_len - length;
 513                         list[iod->npages++] = prp_list;
 514                         prp_list[0] = old_prp_list[i - 1];
 515                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 516                         i = 1;
 517                 }
 518                 prp_list[i++] = cpu_to_le64(dma_addr);
 519                 dma_len -= page_size;
 520                 dma_addr += page_size;
 521                 length -= page_size;
 522                 if (length <= 0)
 523                         break;
 524                 if (dma_len > 0)
 525                         continue;
 526                 BUG_ON(dma_len < 0);
 527                 sg = sg_next(sg);
 528                 dma_addr = sg_dma_address(sg);
 529                 dma_len = sg_dma_len(sg);
 530         }
 531
 532         return total_len;
 533 }
 534
 535 /*
 536  * We reuse the small pool to allocate the 16-byte range here as it is not
 537  * worth having a special pool for these or additional cases to handle freeing
 538  * the iod.
 539  */
 540 static void nvme_submit_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 541                 struct request *req, struct nvme_iod *iod)
 542 {
 543         struct nvme_dsm_range *range =
 544                                 (struct nvme_dsm_range *)iod_list(iod)[0];
 545         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 546
 547         range->cattr = cpu_to_le32(0);
 548         range->nlb = cpu_to_le32(blk_rq_bytes(req) >> ns->lba_shift);
 549         range->slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
 550
 551         memset(cmnd, 0, sizeof(*cmnd));
 552         cmnd->dsm.opcode = nvme_cmd_dsm;
 553         cmnd->dsm.command_id = req->tag;
 554         cmnd->dsm.nsid = cpu_to_le32(ns->ns_id);
 555         cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
 556         cmnd->dsm.nr = 0;
 557         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 558
 559         if (++nvmeq->sq_tail == nvmeq->q_depth)
 560                 nvmeq->sq_tail = 0;
 561         writel(nvmeq->sq_tail, nvmeq->q_db);
 562 }
 563
 564 static void nvme_submit_flush(struct nvme_queue *nvmeq, struct nvme_ns *ns,
 565                                                                 int cmdid)
 566 {
 567         struct nvme_command *cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 568
 569         memset(cmnd, 0, sizeof(*cmnd));
 570         cmnd->common.opcode = nvme_cmd_flush;
 571         cmnd->common.command_id = cmdid;
 572         cmnd->common.nsid = cpu_to_le32(ns->ns_id);
 573
 574         if (++nvmeq->sq_tail == nvmeq->q_depth)
 575                 nvmeq->sq_tail = 0;
 576         writel(nvmeq->sq_tail, nvmeq->q_db);
 577 }
 578
 579 static int nvme_submit_iod(struct nvme_queue *nvmeq, struct nvme_iod *iod,
 580                                                         struct nvme_ns *ns)
 581 {
 582         struct request *req = iod->private;
 583         struct nvme_command *cmnd;
 584         u16 control = 0;
 585         u32 dsmgmt = 0;
 586
 587         if (req->cmd_flags & REQ_FUA)
 588                 control |= NVME_RW_FUA;
 589         if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 590                 control |= NVME_RW_LR;
 591
 592         if (req->cmd_flags & REQ_RAHEAD)
 593                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 594
 595         cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
 596         memset(cmnd, 0, sizeof(*cmnd));
 597
 598         cmnd->rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
 599         cmnd->rw.command_id = req->tag;
 600         cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
 601         cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 602         cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
 603         cmnd->rw.slba = cpu_to_le64(nvme_block_nr(ns, blk_rq_pos(req)));
 604         cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 605         cmnd->rw.control = cpu_to_le16(control);
 606         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 607
 608         if (++nvmeq->sq_tail == nvmeq->q_depth)
 609                 nvmeq->sq_tail = 0;
 610         writel(nvmeq->sq_tail, nvmeq->q_db);
 611
 612         return 0;
 613 }
 614
 615 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 616                          const struct blk_mq_queue_data *bd)
 617 {
 618         struct nvme_ns *ns = hctx->queue->queuedata;
 619         struct nvme_queue *nvmeq = hctx->driver_data;
 620         struct request *req = bd->rq;
 621         struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
 622         struct nvme_iod *iod;
 623         int psegs = req->nr_phys_segments;
 624         int result = BLK_MQ_RQ_QUEUE_BUSY;
 625         enum dma_data_direction dma_dir;
 626         unsigned size = !(req->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(req) :
 627                                                 sizeof(struct nvme_dsm_range);
 628
 629         /*
 630          * Requeued IO has already been prepped
 631          */
 632         iod = req->special;
 633         if (iod)
 634                 goto submit_iod;
 635
 636         iod = nvme_alloc_iod(psegs, size, ns->dev, GFP_ATOMIC);
 637         if (!iod)
 638                 return result;
 639
 640         iod->private = req;
 641         req->special = iod;
 642
 643         nvme_set_info(cmd, iod, req_completion);
 644
 645         if (req->cmd_flags & REQ_DISCARD) {
 646                 void *range;
 647                 /*
 648                  * We reuse the small pool to allocate the 16-byte range here
 649                  * as it is not worth having a special pool for these or
 650                  * additional cases to handle freeing the iod.
 651                  */
 652                 range = dma_pool_alloc(nvmeq->dev->prp_small_pool,
 653                                                 GFP_ATOMIC,
 654                                                 &iod->first_dma);
 655                 if (!range)
 656                         goto finish_cmd;
 657                 iod_list(iod)[0] = (__le64 *)range;
 658                 iod->npages = 0;
 659         } else if (psegs) {
 660                 dma_dir = rq_data_dir(req) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
 661
 662                 sg_init_table(iod->sg, psegs);
 663                 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 664                 if (!iod->nents) {
 665                         result = BLK_MQ_RQ_QUEUE_ERROR;
 666                         goto finish_cmd;
 667                 }
 668
 669                 if (!dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir))
 670                         goto finish_cmd;
 671
 672                 if (blk_rq_bytes(req) != nvme_setup_prps(nvmeq->dev, iod,
 673                                                 blk_rq_bytes(req), GFP_ATOMIC))
 674                         goto finish_cmd;
 675         }
 676
 677         blk_mq_start_request(req);
 678
 679  submit_iod:
 680         spin_lock_irq(&nvmeq->q_lock);
 681         if (req->cmd_flags & REQ_DISCARD)
 682                 nvme_submit_discard(nvmeq, ns, req, iod);
 683         else if (req->cmd_flags & REQ_FLUSH)
 684                 nvme_submit_flush(nvmeq, ns, req->tag);
 685         else
 686                 nvme_submit_iod(nvmeq, iod, ns);
 687
 688         nvme_process_cq(nvmeq);
 689         spin_unlock_irq(&nvmeq->q_lock);
 690         return BLK_MQ_RQ_QUEUE_OK;
 691
 692  finish_cmd:
 693         nvme_finish_cmd(nvmeq, req->tag, NULL);
 694         nvme_free_iod(nvmeq->dev, iod);
 695         return result;
 696 }
 697
 698 static int nvme_process_cq(struct nvme_queue *nvmeq)
 699 {
 700         u16 head, phase;
 701
 702         head = nvmeq->cq_head;
 703         phase = nvmeq->cq_phase;
 704
 705         for (;;) {
 706                 void *ctx;
 707                 nvme_completion_fn fn;
 708                 struct nvme_completion cqe = nvmeq->cqes[head];
 709                 if ((le16_to_cpu(cqe.status) & 1) != phase)
 710                         break;
 711                 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
 712                 if (++head == nvmeq->q_depth) {
 713                         head = 0;
 714                         phase = !phase;
 715                 }
 716                 ctx = nvme_finish_cmd(nvmeq, cqe.command_id, &fn);
 717                 fn(nvmeq, ctx, &cqe);
 718         }
 719
 720         /* If the controller ignores the cq head doorbell and continuously
 721          * writes to the queue, it is theoretically possible to wrap around
 722          * the queue twice and mistakenly return IRQ_NONE.  Linux only
 723          * requires that 0.1% of your interrupts are handled, so this isn't
 724          * a big problem.
