drivers/nvme/host/pci.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVM Express device driver
   4  * Copyright (c) 2011-2014, Intel Corporation.
   5  */
   6
   7 #include <linux/acpi.h>
   8 #include <linux/aer.h>
   9 #include <linux/async.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/blk-mq.h>
  12 #include <linux/blk-mq-pci.h>
  13 #include <linux/blk-integrity.h>
  14 #include <linux/dmi.h>
  15 #include <linux/init.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/io.h>
  18 #include <linux/memremap.h>
  19 #include <linux/mm.h>
  20 #include <linux/module.h>
  21 #include <linux/mutex.h>
  22 #include <linux/once.h>
  23 #include <linux/pci.h>
  24 #include <linux/suspend.h>
  25 #include <linux/t10-pi.h>
  26 #include <linux/types.h>
  27 #include <linux/io-64-nonatomic-lo-hi.h>
  28 #include <linux/io-64-nonatomic-hi-lo.h>
  29 #include <linux/sed-opal.h>
  30 #include <linux/pci-p2pdma.h>
  31
  32 #include "trace.h"
  33 #include "nvme.h"
  34
  35 #define SQ_SIZE(q)      ((q)->q_depth << (q)->sqes)
  36 #define CQ_SIZE(q)      ((q)->q_depth * sizeof(struct nvme_completion))
  37
  38 #define SGES_PER_PAGE   (PAGE_SIZE / sizeof(struct nvme_sgl_desc))
  39
  40 /*
  41  * These can be higher, but we need to ensure that any command doesn't
  42  * require an sg allocation that needs more than a page of data.
  43  */
  44 #define NVME_MAX_KB_SZ  4096
  45 #define NVME_MAX_SEGS   127
  46
  47 static int use_threaded_interrupts;
  48 module_param(use_threaded_interrupts, int, 0444);
  49
  50 static bool use_cmb_sqes = true;
  51 module_param(use_cmb_sqes, bool, 0444);
  52 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
  53
  54 static unsigned int max_host_mem_size_mb = 128;
  55 module_param(max_host_mem_size_mb, uint, 0444);
  56 MODULE_PARM_DESC(max_host_mem_size_mb,
  57         "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
  58
  59 static unsigned int sgl_threshold = SZ_32K;
  60 module_param(sgl_threshold, uint, 0644);
  61 MODULE_PARM_DESC(sgl_threshold,
  62                 "Use SGLs when average request segment size is larger or equal to "
  63                 "this size. Use 0 to disable SGLs.");
  64
  65 #define NVME_PCI_MIN_QUEUE_SIZE 2
  66 #define NVME_PCI_MAX_QUEUE_SIZE 4095
  67 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
  68 static const struct kernel_param_ops io_queue_depth_ops = {
  69         .set = io_queue_depth_set,
  70         .get = param_get_uint,
  71 };
  72
  73 static unsigned int io_queue_depth = 1024;
  74 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
  75 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2 and < 4096");
  76
  77 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
  78 {
  79         unsigned int n;
  80         int ret;
  81
  82         ret = kstrtouint(val, 10, &n);
  83         if (ret != 0 || n > num_possible_cpus())
  84                 return -EINVAL;
  85         return param_set_uint(val, kp);
  86 }
  87
  88 static const struct kernel_param_ops io_queue_count_ops = {
  89         .set = io_queue_count_set,
  90         .get = param_get_uint,
  91 };
  92
  93 static unsigned int write_queues;
  94 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
  95 MODULE_PARM_DESC(write_queues,
  96         "Number of queues to use for writes. If not set, reads and writes "
  97         "will share a queue set.");
  98
  99 static unsigned int poll_queues;
 100 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
 101 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
 102
 103 static bool noacpi;
 104 module_param(noacpi, bool, 0444);
 105 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
 106
 107 struct nvme_dev;
 108 struct nvme_queue;
 109
 110 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
 111 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
 112
 113 /*
 114  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
 115  */
 116 struct nvme_dev {
 117         struct nvme_queue *queues;
 118         struct blk_mq_tag_set tagset;
 119         struct blk_mq_tag_set admin_tagset;
 120         u32 __iomem *dbs;
 121         struct device *dev;
 122         struct dma_pool *prp_page_pool;
 123         struct dma_pool *prp_small_pool;
 124         unsigned online_queues;
 125         unsigned max_qid;
 126         unsigned io_queues[HCTX_MAX_TYPES];
 127         unsigned int num_vecs;
 128         u32 q_depth;
 129         int io_sqes;
 130         u32 db_stride;
 131         void __iomem *bar;
 132         unsigned long bar_mapped_size;
 133         struct work_struct remove_work;
 134         struct mutex shutdown_lock;
 135         bool subsystem;
 136         u64 cmb_size;
 137         bool cmb_use_sqes;
 138         u32 cmbsz;
 139         u32 cmbloc;
 140         struct nvme_ctrl ctrl;
 141         u32 last_ps;
 142         bool hmb;
 143
 144         mempool_t *iod_mempool;
 145
 146         /* shadow doorbell buffer support: */
 147         u32 *dbbuf_dbs;
 148         dma_addr_t dbbuf_dbs_dma_addr;
 149         u32 *dbbuf_eis;
 150         dma_addr_t dbbuf_eis_dma_addr;
 151
 152         /* host memory buffer support: */
 153         u64 host_mem_size;
 154         u32 nr_host_mem_descs;
 155         dma_addr_t host_mem_descs_dma;
 156         struct nvme_host_mem_buf_desc *host_mem_descs;
 157         void **host_mem_desc_bufs;
 158         unsigned int nr_allocated_queues;
 159         unsigned int nr_write_queues;
 160         unsigned int nr_poll_queues;
 161
 162         bool attrs_added;
 163 };
 164
 165 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
 166 {
 167         return param_set_uint_minmax(val, kp, NVME_PCI_MIN_QUEUE_SIZE,
 168                         NVME_PCI_MAX_QUEUE_SIZE);
 169 }
 170
 171 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
 172 {
 173         return qid * 2 * stride;
 174 }
 175
 176 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
 177 {
 178         return (qid * 2 + 1) * stride;
 179 }
 180
 181 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 182 {
 183         return container_of(ctrl, struct nvme_dev, ctrl);
 184 }
 185
 186 /*
 187  * An NVM Express queue.  Each device has at least two (one for admin
 188  * commands and one for I/O commands).
 189  */
 190 struct nvme_queue {
 191         struct nvme_dev *dev;
 192         spinlock_t sq_lock;
 193         void *sq_cmds;
 194          /* only used for poll queues: */
 195         spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
 196         struct nvme_completion *cqes;
 197         dma_addr_t sq_dma_addr;
 198         dma_addr_t cq_dma_addr;
 199         u32 __iomem *q_db;
 200         u32 q_depth;
 201         u16 cq_vector;
 202         u16 sq_tail;
 203         u16 last_sq_tail;
 204         u16 cq_head;
 205         u16 qid;
 206         u8 cq_phase;
 207         u8 sqes;
 208         unsigned long flags;
 209 #define NVMEQ_ENABLED           0
 210 #define NVMEQ_SQ_CMB            1
 211 #define NVMEQ_DELETE_ERROR      2
 212 #define NVMEQ_POLLED            3
 213         u32 *dbbuf_sq_db;
 214         u32 *dbbuf_cq_db;
 215         u32 *dbbuf_sq_ei;
 216         u32 *dbbuf_cq_ei;
 217         struct completion delete_done;
 218 };
 219
 220 /*
 221  * The nvme_iod describes the data in an I/O.
 222  *
 223  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
 224  * to the actual struct scatterlist.
 225  */
 226 struct nvme_iod {
 227         struct nvme_request req;
 228         struct nvme_command cmd;
 229         struct nvme_queue *nvmeq;
 230         bool use_sgl;
 231         int aborted;
 232         int npages;             /* In the PRP list. 0 means small pool in use */
 233         int nents;              /* Used in scatterlist */
 234         dma_addr_t first_dma;
 235         unsigned int dma_len;   /* length of single DMA segment mapping */
 236         dma_addr_t meta_dma;
 237         struct scatterlist *sg;
 238 };
 239
 240 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
 241 {
 242         return dev->nr_allocated_queues * 8 * dev->db_stride;
 243 }
 244
 245 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
 246 {
 247         unsigned int mem_size = nvme_dbbuf_size(dev);
 248
 249         if (dev->dbbuf_dbs) {
 250                 /*
 251                  * Clear the dbbuf memory so the driver doesn't observe stale
 252                  * values from the previous instantiation.
 253                  */
 254                 memset(dev->dbbuf_dbs, 0, mem_size);
 255                 memset(dev->dbbuf_eis, 0, mem_size);
 256                 return 0;
 257         }
 258
 259         dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
 260                                             &dev->dbbuf_dbs_dma_addr,
 261                                             GFP_KERNEL);
 262         if (!dev->dbbuf_dbs)
 263                 return -ENOMEM;
 264         dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
 265                                             &dev->dbbuf_eis_dma_addr,
 266                                             GFP_KERNEL);
 267         if (!dev->dbbuf_eis) {
 268                 dma_free_coherent(dev->dev, mem_size,
 269                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 270                 dev->dbbuf_dbs = NULL;
 271                 return -ENOMEM;
 272         }
 273
 274         return 0;
 275 }
 276
 277 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
 278 {
 279         unsigned int mem_size = nvme_dbbuf_size(dev);
 280
 281         if (dev->dbbuf_dbs) {
 282                 dma_free_coherent(dev->dev, mem_size,
 283                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 284                 dev->dbbuf_dbs = NULL;
 285         }
 286         if (dev->dbbuf_eis) {
 287                 dma_free_coherent(dev->dev, mem_size,
 288                                   dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
 289                 dev->dbbuf_eis = NULL;
 290         }
 291 }
 292
 293 static void nvme_dbbuf_init(struct nvme_dev *dev,
 294                             struct nvme_queue *nvmeq, int qid)
 295 {
 296         if (!dev->dbbuf_dbs || !qid)
 297                 return;
 298
 299         nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
 300         nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
 301         nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
 302         nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
 303 }
 304
 305 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
 306 {
 307         if (!nvmeq->qid)
 308                 return;
 309
 310         nvmeq->dbbuf_sq_db = NULL;
 311         nvmeq->dbbuf_cq_db = NULL;
 312         nvmeq->dbbuf_sq_ei = NULL;
 313         nvmeq->dbbuf_cq_ei = NULL;
 314 }
 315
 316 static void nvme_dbbuf_set(struct nvme_dev *dev)
 317 {
 318         struct nvme_command c = { };
 319         unsigned int i;
 320
 321         if (!dev->dbbuf_dbs)
 322                 return;
 323
 324         c.dbbuf.opcode = nvme_admin_dbbuf;
 325         c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
 326         c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
 327
 328         if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
 329                 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
 330                 /* Free memory and continue on */
 331                 nvme_dbbuf_dma_free(dev);
 332
 333                 for (i = 1; i <= dev->online_queues; i++)
 334                         nvme_dbbuf_free(&dev->queues[i]);
 335         }
 336 }
 337
 338 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
 339 {
 340         return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
 341 }
 342
 343 /* Update dbbuf and return true if an MMIO is required */
 344 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
 345                                               volatile u32 *dbbuf_ei)
 346 {
 347         if (dbbuf_db) {
 348                 u16 old_value;
 349
 350                 /*
 351                  * Ensure that the queue is written before updating
 352                  * the doorbell in memory
 353                  */
 354                 wmb();
 355
 356                 old_value = *dbbuf_db;
 357                 *dbbuf_db = value;
 358
 359                 /*
 360                  * Ensure that the doorbell is updated before reading the event
 361                  * index from memory.  The controller needs to provide similar
 362                  * ordering to ensure the envent index is updated before reading
 363                  * the doorbell.
 364                  */
 365                 mb();
 366
 367                 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
 368                         return false;
 369         }
 370
 371         return true;
 372 }
 373
 374 /*
 375  * Will slightly overestimate the number of pages needed.  This is OK
 376  * as it only leads to a small amount of wasted memory for the lifetime of
 377  * the I/O.
 378  */
 379 static int nvme_pci_npages_prp(void)
 380 {
 381         unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
 382                                       NVME_CTRL_PAGE_SIZE);
 383         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 384 }
 385
 386 /*
 387  * Calculates the number of pages needed for the SGL segments. For example a 4k
 388  * page can accommodate 256 SGL descriptors.
