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/aer.h>
   8 #include <linux/async.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/blk-mq.h>
  11 #include <linux/blk-mq-pci.h>
  12 #include <linux/dmi.h>
  13 #include <linux/init.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/io.h>
  16 #include <linux/mm.h>
  17 #include <linux/module.h>
  18 #include <linux/mutex.h>
  19 #include <linux/once.h>
  20 #include <linux/pci.h>
  21 #include <linux/suspend.h>
  22 #include <linux/t10-pi.h>
  23 #include <linux/types.h>
  24 #include <linux/io-64-nonatomic-lo-hi.h>
  25 #include <linux/sed-opal.h>
  26 #include <linux/pci-p2pdma.h>
  27
  28 #include "trace.h"
  29 #include "nvme.h"
  30
  31 #define SQ_SIZE(q)      ((q)->q_depth << (q)->sqes)
  32 #define CQ_SIZE(q)      ((q)->q_depth * sizeof(struct nvme_completion))
  33
  34 #define SGES_PER_PAGE   (PAGE_SIZE / sizeof(struct nvme_sgl_desc))
  35
  36 /*
  37  * These can be higher, but we need to ensure that any command doesn't
  38  * require an sg allocation that needs more than a page of data.
  39  */
  40 #define NVME_MAX_KB_SZ  4096
  41 #define NVME_MAX_SEGS   127
  42
  43 static int use_threaded_interrupts;
  44 module_param(use_threaded_interrupts, int, 0);
  45
  46 static bool use_cmb_sqes = true;
  47 module_param(use_cmb_sqes, bool, 0444);
  48 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
  49
  50 static unsigned int max_host_mem_size_mb = 128;
  51 module_param(max_host_mem_size_mb, uint, 0444);
  52 MODULE_PARM_DESC(max_host_mem_size_mb,
  53         "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
  54
  55 static unsigned int sgl_threshold = SZ_32K;
  56 module_param(sgl_threshold, uint, 0644);
  57 MODULE_PARM_DESC(sgl_threshold,
  58                 "Use SGLs when average request segment size is larger or equal to "
  59                 "this size. Use 0 to disable SGLs.");
  60
  61 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
  62 static const struct kernel_param_ops io_queue_depth_ops = {
  63         .set = io_queue_depth_set,
  64         .get = param_get_int,
  65 };
  66
  67 static int io_queue_depth = 1024;
  68 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
  69 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
  70
  71 static unsigned int write_queues;
  72 module_param(write_queues, uint, 0644);
  73 MODULE_PARM_DESC(write_queues,
  74         "Number of queues to use for writes. If not set, reads and writes "
  75         "will share a queue set.");
  76
  77 static unsigned int poll_queues;
  78 module_param(poll_queues, uint, 0644);
  79 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
  80
  81 struct nvme_dev;
  82 struct nvme_queue;
  83
  84 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
  85 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
  86
  87 /*
  88  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
  89  */
  90 struct nvme_dev {
  91         struct nvme_queue *queues;
  92         struct blk_mq_tag_set tagset;
  93         struct blk_mq_tag_set admin_tagset;
  94         u32 __iomem *dbs;
  95         struct device *dev;
  96         struct dma_pool *prp_page_pool;
  97         struct dma_pool *prp_small_pool;
  98         unsigned online_queues;
  99         unsigned max_qid;
 100         unsigned io_queues[HCTX_MAX_TYPES];
 101         unsigned int num_vecs;
 102         int q_depth;
 103         int io_sqes;
 104         u32 db_stride;
 105         void __iomem *bar;
 106         unsigned long bar_mapped_size;
 107         struct work_struct remove_work;
 108         struct mutex shutdown_lock;
 109         bool subsystem;
 110         u64 cmb_size;
 111         bool cmb_use_sqes;
 112         u32 cmbsz;
 113         u32 cmbloc;
 114         struct nvme_ctrl ctrl;
 115         u32 last_ps;
 116
 117         mempool_t *iod_mempool;
 118
 119         /* shadow doorbell buffer support: */
 120         u32 *dbbuf_dbs;
 121         dma_addr_t dbbuf_dbs_dma_addr;
 122         u32 *dbbuf_eis;
 123         dma_addr_t dbbuf_eis_dma_addr;
 124
 125         /* host memory buffer support: */
 126         u64 host_mem_size;
 127         u32 nr_host_mem_descs;
 128         dma_addr_t host_mem_descs_dma;
 129         struct nvme_host_mem_buf_desc *host_mem_descs;
 130         void **host_mem_desc_bufs;
 131 };
 132
 133 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
 134 {
 135         int n = 0, ret;
 136
 137         ret = kstrtoint(val, 10, &n);
 138         if (ret != 0 || n < 2)
 139                 return -EINVAL;
 140
 141         return param_set_int(val, kp);
 142 }
 143
 144 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
 145 {
 146         return qid * 2 * stride;
 147 }
 148
 149 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
 150 {
 151         return (qid * 2 + 1) * stride;
 152 }
 153
 154 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 155 {
 156         return container_of(ctrl, struct nvme_dev, ctrl);
 157 }
 158
 159 /*
 160  * An NVM Express queue.  Each device has at least two (one for admin
 161  * commands and one for I/O commands).