 725          */
 726         if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
 727                 return 0;
 728
 729         writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 730         nvmeq->cq_head = head;
 731         nvmeq->cq_phase = phase;
 732
 733         nvmeq->cqe_seen = 1;
 734         return 1;
 735 }
 736
 737 /* Admin queue isn't initialized as a request queue. If at some point this
 738  * happens anyway, make sure to notify the user */
 739 static int nvme_admin_queue_rq(struct blk_mq_hw_ctx *hctx,
 740                                const struct blk_mq_queue_data *bd)
 741 {
 742         WARN_ON_ONCE(1);
 743         return BLK_MQ_RQ_QUEUE_ERROR;
 744 }
 745
 746 static irqreturn_t nvme_irq(int irq, void *data)
 747 {
 748         irqreturn_t result;
 749         struct nvme_queue *nvmeq = data;
 750         spin_lock(&nvmeq->q_lock);
 751         nvme_process_cq(nvmeq);
 752         result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
 753         nvmeq->cqe_seen = 0;
 754         spin_unlock(&nvmeq->q_lock);
 755         return result;
 756 }
 757
 758 static irqreturn_t nvme_irq_check(int irq, void *data)
 759 {
 760         struct nvme_queue *nvmeq = data;
 761         struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
 762         if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
 763                 return IRQ_NONE;
 764         return IRQ_WAKE_THREAD;
 765 }
 766
 767 static void nvme_abort_cmd_info(struct nvme_queue *nvmeq, struct nvme_cmd_info *
 768                                                                 cmd_info)
 769 {
 770         spin_lock_irq(&nvmeq->q_lock);
 771         cancel_cmd_info(cmd_info, NULL);
 772         spin_unlock_irq(&nvmeq->q_lock);
 773 }
 774
 775 struct sync_cmd_info {
 776         struct task_struct *task;
 777         u32 result;
 778         int status;
 779 };
 780
 781 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
 782                                                 struct nvme_completion *cqe)
 783 {
 784         struct sync_cmd_info *cmdinfo = ctx;
 785         cmdinfo->result = le32_to_cpup(&cqe->result);
 786         cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
 787         wake_up_process(cmdinfo->task);
 788 }
 789
 790 /*
 791  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 792  * if the result is positive, it's an NVM Express status code
 793  */
 794 static int nvme_submit_sync_cmd(struct request *req, struct nvme_command *cmd,
 795                                                 u32 *result, unsigned timeout)
 796 {
 797         int ret;
 798         struct sync_cmd_info cmdinfo;
 799         struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
 800         struct nvme_queue *nvmeq = cmd_rq->nvmeq;
 801
 802         cmdinfo.task = current;
 803         cmdinfo.status = -EINTR;
 804
 805         cmd->common.command_id = req->tag;
 806
 807         nvme_set_info(cmd_rq, &cmdinfo, sync_completion);
 808
 809         set_current_state(TASK_KILLABLE);
 810         ret = nvme_submit_cmd(nvmeq, cmd);
 811         if (ret) {
 812                 nvme_finish_cmd(nvmeq, req->tag, NULL);
 813                 set_current_state(TASK_RUNNING);
 814         }
 815         schedule_timeout(timeout);
 816
 817         if (cmdinfo.status == -EINTR) {
 818                 nvme_abort_cmd_info(nvmeq, blk_mq_rq_to_pdu(req));
 819                 return -EINTR;
 820         }
 821
 822         if (result)
 823                 *result = cmdinfo.result;
 824
 825         return cmdinfo.status;
 826 }
 827
 828 static int nvme_submit_async_admin_req(struct nvme_dev *dev)
 829 {
 830         struct nvme_queue *nvmeq = dev->queues[0];
 831         struct nvme_command c;
 832         struct nvme_cmd_info *cmd_info;
 833         struct request *req;
 834
 835         req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC, false);
 836         if (IS_ERR(req))
 837                 return PTR_ERR(req);
 838
 839         cmd_info = blk_mq_rq_to_pdu(req);
 840         nvme_set_info(cmd_info, req, async_req_completion);
 841
 842         memset(&c, 0, sizeof(c));
 843         c.common.opcode = nvme_admin_async_event;
 844         c.common.command_id = req->tag;
 845
 846         return __nvme_submit_cmd(nvmeq, &c);
 847 }
 848
 849 static int nvme_submit_admin_async_cmd(struct nvme_dev *dev,
 850                         struct nvme_command *cmd,
 851                         struct async_cmd_info *cmdinfo, unsigned timeout)
 852 {
 853         struct nvme_queue *nvmeq = dev->queues[0];
 854         struct request *req;
 855         struct nvme_cmd_info *cmd_rq;
 856
 857         req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
 858         if (IS_ERR(req))
 859                 return PTR_ERR(req);
 860
 861         req->timeout = timeout;
 862         cmd_rq = blk_mq_rq_to_pdu(req);
 863         cmdinfo->req = req;
 864         nvme_set_info(cmd_rq, cmdinfo, async_completion);
 865         cmdinfo->status = -EINTR;
 866
 867         cmd->common.command_id = req->tag;
 868
 869         return nvme_submit_cmd(nvmeq, cmd);
 870 }
 871
 872 static int __nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 873                                                 u32 *result, unsigned timeout)
 874 {
 875         int res;
 876         struct request *req;
 877
 878         req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_KERNEL, false);
 879         if (!req)
 880                 return -ENOMEM;
 881         res = nvme_submit_sync_cmd(req, cmd, result, timeout);
 882         blk_mq_free_request(req);
 883         return res;
 884 }
 885
 886 int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 887                                                                 u32 *result)
 888 {
 889         return __nvme_submit_admin_cmd(dev, cmd, result, ADMIN_TIMEOUT);
 890 }
 891
 892 int nvme_submit_io_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
 893                                         struct nvme_command *cmd, u32 *result)
 894 {
 895         int res;
 896         struct request *req;
 897
 898         req = blk_mq_alloc_request(ns->queue, WRITE, (GFP_KERNEL|__GFP_WAIT),
 899                                                                         false);
 900         if (!req)
 901                 return -ENOMEM;
 902         res = nvme_submit_sync_cmd(req, cmd, result, NVME_IO_TIMEOUT);
 903         blk_mq_free_request(req);
 904         return res;
 905 }
 906
 907 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
 908 {
 909         struct nvme_command c;
 910
 911         memset(&c, 0, sizeof(c));
 912         c.delete_queue.opcode = opcode;
 913         c.delete_queue.qid = cpu_to_le16(id);
 914
 915         return nvme_submit_admin_cmd(dev, &c, NULL);
 916 }
 917
 918 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
 919                                                 struct nvme_queue *nvmeq)
 920 {
 921         struct nvme_command c;
 922         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
 923
 924         memset(&c, 0, sizeof(c));
 925         c.create_cq.opcode = nvme_admin_create_cq;
 926         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
 927         c.create_cq.cqid = cpu_to_le16(qid);
 928         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 929         c.create_cq.cq_flags = cpu_to_le16(flags);
 930         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
 931
 932         return nvme_submit_admin_cmd(dev, &c, NULL);
 933 }
 934
 935 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
 936                                                 struct nvme_queue *nvmeq)
 937 {
 938         struct nvme_command c;
 939         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
 940
 941         memset(&c, 0, sizeof(c));
 942         c.create_sq.opcode = nvme_admin_create_sq;
 943         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
 944         c.create_sq.sqid = cpu_to_le16(qid);
 945         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 946         c.create_sq.sq_flags = cpu_to_le16(flags);
 947         c.create_sq.cqid = cpu_to_le16(qid);
 948
 949         return nvme_submit_admin_cmd(dev, &c, NULL);
 950 }
 951
 952 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
 953 {
 954         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
 955 }
 956
 957 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
 958 {
 959         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
 960 }
 961
 962 int nvme_identify(struct nvme_dev *dev, unsigned nsid, unsigned cns,
 963                                                         dma_addr_t dma_addr)
 964 {
 965         struct nvme_command c;
 966
 967         memset(&c, 0, sizeof(c));
 968         c.identify.opcode = nvme_admin_identify;
 969         c.identify.nsid = cpu_to_le32(nsid);
 970         c.identify.prp1 = cpu_to_le64(dma_addr);
 971         c.identify.cns = cpu_to_le32(cns);
 972
 973         return nvme_submit_admin_cmd(dev, &c, NULL);
 974 }
 975
 976 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
 977                                         dma_addr_t dma_addr, u32 *result)
 978 {
 979         struct nvme_command c;
 980
 981         memset(&c, 0, sizeof(c));
 982         c.features.opcode = nvme_admin_get_features;
 983         c.features.nsid = cpu_to_le32(nsid);
 984         c.features.prp1 = cpu_to_le64(dma_addr);
 985         c.features.fid = cpu_to_le32(fid);
 986
 987         return nvme_submit_admin_cmd(dev, &c, result);
 988 }
 989
 990 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
 991                                         dma_addr_t dma_addr, u32 *result)
 992 {
 993         struct nvme_command c;
 994
 995         memset(&c, 0, sizeof(c));
 996         c.features.opcode = nvme_admin_set_features;
 997         c.features.prp1 = cpu_to_le64(dma_addr);
 998         c.features.fid = cpu_to_le32(fid);
 999         c.features.dword11 = cpu_to_le32(dword11);
1000
1001         return nvme_submit_admin_cmd(dev, &c, result);
1002 }
1003
1004 /**
1005  * nvme_abort_req - Attempt aborting a request
1006  *
1007  * Schedule controller reset if the command was already aborted once before and
1008  * still hasn't been returned to the driver, or if this is the admin queue.
1009  */
1010 static void nvme_abort_req(struct request *req)
1011 {
1012         struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
1013         struct nvme_queue *nvmeq = cmd_rq->nvmeq;
1014         struct nvme_dev *dev = nvmeq->dev;
1015         struct request *abort_req;
1016         struct nvme_cmd_info *abort_cmd;
1017         struct nvme_command cmd;
1018
1019         if (!nvmeq->qid || cmd_rq->aborted) {
1020                 if (work_busy(&dev->reset_work))
1021                         return;
1022                 list_del_init(&dev->node);
1023                 dev_warn(&dev->pci_dev->dev,
1024                         "I/O %d QID %d timeout, reset controller\n",
1025                                                         req->tag, nvmeq->qid);
1026                 dev->reset_workfn = nvme_reset_failed_dev;
1027                 queue_work(nvme_workq, &dev->reset_work);
1028                 return;
1029         }
1030
1031         if (!dev->abort_limit)
1032                 return;
1033
1034         abort_req = blk_mq_alloc_request(dev->admin_q, WRITE, GFP_ATOMIC,
1035                                                                         false);
1036         if (IS_ERR(abort_req))
1037                 return;
1038
1039         abort_cmd = blk_mq_rq_to_pdu(abort_req);
1040         nvme_set_info(abort_cmd, abort_req, abort_completion);
1041
1042         memset(&cmd, 0, sizeof(cmd));
1043         cmd.abort.opcode = nvme_admin_abort_cmd;
1044         cmd.abort.cid = req->tag;
1045         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1046         cmd.abort.command_id = abort_req->tag;
1047
1048         --dev->abort_limit;
1049         cmd_rq->aborted = 1;
1050
1051         dev_warn(nvmeq->q_dmadev, "Aborting I/O %d QID %d\n", req->tag,
1052                                                         nvmeq->qid);
1053         if (nvme_submit_cmd(dev->queues[0], &cmd) < 0) {
1054                 dev_warn(nvmeq->q_dmadev,
1055                                 "Could not abort I/O %d QID %d",
1056                                 req->tag, nvmeq->qid);
1057                 blk_mq_free_request(req);
1058         }
1059 }
1060
1061 static void nvme_cancel_queue_ios(struct blk_mq_hw_ctx *hctx,
1062                                 struct request *req, void *data, bool reserved)
1063 {
1064         struct nvme_queue *nvmeq = data;
1065         void *ctx;
1066         nvme_completion_fn fn;
1067         struct nvme_cmd_info *cmd;
1068         static struct nvme_completion cqe = {
1069                 .status = cpu_to_le16(NVME_SC_ABORT_REQ << 1),
1070         };
1071
1072         cmd = blk_mq_rq_to_pdu(req);
1073
1074         if (cmd->ctx == CMD_CTX_CANCELLED)
1075                 return;
1076
1077         dev_warn(nvmeq->q_dmadev, "Cancelling I/O %d QID %d\n",
1078                                                 req->tag, nvmeq->qid);
1079         ctx = cancel_cmd_info(cmd, &fn);
1080         fn(nvmeq, ctx, &cqe);
1081 }
1082
1083 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1084 {
1085         struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
1086         struct nvme_queue *nvmeq = cmd->nvmeq;
1087
1088         dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
1089                                                         nvmeq->qid);
1090         if (nvmeq->dev->initialized)
1091                 nvme_abort_req(req);
1092
1093         /*
1094          * The aborted req will be completed on receiving the abort req.
1095          * We enable the timer again. If hit twice, it'll cause a device reset,
1096          * as the device then is in a faulty state.