 389  */
 390 static int nvme_pci_npages_sgl(void)
 391 {
 392         return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
 393                         PAGE_SIZE);
 394 }
 395
 396 static size_t nvme_pci_iod_alloc_size(void)
 397 {
 398         size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
 399
 400         return sizeof(__le64 *) * npages +
 401                 sizeof(struct scatterlist) * NVME_MAX_SEGS;
 402 }
 403
 404 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 405                                 unsigned int hctx_idx)
 406 {
 407         struct nvme_dev *dev = data;
 408         struct nvme_queue *nvmeq = &dev->queues[0];
 409
 410         WARN_ON(hctx_idx != 0);
 411         WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
 412
 413         hctx->driver_data = nvmeq;
 414         return 0;
 415 }
 416
 417 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 418                           unsigned int hctx_idx)
 419 {
 420         struct nvme_dev *dev = data;
 421         struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
 422
 423         WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
 424         hctx->driver_data = nvmeq;
 425         return 0;
 426 }
 427
 428 static int nvme_pci_init_request(struct blk_mq_tag_set *set,
 429                 struct request *req, unsigned int hctx_idx,
 430                 unsigned int numa_node)
 431 {
 432         struct nvme_dev *dev = set->driver_data;
 433         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 434         int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
 435         struct nvme_queue *nvmeq = &dev->queues[queue_idx];
 436
 437         BUG_ON(!nvmeq);
 438         iod->nvmeq = nvmeq;
 439
 440         nvme_req(req)->ctrl = &dev->ctrl;
 441         nvme_req(req)->cmd = &iod->cmd;
 442         return 0;
 443 }
 444
 445 static int queue_irq_offset(struct nvme_dev *dev)
 446 {
 447         /* if we have more than 1 vec, admin queue offsets us by 1 */
 448         if (dev->num_vecs > 1)
 449                 return 1;
 450
 451         return 0;
 452 }
 453
 454 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
 455 {
 456         struct nvme_dev *dev = set->driver_data;
 457         int i, qoff, offset;
 458
 459         offset = queue_irq_offset(dev);
 460         for (i = 0, qoff = 0; i < set->nr_maps; i++) {
 461                 struct blk_mq_queue_map *map = &set->map[i];
 462
 463                 map->nr_queues = dev->io_queues[i];
 464                 if (!map->nr_queues) {
 465                         BUG_ON(i == HCTX_TYPE_DEFAULT);
 466                         continue;
 467                 }
 468
 469                 /*
 470                  * The poll queue(s) doesn't have an IRQ (and hence IRQ
 471                  * affinity), so use the regular blk-mq cpu mapping
 472                  */
 473                 map->queue_offset = qoff;
 474                 if (i != HCTX_TYPE_POLL && offset)
 475                         blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
 476                 else
 477                         blk_mq_map_queues(map);
 478                 qoff += map->nr_queues;
 479                 offset += map->nr_queues;
 480         }
 481
 482         return 0;
 483 }
 484
 485 /*
 486  * Write sq tail if we are asked to, or if the next command would wrap.
 487  */
 488 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
 489 {
 490         if (!write_sq) {
 491                 u16 next_tail = nvmeq->sq_tail + 1;
 492
 493                 if (next_tail == nvmeq->q_depth)
 494                         next_tail = 0;
 495                 if (next_tail != nvmeq->last_sq_tail)
 496                         return;
 497         }
 498
 499         if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
 500                         nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
 501                 writel(nvmeq->sq_tail, nvmeq->q_db);
 502         nvmeq->last_sq_tail = nvmeq->sq_tail;
 503 }
 504
 505 static inline void nvme_sq_copy_cmd(struct nvme_queue *nvmeq,
 506                                     struct nvme_command *cmd)
 507 {
 508         memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 509                 absolute_pointer(cmd), sizeof(*cmd));
 510         if (++nvmeq->sq_tail == nvmeq->q_depth)
 511                 nvmeq->sq_tail = 0;
 512 }
 513
 514 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
 515 {
 516         struct nvme_queue *nvmeq = hctx->driver_data;
 517
 518         spin_lock(&nvmeq->sq_lock);
 519         if (nvmeq->sq_tail != nvmeq->last_sq_tail)
 520                 nvme_write_sq_db(nvmeq, true);
 521         spin_unlock(&nvmeq->sq_lock);
 522 }
 523
 524 static void **nvme_pci_iod_list(struct request *req)
 525 {
 526         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 527         return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
 528 }
 529
 530 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
 531 {
 532         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 533         int nseg = blk_rq_nr_phys_segments(req);
 534         unsigned int avg_seg_size;
 535
 536         avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
 537
 538         if (!nvme_ctrl_sgl_supported(&dev->ctrl))
 539                 return false;
 540         if (!iod->nvmeq->qid)
 541                 return false;
 542         if (!sgl_threshold || avg_seg_size < sgl_threshold)
 543                 return false;
 544         return true;
 545 }
 546
 547 static void nvme_free_prps(struct nvme_dev *dev, struct request *req)
 548 {
 549         const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
 550         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 551         dma_addr_t dma_addr = iod->first_dma;
 552         int i;
 553
 554         for (i = 0; i < iod->npages; i++) {
 555                 __le64 *prp_list = nvme_pci_iod_list(req)[i];
 556                 dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
 557
 558                 dma_pool_free(dev->prp_page_pool, prp_list, dma_addr);
 559                 dma_addr = next_dma_addr;
 560         }
 561 }
 562
 563 static void nvme_free_sgls(struct nvme_dev *dev, struct request *req)
 564 {
 565         const int last_sg = SGES_PER_PAGE - 1;
 566         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 567         dma_addr_t dma_addr = iod->first_dma;
 568         int i;
 569
 570         for (i = 0; i < iod->npages; i++) {
 571                 struct nvme_sgl_desc *sg_list = nvme_pci_iod_list(req)[i];
 572                 dma_addr_t next_dma_addr = le64_to_cpu((sg_list[last_sg]).addr);
 573
 574                 dma_pool_free(dev->prp_page_pool, sg_list, dma_addr);
 575                 dma_addr = next_dma_addr;
 576         }
 577 }
 578
 579 static void nvme_unmap_sg(struct nvme_dev *dev, struct request *req)
 580 {
 581         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 582
 583         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 584                 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
 585                                     rq_dma_dir(req));
 586         else
 587                 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
 588 }
 589
 590 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 591 {
 592         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 593
 594         if (iod->dma_len) {
 595                 dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len,
 596                                rq_dma_dir(req));
 597                 return;
 598         }
 599
 600         WARN_ON_ONCE(!iod->nents);
 601
 602         nvme_unmap_sg(dev, req);
 603         if (iod->npages == 0)
 604                 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
 605                               iod->first_dma);
 606         else if (iod->use_sgl)
 607                 nvme_free_sgls(dev, req);
 608         else
 609                 nvme_free_prps(dev, req);
 610         mempool_free(iod->sg, dev->iod_mempool);
 611 }
 612
 613 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
 614 {
 615         int i;
 616         struct scatterlist *sg;
 617
 618         for_each_sg(sgl, sg, nents, i) {
 619                 dma_addr_t phys = sg_phys(sg);
 620                 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
 621                         "dma_address:%pad dma_length:%d\n",
 622                         i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
 623                         sg_dma_len(sg));
 624         }
 625 }
 626
 627 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
 628                 struct request *req, struct nvme_rw_command *cmnd)
 629 {
 630         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 631         struct dma_pool *pool;
 632         int length = blk_rq_payload_bytes(req);
 633         struct scatterlist *sg = iod->sg;
 634         int dma_len = sg_dma_len(sg);
 635         u64 dma_addr = sg_dma_address(sg);
 636         int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
 637         __le64 *prp_list;
 638         void **list = nvme_pci_iod_list(req);
 639         dma_addr_t prp_dma;
 640         int nprps, i;
 641
 642         length -= (NVME_CTRL_PAGE_SIZE - offset);
 643         if (length <= 0) {
 644                 iod->first_dma = 0;
 645                 goto done;
 646         }
 647
 648         dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
 649         if (dma_len) {
 650                 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
 651         } else {
 652                 sg = sg_next(sg);
 653                 dma_addr = sg_dma_address(sg);
 654                 dma_len = sg_dma_len(sg);
 655         }
 656
 657         if (length <= NVME_CTRL_PAGE_SIZE) {
 658                 iod->first_dma = dma_addr;
 659                 goto done;
 660         }
 661
 662         nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
 663         if (nprps <= (256 / 8)) {
 664                 pool = dev->prp_small_pool;
 665                 iod->npages = 0;
 666         } else {
 667                 pool = dev->prp_page_pool;
 668                 iod->npages = 1;
 669         }
 670
 671         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 672         if (!prp_list) {
 673                 iod->first_dma = dma_addr;
 674                 iod->npages = -1;
 675                 return BLK_STS_RESOURCE;
 676         }
 677         list[0] = prp_list;
 678         iod->first_dma = prp_dma;
 679         i = 0;
 680         for (;;) {
 681                 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
 682                         __le64 *old_prp_list = prp_list;
 683                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 684                         if (!prp_list)
 685                                 goto free_prps;
 686                         list[iod->npages++] = prp_list;
 687                         prp_list[0] = old_prp_list[i - 1];
 688                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 689                         i = 1;
 690                 }
 691                 prp_list[i++] = cpu_to_le64(dma_addr);
 692                 dma_len -= NVME_CTRL_PAGE_SIZE;
 693                 dma_addr += NVME_CTRL_PAGE_SIZE;
 694                 length -= NVME_CTRL_PAGE_SIZE;
 695                 if (length <= 0)
 696                         break;
 697                 if (dma_len > 0)
 698                         continue;
 699                 if (unlikely(dma_len < 0))
 700                         goto bad_sgl;
 701                 sg = sg_next(sg);
 702                 dma_addr = sg_dma_address(sg);
 703                 dma_len = sg_dma_len(sg);
 704         }
 705 done:
 706         cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 707         cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
 708         return BLK_STS_OK;
 709 free_prps:
 710         nvme_free_prps(dev, req);
 711         return BLK_STS_RESOURCE;
 712 bad_sgl:
 713         WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
 714                         "Invalid SGL for payload:%d nents:%d\n",
 715                         blk_rq_payload_bytes(req), iod->nents);
 716         return BLK_STS_IOERR;
 717 }
 718
 719 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
 720                 struct scatterlist *sg)
 721 {
 722         sge->addr = cpu_to_le64(sg_dma_address(sg));
 723         sge->length = cpu_to_le32(sg_dma_len(sg));
 724         sge->type = NVME_SGL_FMT_DATA_DESC << 4;
 725 }
 726
 727 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
 728                 dma_addr_t dma_addr, int entries)
 729 {
 730         sge->addr = cpu_to_le64(dma_addr);
 731         if (entries < SGES_PER_PAGE) {
 732                 sge->length = cpu_to_le32(entries * sizeof(*sge));
 733                 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
 734         } else {
 735                 sge->length = cpu_to_le32(PAGE_SIZE);
 736                 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
 737         }
 738 }
 739
 740 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
 741                 struct request *req, struct nvme_rw_command *cmd, int entries)
 742 {
 743         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 744         struct dma_pool *pool;
 745         struct nvme_sgl_desc *sg_list;
 746         struct scatterlist *sg = iod->sg;
 747         dma_addr_t sgl_dma;
 748         int i = 0;
 749
 750         /* setting the transfer type as SGL */
 751         cmd->flags = NVME_CMD_SGL_METABUF;
 752
 753         if (entries == 1) {
 754                 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
 755                 return BLK_STS_OK;
 756         }
 757
 758         if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
 759                 pool = dev->prp_small_pool;
 760                 iod->npages = 0;
 761         } else {
 762                 pool = dev->prp_page_pool;
 763                 iod->npages = 1;
 764         }
 765
 766         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 767         if (!sg_list) {
 768                 iod->npages = -1;
 769                 return BLK_STS_RESOURCE;
 770         }
 771
 772         nvme_pci_iod_list(req)[0] = sg_list;
 773         iod->first_dma = sgl_dma;
 774
 775         nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
 776
 777         do {
 778                 if (i == SGES_PER_PAGE) {
 779                         struct nvme_sgl_desc *old_sg_desc = sg_list;
 780                         struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
 781
 782                         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 783                         if (!sg_list)
 784                                 goto free_sgls;
 785
 786                         i = 0;
 787                         nvme_pci_iod_list(req)[iod->npages++] = sg_list;
 788                         sg_list[i++] = *link;
 789                         nvme_pci_sgl_set_seg(link, sgl_dma, entries);
 790                 }
 791
 792                 nvme_pci_sgl_set_data(&sg_list[i++], sg);
 793                 sg = sg_next(sg);
 794         } while (--entries > 0);
 795
 796         return BLK_STS_OK;
 797 free_sgls:
 798         nvme_free_sgls(dev, req);
 799         return BLK_STS_RESOURCE;
 800 }
 801
 802 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
 803                 struct request *req, struct nvme_rw_command *cmnd,
 804                 struct bio_vec *bv)
 805 {
 806         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 807         unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
 808         unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
 809
 810         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 811         if (dma_mapping_error(dev->dev, iod->first_dma))
 812                 return BLK_STS_RESOURCE;
 813         iod->dma_len = bv->bv_len;
 814
 815         cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
 816         if (bv->bv_len > first_prp_len)
 817                 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
 818         return BLK_STS_OK;
 819 }
 820
 821 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
 822                 struct request *req, struct nvme_rw_command *cmnd,
 823                 struct bio_vec *bv)
 824 {
 825         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 826
 827         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 828         if (dma_mapping_error(dev->dev, iod->first_dma))
 829                 return BLK_STS_RESOURCE;
 830         iod->dma_len = bv->bv_len;
 831
 832         cmnd->flags = NVME_CMD_SGL_METABUF;
 833         cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
 834         cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
 835         cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
 836         return BLK_STS_OK;
 837 }
 838
 839 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 840                 struct nvme_command *cmnd)
 841 {
 842         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 843         blk_status_t ret = BLK_STS_RESOURCE;
 844         int nr_mapped;
 845
 846         if (blk_rq_nr_phys_segments(req) == 1) {
 847                 struct bio_vec bv = req_bvec(req);
 848
 849                 if (!is_pci_p2pdma_page(bv.bv_page)) {
 850                         if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
 851                                 return nvme_setup_prp_simple(dev, req,
 852                                                              &cmnd->rw, &bv);
 853
 854                         if (iod->nvmeq->qid && sgl_threshold &&
 855                             nvme_ctrl_sgl_supported(&dev->ctrl))
 856                                 return nvme_setup_sgl_simple(dev, req,
 857                                                              &cmnd->rw, &bv);
 858                 }
 859         }
 860
 861         iod->dma_len = 0;
 862         iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
 863         if (!iod->sg)
 864                 return BLK_STS_RESOURCE;
 865         sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
 866         iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 867         if (!iod->nents)
 868                 goto out_free_sg;
 869
 870         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 871                 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
 872                                 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
 873         else
 874                 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
 875                                              rq_dma_dir(req), DMA_ATTR_NO_WARN);
 876         if (!nr_mapped)
 877                 goto out_free_sg;
 878
 879         iod->use_sgl = nvme_pci_use_sgls(dev, req);
 880         if (iod->use_sgl)
 881                 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
 882         else
 883                 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
 884         if (ret != BLK_STS_OK)
 885                 goto out_unmap_sg;
 886         return BLK_STS_OK;
 887
 888 out_unmap_sg:
 889         nvme_unmap_sg(dev, req);
 890 out_free_sg:
 891         mempool_free(iod->sg, dev->iod_mempool);
 892         return ret;
 893 }
 894
 895 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
 896                 struct nvme_command *cmnd)
 897 {
 898         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 899
 900         iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
 901                         rq_dma_dir(req), 0);
 902         if (dma_mapping_error(dev->dev, iod->meta_dma))
 903                 return BLK_STS_IOERR;
 904         cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
 905         return BLK_STS_OK;
 906 }
 907
 908 static blk_status_t nvme_prep_rq(struct nvme_dev *dev, struct request *req)
 909 {
 910         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 911         blk_status_t ret;
 912
 913         iod->aborted = 0;
 914         iod->npages = -1;
 915         iod->nents = 0;
 916
 917         ret = nvme_setup_cmd(req->q->queuedata, req);
 918         if (ret)
 919                 return ret;
 920
 921         if (blk_rq_nr_phys_segments(req)) {
 922                 ret = nvme_map_data(dev, req, &iod->cmd);
 923                 if (ret)
 924                         goto out_free_cmd;
 925         }
 926
 927         if (blk_integrity_rq(req)) {
 928                 ret = nvme_map_metadata(dev, req, &iod->cmd);
 929                 if (ret)
 930                         goto out_unmap_data;
 931         }
 932
 933         blk_mq_start_request(req);
 934         return BLK_STS_OK;
 935 out_unmap_data:
 936         nvme_unmap_data(dev, req);
 937 out_free_cmd:
 938         nvme_cleanup_cmd(req);
 939         return ret;
 940 }
 941
 942 /*
 943  * NOTE: ns is NULL when called on the admin queue.