 162  */
 163 struct nvme_queue {
 164         struct nvme_dev *dev;
 165         spinlock_t sq_lock;
 166         void *sq_cmds;
 167          /* only used for poll queues: */
 168         spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
 169         volatile struct nvme_completion *cqes;
 170         dma_addr_t sq_dma_addr;
 171         dma_addr_t cq_dma_addr;
 172         u32 __iomem *q_db;
 173         u16 q_depth;
 174         u16 cq_vector;
 175         u16 sq_tail;
 176         u16 last_sq_tail;
 177         u16 cq_head;
 178         u16 qid;
 179         u8 cq_phase;
 180         u8 sqes;
 181         unsigned long flags;
 182 #define NVMEQ_ENABLED           0
 183 #define NVMEQ_SQ_CMB            1
 184 #define NVMEQ_DELETE_ERROR      2
 185 #define NVMEQ_POLLED            3
 186         u32 *dbbuf_sq_db;
 187         u32 *dbbuf_cq_db;
 188         u32 *dbbuf_sq_ei;
 189         u32 *dbbuf_cq_ei;
 190         struct completion delete_done;
 191 };
 192
 193 /*
 194  * The nvme_iod describes the data in an I/O.
 195  *
 196  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
 197  * to the actual struct scatterlist.
 198  */
 199 struct nvme_iod {
 200         struct nvme_request req;
 201         struct nvme_queue *nvmeq;
 202         bool use_sgl;
 203         int aborted;
 204         int npages;             /* In the PRP list. 0 means small pool in use */
 205         int nents;              /* Used in scatterlist */
 206         dma_addr_t first_dma;
 207         unsigned int dma_len;   /* length of single DMA segment mapping */
 208         dma_addr_t meta_dma;
 209         struct scatterlist *sg;
 210 };
 211
 212 static unsigned int max_io_queues(void)
 213 {
 214         return num_possible_cpus() + write_queues + poll_queues;
 215 }
 216
 217 static unsigned int max_queue_count(void)
 218 {
 219         /* IO queues + admin queue */
 220         return 1 + max_io_queues();
 221 }
 222
 223 static inline unsigned int nvme_dbbuf_size(u32 stride)
 224 {
 225         return (max_queue_count() * 8 * stride);
 226 }
 227
 228 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
 229 {
 230         unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
 231
 232         if (dev->dbbuf_dbs)
 233                 return 0;
 234
 235         dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
 236                                             &dev->dbbuf_dbs_dma_addr,
 237                                             GFP_KERNEL);
 238         if (!dev->dbbuf_dbs)
 239                 return -ENOMEM;
 240         dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
 241                                             &dev->dbbuf_eis_dma_addr,
 242                                             GFP_KERNEL);
 243         if (!dev->dbbuf_eis) {
 244                 dma_free_coherent(dev->dev, mem_size,
 245                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 246                 dev->dbbuf_dbs = NULL;
 247                 return -ENOMEM;
 248         }
 249
 250         return 0;
 251 }
 252
 253 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
 254 {
 255         unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
 256
 257         if (dev->dbbuf_dbs) {
 258                 dma_free_coherent(dev->dev, mem_size,
 259                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 260                 dev->dbbuf_dbs = NULL;
 261         }
 262         if (dev->dbbuf_eis) {
 263                 dma_free_coherent(dev->dev, mem_size,
 264                                   dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
 265                 dev->dbbuf_eis = NULL;
 266         }
 267 }
 268
 269 static void nvme_dbbuf_init(struct nvme_dev *dev,
 270                             struct nvme_queue *nvmeq, int qid)
 271 {
 272         if (!dev->dbbuf_dbs || !qid)
 273                 return;
 274
 275         nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
 276         nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
 277         nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
 278         nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
 279 }
 280
 281 static void nvme_dbbuf_set(struct nvme_dev *dev)
 282 {
 283         struct nvme_command c;
 284
 285         if (!dev->dbbuf_dbs)
 286                 return;
 287
 288         memset(&c, 0, sizeof(c));
 289         c.dbbuf.opcode = nvme_admin_dbbuf;
 290         c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
 291         c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
 292
 293         if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
 294                 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
 295                 /* Free memory and continue on */
 296                 nvme_dbbuf_dma_free(dev);
 297         }
 298 }
 299
 300 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
 301 {
 302         return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
 303 }
 304
 305 /* Update dbbuf and return true if an MMIO is required */