1097          */
1098         return BLK_EH_RESET_TIMER;
1099 }
1100
1101 static void nvme_free_queue(struct nvme_queue *nvmeq)
1102 {
1103         dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1104                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1105         dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1106                                         nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1107         kfree(nvmeq);
1108 }
1109
1110 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1111 {
1112         LLIST_HEAD(q_list);
1113         struct nvme_queue *nvmeq, *next;
1114         struct llist_node *entry;
1115         int i;
1116
1117         for (i = dev->queue_count - 1; i >= lowest; i--) {
1118                 struct nvme_queue *nvmeq = dev->queues[i];
1119                 llist_add(&nvmeq->node, &q_list);
1120                 dev->queue_count--;
1121                 dev->queues[i] = NULL;
1122         }
1123         synchronize_rcu();
1124         entry = llist_del_all(&q_list);
1125         llist_for_each_entry_safe(nvmeq, next, entry, node)
1126                 nvme_free_queue(nvmeq);
1127 }
1128
1129 /**
1130  * nvme_suspend_queue - put queue into suspended state
1131  * @nvmeq - queue to suspend
1132  */
1133 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1134 {
1135         int vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1136
1137         spin_lock_irq(&nvmeq->q_lock);
1138         nvmeq->dev->online_queues--;
1139         spin_unlock_irq(&nvmeq->q_lock);
1140
1141         irq_set_affinity_hint(vector, NULL);
1142         free_irq(vector, nvmeq);
1143
1144         return 0;
1145 }
1146
1147 static void nvme_clear_queue(struct nvme_queue *nvmeq)
1148 {
1149         struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
1150
1151         spin_lock_irq(&nvmeq->q_lock);
1152         nvme_process_cq(nvmeq);
1153         if (hctx && hctx->tags)
1154                 blk_mq_tag_busy_iter(hctx, nvme_cancel_queue_ios, nvmeq);
1155         spin_unlock_irq(&nvmeq->q_lock);
1156 }
1157
1158 static void nvme_disable_queue(struct nvme_dev *dev, int qid)
1159 {
1160         struct nvme_queue *nvmeq = dev->queues[qid];
1161
1162         if (!nvmeq)
1163                 return;
1164         if (nvme_suspend_queue(nvmeq))
1165                 return;
1166
1167         /* Don't tell the adapter to delete the admin queue.
1168          * Don't tell a removed adapter to delete IO queues. */
1169         if (qid && readl(&dev->bar->csts) != -1) {
1170                 adapter_delete_sq(dev, qid);
1171                 adapter_delete_cq(dev, qid);
1172         }
1173         nvme_clear_queue(nvmeq);
1174 }
1175
1176 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1177                                                         int depth, int vector)
1178 {
1179         struct device *dmadev = &dev->pci_dev->dev;
1180         struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1181         if (!nvmeq)
1182                 return NULL;
1183
1184         nvmeq->cqes = dma_zalloc_coherent(dmadev, CQ_SIZE(depth),
1185                                           &nvmeq->cq_dma_addr, GFP_KERNEL);
1186         if (!nvmeq->cqes)
1187                 goto free_nvmeq;
1188
1189         nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
1190                                         &nvmeq->sq_dma_addr, GFP_KERNEL);
1191         if (!nvmeq->sq_cmds)
1192                 goto free_cqdma;
1193
1194         nvmeq->q_dmadev = dmadev;
1195         nvmeq->dev = dev;
1196         snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1197                         dev->instance, qid);
1198         spin_lock_init(&nvmeq->q_lock);
1199         nvmeq->cq_head = 0;
1200         nvmeq->cq_phase = 1;
1201         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1202         nvmeq->q_depth = depth;
1203         nvmeq->cq_vector = vector;
1204         nvmeq->qid = qid;
1205         dev->queue_count++;
1206         dev->queues[qid] = nvmeq;
1207
1208         return nvmeq;
1209
1210  free_cqdma:
1211         dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1212                                                         nvmeq->cq_dma_addr);
1213  free_nvmeq:
1214         kfree(nvmeq);
1215         return NULL;
1216 }
1217
1218 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1219                                                         const char *name)
1220 {
1221         if (use_threaded_interrupts)
1222                 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1223                                         nvme_irq_check, nvme_irq, IRQF_SHARED,
1224                                         name, nvmeq);
1225         return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1226                                 IRQF_SHARED, name, nvmeq);
1227 }
1228
1229 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1230 {
1231         struct nvme_dev *dev = nvmeq->dev;
1232
1233         spin_lock_irq(&nvmeq->q_lock);
1234         nvmeq->sq_tail = 0;
1235         nvmeq->cq_head = 0;
1236         nvmeq->cq_phase = 1;
1237         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1238         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1239         dev->online_queues++;
1240         spin_unlock_irq(&nvmeq->q_lock);
1241 }
1242
1243 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1244 {
1245         struct nvme_dev *dev = nvmeq->dev;
1246         int result;
1247
1248         result = adapter_alloc_cq(dev, qid, nvmeq);
1249         if (result < 0)
1250                 return result;
1251
1252         result = adapter_alloc_sq(dev, qid, nvmeq);
1253         if (result < 0)
1254                 goto release_cq;
1255
1256         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1257         if (result < 0)
1258                 goto release_sq;
1259
1260         nvme_init_queue(nvmeq, qid);
1261         return result;
1262
1263  release_sq:
1264         adapter_delete_sq(dev, qid);
1265  release_cq:
1266         adapter_delete_cq(dev, qid);
1267         return result;
1268 }
1269
1270 static int nvme_wait_ready(struct nvme_dev *dev, u64 cap, bool enabled)
1271 {
1272         unsigned long timeout;
1273         u32 bit = enabled ? NVME_CSTS_RDY : 0;
1274
1275         timeout = ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
1276
1277         while ((readl(&dev->bar->csts) & NVME_CSTS_RDY) != bit) {
1278                 msleep(100);
1279                 if (fatal_signal_pending(current))
1280                         return -EINTR;
1281                 if (time_after(jiffies, timeout)) {
1282                         dev_err(&dev->pci_dev->dev,
1283                                 "Device not ready; aborting %s\n", enabled ?
1284                                                 "initialisation" : "reset");
1285                         return -ENODEV;
1286                 }
1287         }
1288
1289         return 0;
1290 }
1291
1292 /*
1293  * If the device has been passed off to us in an enabled state, just clear
1294  * the enabled bit.  The spec says we should set the 'shutdown notification
1295  * bits', but doing so may cause the device to complete commands to the
1296  * admin queue ... and we don't know what memory that might be pointing at!
1297  */
1298 static int nvme_disable_ctrl(struct nvme_dev *dev, u64 cap)
1299 {
1300         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1301         dev->ctrl_config &= ~NVME_CC_ENABLE;
1302         writel(dev->ctrl_config, &dev->bar->cc);
1303
1304         return nvme_wait_ready(dev, cap, false);
1305 }
1306
1307 static int nvme_enable_ctrl(struct nvme_dev *dev, u64 cap)
1308 {
1309         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1310         dev->ctrl_config |= NVME_CC_ENABLE;
1311         writel(dev->ctrl_config, &dev->bar->cc);
1312
1313         return nvme_wait_ready(dev, cap, true);
1314 }
1315
1316 static int nvme_shutdown_ctrl(struct nvme_dev *dev)
1317 {
1318         unsigned long timeout;
1319
1320         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
1321         dev->ctrl_config |= NVME_CC_SHN_NORMAL;
1322
1323         writel(dev->ctrl_config, &dev->bar->cc);
1324
1325         timeout = SHUTDOWN_TIMEOUT + jiffies;
1326         while ((readl(&dev->bar->csts) & NVME_CSTS_SHST_MASK) !=
1327                                                         NVME_CSTS_SHST_CMPLT) {
1328                 msleep(100);
1329                 if (fatal_signal_pending(current))
1330                         return -EINTR;
1331                 if (time_after(jiffies, timeout)) {
1332                         dev_err(&dev->pci_dev->dev,
1333                                 "Device shutdown incomplete; abort shutdown\n");
1334                         return -ENODEV;
1335                 }
1336         }
1337
1338         return 0;
1339 }
1340
1341 static struct blk_mq_ops nvme_mq_admin_ops = {
1342         .queue_rq       = nvme_admin_queue_rq,
1343         .map_queue      = blk_mq_map_queue,
1344         .init_hctx      = nvme_admin_init_hctx,
1345         .exit_hctx      = nvme_exit_hctx,
1346         .init_request   = nvme_admin_init_request,
1347         .timeout        = nvme_timeout,
1348 };
1349
1350 static struct blk_mq_ops nvme_mq_ops = {
1351         .queue_rq       = nvme_queue_rq,
1352         .map_queue      = blk_mq_map_queue,
1353         .init_hctx      = nvme_init_hctx,
1354         .exit_hctx      = nvme_exit_hctx,
1355         .init_request   = nvme_init_request,
1356         .timeout        = nvme_timeout,
1357 };
1358
1359 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1360 {
1361         if (!dev->admin_q) {
1362                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1363                 dev->admin_tagset.nr_hw_queues = 1;
1364                 dev->admin_tagset.queue_depth = NVME_AQ_DEPTH - 1;
1365                 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1366                 dev->admin_tagset.numa_node = dev_to_node(&dev->pci_dev->dev);
1367                 dev->admin_tagset.cmd_size = sizeof(struct nvme_cmd_info);
1368                 dev->admin_tagset.driver_data = dev;
1369
1370                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1371                         return -ENOMEM;
1372
1373                 dev->admin_q = blk_mq_init_queue(&dev->admin_tagset);
1374                 if (!dev->admin_q) {
1375                         blk_mq_free_tag_set(&dev->admin_tagset);
1376                         return -ENOMEM;
1377                 }
1378         }
1379
1380         return 0;
1381 }
1382
1383 static void nvme_free_admin_tags(struct nvme_dev *dev)
1384 {
1385         if (dev->admin_q)
1386                 blk_mq_free_tag_set(&dev->admin_tagset);
1387 }
1388
1389 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1390 {
1391         int result;
1392         u32 aqa;
1393         u64 cap = readq(&dev->bar->cap);
1394         struct nvme_queue *nvmeq;
1395         unsigned page_shift = PAGE_SHIFT;
1396         unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
1397         unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
1398
1399         if (page_shift < dev_page_min) {
1400                 dev_err(&dev->pci_dev->dev,
1401                                 "Minimum device page size (%u) too large for "
1402                                 "host (%u)\n", 1 << dev_page_min,
1403                                 1 << page_shift);
1404                 return -ENODEV;
1405         }
1406         if (page_shift > dev_page_max) {
1407                 dev_info(&dev->pci_dev->dev,
1408                                 "Device maximum page size (%u) smaller than "
1409                                 "host (%u); enabling work-around\n",
1410                                 1 << dev_page_max, 1 << page_shift);