 944  */
 945 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 946                          const struct blk_mq_queue_data *bd)
 947 {
 948         struct nvme_queue *nvmeq = hctx->driver_data;
 949         struct nvme_dev *dev = nvmeq->dev;
 950         struct request *req = bd->rq;
 951         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 952         blk_status_t ret;
 953
 954         /*
 955          * We should not need to do this, but we're still using this to
 956          * ensure we can drain requests on a dying queue.
 957          */
 958         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 959                 return BLK_STS_IOERR;
 960
 961         if (unlikely(!nvme_check_ready(&dev->ctrl, req, true)))
 962                 return nvme_fail_nonready_command(&dev->ctrl, req);
 963
 964         ret = nvme_prep_rq(dev, req);
 965         if (unlikely(ret))
 966                 return ret;
 967         spin_lock(&nvmeq->sq_lock);
 968         nvme_sq_copy_cmd(nvmeq, &iod->cmd);
 969         nvme_write_sq_db(nvmeq, bd->last);
 970         spin_unlock(&nvmeq->sq_lock);
 971         return BLK_STS_OK;
 972 }
 973
 974 static void nvme_submit_cmds(struct nvme_queue *nvmeq, struct request **rqlist)
 975 {
 976         spin_lock(&nvmeq->sq_lock);
 977         while (!rq_list_empty(*rqlist)) {
 978                 struct request *req = rq_list_pop(rqlist);
 979                 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 980
 981                 nvme_sq_copy_cmd(nvmeq, &iod->cmd);
 982         }
 983         nvme_write_sq_db(nvmeq, true);
 984         spin_unlock(&nvmeq->sq_lock);
 985 }
 986
 987 static bool nvme_prep_rq_batch(struct nvme_queue *nvmeq, struct request *req)
 988 {
 989         /*
 990          * We should not need to do this, but we're still using this to
 991          * ensure we can drain requests on a dying queue.
 992          */
 993         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 994                 return false;
 995         if (unlikely(!nvme_check_ready(&nvmeq->dev->ctrl, req, true)))
 996                 return false;
 997
 998         req->mq_hctx->tags->rqs[req->tag] = req;
 999         return nvme_prep_rq(nvmeq->dev, req) == BLK_STS_OK;
1000 }
1001
1002 static void nvme_queue_rqs(struct request **rqlist)
1003 {
1004         struct request *req, *next, *prev = NULL;
1005         struct request *requeue_list = NULL;
1006
1007         rq_list_for_each_safe(rqlist, req, next) {
1008                 struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1009
1010                 if (!nvme_prep_rq_batch(nvmeq, req)) {
1011                         /* detach 'req' and add to remainder list */
1012                         rq_list_move(rqlist, &requeue_list, req, prev);
1013
1014                         req = prev;
1015                         if (!req)
1016                                 continue;
1017                 }
1018
1019                 if (!next || req->mq_hctx != next->mq_hctx) {
1020                         /* detach rest of list, and submit */
1021                         req->rq_next = NULL;
1022                         nvme_submit_cmds(nvmeq, rqlist);
1023                         *rqlist = next;
1024                         prev = NULL;
1025                 } else
1026                         prev = req;
1027         }
1028
1029         *rqlist = requeue_list;
1030 }
1031
1032 static __always_inline void nvme_pci_unmap_rq(struct request *req)
1033 {
1034         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1035         struct nvme_dev *dev = iod->nvmeq->dev;
1036
1037         if (blk_integrity_rq(req))
1038                 dma_unmap_page(dev->dev, iod->meta_dma,
1039                                rq_integrity_vec(req)->bv_len, rq_data_dir(req));
1040         if (blk_rq_nr_phys_segments(req))
1041                 nvme_unmap_data(dev, req);
1042 }
1043
1044 static void nvme_pci_complete_rq(struct request *req)
1045 {
1046         nvme_pci_unmap_rq(req);
1047         nvme_complete_rq(req);
1048 }
1049
1050 static void nvme_pci_complete_batch(struct io_comp_batch *iob)
1051 {
1052         nvme_complete_batch(iob, nvme_pci_unmap_rq);
1053 }
1054
1055 /* We read the CQE phase first to check if the rest of the entry is valid */
1056 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
1057 {
1058         struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
1059
1060         return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
1061 }
1062
1063 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
1064 {
1065         u16 head = nvmeq->cq_head;
1066
1067         if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
1068                                               nvmeq->dbbuf_cq_ei))
1069                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
1070 }
1071
1072 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
1073 {
1074         if (!nvmeq->qid)
1075                 return nvmeq->dev->admin_tagset.tags[0];
1076         return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
1077 }
1078
1079 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
1080                                    struct io_comp_batch *iob, u16 idx)
1081 {
1082         struct nvme_completion *cqe = &nvmeq->cqes[idx];
1083         __u16 command_id = READ_ONCE(cqe->command_id);
1084         struct request *req;
1085
1086         /*
1087          * AEN requests are special as they don't time out and can
1088          * survive any kind of queue freeze and often don't respond to
1089          * aborts.  We don't even bother to allocate a struct request
1090          * for them but rather special case them here.
1091          */
1092         if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1093                 nvme_complete_async_event(&nvmeq->dev->ctrl,
1094                                 cqe->status, &cqe->result);
1095                 return;
1096         }
1097
1098         req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id);
1099         if (unlikely(!req)) {
1100                 dev_warn(nvmeq->dev->ctrl.device,
1101                         "invalid id %d completed on queue %d\n",
1102                         command_id, le16_to_cpu(cqe->sq_id));
1103                 return;
1104         }
1105
1106         trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
1107         if (!nvme_try_complete_req(req, cqe->status, cqe->result) &&
1108             !blk_mq_add_to_batch(req, iob, nvme_req(req)->status,
1109                                         nvme_pci_complete_batch))
1110                 nvme_pci_complete_rq(req);
1111 }
1112
1113 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1114 {
1115         u32 tmp = nvmeq->cq_head + 1;
1116
1117         if (tmp == nvmeq->q_depth) {
1118                 nvmeq->cq_head = 0;
1119                 nvmeq->cq_phase ^= 1;
1120         } else {
1121                 nvmeq->cq_head = tmp;
1122         }
1123 }
1124
1125 static inline int nvme_poll_cq(struct nvme_queue *nvmeq,
1126                                struct io_comp_batch *iob)
1127 {
1128         int found = 0;
1129
1130         while (nvme_cqe_pending(nvmeq)) {
1131                 found++;
1132                 /*
1133                  * load-load control dependency between phase and the rest of
1134                  * the cqe requires a full read memory barrier
1135                  */
1136                 dma_rmb();
1137                 nvme_handle_cqe(nvmeq, iob, nvmeq->cq_head);
1138                 nvme_update_cq_head(nvmeq);
1139         }
1140
1141         if (found)
1142                 nvme_ring_cq_doorbell(nvmeq);
1143         return found;
1144 }
1145
1146 static irqreturn_t nvme_irq(int irq, void *data)
1147 {
1148         struct nvme_queue *nvmeq = data;
1149         DEFINE_IO_COMP_BATCH(iob);
1150
1151         if (nvme_poll_cq(nvmeq, &iob)) {
1152                 if (!rq_list_empty(iob.req_list))
1153                         nvme_pci_complete_batch(&iob);
1154                 return IRQ_HANDLED;
1155         }
1156         return IRQ_NONE;
1157 }
1158
1159 static irqreturn_t nvme_irq_check(int irq, void *data)
1160 {
1161         struct nvme_queue *nvmeq = data;
1162
1163         if (nvme_cqe_pending(nvmeq))
1164                 return IRQ_WAKE_THREAD;
1165         return IRQ_NONE;
1166 }
1167
1168 /*
1169  * Poll for completions for any interrupt driven queue
1170  * Can be called from any context.
1171  */
1172 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1173 {
1174         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1175
1176         WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1177
1178         disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1179         nvme_poll_cq(nvmeq, NULL);
1180         enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1181 }
1182
1183 static int nvme_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1184 {
1185         struct nvme_queue *nvmeq = hctx->driver_data;
1186         bool found;
1187
1188         if (!nvme_cqe_pending(nvmeq))
1189                 return 0;
1190
1191         spin_lock(&nvmeq->cq_poll_lock);
1192         found = nvme_poll_cq(nvmeq, iob);
1193         spin_unlock(&nvmeq->cq_poll_lock);
1194
1195         return found;
1196 }
1197
1198 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1199 {
1200         struct nvme_dev *dev = to_nvme_dev(ctrl);
1201         struct nvme_queue *nvmeq = &dev->queues[0];
1202         struct nvme_command c = { };
1203
1204         c.common.opcode = nvme_admin_async_event;
1205         c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1206
1207         spin_lock(&nvmeq->sq_lock);
1208         nvme_sq_copy_cmd(nvmeq, &c);
1209         nvme_write_sq_db(nvmeq, true);
1210         spin_unlock(&nvmeq->sq_lock);
1211 }
1212
1213 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1214 {
1215         struct nvme_command c = { };
1216
1217         c.delete_queue.opcode = opcode;
1218         c.delete_queue.qid = cpu_to_le16(id);
1219
1220         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1221 }
1222
1223 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1224                 struct nvme_queue *nvmeq, s16 vector)
1225 {
1226         struct nvme_command c = { };
1227         int flags = NVME_QUEUE_PHYS_CONTIG;
1228
1229         if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1230                 flags |= NVME_CQ_IRQ_ENABLED;
1231
1232         /*
1233          * Note: we (ab)use the fact that the prp fields survive if no data
1234          * is attached to the request.
1235          */
1236         c.create_cq.opcode = nvme_admin_create_cq;
1237         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1238         c.create_cq.cqid = cpu_to_le16(qid);
1239         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1240         c.create_cq.cq_flags = cpu_to_le16(flags);
1241         c.create_cq.irq_vector = cpu_to_le16(vector);
1242
1243         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1244 }
1245
1246 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1247                                                 struct nvme_queue *nvmeq)
1248 {
1249         struct nvme_ctrl *ctrl = &dev->ctrl;
1250         struct nvme_command c = { };
1251         int flags = NVME_QUEUE_PHYS_CONTIG;
1252
1253         /*
1254          * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1255          * set. Since URGENT priority is zeroes, it makes all queues
1256          * URGENT.
1257          */
1258         if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1259                 flags |= NVME_SQ_PRIO_MEDIUM;
1260
1261         /*
1262          * Note: we (ab)use the fact that the prp fields survive if no data
1263          * is attached to the request.
1264          */
1265         c.create_sq.opcode = nvme_admin_create_sq;
1266         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1267         c.create_sq.sqid = cpu_to_le16(qid);
1268         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1269         c.create_sq.sq_flags = cpu_to_le16(flags);
1270         c.create_sq.cqid = cpu_to_le16(qid);
1271
1272         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1273 }
1274
1275 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1276 {
1277         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1278 }
1279
1280 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1281 {
1282         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1283 }
1284
1285 static void abort_endio(struct request *req, blk_status_t error)
1286 {
1287         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1288         struct nvme_queue *nvmeq = iod->nvmeq;
1289
1290         dev_warn(nvmeq->dev->ctrl.device,
1291                  "Abort status: 0x%x", nvme_req(req)->status);
1292         atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1293         blk_mq_free_request(req);
1294 }
1295
1296 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1297 {
1298         /* If true, indicates loss of adapter communication, possibly by a
1299          * NVMe Subsystem reset.
1300          */
1301         bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1302
1303         /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1304         switch (dev->ctrl.state) {
1305         case NVME_CTRL_RESETTING:
1306         case NVME_CTRL_CONNECTING:
1307                 return false;
1308         default:
1309                 break;
1310         }
1311
1312         /* We shouldn't reset unless the controller is on fatal error state
1313          * _or_ if we lost the communication with it.