 306 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
 307                                               volatile u32 *dbbuf_ei)
 308 {
 309         if (dbbuf_db) {
 310                 u16 old_value;
 311
 312                 /*
 313                  * Ensure that the queue is written before updating
 314                  * the doorbell in memory
 315                  */
 316                 wmb();
 317
 318                 old_value = *dbbuf_db;
 319                 *dbbuf_db = value;
 320
 321                 /*
 322                  * Ensure that the doorbell is updated before reading the event
 323                  * index from memory.  The controller needs to provide similar
 324                  * ordering to ensure the envent index is updated before reading
 325                  * the doorbell.
 326                  */
 327                 mb();
 328
 329                 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
 330                         return false;
 331         }
 332
 333         return true;
 334 }
 335
 336 /*
 337  * Will slightly overestimate the number of pages needed.  This is OK
 338  * as it only leads to a small amount of wasted memory for the lifetime of
 339  * the I/O.
 340  */
 341 static int nvme_npages(unsigned size, struct nvme_dev *dev)
 342 {
 343         unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
 344                                       dev->ctrl.page_size);
 345         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 346 }
 347
 348 /*
 349  * Calculates the number of pages needed for the SGL segments. For example a 4k
 350  * page can accommodate 256 SGL descriptors.
 351  */
 352 static int nvme_pci_npages_sgl(unsigned int num_seg)
 353 {
 354         return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
 355 }
 356
 357 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
 358                 unsigned int size, unsigned int nseg, bool use_sgl)
 359 {
 360         size_t alloc_size;
 361
 362         if (use_sgl)
 363                 alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
 364         else
 365                 alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
 366
 367         return alloc_size + sizeof(struct scatterlist) * nseg;
 368 }
 369
 370 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 371                                 unsigned int hctx_idx)
 372 {
 373         struct nvme_dev *dev = data;
 374         struct nvme_queue *nvmeq = &dev->queues[0];
 375
 376         WARN_ON(hctx_idx != 0);
 377         WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
 378
 379         hctx->driver_data = nvmeq;
 380         return 0;
 381 }
 382
 383 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 384                           unsigned int hctx_idx)
 385 {
 386         struct nvme_dev *dev = data;
 387         struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
 388
 389         WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
 390         hctx->driver_data = nvmeq;
 391         return 0;
 392 }
 393
 394 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
 395                 unsigned int hctx_idx, unsigned int numa_node)
 396 {
 397         struct nvme_dev *dev = set->driver_data;
 398         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 399         int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
 400         struct nvme_queue *nvmeq = &dev->queues[queue_idx];
 401
 402         BUG_ON(!nvmeq);
 403         iod->nvmeq = nvmeq;
 404
 405         nvme_req(req)->ctrl = &dev->ctrl;
 406         return 0;
 407 }
 408
 409 static int queue_irq_offset(struct nvme_dev *dev)
 410 {
 411         /* if we have more than 1 vec, admin queue offsets us by 1 */
 412         if (dev->num_vecs > 1)
 413                 return 1;
 414
 415         return 0;
 416 }
 417
 418 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
 419 {
 420         struct nvme_dev *dev = set->driver_data;
 421         int i, qoff, offset;
 422
 423         offset = queue_irq_offset(dev);
 424         for (i = 0, qoff = 0; i < set->nr_maps; i++) {
 425                 struct blk_mq_queue_map *map = &set->map[i];
 426
 427                 map->nr_queues = dev->io_queues[i];
 428                 if (!map->nr_queues) {
 429                         BUG_ON(i == HCTX_TYPE_DEFAULT);
 430                         continue;
 431                 }
 432
 433                 /*
 434                  * The poll queue(s) doesn't have an IRQ (and hence IRQ
 435                  * affinity), so use the regular blk-mq cpu mapping
 436                  */
 437                 map->queue_offset = qoff;
 438                 if (i != HCTX_TYPE_POLL && offset)
 439                         blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
 440                 else
 441                         blk_mq_map_queues(map);
 442                 qoff += map->nr_queues;
 443                 offset += map->nr_queues;
 444         }
 445
 446         return 0;
 447 }
 448
 449 /*
 450  * Write sq tail if we are asked to, or if the next command would wrap.