1411                 page_shift = dev_page_max;
1412         }
1413
1414         result = nvme_disable_ctrl(dev, cap);
1415         if (result < 0)
1416                 return result;
1417
1418         nvmeq = dev->queues[0];
1419         if (!nvmeq) {
1420                 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH, 0);
1421                 if (!nvmeq)
1422                         return -ENOMEM;
1423         }
1424
1425         aqa = nvmeq->q_depth - 1;
1426         aqa |= aqa << 16;
1427
1428         dev->page_size = 1 << page_shift;
1429
1430         dev->ctrl_config = NVME_CC_CSS_NVM;
1431         dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
1432         dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
1433         dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
1434
1435         writel(aqa, &dev->bar->aqa);
1436         writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
1437         writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
1438
1439         result = nvme_enable_ctrl(dev, cap);
1440         if (result)
1441                 goto free_nvmeq;
1442
1443         result = nvme_alloc_admin_tags(dev);
1444         if (result)
1445                 goto free_nvmeq;
1446
1447         result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1448         if (result)
1449                 goto free_tags;
1450
1451         return result;
1452
1453  free_tags:
1454         nvme_free_admin_tags(dev);
1455  free_nvmeq:
1456         nvme_free_queues(dev, 0);
1457         return result;
1458 }
1459
1460 struct nvme_iod *nvme_map_user_pages(struct nvme_dev *dev, int write,
1461                                 unsigned long addr, unsigned length)
1462 {
1463         int i, err, count, nents, offset;
1464         struct scatterlist *sg;
1465         struct page **pages;
1466         struct nvme_iod *iod;
1467
1468         if (addr & 3)
1469                 return ERR_PTR(-EINVAL);
1470         if (!length || length > INT_MAX - PAGE_SIZE)
1471                 return ERR_PTR(-EINVAL);
1472
1473         offset = offset_in_page(addr);
1474         count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
1475         pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
1476         if (!pages)
1477                 return ERR_PTR(-ENOMEM);
1478
1479         err = get_user_pages_fast(addr, count, 1, pages);
1480         if (err < count) {
1481                 count = err;
1482                 err = -EFAULT;
1483                 goto put_pages;
1484         }
1485
1486         err = -ENOMEM;
1487         iod = nvme_alloc_iod(count, length, dev, GFP_KERNEL);
1488         if (!iod)
1489                 goto put_pages;
1490
1491         sg = iod->sg;
1492         sg_init_table(sg, count);
1493         for (i = 0; i < count; i++) {
1494                 sg_set_page(&sg[i], pages[i],
1495                             min_t(unsigned, length, PAGE_SIZE - offset),
1496                             offset);
1497                 length -= (PAGE_SIZE - offset);
1498                 offset = 0;
1499         }
1500         sg_mark_end(&sg[i - 1]);
1501         iod->nents = count;
1502
1503         nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
1504                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1505         if (!nents)
1506                 goto free_iod;
1507
1508         kfree(pages);
1509         return iod;
1510
1511  free_iod:
1512         kfree(iod);
1513  put_pages:
1514         for (i = 0; i < count; i++)
1515                 put_page(pages[i]);
1516         kfree(pages);
1517         return ERR_PTR(err);
1518 }
1519
1520 void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
1521                         struct nvme_iod *iod)
1522 {
1523         int i;
1524
1525         dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
1526                                 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1527
1528         for (i = 0; i < iod->nents; i++)
1529                 put_page(sg_page(&iod->sg[i]));
1530 }
1531
1532 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1533 {
1534         struct nvme_dev *dev = ns->dev;
1535         struct nvme_user_io io;
1536         struct nvme_command c;
1537         unsigned length, meta_len;
1538         int status, i;
1539         struct nvme_iod *iod, *meta_iod = NULL;
1540         dma_addr_t meta_dma_addr;
1541         void *meta, *uninitialized_var(meta_mem);
1542
1543         if (copy_from_user(&io, uio, sizeof(io)))
1544                 return -EFAULT;
1545         length = (io.nblocks + 1) << ns->lba_shift;
1546         meta_len = (io.nblocks + 1) * ns->ms;
1547
1548         if (meta_len && ((io.metadata & 3) || !io.metadata))
1549                 return -EINVAL;
1550
1551         switch (io.opcode) {
1552         case nvme_cmd_write:
1553         case nvme_cmd_read:
1554         case nvme_cmd_compare:
1555                 iod = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length);
1556                 break;
1557         default:
1558                 return -EINVAL;
1559         }
1560
1561         if (IS_ERR(iod))
1562                 return PTR_ERR(iod);
1563
1564         memset(&c, 0, sizeof(c));
1565         c.rw.opcode = io.opcode;
1566         c.rw.flags = io.flags;
1567         c.rw.nsid = cpu_to_le32(ns->ns_id);
1568         c.rw.slba = cpu_to_le64(io.slba);
1569         c.rw.length = cpu_to_le16(io.nblocks);
1570         c.rw.control = cpu_to_le16(io.control);
1571         c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1572         c.rw.reftag = cpu_to_le32(io.reftag);
1573         c.rw.apptag = cpu_to_le16(io.apptag);
1574         c.rw.appmask = cpu_to_le16(io.appmask);
1575
1576         if (meta_len) {
1577                 meta_iod = nvme_map_user_pages(dev, io.opcode & 1, io.metadata,
1578                                                                 meta_len);
1579                 if (IS_ERR(meta_iod)) {
1580                         status = PTR_ERR(meta_iod);
1581                         meta_iod = NULL;
1582                         goto unmap;
1583                 }
1584
1585                 meta_mem = dma_alloc_coherent(&dev->pci_dev->dev, meta_len,
1586                                                 &meta_dma_addr, GFP_KERNEL);
1587                 if (!meta_mem) {
1588                         status = -ENOMEM;
1589                         goto unmap;
1590                 }
1591
1592                 if (io.opcode & 1) {
1593                         int meta_offset = 0;
1594
1595                         for (i = 0; i < meta_iod->nents; i++) {
1596                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1597                                                 meta_iod->sg[i].offset;
1598                                 memcpy(meta_mem + meta_offset, meta,
1599                                                 meta_iod->sg[i].length);
1600                                 kunmap_atomic(meta);
1601                                 meta_offset += meta_iod->sg[i].length;
1602                         }
1603                 }
1604
1605                 c.rw.metadata = cpu_to_le64(meta_dma_addr);
1606         }
1607
1608         length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1609         c.rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1610         c.rw.prp2 = cpu_to_le64(iod->first_dma);
1611
1612         if (length != (io.nblocks + 1) << ns->lba_shift)
1613                 status = -ENOMEM;
1614         else
1615                 status = nvme_submit_io_cmd(dev, ns, &c, NULL);
1616
1617         if (meta_len) {
1618                 if (status == NVME_SC_SUCCESS && !(io.opcode & 1)) {
1619                         int meta_offset = 0;
1620
1621                         for (i = 0; i < meta_iod->nents; i++) {
1622                                 meta = kmap_atomic(sg_page(&meta_iod->sg[i])) +
1623                                                 meta_iod->sg[i].offset;
1624                                 memcpy(meta, meta_mem + meta_offset,
1625                                                 meta_iod->sg[i].length);
1626                                 kunmap_atomic(meta);
1627                                 meta_offset += meta_iod->sg[i].length;
1628                         }
1629                 }
1630
1631                 dma_free_coherent(&dev->pci_dev->dev, meta_len, meta_mem,
1632                                                                 meta_dma_addr);
1633         }
1634
1635  unmap:
1636         nvme_unmap_user_pages(dev, io.opcode & 1, iod);
1637         nvme_free_iod(dev, iod);
1638
1639         if (meta_iod) {
1640                 nvme_unmap_user_pages(dev, io.opcode & 1, meta_iod);
1641                 nvme_free_iod(dev, meta_iod);
1642         }
1643
1644         return status;
1645 }
1646
1647 static int nvme_user_cmd(struct nvme_dev *dev, struct nvme_ns *ns,
1648                         struct nvme_passthru_cmd __user *ucmd)
1649 {
1650         struct nvme_passthru_cmd cmd;
1651         struct nvme_command c;
1652         int status, length;
1653         struct nvme_iod *uninitialized_var(iod);
1654         unsigned timeout;
1655
1656         if (!capable(CAP_SYS_ADMIN))
1657                 return -EACCES;
1658         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1659                 return -EFAULT;
1660
1661         memset(&c, 0, sizeof(c));
1662         c.common.opcode = cmd.opcode;
1663         c.common.flags = cmd.flags;
1664         c.common.nsid = cpu_to_le32(cmd.nsid);
1665         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1666         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1667         c.common.cdw10[0] = cpu_to_le32(cmd.cdw10);
1668         c.common.cdw10[1] = cpu_to_le32(cmd.cdw11);
1669         c.common.cdw10[2] = cpu_to_le32(cmd.cdw12);
1670         c.common.cdw10[3] = cpu_to_le32(cmd.cdw13);
1671         c.common.cdw10[4] = cpu_to_le32(cmd.cdw14);
1672         c.common.cdw10[5] = cpu_to_le32(cmd.cdw15);
1673
1674         length = cmd.data_len;
1675         if (cmd.data_len) {
1676                 iod = nvme_map_user_pages(dev, cmd.opcode & 1, cmd.addr,
1677                                                                 length);
1678                 if (IS_ERR(iod))
1679                         return PTR_ERR(iod);
1680                 length = nvme_setup_prps(dev, iod, length, GFP_KERNEL);
1681                 c.common.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
1682                 c.common.prp2 = cpu_to_le64(iod->first_dma);
1683         }
1684
1685         timeout = cmd.timeout_ms ? msecs_to_jiffies(cmd.timeout_ms) :
1686                                                                 ADMIN_TIMEOUT;
1687
1688         if (length != cmd.data_len)
1689                 status = -ENOMEM;
1690         else if (ns) {
1691                 struct request *req;
1692
1693                 req = blk_mq_alloc_request(ns->queue, WRITE,
1694                                                 (GFP_KERNEL|__GFP_WAIT), false);
1695                 if (!req)
1696                         status = -ENOMEM;
1697                 else {
1698                         status = nvme_submit_sync_cmd(req, &c, &cmd.result,
1699                                                                 timeout);
1700                         blk_mq_free_request(req);
1701                 }
1702         } else
1703                 status = __nvme_submit_admin_cmd(dev, &c, &cmd.result, timeout);
1704
1705         if (cmd.data_len) {
1706                 nvme_unmap_user_pages(dev, cmd.opcode & 1, iod);
1707                 nvme_free_iod(dev, iod);
1708         }
1709
1710         if ((status >= 0) && copy_to_user(&ucmd->result, &cmd.result,
1711                                                         sizeof(cmd.result)))