1314          */
1315         if (!(csts & NVME_CSTS_CFS) && !nssro)
1316                 return false;
1317
1318         return true;
1319 }
1320
1321 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1322 {
1323         /* Read a config register to help see what died. */
1324         u16 pci_status;
1325         int result;
1326
1327         result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1328                                       &pci_status);
1329         if (result == PCIBIOS_SUCCESSFUL)
1330                 dev_warn(dev->ctrl.device,
1331                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1332                          csts, pci_status);
1333         else
1334                 dev_warn(dev->ctrl.device,
1335                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1336                          csts, result);
1337 }
1338
1339 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1340 {
1341         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1342         struct nvme_queue *nvmeq = iod->nvmeq;
1343         struct nvme_dev *dev = nvmeq->dev;
1344         struct request *abort_req;
1345         struct nvme_command cmd = { };
1346         u32 csts = readl(dev->bar + NVME_REG_CSTS);
1347
1348         /* If PCI error recovery process is happening, we cannot reset or
1349          * the recovery mechanism will surely fail.
1350          */
1351         mb();
1352         if (pci_channel_offline(to_pci_dev(dev->dev)))
1353                 return BLK_EH_RESET_TIMER;
1354
1355         /*
1356          * Reset immediately if the controller is failed
1357          */
1358         if (nvme_should_reset(dev, csts)) {
1359                 nvme_warn_reset(dev, csts);
1360                 nvme_dev_disable(dev, false);
1361                 nvme_reset_ctrl(&dev->ctrl);
1362                 return BLK_EH_DONE;
1363         }
1364
1365         /*
1366          * Did we miss an interrupt?
1367          */
1368         if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1369                 nvme_poll(req->mq_hctx, NULL);
1370         else
1371                 nvme_poll_irqdisable(nvmeq);
1372
1373         if (blk_mq_request_completed(req)) {
1374                 dev_warn(dev->ctrl.device,
1375                          "I/O %d QID %d timeout, completion polled\n",
1376                          req->tag, nvmeq->qid);
1377                 return BLK_EH_DONE;
1378         }
1379
1380         /*
1381          * Shutdown immediately if controller times out while starting. The
1382          * reset work will see the pci device disabled when it gets the forced
1383          * cancellation error. All outstanding requests are completed on
1384          * shutdown, so we return BLK_EH_DONE.
1385          */
1386         switch (dev->ctrl.state) {
1387         case NVME_CTRL_CONNECTING:
1388                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1389                 fallthrough;
1390         case NVME_CTRL_DELETING:
1391                 dev_warn_ratelimited(dev->ctrl.device,
1392                          "I/O %d QID %d timeout, disable controller\n",
1393                          req->tag, nvmeq->qid);
1394                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1395                 nvme_dev_disable(dev, true);
1396                 return BLK_EH_DONE;
1397         case NVME_CTRL_RESETTING:
1398                 return BLK_EH_RESET_TIMER;
1399         default:
1400                 break;
1401         }
1402
1403         /*
1404          * Shutdown the controller immediately and schedule a reset if the
1405          * command was already aborted once before and still hasn't been
1406          * returned to the driver, or if this is the admin queue.
1407          */
1408         if (!nvmeq->qid || iod->aborted) {
1409                 dev_warn(dev->ctrl.device,
1410                          "I/O %d QID %d timeout, reset controller\n",
1411                          req->tag, nvmeq->qid);
1412                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1413                 nvme_dev_disable(dev, false);
1414                 nvme_reset_ctrl(&dev->ctrl);
1415
1416                 return BLK_EH_DONE;
1417         }
1418
1419         if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1420                 atomic_inc(&dev->ctrl.abort_limit);
1421                 return BLK_EH_RESET_TIMER;
1422         }
1423         iod->aborted = 1;
1424
1425         cmd.abort.opcode = nvme_admin_abort_cmd;
1426         cmd.abort.cid = nvme_cid(req);
1427         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1428
1429         dev_warn(nvmeq->dev->ctrl.device,
1430                 "I/O %d QID %d timeout, aborting\n",
1431                  req->tag, nvmeq->qid);
1432
1433         abort_req = blk_mq_alloc_request(dev->ctrl.admin_q, nvme_req_op(&cmd),
1434                                          BLK_MQ_REQ_NOWAIT);
1435         if (IS_ERR(abort_req)) {
1436                 atomic_inc(&dev->ctrl.abort_limit);
1437                 return BLK_EH_RESET_TIMER;
1438         }
1439         nvme_init_request(abort_req, &cmd);
1440
1441         abort_req->end_io_data = NULL;
1442         blk_execute_rq_nowait(abort_req, false, abort_endio);
1443
1444         /*
1445          * The aborted req will be completed on receiving the abort req.
1446          * We enable the timer again. If hit twice, it'll cause a device reset,
1447          * as the device then is in a faulty state.
1448          */
1449         return BLK_EH_RESET_TIMER;
1450 }
1451
1452 static void nvme_free_queue(struct nvme_queue *nvmeq)
1453 {
1454         dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1455                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1456         if (!nvmeq->sq_cmds)
1457                 return;
1458
1459         if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1460                 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1461                                 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1462         } else {
1463                 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1464                                 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1465         }
1466 }
1467
1468 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1469 {
1470         int i;
1471
1472         for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1473                 dev->ctrl.queue_count--;
1474                 nvme_free_queue(&dev->queues[i]);
1475         }
1476 }
1477
1478 /**
1479  * nvme_suspend_queue - put queue into suspended state
1480  * @nvmeq: queue to suspend
1481  */
1482 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1483 {
1484         if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1485                 return 1;
1486
1487         /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1488         mb();
1489
1490         nvmeq->dev->online_queues--;
1491         if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1492                 nvme_stop_admin_queue(&nvmeq->dev->ctrl);
1493         if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1494                 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1495         return 0;
1496 }
1497
1498 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1499 {
1500         int i;
1501
1502         for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1503                 nvme_suspend_queue(&dev->queues[i]);
1504 }
1505
1506 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1507 {
1508         struct nvme_queue *nvmeq = &dev->queues[0];
1509
1510         if (shutdown)
1511                 nvme_shutdown_ctrl(&dev->ctrl);
1512         else
1513                 nvme_disable_ctrl(&dev->ctrl);
1514
1515         nvme_poll_irqdisable(nvmeq);
1516 }
1517
1518 /*
1519  * Called only on a device that has been disabled and after all other threads
1520  * that can check this device's completion queues have synced, except
1521  * nvme_poll(). This is the last chance for the driver to see a natural
1522  * completion before nvme_cancel_request() terminates all incomplete requests.
1523  */
1524 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1525 {
1526         int i;
1527
1528         for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1529                 spin_lock(&dev->queues[i].cq_poll_lock);
1530                 nvme_poll_cq(&dev->queues[i], NULL);
1531                 spin_unlock(&dev->queues[i].cq_poll_lock);
1532         }
1533 }
1534
1535 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1536                                 int entry_size)
1537 {
1538         int q_depth = dev->q_depth;
1539         unsigned q_size_aligned = roundup(q_depth * entry_size,
1540                                           NVME_CTRL_PAGE_SIZE);
1541
1542         if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1543                 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1544
1545                 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1546                 q_depth = div_u64(mem_per_q, entry_size);
1547
1548                 /*
1549                  * Ensure the reduced q_depth is above some threshold where it
1550                  * would be better to map queues in system memory with the
1551                  * original depth
1552                  */
1553                 if (q_depth < 64)
1554                         return -ENOMEM;
1555         }
1556
1557         return q_depth;
1558 }
1559
1560 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1561                                 int qid)
1562 {
1563         struct pci_dev *pdev = to_pci_dev(dev->dev);
1564
1565         if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1566                 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1567                 if (nvmeq->sq_cmds) {
1568                         nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1569                                                         nvmeq->sq_cmds);
1570                         if (nvmeq->sq_dma_addr) {
1571                                 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1572                                 return 0;
1573                         }
1574
1575                         pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1576                 }
1577         }
1578
1579         nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1580                                 &nvmeq->sq_dma_addr, GFP_KERNEL);
1581         if (!nvmeq->sq_cmds)
1582                 return -ENOMEM;
1583         return 0;
1584 }
1585
1586 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1587 {
1588         struct nvme_queue *nvmeq = &dev->queues[qid];
1589
1590         if (dev->ctrl.queue_count > qid)
1591                 return 0;
1592
1593         nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1594         nvmeq->q_depth = depth;
1595         nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1596                                          &nvmeq->cq_dma_addr, GFP_KERNEL);
1597         if (!nvmeq->cqes)
1598                 goto free_nvmeq;
1599
1600         if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1601                 goto free_cqdma;
1602
1603         nvmeq->dev = dev;
1604         spin_lock_init(&nvmeq->sq_lock);
1605         spin_lock_init(&nvmeq->cq_poll_lock);
1606         nvmeq->cq_head = 0;
1607         nvmeq->cq_phase = 1;
1608         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1609         nvmeq->qid = qid;
1610         dev->ctrl.queue_count++;
1611
1612         return 0;
1613
1614  free_cqdma:
1615         dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1616                           nvmeq->cq_dma_addr);
1617  free_nvmeq:
1618         return -ENOMEM;
1619 }
1620
1621 static int queue_request_irq(struct nvme_queue *nvmeq)
1622 {
1623         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1624         int nr = nvmeq->dev->ctrl.instance;
1625
1626         if (use_threaded_interrupts) {
1627                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1628                                 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1629         } else {
1630                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1631                                 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1632         }
1633 }
1634
1635 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1636 {
1637         struct nvme_dev *dev = nvmeq->dev;
1638
1639         nvmeq->sq_tail = 0;
1640         nvmeq->last_sq_tail = 0;
1641         nvmeq->cq_head = 0;
1642         nvmeq->cq_phase = 1;
1643         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1644         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1645         nvme_dbbuf_init(dev, nvmeq, qid);
1646         dev->online_queues++;
1647         wmb(); /* ensure the first interrupt sees the initialization */
1648 }
1649
1650 /*
1651  * Try getting shutdown_lock while setting up IO queues.
1652  */
1653 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev)
1654 {
1655         /*
1656          * Give up if the lock is being held by nvme_dev_disable.
1657          */
1658         if (!mutex_trylock(&dev->shutdown_lock))
1659                 return -ENODEV;
1660
1661         /*
1662          * Controller is in wrong state, fail early.
1663          */
1664         if (dev->ctrl.state != NVME_CTRL_CONNECTING) {
1665                 mutex_unlock(&dev->shutdown_lock);
1666                 return -ENODEV;
1667         }
1668
1669         return 0;
1670 }
1671
1672 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1673 {
1674         struct nvme_dev *dev = nvmeq->dev;
1675         int result;
1676         u16 vector = 0;
1677
1678         clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1679
1680         /*
1681          * A queue's vector matches the queue identifier unless the controller
1682          * has only one vector available.
1683          */
1684         if (!polled)
1685                 vector = dev->num_vecs == 1 ? 0 : qid;
1686         else
1687                 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1688
1689         result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1690         if (result)
1691                 return result;
1692
1693         result = adapter_alloc_sq(dev, qid, nvmeq);
1694         if (result < 0)
1695                 return result;
1696         if (result)
1697                 goto release_cq;
1698
1699         nvmeq->cq_vector = vector;
1700
1701         result = nvme_setup_io_queues_trylock(dev);
1702         if (result)
1703                 return result;
1704         nvme_init_queue(nvmeq, qid);
1705         if (!polled) {
1706                 result = queue_request_irq(nvmeq);
1707                 if (result < 0)
1708                         goto release_sq;
1709         }
1710
1711         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1712         mutex_unlock(&dev->shutdown_lock);
1713         return result;
1714
1715 release_sq:
1716         dev->online_queues--;
1717         mutex_unlock(&dev->shutdown_lock);
1718         adapter_delete_sq(dev, qid);
1719 release_cq:
1720         adapter_delete_cq(dev, qid);
1721         return result;
1722 }
1723
1724 static const struct blk_mq_ops nvme_mq_admin_ops = {
1725         .queue_rq       = nvme_queue_rq,
1726         .complete       = nvme_pci_complete_rq,
1727         .init_hctx      = nvme_admin_init_hctx,
1728         .init_request   = nvme_pci_init_request,
1729         .timeout        = nvme_timeout,
1730 };
1731
1732 static const struct blk_mq_ops nvme_mq_ops = {
1733         .queue_rq       = nvme_queue_rq,
1734         .queue_rqs      = nvme_queue_rqs,
1735         .complete       = nvme_pci_complete_rq,
1736         .commit_rqs     = nvme_commit_rqs,
1737         .init_hctx      = nvme_init_hctx,
1738         .init_request   = nvme_pci_init_request,
1739         .map_queues     = nvme_pci_map_queues,
1740         .timeout        = nvme_timeout,
1741         .poll           = nvme_poll,
1742 };
1743
1744 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1745 {
1746         if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1747                 /*
1748                  * If the controller was reset during removal, it's possible
1749                  * user requests may be waiting on a stopped queue. Start the
1750                  * queue to flush these to completion.