 451  */
 452 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
 453 {
 454         if (!write_sq) {
 455                 u16 next_tail = nvmeq->sq_tail + 1;
 456
 457                 if (next_tail == nvmeq->q_depth)
 458                         next_tail = 0;
 459                 if (next_tail != nvmeq->last_sq_tail)
 460                         return;
 461         }
 462
 463         if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
 464                         nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
 465                 writel(nvmeq->sq_tail, nvmeq->q_db);
 466         nvmeq->last_sq_tail = nvmeq->sq_tail;
 467 }
 468
 469 /**
 470  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 471  * @nvmeq: The queue to use
 472  * @cmd: The command to send
 473  * @write_sq: whether to write to the SQ doorbell
 474  */
 475 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
 476                             bool write_sq)
 477 {
 478         spin_lock(&nvmeq->sq_lock);
 479         memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 480                cmd, sizeof(*cmd));
 481         if (++nvmeq->sq_tail == nvmeq->q_depth)
 482                 nvmeq->sq_tail = 0;
 483         nvme_write_sq_db(nvmeq, write_sq);
 484         spin_unlock(&nvmeq->sq_lock);
 485 }
 486
 487 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
 488 {
 489         struct nvme_queue *nvmeq = hctx->driver_data;
 490
 491         spin_lock(&nvmeq->sq_lock);
 492         if (nvmeq->sq_tail != nvmeq->last_sq_tail)
 493                 nvme_write_sq_db(nvmeq, true);
 494         spin_unlock(&nvmeq->sq_lock);
 495 }
 496
 497 static void **nvme_pci_iod_list(struct request *req)
 498 {
 499         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 500         return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
 501 }
 502
 503 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
 504 {
 505         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 506         int nseg = blk_rq_nr_phys_segments(req);
 507         unsigned int avg_seg_size;
 508
 509         if (nseg == 0)
 510                 return false;
 511
 512         avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
 513
 514         if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
 515                 return false;
 516         if (!iod->nvmeq->qid)
 517                 return false;
 518         if (!sgl_threshold || avg_seg_size < sgl_threshold)
 519                 return false;
 520         return true;
 521 }
 522
 523 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 524 {
 525         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 526         const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
 527         dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
 528         int i;
 529
 530         if (iod->dma_len) {
 531                 dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
 532                                rq_dma_dir(req));
 533                 return;
 534         }
 535
 536         WARN_ON_ONCE(!iod->nents);
 537
 538         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 539                 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
 540                                     rq_dma_dir(req));
 541         else
 542                 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
 543
 544
 545         if (iod->npages == 0)
 546                 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
 547                         dma_addr);
 548
 549         for (i = 0; i < iod->npages; i++) {
 550                 void *addr = nvme_pci_iod_list(req)[i];
 551
 552                 if (iod->use_sgl) {
 553                         struct nvme_sgl_desc *sg_list = addr;
 554
 555                         next_dma_addr =
 556                             le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
 557                 } else {
 558                         __le64 *prp_list = addr;
 559
 560                         next_dma_addr = le64_to_cpu(prp_list[last_prp]);
 561                 }
 562
 563                 dma_pool_free(dev->prp_page_pool, addr, dma_addr);
 564                 dma_addr = next_dma_addr;
 565         }
 566
 567         mempool_free(iod->sg, dev->iod_mempool);
 568 }
 569
 570 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
 571 {
 572         int i;
 573         struct scatterlist *sg;
 574
 575         for_each_sg(sgl, sg, nents, i) {
 576                 dma_addr_t phys = sg_phys(sg);
 577                 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
 578                         "dma_address:%pad dma_length:%d\n",
 579                         i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
 580                         sg_dma_len(sg));
 581         }
 582 }
 583
 584 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
 585                 struct request *req, struct nvme_rw_command *cmnd)
 586 {
 587         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 588         struct dma_pool *pool;
 589         int length = blk_rq_payload_bytes(req);
 590         struct scatterlist *sg = iod->sg;
 591         int dma_len = sg_dma_len(sg);
 592         u64 dma_addr = sg_dma_address(sg);
 593         u32 page_size = dev->ctrl.page_size;
 594         int offset = dma_addr & (page_size - 1);
 595         __le64 *prp_list;
 596         void **list = nvme_pci_iod_list(req);