1712                 status = -EFAULT;
1713
1714         return status;
1715 }
1716
1717 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1718                                                         unsigned long arg)
1719 {
1720         struct nvme_ns *ns = bdev->bd_disk->private_data;
1721
1722         switch (cmd) {
1723         case NVME_IOCTL_ID:
1724                 force_successful_syscall_return();
1725                 return ns->ns_id;
1726         case NVME_IOCTL_ADMIN_CMD:
1727                 return nvme_user_cmd(ns->dev, NULL, (void __user *)arg);
1728         case NVME_IOCTL_IO_CMD:
1729                 return nvme_user_cmd(ns->dev, ns, (void __user *)arg);
1730         case NVME_IOCTL_SUBMIT_IO:
1731                 return nvme_submit_io(ns, (void __user *)arg);
1732         case SG_GET_VERSION_NUM:
1733                 return nvme_sg_get_version_num((void __user *)arg);
1734         case SG_IO:
1735                 return nvme_sg_io(ns, (void __user *)arg);
1736         default:
1737                 return -ENOTTY;
1738         }
1739 }
1740
1741 #ifdef CONFIG_COMPAT
1742 static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
1743                                         unsigned int cmd, unsigned long arg)
1744 {
1745         switch (cmd) {
1746         case SG_IO:
1747                 return -ENOIOCTLCMD;
1748         }
1749         return nvme_ioctl(bdev, mode, cmd, arg);
1750 }
1751 #else
1752 #define nvme_compat_ioctl       NULL
1753 #endif
1754
1755 static int nvme_open(struct block_device *bdev, fmode_t mode)
1756 {
1757         int ret = 0;
1758         struct nvme_ns *ns;
1759
1760         spin_lock(&dev_list_lock);
1761         ns = bdev->bd_disk->private_data;
1762         if (!ns)
1763                 ret = -ENXIO;
1764         else if (!kref_get_unless_zero(&ns->dev->kref))
1765                 ret = -ENXIO;
1766         spin_unlock(&dev_list_lock);
1767
1768         return ret;
1769 }
1770
1771 static void nvme_free_dev(struct kref *kref);
1772
1773 static void nvme_release(struct gendisk *disk, fmode_t mode)
1774 {
1775         struct nvme_ns *ns = disk->private_data;
1776         struct nvme_dev *dev = ns->dev;
1777
1778         kref_put(&dev->kref, nvme_free_dev);
1779 }
1780
1781 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
1782 {
1783         /* some standard values */
1784         geo->heads = 1 << 6;
1785         geo->sectors = 1 << 5;
1786         geo->cylinders = get_capacity(bd->bd_disk) >> 11;
1787         return 0;
1788 }
1789
1790 static int nvme_revalidate_disk(struct gendisk *disk)
1791 {
1792         struct nvme_ns *ns = disk->private_data;
1793         struct nvme_dev *dev = ns->dev;
1794         struct nvme_id_ns *id;
1795         dma_addr_t dma_addr;
1796         int lbaf;
1797
1798         id = dma_alloc_coherent(&dev->pci_dev->dev, 4096, &dma_addr,
1799                                                                 GFP_KERNEL);
1800         if (!id) {
1801                 dev_warn(&dev->pci_dev->dev, "%s: Memory alocation failure\n",
1802                                                                 __func__);
1803                 return 0;
1804         }
1805
1806         if (nvme_identify(dev, ns->ns_id, 0, dma_addr))
1807                 goto free;
1808
1809         lbaf = id->flbas & 0xf;
1810         ns->lba_shift = id->lbaf[lbaf].ds;
1811
1812         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1813         set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1814  free:
1815         dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1816         return 0;
1817 }
1818
1819 static const struct block_device_operations nvme_fops = {
1820         .owner          = THIS_MODULE,
1821         .ioctl          = nvme_ioctl,
1822         .compat_ioctl   = nvme_compat_ioctl,
1823         .open           = nvme_open,
1824         .release        = nvme_release,
1825         .getgeo         = nvme_getgeo,
1826         .revalidate_disk= nvme_revalidate_disk,
1827 };
1828
1829 static int nvme_kthread(void *data)
1830 {
1831         struct nvme_dev *dev, *next;
1832
1833         while (!kthread_should_stop()) {
1834                 set_current_state(TASK_INTERRUPTIBLE);
1835                 spin_lock(&dev_list_lock);
1836                 list_for_each_entry_safe(dev, next, &dev_list, node) {
1837                         int i;
1838                         if (readl(&dev->bar->csts) & NVME_CSTS_CFS &&
1839                                                         dev->initialized) {
1840                                 if (work_busy(&dev->reset_work))
1841                                         continue;
1842                                 list_del_init(&dev->node);
1843                                 dev_warn(&dev->pci_dev->dev,
1844                                         "Failed status: %x, reset controller\n",
1845                                         readl(&dev->bar->csts));
1846                                 dev->reset_workfn = nvme_reset_failed_dev;
1847                                 queue_work(nvme_workq, &dev->reset_work);
1848                                 continue;
1849                         }
1850                         for (i = 0; i < dev->queue_count; i++) {
1851                                 struct nvme_queue *nvmeq = dev->queues[i];
1852                                 if (!nvmeq)
1853                                         continue;
1854                                 spin_lock_irq(&nvmeq->q_lock);
1855                                 nvme_process_cq(nvmeq);
1856
1857                                 while ((i == 0) && (dev->event_limit > 0)) {
1858                                         if (nvme_submit_async_admin_req(dev))
1859                                                 break;
1860                                         dev->event_limit--;
1861                                 }
1862                                 spin_unlock_irq(&nvmeq->q_lock);
1863                         }
1864                 }
1865                 spin_unlock(&dev_list_lock);
1866                 schedule_timeout(round_jiffies_relative(HZ));
1867         }
1868         return 0;
1869 }
1870
1871 static void nvme_config_discard(struct nvme_ns *ns)
1872 {
1873         u32 logical_block_size = queue_logical_block_size(ns->queue);
1874         ns->queue->limits.discard_zeroes_data = 0;
1875         ns->queue->limits.discard_alignment = logical_block_size;
1876         ns->queue->limits.discard_granularity = logical_block_size;
1877         ns->queue->limits.max_discard_sectors = 0xffffffff;
1878         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, ns->queue);
1879 }
1880
1881 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
1882                         struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1883 {
1884         struct nvme_ns *ns;
1885         struct gendisk *disk;
1886         int node = dev_to_node(&dev->pci_dev->dev);
1887         int lbaf;
1888
1889         if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1890                 return NULL;
1891
1892         ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
1893         if (!ns)
1894                 return NULL;
1895         ns->queue = blk_mq_init_queue(&dev->tagset);
1896         if (IS_ERR(ns->queue))
1897                 goto out_free_ns;
1898         queue_flag_set_unlocked(QUEUE_FLAG_NOMERGES, ns->queue);
1899         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, ns->queue);
1900         queue_flag_set_unlocked(QUEUE_FLAG_SG_GAPS, ns->queue);
1901         ns->dev = dev;
1902         ns->queue->queuedata = ns;
1903
1904         disk = alloc_disk_node(0, node);
1905         if (!disk)
1906                 goto out_free_queue;
1907
1908         ns->ns_id = nsid;
1909         ns->disk = disk;
1910         lbaf = id->flbas & 0xf;
1911         ns->lba_shift = id->lbaf[lbaf].ds;
1912         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1913         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
1914         if (dev->max_hw_sectors)
1915                 blk_queue_max_hw_sectors(ns->queue, dev->max_hw_sectors);
1916         if (dev->stripe_size)
1917                 blk_queue_chunk_sectors(ns->queue, dev->stripe_size >> 9);
1918         if (dev->vwc & NVME_CTRL_VWC_PRESENT)
1919                 blk_queue_flush(ns->queue, REQ_FLUSH | REQ_FUA);
1920
1921         disk->major = nvme_major;
1922         disk->first_minor = 0;
1923         disk->fops = &nvme_fops;
1924         disk->private_data = ns;
1925         disk->queue = ns->queue;
1926         disk->driverfs_dev = &dev->pci_dev->dev;
1927         disk->flags = GENHD_FL_EXT_DEVT;
1928         sprintf(disk->disk_name, "nvme%dn%d", dev->instance, nsid);
1929         set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1930
1931         if (dev->oncs & NVME_CTRL_ONCS_DSM)
1932                 nvme_config_discard(ns);
1933
1934         return ns;
1935
1936  out_free_queue:
1937         blk_cleanup_queue(ns->queue);
1938  out_free_ns:
1939         kfree(ns);
1940         return NULL;
1941 }
1942
1943 static void nvme_create_io_queues(struct nvme_dev *dev)
1944 {
1945         unsigned i;
1946
1947         for (i = dev->queue_count; i <= dev->max_qid; i++)
1948                 if (!nvme_alloc_queue(dev, i, dev->q_depth, i - 1))
1949                         break;
1950
1951         for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
1952                 if (nvme_create_queue(dev->queues[i], i))
1953                         break;
1954 }
1955
1956 static int set_queue_count(struct nvme_dev *dev, int count)
1957 {
1958         int status;
1959         u32 result;
1960         u32 q_count = (count - 1) | ((count - 1) << 16);
1961
1962         status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES, q_count, 0,
1963                                                                 &result);
1964         if (status < 0)
1965                 return status;
1966         if (status > 0) {
1967                 dev_err(&dev->pci_dev->dev, "Could not set queue count (%d)\n",
1968                                                                         status);
1969                 return 0;
1970         }
1971         return min(result & 0xffff, result >> 16) + 1;
1972 }
1973
1974 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1975 {
1976         return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1977 }
1978
1979 static int nvme_setup_io_queues(struct nvme_dev *dev)
1980 {
1981         struct nvme_queue *adminq = dev->queues[0];
1982         struct pci_dev *pdev = dev->pci_dev;
1983         int result, i, vecs, nr_io_queues, size;
1984
1985         nr_io_queues = num_possible_cpus();
1986         result = set_queue_count(dev, nr_io_queues);
1987         if (result <= 0)
1988                 return result;
1989         if (result < nr_io_queues)
1990                 nr_io_queues = result;
1991
1992         size = db_bar_size(dev, nr_io_queues);
1993         if (size > 8192) {
1994                 iounmap(dev->bar);
1995                 do {
1996                         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1997                         if (dev->bar)
1998                                 break;
1999                         if (!--nr_io_queues)
2000                                 return -ENOMEM;
2001                         size = db_bar_size(dev, nr_io_queues);
2002                 } while (1);
2003                 dev->dbs = ((void __iomem *)dev->bar) + 4096;
2004                 adminq->q_db = dev->dbs;
2005         }
2006
2007         /* Deregister the admin queue's interrupt */
2008         free_irq(dev->entry[0].vector, adminq);
2009
2010         /*
2011          * If we enable msix early due to not intx, disable it again before
2012          * setting up the full range we need.