1751                  */
1752                 nvme_start_admin_queue(&dev->ctrl);
1753                 blk_cleanup_queue(dev->ctrl.admin_q);
1754                 blk_mq_free_tag_set(&dev->admin_tagset);
1755         }
1756 }
1757
1758 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1759 {
1760         if (!dev->ctrl.admin_q) {
1761                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1762                 dev->admin_tagset.nr_hw_queues = 1;
1763
1764                 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1765                 dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1766                 dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1767                 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1768                 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1769                 dev->admin_tagset.driver_data = dev;
1770
1771                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1772                         return -ENOMEM;
1773                 dev->ctrl.admin_tagset = &dev->admin_tagset;
1774
1775                 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1776                 if (IS_ERR(dev->ctrl.admin_q)) {
1777                         blk_mq_free_tag_set(&dev->admin_tagset);
1778                         return -ENOMEM;
1779                 }
1780                 if (!blk_get_queue(dev->ctrl.admin_q)) {
1781                         nvme_dev_remove_admin(dev);
1782                         dev->ctrl.admin_q = NULL;
1783                         return -ENODEV;
1784                 }
1785         } else
1786                 nvme_start_admin_queue(&dev->ctrl);
1787
1788         return 0;
1789 }
1790
1791 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1792 {
1793         return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1794 }
1795
1796 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1797 {
1798         struct pci_dev *pdev = to_pci_dev(dev->dev);
1799
1800         if (size <= dev->bar_mapped_size)
1801                 return 0;
1802         if (size > pci_resource_len(pdev, 0))
1803                 return -ENOMEM;
1804         if (dev->bar)
1805                 iounmap(dev->bar);
1806         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1807         if (!dev->bar) {
1808                 dev->bar_mapped_size = 0;
1809                 return -ENOMEM;
1810         }
1811         dev->bar_mapped_size = size;
1812         dev->dbs = dev->bar + NVME_REG_DBS;
1813
1814         return 0;
1815 }
1816
1817 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1818 {
1819         int result;
1820         u32 aqa;
1821         struct nvme_queue *nvmeq;
1822
1823         result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1824         if (result < 0)
1825                 return result;
1826
1827         dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1828                                 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1829
1830         if (dev->subsystem &&
1831             (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1832                 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1833
1834         result = nvme_disable_ctrl(&dev->ctrl);
1835         if (result < 0)
1836                 return result;
1837
1838         result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1839         if (result)
1840                 return result;
1841
1842         dev->ctrl.numa_node = dev_to_node(dev->dev);
1843
1844         nvmeq = &dev->queues[0];
1845         aqa = nvmeq->q_depth - 1;
1846         aqa |= aqa << 16;
1847
1848         writel(aqa, dev->bar + NVME_REG_AQA);
1849         lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1850         lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1851
1852         result = nvme_enable_ctrl(&dev->ctrl);
1853         if (result)
1854                 return result;
1855
1856         nvmeq->cq_vector = 0;
1857         nvme_init_queue(nvmeq, 0);
1858         result = queue_request_irq(nvmeq);
1859         if (result) {
1860                 dev->online_queues--;
1861                 return result;
1862         }
1863
1864         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1865         return result;
1866 }
1867
1868 static int nvme_create_io_queues(struct nvme_dev *dev)
1869 {
1870         unsigned i, max, rw_queues;
1871         int ret = 0;
1872
1873         for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1874                 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1875                         ret = -ENOMEM;
1876                         break;
1877                 }
1878         }
1879
1880         max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1881         if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1882                 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1883                                 dev->io_queues[HCTX_TYPE_READ];
1884         } else {
1885                 rw_queues = max;
1886         }
1887
1888         for (i = dev->online_queues; i <= max; i++) {
1889                 bool polled = i > rw_queues;
1890
1891                 ret = nvme_create_queue(&dev->queues[i], i, polled);
1892                 if (ret)
1893                         break;
1894         }
1895
1896         /*
1897          * Ignore failing Create SQ/CQ commands, we can continue with less
1898          * than the desired amount of queues, and even a controller without
1899          * I/O queues can still be used to issue admin commands.  This might
1900          * be useful to upgrade a buggy firmware for example.
1901          */
1902         return ret >= 0 ? 0 : ret;
1903 }
1904
1905 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1906 {
1907         u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1908
1909         return 1ULL << (12 + 4 * szu);
1910 }
1911
1912 static u32 nvme_cmb_size(struct nvme_dev *dev)
1913 {
1914         return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1915 }
1916
1917 static void nvme_map_cmb(struct nvme_dev *dev)
1918 {
1919         u64 size, offset;
1920         resource_size_t bar_size;
1921         struct pci_dev *pdev = to_pci_dev(dev->dev);
1922         int bar;
1923
1924         if (dev->cmb_size)
1925                 return;
1926
1927         if (NVME_CAP_CMBS(dev->ctrl.cap))
1928                 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1929
1930         dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1931         if (!dev->cmbsz)
1932                 return;
1933         dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1934
1935         size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1936         offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1937         bar = NVME_CMB_BIR(dev->cmbloc);
1938         bar_size = pci_resource_len(pdev, bar);
1939
1940         if (offset > bar_size)
1941                 return;
1942
1943         /*
1944          * Tell the controller about the host side address mapping the CMB,
1945          * and enable CMB decoding for the NVMe 1.4+ scheme:
1946          */
1947         if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1948                 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1949                              (pci_bus_address(pdev, bar) + offset),
1950                              dev->bar + NVME_REG_CMBMSC);
1951         }
1952
1953         /*
1954          * Controllers may support a CMB size larger than their BAR,
1955          * for example, due to being behind a bridge. Reduce the CMB to
1956          * the reported size of the BAR
1957          */
1958         if (size > bar_size - offset)
1959                 size = bar_size - offset;
1960
1961         if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1962                 dev_warn(dev->ctrl.device,
1963                          "failed to register the CMB\n");
1964                 return;
1965         }
1966
1967         dev->cmb_size = size;
1968         dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1969
1970         if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1971                         (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1972                 pci_p2pmem_publish(pdev, true);
1973 }
1974
1975 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1976 {
1977         u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1978         u64 dma_addr = dev->host_mem_descs_dma;
1979         struct nvme_command c = { };
1980         int ret;
1981
1982         c.features.opcode       = nvme_admin_set_features;
1983         c.features.fid          = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1984         c.features.dword11      = cpu_to_le32(bits);
1985         c.features.dword12      = cpu_to_le32(host_mem_size);
1986         c.features.dword13      = cpu_to_le32(lower_32_bits(dma_addr));
1987         c.features.dword14      = cpu_to_le32(upper_32_bits(dma_addr));
1988         c.features.dword15      = cpu_to_le32(dev->nr_host_mem_descs);
1989
1990         ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1991         if (ret) {
1992                 dev_warn(dev->ctrl.device,
1993                          "failed to set host mem (err %d, flags %#x).\n",
1994                          ret, bits);
1995         } else
1996                 dev->hmb = bits & NVME_HOST_MEM_ENABLE;
1997
1998         return ret;
1999 }
2000
2001 static void nvme_free_host_mem(struct nvme_dev *dev)
2002 {
2003         int i;
2004
2005         for (i = 0; i < dev->nr_host_mem_descs; i++) {
2006                 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
2007                 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
2008
2009                 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
2010                                le64_to_cpu(desc->addr),
2011                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2012         }
2013
2014         kfree(dev->host_mem_desc_bufs);
2015         dev->host_mem_desc_bufs = NULL;
2016         dma_free_coherent(dev->dev,
2017                         dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
2018                         dev->host_mem_descs, dev->host_mem_descs_dma);
2019         dev->host_mem_descs = NULL;
2020         dev->nr_host_mem_descs = 0;
2021 }
2022
2023 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
2024                 u32 chunk_size)
2025 {
2026         struct nvme_host_mem_buf_desc *descs;
2027         u32 max_entries, len;
2028         dma_addr_t descs_dma;
2029         int i = 0;
2030         void **bufs;
2031         u64 size, tmp;
2032
2033         tmp = (preferred + chunk_size - 1);
2034         do_div(tmp, chunk_size);
2035         max_entries = tmp;
2036
2037         if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
2038                 max_entries = dev->ctrl.hmmaxd;
2039
2040         descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
2041                                    &descs_dma, GFP_KERNEL);
2042         if (!descs)
2043                 goto out;
2044
2045         bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
2046         if (!bufs)
2047                 goto out_free_descs;
2048
2049         for (size = 0; size < preferred && i < max_entries; size += len) {
2050                 dma_addr_t dma_addr;
2051
2052                 len = min_t(u64, chunk_size, preferred - size);
2053                 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
2054                                 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2055                 if (!bufs[i])
2056                         break;
2057
2058                 descs[i].addr = cpu_to_le64(dma_addr);
2059                 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
2060                 i++;
2061         }
2062
2063         if (!size)
2064                 goto out_free_bufs;
2065
2066         dev->nr_host_mem_descs = i;
2067         dev->host_mem_size = size;
2068         dev->host_mem_descs = descs;
2069         dev->host_mem_descs_dma = descs_dma;
2070         dev->host_mem_desc_bufs = bufs;
2071         return 0;
2072
2073 out_free_bufs:
2074         while (--i >= 0) {
2075                 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
2076
2077                 dma_free_attrs(dev->dev, size, bufs[i],
2078                                le64_to_cpu(descs[i].addr),
2079                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2080         }
2081
2082         kfree(bufs);
2083 out_free_descs:
2084         dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
2085                         descs_dma);
2086 out:
2087         dev->host_mem_descs = NULL;
2088         return -ENOMEM;
2089 }
2090
2091 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2092 {
2093         u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2094         u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2095         u64 chunk_size;
2096
2097         /* start big and work our way down */
2098         for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2099                 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
2100                         if (!min || dev->host_mem_size >= min)
2101                                 return 0;
2102                         nvme_free_host_mem(dev);
2103                 }
2104         }
2105
2106         return -ENOMEM;
2107 }
2108
2109 static int nvme_setup_host_mem(struct nvme_dev *dev)
2110 {
2111         u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2112         u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2113         u64 min = (u64)dev->ctrl.hmmin * 4096;
2114         u32 enable_bits = NVME_HOST_MEM_ENABLE;
2115         int ret;
2116
2117         preferred = min(preferred, max);
2118         if (min > max) {
2119                 dev_warn(dev->ctrl.device,
2120                         "min host memory (%lld MiB) above limit (%d MiB).\n",
2121                         min >> ilog2(SZ_1M), max_host_mem_size_mb);
2122                 nvme_free_host_mem(dev);
2123                 return 0;
2124         }
2125
2126         /*
2127          * If we already have a buffer allocated check if we can reuse it.
2128          */
2129         if (dev->host_mem_descs) {
2130                 if (dev->host_mem_size >= min)
2131                         enable_bits |= NVME_HOST_MEM_RETURN;
2132                 else
2133                         nvme_free_host_mem(dev);
2134         }
2135
2136         if (!dev->host_mem_descs) {
2137                 if (nvme_alloc_host_mem(dev, min, preferred)) {
2138                         dev_warn(dev->ctrl.device,
2139                                 "failed to allocate host memory buffer.\n");
2140                         return 0; /* controller must work without HMB */
2141                 }
2142
2143                 dev_info(dev->ctrl.device,
2144                         "allocated %lld MiB host memory buffer.\n",
2145                         dev->host_mem_size >> ilog2(SZ_1M));
2146         }
2147
2148         ret = nvme_set_host_mem(dev, enable_bits);
2149         if (ret)
2150                 nvme_free_host_mem(dev);
2151         return ret;
2152 }
2153
2154 static ssize_t cmb_show(struct device *dev, struct device_attribute *attr,
2155                 char *buf)
2156 {
2157         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2158
2159         return sysfs_emit(buf, "cmbloc : x%08x\ncmbsz  : x%08x\n",
2160                        ndev->cmbloc, ndev->cmbsz);
2161 }
2162 static DEVICE_ATTR_RO(cmb);
2163
2164 static ssize_t cmbloc_show(struct device *dev, struct device_attribute *attr,
2165                 char *buf)
2166 {
2167         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2168
2169         return sysfs_emit(buf, "%u\n", ndev->cmbloc);
2170 }
2171 static DEVICE_ATTR_RO(cmbloc);
2172
2173 static ssize_t cmbsz_show(struct device *dev, struct device_attribute *attr,
2174                 char *buf)
2175 {
2176         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2177
2178         return sysfs_emit(buf, "%u\n", ndev->cmbsz);
2179 }
2180 static DEVICE_ATTR_RO(cmbsz);
2181
2182 static ssize_t hmb_show(struct device *dev, struct device_attribute *attr,
2183                         char *buf)
2184 {
2185         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2186
2187         return sysfs_emit(buf, "%d\n", ndev->hmb);
2188 }
2189
2190 static ssize_t hmb_store(struct device *dev, struct device_attribute *attr,
2191                          const char *buf, size_t count)
2192 {
2193         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2194         bool new;
2195         int ret;
2196
2197         if (strtobool(buf, &new) < 0)
2198                 return -EINVAL;
2199
2200         if (new == ndev->hmb)
2201                 return count;
2202
2203         if (new) {
2204                 ret = nvme_setup_host_mem(ndev);
2205         } else {
2206                 ret = nvme_set_host_mem(ndev, 0);
2207                 if (!ret)
2208                         nvme_free_host_mem(ndev);
2209         }
2210
2211         if (ret < 0)
2212                 return ret;
2213
2214         return count;
2215 }
2216 static DEVICE_ATTR_RW(hmb);
2217
2218 static umode_t nvme_pci_attrs_are_visible(struct kobject *kobj,
2219                 struct attribute *a, int n)
2220 {
2221         struct nvme_ctrl *ctrl =
2222                 dev_get_drvdata(container_of(kobj, struct device, kobj));
2223         struct nvme_dev *dev = to_nvme_dev(ctrl);
2224
2225         if (a == &dev_attr_cmb.attr ||
2226             a == &dev_attr_cmbloc.attr ||
2227             a == &dev_attr_cmbsz.attr) {
2228                 if (!dev->cmbsz)
2229                         return 0;
2230         }
2231         if (a == &dev_attr_hmb.attr && !ctrl->hmpre)
2232                 return 0;
2233
2234         return a->mode;
2235 }
2236
2237 static struct attribute *nvme_pci_attrs[] = {
2238         &dev_attr_cmb.attr,
2239         &dev_attr_cmbloc.attr,
2240         &dev_attr_cmbsz.attr,
2241         &dev_attr_hmb.attr,
2242         NULL,
2243 };
2244
2245 static const struct attribute_group nvme_pci_attr_group = {
2246         .attrs          = nvme_pci_attrs,
2247         .is_visible     = nvme_pci_attrs_are_visible,
2248 };
2249
2250 /*
2251  * nirqs is the number of interrupts available for write and read
2252  * queues. The core already reserved an interrupt for the admin queue.