 597         dma_addr_t prp_dma;
 598         int nprps, i;
 599
 600         length -= (page_size - offset);
 601         if (length <= 0) {
 602                 iod->first_dma = 0;
 603                 goto done;
 604         }
 605
 606         dma_len -= (page_size - offset);
 607         if (dma_len) {
 608                 dma_addr += (page_size - offset);
 609         } else {
 610                 sg = sg_next(sg);
 611                 dma_addr = sg_dma_address(sg);
 612                 dma_len = sg_dma_len(sg);
 613         }
 614
 615         if (length <= page_size) {
 616                 iod->first_dma = dma_addr;
 617                 goto done;
 618         }
 619
 620         nprps = DIV_ROUND_UP(length, page_size);
 621         if (nprps <= (256 / 8)) {
 622                 pool = dev->prp_small_pool;
 623                 iod->npages = 0;
 624         } else {
 625                 pool = dev->prp_page_pool;
 626                 iod->npages = 1;
 627         }
 628
 629         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 630         if (!prp_list) {
 631                 iod->first_dma = dma_addr;
 632                 iod->npages = -1;
 633                 return BLK_STS_RESOURCE;
 634         }
 635         list[0] = prp_list;
 636         iod->first_dma = prp_dma;
 637         i = 0;
 638         for (;;) {
 639                 if (i == page_size >> 3) {
 640                         __le64 *old_prp_list = prp_list;
 641                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 642                         if (!prp_list)
 643                                 return BLK_STS_RESOURCE;
 644                         list[iod->npages++] = prp_list;
 645                         prp_list[0] = old_prp_list[i - 1];
 646                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 647                         i = 1;
 648                 }
 649                 prp_list[i++] = cpu_to_le64(dma_addr);
 650                 dma_len -= page_size;
 651                 dma_addr += page_size;
 652                 length -= page_size;
 653                 if (length <= 0)
 654                         break;
 655                 if (dma_len > 0)
 656                         continue;
 657                 if (unlikely(dma_len < 0))
 658                         goto bad_sgl;
 659                 sg = sg_next(sg);
 660                 dma_addr = sg_dma_address(sg);
 661                 dma_len = sg_dma_len(sg);
 662         }
 663
 664 done:
 665         cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 666         cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
 667
 668         return BLK_STS_OK;
 669
 670  bad_sgl:
 671         WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
 672                         "Invalid SGL for payload:%d nents:%d\n",
 673                         blk_rq_payload_bytes(req), iod->nents);
 674         return BLK_STS_IOERR;
 675 }
 676
 677 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
 678                 struct scatterlist *sg)
 679 {
 680         sge->addr = cpu_to_le64(sg_dma_address(sg));
 681         sge->length = cpu_to_le32(sg_dma_len(sg));
 682         sge->type = NVME_SGL_FMT_DATA_DESC << 4;
 683 }
 684
 685 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
 686                 dma_addr_t dma_addr, int entries)
 687 {
 688         sge->addr = cpu_to_le64(dma_addr);
 689         if (entries < SGES_PER_PAGE) {
 690                 sge->length = cpu_to_le32(entries * sizeof(*sge));
 691                 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
 692         } else {
 693                 sge->length = cpu_to_le32(PAGE_SIZE);
 694                 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
 695         }
 696 }
 697
 698 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
 699                 struct request *req, struct nvme_rw_command *cmd, int entries)
 700 {
 701         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 702         struct dma_pool *pool;
 703         struct nvme_sgl_desc *sg_list;
 704         struct scatterlist *sg = iod->sg;
 705         dma_addr_t sgl_dma;
 706         int i = 0;
 707
 708         /* setting the transfer type as SGL */
 709         cmd->flags = NVME_CMD_SGL_METABUF;
 710
 711         if (entries == 1) {
 712                 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
 713                 return BLK_STS_OK;
 714         }
 715
 716         if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
 717                 pool = dev->prp_small_pool;
 718                 iod->npages = 0;
 719         } else {
 720                 pool = dev->prp_page_pool;
 721                 iod->npages = 1;
 722         }
 723
 724         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 725         if (!sg_list) {
 726                 iod->npages = -1;
 727                 return BLK_STS_RESOURCE;
 728         }
 729
 730         nvme_pci_iod_list(req)[0] = sg_list;
 731         iod->first_dma = sgl_dma;
 732
 733         nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
 734
 735         do {
 736                 if (i == SGES_PER_PAGE) {
 737                         struct nvme_sgl_desc *old_sg_desc = sg_list;
 738                         struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
 739
 740                         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 741                         if (!sg_list)