2013          */
2014         if (!pdev->irq)
2015                 pci_disable_msix(pdev);
2016
2017         for (i = 0; i < nr_io_queues; i++)
2018                 dev->entry[i].entry = i;
2019         vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
2020         if (vecs < 0) {
2021                 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
2022                 if (vecs < 0) {
2023                         vecs = 1;
2024                 } else {
2025                         for (i = 0; i < vecs; i++)
2026                                 dev->entry[i].vector = i + pdev->irq;
2027                 }
2028         }
2029
2030         /*
2031          * Should investigate if there's a performance win from allocating
2032          * more queues than interrupt vectors; it might allow the submission
2033          * path to scale better, even if the receive path is limited by the
2034          * number of interrupts.
2035          */
2036         nr_io_queues = vecs;
2037         dev->max_qid = nr_io_queues;
2038
2039         result = queue_request_irq(dev, adminq, adminq->irqname);
2040         if (result)
2041                 goto free_queues;
2042
2043         /* Free previously allocated queues that are no longer usable */
2044         nvme_free_queues(dev, nr_io_queues + 1);
2045         nvme_create_io_queues(dev);
2046
2047         return 0;
2048
2049  free_queues:
2050         nvme_free_queues(dev, 1);
2051         return result;
2052 }
2053
2054 /*
2055  * Return: error value if an error occurred setting up the queues or calling
2056  * Identify Device.  0 if these succeeded, even if adding some of the
2057  * namespaces failed.  At the moment, these failures are silent.  TBD which
2058  * failures should be reported.
2059  */
2060 static int nvme_dev_add(struct nvme_dev *dev)
2061 {
2062         struct pci_dev *pdev = dev->pci_dev;
2063         int res;
2064         unsigned nn, i;
2065         struct nvme_ns *ns;
2066         struct nvme_id_ctrl *ctrl;
2067         struct nvme_id_ns *id_ns;
2068         void *mem;
2069         dma_addr_t dma_addr;
2070         int shift = NVME_CAP_MPSMIN(readq(&dev->bar->cap)) + 12;
2071
2072         mem = dma_alloc_coherent(&pdev->dev, 8192, &dma_addr, GFP_KERNEL);
2073         if (!mem)
2074                 return -ENOMEM;
2075
2076         res = nvme_identify(dev, 0, 1, dma_addr);
2077         if (res) {
2078                 dev_err(&pdev->dev, "Identify Controller failed (%d)\n", res);
2079                 res = -EIO;
2080                 goto out;
2081         }
2082
2083         ctrl = mem;
2084         nn = le32_to_cpup(&ctrl->nn);
2085         dev->oncs = le16_to_cpup(&ctrl->oncs);
2086         dev->abort_limit = ctrl->acl + 1;
2087         dev->vwc = ctrl->vwc;
2088         dev->event_limit = min(ctrl->aerl + 1, 8);
2089         memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
2090         memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
2091         memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
2092         if (ctrl->mdts)
2093                 dev->max_hw_sectors = 1 << (ctrl->mdts + shift - 9);
2094         if ((pdev->vendor == PCI_VENDOR_ID_INTEL) &&
2095                         (pdev->device == 0x0953) && ctrl->vs[3]) {
2096                 unsigned int max_hw_sectors;
2097
2098                 dev->stripe_size = 1 << (ctrl->vs[3] + shift);
2099                 max_hw_sectors = dev->stripe_size >> (shift - 9);
2100                 if (dev->max_hw_sectors) {
2101                         dev->max_hw_sectors = min(max_hw_sectors,
2102                                                         dev->max_hw_sectors);
2103                 } else
2104                         dev->max_hw_sectors = max_hw_sectors;
2105         }
2106
2107         dev->tagset.ops = &nvme_mq_ops;
2108         dev->tagset.nr_hw_queues = dev->online_queues - 1;
2109         dev->tagset.timeout = NVME_IO_TIMEOUT;
2110         dev->tagset.numa_node = dev_to_node(&dev->pci_dev->dev);
2111         dev->tagset.queue_depth =
2112                                 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2113         dev->tagset.cmd_size = sizeof(struct nvme_cmd_info);
2114         dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2115         dev->tagset.driver_data = dev;
2116
2117         if (blk_mq_alloc_tag_set(&dev->tagset))
2118                 goto out;
2119
2120         id_ns = mem;
2121         for (i = 1; i <= nn; i++) {
2122                 res = nvme_identify(dev, i, 0, dma_addr);
2123                 if (res)
2124                         continue;
2125
2126                 if (id_ns->ncap == 0)
2127                         continue;
2128
2129                 res = nvme_get_features(dev, NVME_FEAT_LBA_RANGE, i,
2130                                                         dma_addr + 4096, NULL);
2131                 if (res)
2132                         memset(mem + 4096, 0, 4096);
2133
2134                 ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
2135                 if (ns)
2136                         list_add_tail(&ns->list, &dev->namespaces);
2137         }
2138         list_for_each_entry(ns, &dev->namespaces, list)
2139                 add_disk(ns->disk);
2140         res = 0;
2141
2142  out:
2143         dma_free_coherent(&dev->pci_dev->dev, 8192, mem, dma_addr);
2144         return res;
2145 }
2146
2147 static int nvme_dev_map(struct nvme_dev *dev)
2148 {
2149         u64 cap;
2150         int bars, result = -ENOMEM;
2151         struct pci_dev *pdev = dev->pci_dev;
2152
2153         if (pci_enable_device_mem(pdev))
2154                 return result;
2155
2156         dev->entry[0].vector = pdev->irq;
2157         pci_set_master(pdev);
2158         bars = pci_select_bars(pdev, IORESOURCE_MEM);
2159         if (!bars)
2160                 goto disable_pci;
2161
2162         if (pci_request_selected_regions(pdev, bars, "nvme"))
2163                 goto disable_pci;
2164
2165         if (dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64)) &&
2166             dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32)))
2167                 goto disable;
2168
2169         dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
2170         if (!dev->bar)
2171                 goto disable;
2172
2173         if (readl(&dev->bar->csts) == -1) {
2174                 result = -ENODEV;
2175                 goto unmap;
2176         }
2177
2178         /*
2179          * Some devices don't advertse INTx interrupts, pre-enable a single
2180          * MSIX vec for setup. We'll adjust this later.
2181          */
2182         if (!pdev->irq) {
2183                 result = pci_enable_msix(pdev, dev->entry, 1);
2184                 if (result < 0)
2185                         goto unmap;
2186         }
2187
2188         cap = readq(&dev->bar->cap);
2189         dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
2190         dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
2191         dev->dbs = ((void __iomem *)dev->bar) + 4096;
2192
2193         return 0;
2194
2195  unmap:
2196         iounmap(dev->bar);
2197         dev->bar = NULL;
2198  disable:
2199         pci_release_regions(pdev);
2200  disable_pci:
2201         pci_disable_device(pdev);
2202         return result;
2203 }
2204
2205 static void nvme_dev_unmap(struct nvme_dev *dev)
2206 {
2207         if (dev->pci_dev->msi_enabled)
2208                 pci_disable_msi(dev->pci_dev);
2209         else if (dev->pci_dev->msix_enabled)
2210                 pci_disable_msix(dev->pci_dev);
2211
2212         if (dev->bar) {
2213                 iounmap(dev->bar);
2214                 dev->bar = NULL;
2215                 pci_release_regions(dev->pci_dev);
2216         }
2217
2218         if (pci_is_enabled(dev->pci_dev))
2219                 pci_disable_device(dev->pci_dev);
2220 }
2221
2222 struct nvme_delq_ctx {
2223         struct task_struct *waiter;
2224         struct kthread_worker *worker;
2225         atomic_t refcount;
2226 };
2227
2228 static void nvme_wait_dq(struct nvme_delq_ctx *dq, struct nvme_dev *dev)
2229 {
2230         dq->waiter = current;
2231         mb();
2232
2233         for (;;) {
2234                 set_current_state(TASK_KILLABLE);
2235                 if (!atomic_read(&dq->refcount))
2236                         break;
2237                 if (!schedule_timeout(ADMIN_TIMEOUT) ||
2238                                         fatal_signal_pending(current)) {
2239                         set_current_state(TASK_RUNNING);
2240
2241                         nvme_disable_ctrl(dev, readq(&dev->bar->cap));
2242                         nvme_disable_queue(dev, 0);
2243
2244                         send_sig(SIGKILL, dq->worker->task, 1);
2245                         flush_kthread_worker(dq->worker);
2246                         return;
2247                 }
2248         }
2249         set_current_state(TASK_RUNNING);
2250 }
2251
2252 static void nvme_put_dq(struct nvme_delq_ctx *dq)
2253 {
2254         atomic_dec(&dq->refcount);
2255         if (dq->waiter)
2256                 wake_up_process(dq->waiter);
2257 }
2258
2259 static struct nvme_delq_ctx *nvme_get_dq(struct nvme_delq_ctx *dq)
2260 {
2261         atomic_inc(&dq->refcount);
2262         return dq;
2263 }
2264
2265 static void nvme_del_queue_end(struct nvme_queue *nvmeq)
2266 {
2267         struct nvme_delq_ctx *dq = nvmeq->cmdinfo.ctx;
2268
2269         nvme_clear_queue(nvmeq);
2270         nvme_put_dq(dq);
2271 }
2272
2273 static int adapter_async_del_queue(struct nvme_queue *nvmeq, u8 opcode,
2274                                                 kthread_work_func_t fn)
2275 {
2276         struct nvme_command c;
2277
2278         memset(&c, 0, sizeof(c));
2279         c.delete_queue.opcode = opcode;
2280         c.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2281
2282         init_kthread_work(&nvmeq->cmdinfo.work, fn);
2283         return nvme_submit_admin_async_cmd(nvmeq->dev, &c, &nvmeq->cmdinfo,
2284                                                                 ADMIN_TIMEOUT);
2285 }
2286
2287 static void nvme_del_cq_work_handler(struct kthread_work *work)
2288 {