2253  */
2254 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2255 {
2256         struct nvme_dev *dev = affd->priv;
2257         unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2258
2259         /*
2260          * If there is no interrupt available for queues, ensure that
2261          * the default queue is set to 1. The affinity set size is
2262          * also set to one, but the irq core ignores it for this case.
2263          *
2264          * If only one interrupt is available or 'write_queue' == 0, combine
2265          * write and read queues.
2266          *
2267          * If 'write_queues' > 0, ensure it leaves room for at least one read
2268          * queue.
2269          */
2270         if (!nrirqs) {
2271                 nrirqs = 1;
2272                 nr_read_queues = 0;
2273         } else if (nrirqs == 1 || !nr_write_queues) {
2274                 nr_read_queues = 0;
2275         } else if (nr_write_queues >= nrirqs) {
2276                 nr_read_queues = 1;
2277         } else {
2278                 nr_read_queues = nrirqs - nr_write_queues;
2279         }
2280
2281         dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2282         affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2283         dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2284         affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2285         affd->nr_sets = nr_read_queues ? 2 : 1;
2286 }
2287
2288 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2289 {
2290         struct pci_dev *pdev = to_pci_dev(dev->dev);
2291         struct irq_affinity affd = {
2292                 .pre_vectors    = 1,
2293                 .calc_sets      = nvme_calc_irq_sets,
2294                 .priv           = dev,
2295         };
2296         unsigned int irq_queues, poll_queues;
2297
2298         /*
2299          * Poll queues don't need interrupts, but we need at least one I/O queue
2300          * left over for non-polled I/O.
2301          */
2302         poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2303         dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2304
2305         /*
2306          * Initialize for the single interrupt case, will be updated in
2307          * nvme_calc_irq_sets().
2308          */
2309         dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2310         dev->io_queues[HCTX_TYPE_READ] = 0;
2311
2312         /*
2313          * We need interrupts for the admin queue and each non-polled I/O queue,
2314          * but some Apple controllers require all queues to use the first
2315          * vector.
2316          */
2317         irq_queues = 1;
2318         if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2319                 irq_queues += (nr_io_queues - poll_queues);
2320         return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2321                               PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2322 }
2323
2324 static void nvme_disable_io_queues(struct nvme_dev *dev)
2325 {
2326         if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2327                 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2328 }
2329
2330 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2331 {
2332         /*
2333          * If tags are shared with admin queue (Apple bug), then
2334          * make sure we only use one IO queue.
2335          */
2336         if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2337                 return 1;
2338         return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2339 }
2340
2341 static int nvme_setup_io_queues(struct nvme_dev *dev)
2342 {
2343         struct nvme_queue *adminq = &dev->queues[0];
2344         struct pci_dev *pdev = to_pci_dev(dev->dev);
2345         unsigned int nr_io_queues;
2346         unsigned long size;
2347         int result;
2348
2349         /*
2350          * Sample the module parameters once at reset time so that we have
2351          * stable values to work with.
2352          */
2353         dev->nr_write_queues = write_queues;
2354         dev->nr_poll_queues = poll_queues;
2355
2356         nr_io_queues = dev->nr_allocated_queues - 1;
2357         result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2358         if (result < 0)
2359                 return result;
2360
2361         if (nr_io_queues == 0)
2362                 return 0;
2363
2364         /*
2365          * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions
2366          * from set to unset. If there is a window to it is truely freed,
2367          * pci_free_irq_vectors() jumping into this window will crash.
2368          * And take lock to avoid racing with pci_free_irq_vectors() in
2369          * nvme_dev_disable() path.
2370          */
2371         result = nvme_setup_io_queues_trylock(dev);
2372         if (result)
2373                 return result;
2374         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2375                 pci_free_irq(pdev, 0, adminq);
2376
2377         if (dev->cmb_use_sqes) {
2378                 result = nvme_cmb_qdepth(dev, nr_io_queues,
2379                                 sizeof(struct nvme_command));
2380                 if (result > 0)
2381                         dev->q_depth = result;
2382                 else
2383                         dev->cmb_use_sqes = false;
2384         }
2385
2386         do {
2387                 size = db_bar_size(dev, nr_io_queues);
2388                 result = nvme_remap_bar(dev, size);
2389                 if (!result)
2390                         break;
2391                 if (!--nr_io_queues) {
2392                         result = -ENOMEM;
2393                         goto out_unlock;
2394                 }
2395         } while (1);
2396         adminq->q_db = dev->dbs;
2397
2398  retry:
2399         /* Deregister the admin queue's interrupt */
2400         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2401                 pci_free_irq(pdev, 0, adminq);
2402
2403         /*
2404          * If we enable msix early due to not intx, disable it again before
2405          * setting up the full range we need.
2406          */
2407         pci_free_irq_vectors(pdev);
2408
2409         result = nvme_setup_irqs(dev, nr_io_queues);
2410         if (result <= 0) {
2411                 result = -EIO;
2412                 goto out_unlock;
2413         }
2414
2415         dev->num_vecs = result;
2416         result = max(result - 1, 1);
2417         dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2418
2419         /*
2420          * Should investigate if there's a performance win from allocating
2421          * more queues than interrupt vectors; it might allow the submission
2422          * path to scale better, even if the receive path is limited by the
2423          * number of interrupts.
2424          */
2425         result = queue_request_irq(adminq);
2426         if (result)
2427                 goto out_unlock;
2428         set_bit(NVMEQ_ENABLED, &adminq->flags);
2429         mutex_unlock(&dev->shutdown_lock);
2430
2431         result = nvme_create_io_queues(dev);
2432         if (result || dev->online_queues < 2)
2433                 return result;
2434
2435         if (dev->online_queues - 1 < dev->max_qid) {
2436                 nr_io_queues = dev->online_queues - 1;
2437                 nvme_disable_io_queues(dev);
2438                 result = nvme_setup_io_queues_trylock(dev);
2439                 if (result)
2440                         return result;
2441                 nvme_suspend_io_queues(dev);
2442                 goto retry;
2443         }
2444         dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2445                                         dev->io_queues[HCTX_TYPE_DEFAULT],
2446                                         dev->io_queues[HCTX_TYPE_READ],
2447                                         dev->io_queues[HCTX_TYPE_POLL]);
2448         return 0;
2449 out_unlock:
2450         mutex_unlock(&dev->shutdown_lock);
2451         return result;
2452 }
2453
2454 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2455 {
2456         struct nvme_queue *nvmeq = req->end_io_data;
2457
2458         blk_mq_free_request(req);
2459         complete(&nvmeq->delete_done);
2460 }
2461
2462 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2463 {
2464         struct nvme_queue *nvmeq = req->end_io_data;
2465
2466         if (error)
2467                 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2468
2469         nvme_del_queue_end(req, error);
2470 }
2471
2472 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2473 {
2474         struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2475         struct request *req;
2476         struct nvme_command cmd = { };
2477
2478         cmd.delete_queue.opcode = opcode;
2479         cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2480
2481         req = blk_mq_alloc_request(q, nvme_req_op(&cmd), BLK_MQ_REQ_NOWAIT);
2482         if (IS_ERR(req))
2483                 return PTR_ERR(req);
2484         nvme_init_request(req, &cmd);
2485
2486         req->end_io_data = nvmeq;
2487
2488         init_completion(&nvmeq->delete_done);
2489         blk_execute_rq_nowait(req, false, opcode == nvme_admin_delete_cq ?
2490                         nvme_del_cq_end : nvme_del_queue_end);
2491         return 0;
2492 }
2493
2494 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2495 {
2496         int nr_queues = dev->online_queues - 1, sent = 0;
2497         unsigned long timeout;
2498
2499  retry:
2500         timeout = NVME_ADMIN_TIMEOUT;
2501         while (nr_queues > 0) {
2502                 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2503                         break;
2504                 nr_queues--;
2505                 sent++;
2506         }
2507         while (sent) {
2508                 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2509
2510                 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2511                                 timeout);
2512                 if (timeout == 0)
2513                         return false;
2514
2515                 sent--;
2516                 if (nr_queues)
2517                         goto retry;
2518         }
2519         return true;
2520 }
2521
2522 static void nvme_dev_add(struct nvme_dev *dev)
2523 {
2524         int ret;
2525
2526         if (!dev->ctrl.tagset) {
2527                 dev->tagset.ops = &nvme_mq_ops;
2528                 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2529                 dev->tagset.nr_maps = 2; /* default + read */
2530                 if (dev->io_queues[HCTX_TYPE_POLL])
2531                         dev->tagset.nr_maps++;
2532                 dev->tagset.timeout = NVME_IO_TIMEOUT;
2533                 dev->tagset.numa_node = dev->ctrl.numa_node;
2534                 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2535                                                 BLK_MQ_MAX_DEPTH) - 1;
2536                 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2537                 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2538                 dev->tagset.driver_data = dev;
2539
2540                 /*
2541                  * Some Apple controllers requires tags to be unique
2542                  * across admin and IO queue, so reserve the first 32
2543                  * tags of the IO queue.
2544                  */
2545                 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2546                         dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2547
2548                 ret = blk_mq_alloc_tag_set(&dev->tagset);
2549                 if (ret) {
2550                         dev_warn(dev->ctrl.device,
2551                                 "IO queues tagset allocation failed %d\n", ret);
2552                         return;
2553                 }
2554                 dev->ctrl.tagset = &dev->tagset;
2555         } else {
2556                 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2557
2558                 /* Free previously allocated queues that are no longer usable */
2559                 nvme_free_queues(dev, dev->online_queues);
2560         }
2561
2562         nvme_dbbuf_set(dev);
2563 }
2564
2565 static int nvme_pci_enable(struct nvme_dev *dev)
2566 {
2567         int result = -ENOMEM;
2568         struct pci_dev *pdev = to_pci_dev(dev->dev);
2569         int dma_address_bits = 64;
2570
2571         if (pci_enable_device_mem(pdev))
2572                 return result;
2573
2574         pci_set_master(pdev);
2575
2576         if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48)
2577                 dma_address_bits = 48;
2578         if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(dma_address_bits)))
2579                 goto disable;
2580
2581         if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2582                 result = -ENODEV;
2583                 goto disable;
2584         }
2585
2586         /*
2587          * Some devices and/or platforms don't advertise or work with INTx
2588          * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2589          * adjust this later.
2590          */
2591         result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2592         if (result < 0)
2593                 return result;
2594
2595         dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2596
2597         dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2598                                 io_queue_depth);
2599         dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2600         dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2601         dev->dbs = dev->bar + 4096;
2602
2603         /*
2604          * Some Apple controllers require a non-standard SQE size.
2605          * Interestingly they also seem to ignore the CC:IOSQES register
2606          * so we don't bother updating it here.
2607          */
2608         if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2609                 dev->io_sqes = 7;
2610         else
2611                 dev->io_sqes = NVME_NVM_IOSQES;
2612
2613         /*
2614          * Temporary fix for the Apple controller found in the MacBook8,1 and
2615          * some MacBook7,1 to avoid controller resets and data loss.
2616          */
2617         if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2618                 dev->q_depth = 2;
2619                 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2620                         "set queue depth=%u to work around controller resets\n",
2621                         dev->q_depth);
2622         } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2623                    (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2624                    NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2625                 dev->q_depth = 64;
2626                 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2627                         "set queue depth=%u\n", dev->q_depth);
2628         }
2629
2630         /*
2631          * Controllers with the shared tags quirk need the IO queue to be
2632          * big enough so that we get 32 tags for the admin queue
2633          */
2634         if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2635             (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2636                 dev->q_depth = NVME_AQ_DEPTH + 2;
2637                 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2638                          dev->q_depth);
2639         }
2640
2641
2642         nvme_map_cmb(dev);
2643
2644         pci_enable_pcie_error_reporting(pdev);
2645         pci_save_state(pdev);
2646         return 0;
2647
2648  disable:
2649         pci_disable_device(pdev);
2650         return result;
2651 }
2652
2653 static void nvme_dev_unmap(struct nvme_dev *dev)
2654 {
2655         if (dev->bar)
2656                 iounmap(dev->bar);
2657         pci_release_mem_regions(to_pci_dev(dev->dev));
2658 }
2659
2660 static void nvme_pci_disable(struct nvme_dev *dev)
2661 {
2662         struct pci_dev *pdev = to_pci_dev(dev->dev);
2663
2664         pci_free_irq_vectors(pdev);
2665
2666         if (pci_is_enabled(pdev)) {
2667                 pci_disable_pcie_error_reporting(pdev);
2668                 pci_disable_device(pdev);
2669         }
2670 }
2671
2672 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2673 {
2674         bool dead = true, freeze = false;
2675         struct pci_dev *pdev = to_pci_dev(dev->dev);
2676
2677         mutex_lock(&dev->shutdown_lock);
2678         if (pci_is_enabled(pdev)) {
2679                 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2680
2681                 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2682                     dev->ctrl.state == NVME_CTRL_RESETTING) {
2683                         freeze = true;
2684                         nvme_start_freeze(&dev->ctrl);
2685                 }
2686                 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2687                         pdev->error_state  != pci_channel_io_normal);
2688         }
2689
2690         /*
2691          * Give the controller a chance to complete all entered requests if
2692          * doing a safe shutdown.
2693          */
2694         if (!dead && shutdown && freeze)
2695                 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2696
2697         nvme_stop_queues(&dev->ctrl);
2698
2699         if (!dead && dev->ctrl.queue_count > 0) {
2700                 nvme_disable_io_queues(dev);
2701                 nvme_disable_admin_queue(dev, shutdown);
2702         }
2703         nvme_suspend_io_queues(dev);
2704         nvme_suspend_queue(&dev->queues[0]);
2705         nvme_pci_disable(dev);
2706         nvme_reap_pending_cqes(dev);
2707
2708         blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2709         blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2710         blk_mq_tagset_wait_completed_request(&dev->tagset);
2711         blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2712
2713         /*
2714          * The driver will not be starting up queues again if shutting down so
2715          * must flush all entered requests to their failed completion to avoid
2716          * deadlocking blk-mq hot-cpu notifier.
2717          */
2718         if (shutdown) {
2719                 nvme_start_queues(&dev->ctrl);
2720                 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2721                         nvme_start_admin_queue(&dev->ctrl);
2722         }
2723         mutex_unlock(&dev->shutdown_lock);
2724 }
2725
2726 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2727 {
2728         if (!nvme_wait_reset(&dev->ctrl))
2729                 return -EBUSY;
2730         nvme_dev_disable(dev, shutdown);
2731         return 0;
2732 }
2733
2734 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2735 {
2736         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2737                                                 NVME_CTRL_PAGE_SIZE,
2738                                                 NVME_CTRL_PAGE_SIZE, 0);
2739         if (!dev->prp_page_pool)
2740                 return -ENOMEM;
2741
2742         /* Optimisation for I/Os between 4k and 128k */
2743         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2744                                                 256, 256, 0);
2745         if (!dev->prp_small_pool) {
2746                 dma_pool_destroy(dev->prp_page_pool);
2747                 return -ENOMEM;
2748         }
2749         return 0;
2750 }
2751
2752 static void nvme_release_prp_pools(struct nvme_dev *dev)
2753 {
2754         dma_pool_destroy(dev->prp_page_pool);
2755         dma_pool_destroy(dev->prp_small_pool);
2756 }
2757
2758 static void nvme_free_tagset(struct nvme_dev *dev)
2759 {
2760         if (dev->tagset.tags)
2761                 blk_mq_free_tag_set(&dev->tagset);
2762         dev->ctrl.tagset = NULL;
2763 }
2764
2765 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2766 {
2767         struct nvme_dev *dev = to_nvme_dev(ctrl);
2768
2769         nvme_dbbuf_dma_free(dev);
2770         nvme_free_tagset(dev);
2771         if (dev->ctrl.admin_q)
2772                 blk_put_queue(dev->ctrl.admin_q);
2773         free_opal_dev(dev->ctrl.opal_dev);
2774         mempool_destroy(dev->iod_mempool);
2775         put_device(dev->dev);
2776         kfree(dev->queues);
2777         kfree(dev);
2778 }
2779
2780 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2781 {
2782         /*
2783          * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2784          * may be holding this pci_dev's device lock.
2785          */
2786         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2787         nvme_get_ctrl(&dev->ctrl);
2788         nvme_dev_disable(dev, false);
2789         nvme_kill_queues(&dev->ctrl);
2790         if (!queue_work(nvme_wq, &dev->remove_work))
2791                 nvme_put_ctrl(&dev->ctrl);
2792 }
2793
2794 static void nvme_reset_work(struct work_struct *work)
2795 {
2796         struct nvme_dev *dev =
2797                 container_of(work, struct nvme_dev, ctrl.reset_work);
2798         bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2799         int result;
2800
2801         if (dev->ctrl.state != NVME_CTRL_RESETTING) {
2802                 dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
2803                          dev->ctrl.state);
2804                 result = -ENODEV;
2805                 goto out;
2806         }
2807
2808         /*
2809          * If we're called to reset a live controller first shut it down before
2810          * moving on.
2811          */
2812         if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2813                 nvme_dev_disable(dev, false);
2814         nvme_sync_queues(&dev->ctrl);
2815
2816         mutex_lock(&dev->shutdown_lock);
2817         result = nvme_pci_enable(dev);
2818         if (result)
2819                 goto out_unlock;
2820
2821         result = nvme_pci_configure_admin_queue(dev);
2822         if (result)
2823                 goto out_unlock;
2824
2825         result = nvme_alloc_admin_tags(dev);
2826         if (result)
2827                 goto out_unlock;
2828
2829         /*
2830          * Limit the max command size to prevent iod->sg allocations going
2831          * over a single page.
2832          */
2833         dev->ctrl.max_hw_sectors = min_t(u32,
2834                 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2835         dev->ctrl.max_segments = NVME_MAX_SEGS;
2836
2837         /*
2838          * Don't limit the IOMMU merged segment size.
2839          */
2840         dma_set_max_seg_size(dev->dev, 0xffffffff);
2841         dma_set_min_align_mask(dev->dev, NVME_CTRL_PAGE_SIZE - 1);
2842
2843         mutex_unlock(&dev->shutdown_lock);
2844
2845         /*
2846          * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2847          * initializing procedure here.
2848          */
2849         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2850                 dev_warn(dev->ctrl.device,
2851                         "failed to mark controller CONNECTING\n");
2852                 result = -EBUSY;
2853                 goto out;
2854         }
2855
2856         /*
2857          * We do not support an SGL for metadata (yet), so we are limited to a
2858          * single integrity segment for the separate metadata pointer.
2859          */
2860         dev->ctrl.max_integrity_segments = 1;
2861
2862         result = nvme_init_ctrl_finish(&dev->ctrl);
2863         if (result)
2864                 goto out;
2865
2866         if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2867                 if (!dev->ctrl.opal_dev)
2868                         dev->ctrl.opal_dev =
2869                                 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2870                 else if (was_suspend)
2871                         opal_unlock_from_suspend(dev->ctrl.opal_dev);
2872         } else {
2873                 free_opal_dev(dev->ctrl.opal_dev);
2874                 dev->ctrl.opal_dev = NULL;
2875         }
2876
2877         if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2878                 result = nvme_dbbuf_dma_alloc(dev);
2879                 if (result)
2880                         dev_warn(dev->dev,
2881                                  "unable to allocate dma for dbbuf\n");
2882         }
2883
2884         if (dev->ctrl.hmpre) {
2885                 result = nvme_setup_host_mem(dev);
2886                 if (result < 0)
2887                         goto out;
2888         }
2889
2890         result = nvme_setup_io_queues(dev);
2891         if (result)
2892                 goto out;
2893
2894         /*
2895          * Keep the controller around but remove all namespaces if we don't have
2896          * any working I/O queue.
2897          */
2898         if (dev->online_queues < 2) {
2899                 dev_warn(dev->ctrl.device, "IO queues not created\n");
2900                 nvme_kill_queues(&dev->ctrl);
2901                 nvme_remove_namespaces(&dev->ctrl);
2902                 nvme_free_tagset(dev);
2903         } else {
2904                 nvme_start_queues(&dev->ctrl);
2905                 nvme_wait_freeze(&dev->ctrl);
2906                 nvme_dev_add(dev);
2907                 nvme_unfreeze(&dev->ctrl);
2908         }
2909
2910         /*
2911          * If only admin queue live, keep it to do further investigation or
2912          * recovery.
2913          */
2914         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2915                 dev_warn(dev->ctrl.device,
2916                         "failed to mark controller live state\n");
2917                 result = -ENODEV;
2918                 goto out;
2919         }
2920
2921         if (!dev->attrs_added && !sysfs_create_group(&dev->ctrl.device->kobj,
2922                         &nvme_pci_attr_group))
2923                 dev->attrs_added = true;
2924
2925         nvme_start_ctrl(&dev->ctrl);
2926         return;
2927
2928  out_unlock:
2929         mutex_unlock(&dev->shutdown_lock);
2930  out:
2931         if (result)
2932                 dev_warn(dev->ctrl.device,
2933                          "Removing after probe failure status: %d\n", result);
2934         nvme_remove_dead_ctrl(dev);
2935 }
2936
2937 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2938 {
2939         struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2940         struct pci_dev *pdev = to_pci_dev(dev->dev);
2941
2942         if (pci_get_drvdata(pdev))
2943                 device_release_driver(&pdev->dev);
2944         nvme_put_ctrl(&dev->ctrl);
2945 }
2946
2947 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2948 {
2949         *val = readl(to_nvme_dev(ctrl)->bar + off);
2950         return 0;
2951 }
2952
2953 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2954 {
2955         writel(val, to_nvme_dev(ctrl)->bar + off);
2956         return 0;
2957 }
2958
2959 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2960 {
2961         *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2962         return 0;
2963 }
2964
2965 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2966 {
2967         struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2968
2969         return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2970 }
2971
2972 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2973         .name                   = "pcie",
2974         .module                 = THIS_MODULE,
2975         .flags                  = NVME_F_METADATA_SUPPORTED |
2976                                   NVME_F_PCI_P2PDMA,
2977         .reg_read32             = nvme_pci_reg_read32,
2978         .reg_write32            = nvme_pci_reg_write32,
2979         .reg_read64             = nvme_pci_reg_read64,
2980         .free_ctrl              = nvme_pci_free_ctrl,
2981         .submit_async_event     = nvme_pci_submit_async_event,
2982         .get_address            = nvme_pci_get_address,
2983 };
2984
2985 static int nvme_dev_map(struct nvme_dev *dev)
2986 {
2987         struct pci_dev *pdev = to_pci_dev(dev->dev);
2988
2989         if (pci_request_mem_regions(pdev, "nvme"))
2990                 return -ENODEV;
2991
2992         if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2993                 goto release;
2994
2995         return 0;
2996   release:
2997         pci_release_mem_regions(pdev);
2998         return -ENODEV;
2999 }
3000
3001 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
3002 {
3003         if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
3004                 /*
3005                  * Several Samsung devices seem to drop off the PCIe bus
3006                  * randomly when APST is on and uses the deepest sleep state.
3007                  * This has been observed on a Samsung "SM951 NVMe SAMSUNG
3008                  * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
3009                  * 950 PRO 256GB", but it seems to be restricted to two Dell
3010                  * laptops.
3011                  */
3012                 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
3013                     (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
3014                      dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
3015                         return NVME_QUIRK_NO_DEEPEST_PS;
3016         } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
3017                 /*
3018                  * Samsung SSD 960 EVO drops off the PCIe bus after system
3019                  * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
3020                  * within few minutes after bootup on a Coffee Lake board -
3021                  * ASUS PRIME Z370-A
3022                  */
3023                 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
3024                     (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
3025                      dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
3026                         return NVME_QUIRK_NO_APST;
3027         } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
3028                     pdev->device == 0xa808 || pdev->device == 0xa809)) ||
3029                    (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
3030                 /*
3031                  * Forcing to use host managed nvme power settings for
3032                  * lowest idle power with quick resume latency on
3033                  * Samsung and Toshiba SSDs based on suspend behavior
3034                  * on Coffee Lake board for LENOVO C640
3035                  */
3036                 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
3037                      dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
3038                         return NVME_QUIRK_SIMPLE_SUSPEND;
3039         }
3040
3041         return 0;
3042 }
3043
3044 static void nvme_async_probe(void *data, async_cookie_t cookie)
3045 {
3046         struct nvme_dev *dev = data;
3047
3048         flush_work(&dev->ctrl.reset_work);
3049         flush_work(&dev->ctrl.scan_work);
3050         nvme_put_ctrl(&dev->ctrl);
3051 }
3052
3053 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3054 {
3055         int node, result = -ENOMEM;
3056         struct nvme_dev *dev;
3057         unsigned long quirks = id->driver_data;
3058         size_t alloc_size;
3059
3060         node = dev_to_node(&pdev->dev);
3061         if (node == NUMA_NO_NODE)
3062                 set_dev_node(&pdev->dev, first_memory_node);
3063
3064         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
3065         if (!dev)
3066                 return -ENOMEM;
3067
3068         dev->nr_write_queues = write_queues;
3069         dev->nr_poll_queues = poll_queues;
3070         dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
3071         dev->queues = kcalloc_node(dev->nr_allocated_queues,
3072                         sizeof(struct nvme_queue), GFP_KERNEL, node);
3073         if (!dev->queues)
3074                 goto free;
3075
3076         dev->dev = get_device(&pdev->dev);
3077         pci_set_drvdata(pdev, dev);
3078
3079         result = nvme_dev_map(dev);
3080         if (result)
3081                 goto put_pci;
3082
3083         INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
3084         INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
3085         mutex_init(&dev->shutdown_lock);
3086
3087         result = nvme_setup_prp_pools(dev);
3088         if (result)
3089                 goto unmap;
3090
3091         quirks |= check_vendor_combination_bug(pdev);
3092
3093         if (!noacpi && acpi_storage_d3(&pdev->dev)) {
3094                 /*
3095                  * Some systems use a bios work around to ask for D3 on
3096                  * platforms that support kernel managed suspend.
3097                  */
3098                 dev_info(&pdev->dev,
3099                          "platform quirk: setting simple suspend\n");
3100                 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
3101         }
3102
3103         /*
3104          * Double check that our mempool alloc size will cover the biggest
3105          * command we support.