2289         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2290                                                         cmdinfo.work);
2291         nvme_del_queue_end(nvmeq);
2292 }
2293
2294 static int nvme_delete_cq(struct nvme_queue *nvmeq)
2295 {
2296         return adapter_async_del_queue(nvmeq, nvme_admin_delete_cq,
2297                                                 nvme_del_cq_work_handler);
2298 }
2299
2300 static void nvme_del_sq_work_handler(struct kthread_work *work)
2301 {
2302         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2303                                                         cmdinfo.work);
2304         int status = nvmeq->cmdinfo.status;
2305
2306         if (!status)
2307                 status = nvme_delete_cq(nvmeq);
2308         if (status)
2309                 nvme_del_queue_end(nvmeq);
2310 }
2311
2312 static int nvme_delete_sq(struct nvme_queue *nvmeq)
2313 {
2314         return adapter_async_del_queue(nvmeq, nvme_admin_delete_sq,
2315                                                 nvme_del_sq_work_handler);
2316 }
2317
2318 static void nvme_del_queue_start(struct kthread_work *work)
2319 {
2320         struct nvme_queue *nvmeq = container_of(work, struct nvme_queue,
2321                                                         cmdinfo.work);
2322         allow_signal(SIGKILL);
2323         if (nvme_delete_sq(nvmeq))
2324                 nvme_del_queue_end(nvmeq);
2325 }
2326
2327 static void nvme_disable_io_queues(struct nvme_dev *dev)
2328 {
2329         int i;
2330         DEFINE_KTHREAD_WORKER_ONSTACK(worker);
2331         struct nvme_delq_ctx dq;
2332         struct task_struct *kworker_task = kthread_run(kthread_worker_fn,
2333                                         &worker, "nvme%d", dev->instance);
2334
2335         if (IS_ERR(kworker_task)) {
2336                 dev_err(&dev->pci_dev->dev,
2337                         "Failed to create queue del task\n");
2338                 for (i = dev->queue_count - 1; i > 0; i--)
2339                         nvme_disable_queue(dev, i);
2340                 return;
2341         }
2342
2343         dq.waiter = NULL;
2344         atomic_set(&dq.refcount, 0);
2345         dq.worker = &worker;
2346         for (i = dev->queue_count - 1; i > 0; i--) {
2347                 struct nvme_queue *nvmeq = dev->queues[i];
2348
2349                 if (nvme_suspend_queue(nvmeq))
2350                         continue;
2351                 nvmeq->cmdinfo.ctx = nvme_get_dq(&dq);
2352                 nvmeq->cmdinfo.worker = dq.worker;
2353                 init_kthread_work(&nvmeq->cmdinfo.work, nvme_del_queue_start);
2354                 queue_kthread_work(dq.worker, &nvmeq->cmdinfo.work);
2355         }
2356         nvme_wait_dq(&dq, dev);
2357         kthread_stop(kworker_task);
2358 }
2359
2360 /*
2361 * Remove the node from the device list and check
2362 * for whether or not we need to stop the nvme_thread.
2363 */
2364 static void nvme_dev_list_remove(struct nvme_dev *dev)
2365 {
2366         struct task_struct *tmp = NULL;
2367
2368         spin_lock(&dev_list_lock);
2369         list_del_init(&dev->node);
2370         if (list_empty(&dev_list) && !IS_ERR_OR_NULL(nvme_thread)) {
2371                 tmp = nvme_thread;
2372                 nvme_thread = NULL;
2373         }
2374         spin_unlock(&dev_list_lock);
2375
2376         if (tmp)
2377                 kthread_stop(tmp);
2378 }
2379
2380 static void nvme_dev_shutdown(struct nvme_dev *dev)
2381 {
2382         int i;
2383         u32 csts = -1;
2384
2385         dev->initialized = 0;
2386         nvme_dev_list_remove(dev);
2387
2388         if (dev->bar)
2389                 csts = readl(&dev->bar->csts);
2390         if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
2391                 for (i = dev->queue_count - 1; i >= 0; i--) {
2392                         struct nvme_queue *nvmeq = dev->queues[i];
2393                         nvme_suspend_queue(nvmeq);
2394                         nvme_clear_queue(nvmeq);
2395                 }
2396         } else {
2397                 nvme_disable_io_queues(dev);
2398                 nvme_shutdown_ctrl(dev);
2399                 nvme_disable_queue(dev, 0);
2400         }
2401         nvme_dev_unmap(dev);
2402 }
2403
2404 static void nvme_dev_remove_admin(struct nvme_dev *dev)
2405 {
2406         if (dev->admin_q && !blk_queue_dying(dev->admin_q))
2407                 blk_cleanup_queue(dev->admin_q);
2408 }
2409
2410 static void nvme_dev_remove(struct nvme_dev *dev)
2411 {
2412         struct nvme_ns *ns;
2413
2414         list_for_each_entry(ns, &dev->namespaces, list) {
2415                 if (ns->disk->flags & GENHD_FL_UP)
2416                         del_gendisk(ns->disk);
2417                 if (!blk_queue_dying(ns->queue))
2418                         blk_cleanup_queue(ns->queue);
2419         }
2420 }
2421
2422 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2423 {
2424         struct device *dmadev = &dev->pci_dev->dev;
2425         dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
2426                                                 PAGE_SIZE, PAGE_SIZE, 0);
2427         if (!dev->prp_page_pool)
2428                 return -ENOMEM;
2429
2430         /* Optimisation for I/Os between 4k and 128k */
2431         dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
2432                                                 256, 256, 0);
2433         if (!dev->prp_small_pool) {
2434                 dma_pool_destroy(dev->prp_page_pool);
2435                 return -ENOMEM;
2436         }
2437         return 0;
2438 }
2439
2440 static void nvme_release_prp_pools(struct nvme_dev *dev)
2441 {
2442         dma_pool_destroy(dev->prp_page_pool);
2443         dma_pool_destroy(dev->prp_small_pool);
2444 }
2445
2446 static DEFINE_IDA(nvme_instance_ida);
2447
2448 static int nvme_set_instance(struct nvme_dev *dev)
2449 {
2450         int instance, error;
2451
2452         do {
2453                 if (!ida_pre_get(&nvme_instance_ida, GFP_KERNEL))
2454                         return -ENODEV;
2455
2456                 spin_lock(&dev_list_lock);
2457                 error = ida_get_new(&nvme_instance_ida, &instance);
2458                 spin_unlock(&dev_list_lock);
2459         } while (error == -EAGAIN);
2460
2461         if (error)
2462                 return -ENODEV;
2463
2464         dev->instance = instance;
2465         return 0;
2466 }
2467
2468 static void nvme_release_instance(struct nvme_dev *dev)
2469 {
2470         spin_lock(&dev_list_lock);
2471         ida_remove(&nvme_instance_ida, dev->instance);
2472         spin_unlock(&dev_list_lock);
2473 }
2474
2475 static void nvme_free_namespaces(struct nvme_dev *dev)
2476 {
2477         struct nvme_ns *ns, *next;
2478
2479         list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
2480                 list_del(&ns->list);
2481
2482                 spin_lock(&dev_list_lock);
2483                 ns->disk->private_data = NULL;
2484                 spin_unlock(&dev_list_lock);
2485
2486                 put_disk(ns->disk);
2487                 kfree(ns);
2488         }
2489 }
2490
2491 static void nvme_free_dev(struct kref *kref)
2492 {
2493         struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
2494
2495         pci_dev_put(dev->pci_dev);
2496         nvme_free_namespaces(dev);
2497         blk_mq_free_tag_set(&dev->tagset);
2498         kfree(dev->queues);
2499         kfree(dev->entry);
2500         kfree(dev);
2501 }
2502
2503 static int nvme_dev_open(struct inode *inode, struct file *f)
2504 {
2505         struct nvme_dev *dev = container_of(f->private_data, struct nvme_dev,
2506                                                                 miscdev);
2507         kref_get(&dev->kref);
2508         f->private_data = dev;
2509         return 0;
2510 }
2511
2512 static int nvme_dev_release(struct inode *inode, struct file *f)
2513 {
2514         struct nvme_dev *dev = f->private_data;
2515         kref_put(&dev->kref, nvme_free_dev);
2516         return 0;
2517 }
2518
2519 static long nvme_dev_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
2520 {
2521         struct nvme_dev *dev = f->private_data;
2522         struct nvme_ns *ns;
2523
2524         switch (cmd) {
2525         case NVME_IOCTL_ADMIN_CMD:
2526                 return nvme_user_cmd(dev, NULL, (void __user *)arg);
2527         case NVME_IOCTL_IO_CMD:
2528                 if (list_empty(&dev->namespaces))
2529                         return -ENOTTY;
2530                 ns = list_first_entry(&dev->namespaces, struct nvme_ns, list);
2531                 return nvme_user_cmd(dev, ns, (void __user *)arg);
2532         default:
2533                 return -ENOTTY;
2534         }
2535 }
2536
2537 static const struct file_operations nvme_dev_fops = {
2538         .owner          = THIS_MODULE,
2539         .open           = nvme_dev_open,
2540         .release        = nvme_dev_release,
2541         .unlocked_ioctl = nvme_dev_ioctl,
2542         .compat_ioctl   = nvme_dev_ioctl,
2543 };
2544
2545 static void nvme_set_irq_hints(struct nvme_dev *dev)
2546 {
2547         struct nvme_queue *nvmeq;
2548         int i;
2549
2550         for (i = 0; i < dev->online_queues; i++) {
2551                 nvmeq = dev->queues[i];
2552
2553                 if (!nvmeq->hctx)
2554                         continue;
2555
2556                 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
2557                                                         nvmeq->hctx->cpumask);
2558         }
2559 }
2560
2561 static int nvme_dev_start(struct nvme_dev *dev)
2562 {
2563         int result;
2564         bool start_thread = false;
2565
2566         result = nvme_dev_map(dev);
2567         if (result)
2568                 return result;
2569
2570         result = nvme_configure_admin_queue(dev);
2571         if (result)
2572                 goto unmap;
2573
2574         spin_lock(&dev_list_lock);
2575         if (list_empty(&dev_list) && IS_ERR_OR_NULL(nvme_thread)) {
2576                 start_thread = true;
2577                 nvme_thread = NULL;
2578         }
2579         list_add(&dev->node, &dev_list);
2580         spin_unlock(&dev_list_lock);
2581
2582         if (start_thread) {