3106          */
3107         alloc_size = nvme_pci_iod_alloc_size();
3108         WARN_ON_ONCE(alloc_size > PAGE_SIZE);
3109
3110         dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
3111                                                 mempool_kfree,
3112                                                 (void *) alloc_size,
3113                                                 GFP_KERNEL, node);
3114         if (!dev->iod_mempool) {
3115                 result = -ENOMEM;
3116                 goto release_pools;
3117         }
3118
3119         result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
3120                         quirks);
3121         if (result)
3122                 goto release_mempool;
3123
3124         dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
3125
3126         nvme_reset_ctrl(&dev->ctrl);
3127         async_schedule(nvme_async_probe, dev);
3128
3129         return 0;
3130
3131  release_mempool:
3132         mempool_destroy(dev->iod_mempool);
3133  release_pools:
3134         nvme_release_prp_pools(dev);
3135  unmap:
3136         nvme_dev_unmap(dev);
3137  put_pci:
3138         put_device(dev->dev);
3139  free:
3140         kfree(dev->queues);
3141         kfree(dev);
3142         return result;
3143 }
3144
3145 static void nvme_reset_prepare(struct pci_dev *pdev)
3146 {
3147         struct nvme_dev *dev = pci_get_drvdata(pdev);
3148
3149         /*
3150          * We don't need to check the return value from waiting for the reset
3151          * state as pci_dev device lock is held, making it impossible to race
3152          * with ->remove().
3153          */
3154         nvme_disable_prepare_reset(dev, false);
3155         nvme_sync_queues(&dev->ctrl);
3156 }
3157
3158 static void nvme_reset_done(struct pci_dev *pdev)
3159 {
3160         struct nvme_dev *dev = pci_get_drvdata(pdev);
3161
3162         if (!nvme_try_sched_reset(&dev->ctrl))
3163                 flush_work(&dev->ctrl.reset_work);
3164 }
3165
3166 static void nvme_shutdown(struct pci_dev *pdev)
3167 {
3168         struct nvme_dev *dev = pci_get_drvdata(pdev);
3169
3170         nvme_disable_prepare_reset(dev, true);
3171 }
3172
3173 static void nvme_remove_attrs(struct nvme_dev *dev)
3174 {
3175         if (dev->attrs_added)
3176                 sysfs_remove_group(&dev->ctrl.device->kobj,
3177                                    &nvme_pci_attr_group);
3178 }
3179
3180 /*
3181  * The driver's remove may be called on a device in a partially initialized
3182  * state. This function must not have any dependencies on the device state in
3183  * order to proceed.
3184  */
3185 static void nvme_remove(struct pci_dev *pdev)
3186 {
3187         struct nvme_dev *dev = pci_get_drvdata(pdev);
3188
3189         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3190         pci_set_drvdata(pdev, NULL);
3191
3192         if (!pci_device_is_present(pdev)) {
3193                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3194                 nvme_dev_disable(dev, true);
3195         }
3196
3197         flush_work(&dev->ctrl.reset_work);
3198         nvme_stop_ctrl(&dev->ctrl);
3199         nvme_remove_namespaces(&dev->ctrl);
3200         nvme_dev_disable(dev, true);
3201         nvme_remove_attrs(dev);
3202         nvme_free_host_mem(dev);
3203         nvme_dev_remove_admin(dev);
3204         nvme_free_queues(dev, 0);
3205         nvme_release_prp_pools(dev);
3206         nvme_dev_unmap(dev);
3207         nvme_uninit_ctrl(&dev->ctrl);
3208 }
3209
3210 #ifdef CONFIG_PM_SLEEP
3211 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3212 {
3213         return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3214 }
3215
3216 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3217 {
3218         return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3219 }
3220
3221 static int nvme_resume(struct device *dev)
3222 {
3223         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3224         struct nvme_ctrl *ctrl = &ndev->ctrl;
3225
3226         if (ndev->last_ps == U32_MAX ||
3227             nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3228                 goto reset;
3229         if (ctrl->hmpre && nvme_setup_host_mem(ndev))
3230                 goto reset;
3231
3232         return 0;
3233 reset:
3234         return nvme_try_sched_reset(ctrl);
3235 }
3236
3237 static int nvme_suspend(struct device *dev)
3238 {
3239         struct pci_dev *pdev = to_pci_dev(dev);
3240         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3241         struct nvme_ctrl *ctrl = &ndev->ctrl;
3242         int ret = -EBUSY;
3243
3244         ndev->last_ps = U32_MAX;
3245
3246         /*
3247          * The platform does not remove power for a kernel managed suspend so
3248          * use host managed nvme power settings for lowest idle power if
3249          * possible. This should have quicker resume latency than a full device
3250          * shutdown.  But if the firmware is involved after the suspend or the
3251          * device does not support any non-default power states, shut down the
3252          * device fully.
3253          *
3254          * If ASPM is not enabled for the device, shut down the device and allow
3255          * the PCI bus layer to put it into D3 in order to take the PCIe link
3256          * down, so as to allow the platform to achieve its minimum low-power
3257          * state (which may not be possible if the link is up).
3258          */
3259         if (pm_suspend_via_firmware() || !ctrl->npss ||
3260             !pcie_aspm_enabled(pdev) ||
3261             (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3262                 return nvme_disable_prepare_reset(ndev, true);
3263
3264         nvme_start_freeze(ctrl);
3265         nvme_wait_freeze(ctrl);
3266         nvme_sync_queues(ctrl);
3267
3268         if (ctrl->state != NVME_CTRL_LIVE)
3269                 goto unfreeze;
3270
3271         /*
3272          * Host memory access may not be successful in a system suspend state,
3273          * but the specification allows the controller to access memory in a
3274          * non-operational power state.
3275          */
3276         if (ndev->hmb) {
3277                 ret = nvme_set_host_mem(ndev, 0);
3278                 if (ret < 0)
3279                         goto unfreeze;
3280         }
3281
3282         ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3283         if (ret < 0)
3284                 goto unfreeze;
3285
3286         /*
3287          * A saved state prevents pci pm from generically controlling the
3288          * device's power. If we're using protocol specific settings, we don't
3289          * want pci interfering.
3290          */
3291         pci_save_state(pdev);
3292
3293         ret = nvme_set_power_state(ctrl, ctrl->npss);
3294         if (ret < 0)
3295                 goto unfreeze;
3296
3297         if (ret) {
3298                 /* discard the saved state */
3299                 pci_load_saved_state(pdev, NULL);
3300
3301                 /*
3302                  * Clearing npss forces a controller reset on resume. The
3303                  * correct value will be rediscovered then.
3304                  */
3305                 ret = nvme_disable_prepare_reset(ndev, true);
3306                 ctrl->npss = 0;
3307         }
3308 unfreeze:
3309         nvme_unfreeze(ctrl);
3310         return ret;
3311 }
3312
3313 static int nvme_simple_suspend(struct device *dev)
3314 {
3315         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3316
3317         return nvme_disable_prepare_reset(ndev, true);
3318 }
3319
3320 static int nvme_simple_resume(struct device *dev)
3321 {
3322         struct pci_dev *pdev = to_pci_dev(dev);
3323         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3324
3325         return nvme_try_sched_reset(&ndev->ctrl);
3326 }
3327
3328 static const struct dev_pm_ops nvme_dev_pm_ops = {
3329         .suspend        = nvme_suspend,
3330         .resume         = nvme_resume,
3331         .freeze         = nvme_simple_suspend,
3332         .thaw           = nvme_simple_resume,
3333         .poweroff       = nvme_simple_suspend,
3334         .restore        = nvme_simple_resume,
3335 };
3336 #endif /* CONFIG_PM_SLEEP */
3337
3338 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3339                                                 pci_channel_state_t state)
3340 {
3341         struct nvme_dev *dev = pci_get_drvdata(pdev);
3342
3343         /*
3344          * A frozen channel requires a reset. When detected, this method will
3345          * shutdown the controller to quiesce. The controller will be restarted
3346          * after the slot reset through driver's slot_reset callback.
3347          */
3348         switch (state) {
3349         case pci_channel_io_normal:
3350                 return PCI_ERS_RESULT_CAN_RECOVER;
3351         case pci_channel_io_frozen:
3352                 dev_warn(dev->ctrl.device,
3353                         "frozen state error detected, reset controller\n");
3354                 nvme_dev_disable(dev, false);
3355                 return PCI_ERS_RESULT_NEED_RESET;
3356         case pci_channel_io_perm_failure:
3357                 dev_warn(dev->ctrl.device,
3358                         "failure state error detected, request disconnect\n");
3359                 return PCI_ERS_RESULT_DISCONNECT;
3360         }
3361         return PCI_ERS_RESULT_NEED_RESET;
3362 }
3363
3364 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3365 {
3366         struct nvme_dev *dev = pci_get_drvdata(pdev);
3367
3368         dev_info(dev->ctrl.device, "restart after slot reset\n");
3369         pci_restore_state(pdev);
3370         nvme_reset_ctrl(&dev->ctrl);
3371         return PCI_ERS_RESULT_RECOVERED;
3372 }
3373
3374 static void nvme_error_resume(struct pci_dev *pdev)
3375 {
3376         struct nvme_dev *dev = pci_get_drvdata(pdev);
3377
3378         flush_work(&dev->ctrl.reset_work);
3379 }
3380
3381 static const struct pci_error_handlers nvme_err_handler = {
3382         .error_detected = nvme_error_detected,
3383         .slot_reset     = nvme_slot_reset,
3384         .resume         = nvme_error_resume,
3385         .reset_prepare  = nvme_reset_prepare,
3386         .reset_done     = nvme_reset_done,
3387 };
3388
3389 static const struct pci_device_id nvme_id_table[] = {
3390         { PCI_VDEVICE(INTEL, 0x0953),   /* Intel 750/P3500/P3600/P3700 */
3391                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3392                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3393         { PCI_VDEVICE(INTEL, 0x0a53),   /* Intel P3520 */
3394                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3395                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3396         { PCI_VDEVICE(INTEL, 0x0a54),   /* Intel P4500/P4600 */
3397                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3398                                 NVME_QUIRK_DEALLOCATE_ZEROES |
3399                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3400         { PCI_VDEVICE(INTEL, 0x0a55),   /* Dell Express Flash P4600 */
3401                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3402                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3403         { PCI_VDEVICE(INTEL, 0xf1a5),   /* Intel 600P/P3100 */
3404                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3405                                 NVME_QUIRK_MEDIUM_PRIO_SQ |
3406                                 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3407                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3408         { PCI_VDEVICE(INTEL, 0xf1a6),   /* Intel 760p/Pro 7600p */
3409                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3410         { PCI_VDEVICE(INTEL, 0x5845),   /* Qemu emulated controller */
3411                 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3412                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3413         { PCI_DEVICE(0x126f, 0x2263),   /* Silicon Motion unidentified */
3414                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
3415         { PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3416                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3417                                 NVME_QUIRK_NO_NS_DESC_LIST, },
3418         { PCI_DEVICE(0x1c58, 0x0003),   /* HGST adapter */
3419                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3420         { PCI_DEVICE(0x1c58, 0x0023),   /* WDC SN200 adapter */
3421                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3422         { PCI_DEVICE(0x1c5f, 0x0540),   /* Memblaze Pblaze4 adapter */
3423                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3424         { PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3425                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3426         { PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3427                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3428                                 NVME_QUIRK_DISABLE_WRITE_ZEROES|
3429                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3430         { PCI_DEVICE(0x1987, 0x5016),   /* Phison E16 */
3431                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3432         { PCI_DEVICE(0x1b4b, 0x1092),   /* Lexar 256 GB SSD */
3433                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3434                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3435         { PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3436                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3437         { PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3438                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3439                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3440         { PCI_DEVICE(0x1c5c, 0x1504),   /* SK Hynix PC400 */
3441                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3442         { PCI_DEVICE(0x15b7, 0x2001),   /*  Sandisk Skyhawk */
3443                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3444         { PCI_DEVICE(0x1d97, 0x2263),   /* SPCC */
3445                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3446         { PCI_DEVICE(0x2646, 0x2262),   /* KINGSTON SKC2000 NVMe SSD */
3447                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3448         { PCI_DEVICE(0x2646, 0x2263),   /* KINGSTON A2000 NVMe SSD  */
3449                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3450         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061),
3451                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3452         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065),
3453                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3454         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061),
3455                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3456         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00),
3457                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3458         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01),
3459                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3460         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02),
3461                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3462         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3463                 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3464         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3465         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3466                 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3467                                 NVME_QUIRK_128_BYTES_SQES |
3468                                 NVME_QUIRK_SHARED_TAGS |
3469                                 NVME_QUIRK_SKIP_CID_GEN },
3470         { PCI_DEVICE(0x144d, 0xa808),   /* Samsung X5 */
3471                 .driver_data =  NVME_QUIRK_DELAY_BEFORE_CHK_RDY|
3472                                 NVME_QUIRK_NO_DEEPEST_PS |
3473                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3474         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3475         { 0, }
3476 };
3477 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3478
3479 static struct pci_driver nvme_driver = {
3480         .name           = "nvme",
3481         .id_table       = nvme_id_table,
3482         .probe          = nvme_probe,
3483         .remove         = nvme_remove,
3484         .shutdown       = nvme_shutdown,
3485 #ifdef CONFIG_PM_SLEEP
3486         .driver         = {
3487                 .pm     = &nvme_dev_pm_ops,
3488         },
3489 #endif
3490         .sriov_configure = pci_sriov_configure_simple,
3491         .err_handler    = &nvme_err_handler,
3492 };
3493
3494 static int __init nvme_init(void)
3495 {
3496         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3497         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3498         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3499         BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3500
3501         return pci_register_driver(&nvme_driver);
3502 }
3503
3504 static void __exit nvme_exit(void)
3505 {
3506         pci_unregister_driver(&nvme_driver);
3507         flush_workqueue(nvme_wq);
3508 }
3509
3510 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3511 MODULE_LICENSE("GPL");
3512 MODULE_VERSION("1.0");
3513 module_init(nvme_init);
3514 module_exit(nvme_exit);