2583                 nvme_thread = kthread_run(nvme_kthread, NULL, "nvme");
2584                 wake_up_all(&nvme_kthread_wait);
2585         } else
2586                 wait_event_killable(nvme_kthread_wait, nvme_thread);
2587
2588         if (IS_ERR_OR_NULL(nvme_thread)) {
2589                 result = nvme_thread ? PTR_ERR(nvme_thread) : -EINTR;
2590                 goto disable;
2591         }
2592
2593         nvme_init_queue(dev->queues[0], 0);
2594
2595         result = nvme_setup_io_queues(dev);
2596         if (result)
2597                 goto disable;
2598
2599         nvme_set_irq_hints(dev);
2600
2601         return result;
2602
2603  disable:
2604         nvme_disable_queue(dev, 0);
2605         nvme_dev_list_remove(dev);
2606  unmap:
2607         nvme_dev_unmap(dev);
2608         return result;
2609 }
2610
2611 static int nvme_remove_dead_ctrl(void *arg)
2612 {
2613         struct nvme_dev *dev = (struct nvme_dev *)arg;
2614         struct pci_dev *pdev = dev->pci_dev;
2615
2616         if (pci_get_drvdata(pdev))
2617                 pci_stop_and_remove_bus_device_locked(pdev);
2618         kref_put(&dev->kref, nvme_free_dev);
2619         return 0;
2620 }
2621
2622 static void nvme_remove_disks(struct work_struct *ws)
2623 {
2624         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2625
2626         nvme_free_queues(dev, 1);
2627         nvme_dev_remove(dev);
2628 }
2629
2630 static int nvme_dev_resume(struct nvme_dev *dev)
2631 {
2632         int ret;
2633
2634         ret = nvme_dev_start(dev);
2635         if (ret)
2636                 return ret;
2637         if (dev->online_queues < 2) {
2638                 spin_lock(&dev_list_lock);
2639                 dev->reset_workfn = nvme_remove_disks;
2640                 queue_work(nvme_workq, &dev->reset_work);
2641                 spin_unlock(&dev_list_lock);
2642         }
2643         dev->initialized = 1;
2644         return 0;
2645 }
2646
2647 static void nvme_dev_reset(struct nvme_dev *dev)
2648 {
2649         nvme_dev_shutdown(dev);
2650         if (nvme_dev_resume(dev)) {
2651                 dev_warn(&dev->pci_dev->dev, "Device failed to resume\n");
2652                 kref_get(&dev->kref);
2653                 if (IS_ERR(kthread_run(nvme_remove_dead_ctrl, dev, "nvme%d",
2654                                                         dev->instance))) {
2655                         dev_err(&dev->pci_dev->dev,
2656                                 "Failed to start controller remove task\n");
2657                         kref_put(&dev->kref, nvme_free_dev);
2658                 }
2659         }
2660 }
2661
2662 static void nvme_reset_failed_dev(struct work_struct *ws)
2663 {
2664         struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
2665         nvme_dev_reset(dev);
2666 }
2667
2668 static void nvme_reset_workfn(struct work_struct *work)
2669 {
2670         struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2671         dev->reset_workfn(work);
2672 }
2673
2674 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2675 {
2676         int node, result = -ENOMEM;
2677         struct nvme_dev *dev;
2678
2679         node = dev_to_node(&pdev->dev);
2680         if (node == NUMA_NO_NODE)
2681                 set_dev_node(&pdev->dev, 0);
2682
2683         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2684         if (!dev)
2685                 return -ENOMEM;
2686         dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
2687                                                         GFP_KERNEL, node);
2688         if (!dev->entry)
2689                 goto free;
2690         dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2691                                                         GFP_KERNEL, node);
2692         if (!dev->queues)
2693                 goto free;
2694
2695         INIT_LIST_HEAD(&dev->namespaces);
2696         dev->reset_workfn = nvme_reset_failed_dev;
2697         INIT_WORK(&dev->reset_work, nvme_reset_workfn);
2698         dev->pci_dev = pci_dev_get(pdev);
2699         pci_set_drvdata(pdev, dev);
2700         result = nvme_set_instance(dev);
2701         if (result)
2702                 goto put_pci;
2703
2704         result = nvme_setup_prp_pools(dev);
2705         if (result)
2706                 goto release;
2707
2708         kref_init(&dev->kref);
2709         result = nvme_dev_start(dev);
2710         if (result)
2711                 goto release_pools;
2712
2713         if (dev->online_queues > 1)
2714                 result = nvme_dev_add(dev);
2715         if (result)
2716                 goto shutdown;
2717
2718         scnprintf(dev->name, sizeof(dev->name), "nvme%d", dev->instance);
2719         dev->miscdev.minor = MISC_DYNAMIC_MINOR;
2720         dev->miscdev.parent = &pdev->dev;
2721         dev->miscdev.name = dev->name;
2722         dev->miscdev.fops = &nvme_dev_fops;
2723         result = misc_register(&dev->miscdev);
2724         if (result)
2725                 goto remove;
2726
2727         nvme_set_irq_hints(dev);
2728
2729         dev->initialized = 1;
2730         return 0;
2731
2732  remove:
2733         nvme_dev_remove(dev);
2734         nvme_dev_remove_admin(dev);
2735         nvme_free_namespaces(dev);
2736  shutdown:
2737         nvme_dev_shutdown(dev);
2738  release_pools:
2739         nvme_free_queues(dev, 0);
2740         nvme_release_prp_pools(dev);
2741  release:
2742         nvme_release_instance(dev);
2743  put_pci:
2744         pci_dev_put(dev->pci_dev);
2745  free:
2746         kfree(dev->queues);
2747         kfree(dev->entry);
2748         kfree(dev);
2749         return result;
2750 }
2751
2752 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2753 {
2754         struct nvme_dev *dev = pci_get_drvdata(pdev);
2755
2756         if (prepare)
2757                 nvme_dev_shutdown(dev);
2758         else
2759                 nvme_dev_resume(dev);
2760 }
2761
2762 static void nvme_shutdown(struct pci_dev *pdev)
2763 {
2764         struct nvme_dev *dev = pci_get_drvdata(pdev);
2765         nvme_dev_shutdown(dev);
2766 }
2767
2768 static void nvme_remove(struct pci_dev *pdev)
2769 {
2770         struct nvme_dev *dev = pci_get_drvdata(pdev);
2771
2772         spin_lock(&dev_list_lock);
2773         list_del_init(&dev->node);
2774         spin_unlock(&dev_list_lock);
2775
2776         pci_set_drvdata(pdev, NULL);
2777         flush_work(&dev->reset_work);
2778         misc_deregister(&dev->miscdev);
2779         nvme_dev_remove(dev);
2780         nvme_dev_shutdown(dev);
2781         nvme_dev_remove_admin(dev);
2782         nvme_free_queues(dev, 0);
2783         nvme_free_admin_tags(dev);
2784         nvme_release_instance(dev);
2785         nvme_release_prp_pools(dev);
2786         kref_put(&dev->kref, nvme_free_dev);
2787 }
2788
2789 /* These functions are yet to be implemented */
2790 #define nvme_error_detected NULL
2791 #define nvme_dump_registers NULL
2792 #define nvme_link_reset NULL
2793 #define nvme_slot_reset NULL
2794 #define nvme_error_resume NULL
2795
2796 #ifdef CONFIG_PM_SLEEP
2797 static int nvme_suspend(struct device *dev)
2798 {
2799         struct pci_dev *pdev = to_pci_dev(dev);
2800         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2801
2802         nvme_dev_shutdown(ndev);
2803         return 0;
2804 }
2805
2806 static int nvme_resume(struct device *dev)
2807 {
2808         struct pci_dev *pdev = to_pci_dev(dev);
2809         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2810
2811         if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
2812                 ndev->reset_workfn = nvme_reset_failed_dev;
2813                 queue_work(nvme_workq, &ndev->reset_work);
2814         }
2815         return 0;
2816 }
2817 #endif
2818
2819 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2820
2821 static const struct pci_error_handlers nvme_err_handler = {
2822         .error_detected = nvme_error_detected,
2823         .mmio_enabled   = nvme_dump_registers,
2824         .link_reset     = nvme_link_reset,
2825         .slot_reset     = nvme_slot_reset,
2826         .resume         = nvme_error_resume,
2827         .reset_notify   = nvme_reset_notify,
2828 };
2829
2830 /* Move to pci_ids.h later */
2831 #define PCI_CLASS_STORAGE_EXPRESS       0x010802
2832
2833 static const struct pci_device_id nvme_id_table[] = {
2834         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2835         { 0, }
2836 };
2837 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2838
2839 static struct pci_driver nvme_driver = {
2840         .name           = "nvme",
2841         .id_table       = nvme_id_table,
2842         .probe          = nvme_probe,
2843         .remove         = nvme_remove,
2844         .shutdown       = nvme_shutdown,
2845         .driver         = {
2846                 .pm     = &nvme_dev_pm_ops,
2847         },
2848         .err_handler    = &nvme_err_handler,
2849 };
2850
2851 static int __init nvme_init(void)
2852 {
2853         int result;
2854
2855         init_waitqueue_head(&nvme_kthread_wait);
2856
2857         nvme_workq = create_singlethread_workqueue("nvme");
2858         if (!nvme_workq)
2859                 return -ENOMEM;
2860
2861         result = register_blkdev(nvme_major, "nvme");
2862         if (result < 0)
2863                 goto kill_workq;
2864         else if (result > 0)
2865                 nvme_major = result;
2866
2867         result = pci_register_driver(&nvme_driver);
2868         if (result)
2869                 goto unregister_blkdev;
2870         return 0;
2871
2872  unregister_blkdev:
2873         unregister_blkdev(nvme_major, "nvme");
2874  kill_workq:
2875         destroy_workqueue(nvme_workq);
2876         return result;
2877 }
2878
2879 static void __exit nvme_exit(void)
2880 {
2881         pci_unregister_driver(&nvme_driver);
2882         unregister_hotcpu_notifier(&nvme_nb);
2883         unregister_blkdev(nvme_major, "nvme");
2884         destroy_workqueue(nvme_workq);
2885         BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
2886         _nvme_check_size();
2887 }
2888
2889 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2890 MODULE_LICENSE("GPL");
2891 MODULE_VERSION("0.9");
2892 module_init(nvme_init);
2893 module_exit(nvme_exit);