 742                                 return BLK_STS_RESOURCE;
 743
 744                         i = 0;
 745                         nvme_pci_iod_list(req)[iod->npages++] = sg_list;
 746                         sg_list[i++] = *link;
 747                         nvme_pci_sgl_set_seg(link, sgl_dma, entries);
 748                 }
 749
 750                 nvme_pci_sgl_set_data(&sg_list[i++], sg);
 751                 sg = sg_next(sg);
 752         } while (--entries > 0);
 753
 754         return BLK_STS_OK;
 755 }
 756
 757 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
 758                 struct request *req, struct nvme_rw_command *cmnd,
 759                 struct bio_vec *bv)
 760 {
 761         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 762         unsigned int offset = bv->bv_offset & (dev->ctrl.page_size - 1);
 763         unsigned int first_prp_len = dev->ctrl.page_size - offset;
 764
 765         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 766         if (dma_mapping_error(dev->dev, iod->first_dma))
 767                 return BLK_STS_RESOURCE;
 768         iod->dma_len = bv->bv_len;
 769
 770         cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
 771         if (bv->bv_len > first_prp_len)
 772                 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
 773         return 0;
 774 }
 775
 776 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
 777                 struct request *req, struct nvme_rw_command *cmnd,
 778                 struct bio_vec *bv)
 779 {
 780         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 781
 782         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 783         if (dma_mapping_error(dev->dev, iod->first_dma))
 784                 return BLK_STS_RESOURCE;
 785         iod->dma_len = bv->bv_len;
 786
 787         cmnd->flags = NVME_CMD_SGL_METABUF;
 788         cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
 789         cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
 790         cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
 791         return 0;
 792 }
 793
 794 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 795                 struct nvme_command *cmnd)
 796 {
 797         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 798         blk_status_t ret = BLK_STS_RESOURCE;
 799         int nr_mapped;
 800
 801         if (blk_rq_nr_phys_segments(req) == 1) {
 802                 struct bio_vec bv = req_bvec(req);
 803
 804                 if (!is_pci_p2pdma_page(bv.bv_page)) {
 805                         if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
 806                                 return nvme_setup_prp_simple(dev, req,
 807                                                              &cmnd->rw, &bv);
 808
 809                         if (iod->nvmeq->qid &&
 810                             dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
 811                                 return nvme_setup_sgl_simple(dev, req,
 812                                                              &cmnd->rw, &bv);
 813                 }
 814         }
 815
 816         iod->dma_len = 0;
 817         iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
 818         if (!iod->sg)
 819                 return BLK_STS_RESOURCE;
 820         sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
 821         iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 822         if (!iod->nents)
 823                 goto out;
 824
 825         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 826                 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
 827                                 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
 828         else
 829                 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
 830                                              rq_dma_dir(req), DMA_ATTR_NO_WARN);
 831         if (!nr_mapped)
 832                 goto out;
 833
 834         iod->use_sgl = nvme_pci_use_sgls(dev, req);
 835         if (iod->use_sgl)
 836                 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
 837         else
 838                 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
 839 out:
 840         if (ret != BLK_STS_OK)
 841                 nvme_unmap_data(dev, req);
 842         return ret;
 843 }
 844
 845 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
 846                 struct nvme_command *cmnd)
 847 {
 848         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 849
 850         iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
 851                         rq_dma_dir(req), 0);
 852         if (dma_mapping_error(dev->dev, iod->meta_dma))
 853                 return BLK_STS_IOERR;
 854         cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
 855         return 0;
 856 }
 857
 858 /*
 859  * NOTE: ns is NULL when called on the admin queue.
 860  */
 861 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 862                          const struct blk_mq_queue_data *bd)
 863 {
 864         struct nvme_ns *ns = hctx->queue->queuedata;
 865         struct nvme_queue *nvmeq = hctx->driver_data;
 866         struct nvme_dev *dev = nvmeq->dev;
 867         struct request *req = bd->rq;
 868         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 869         struct nvme_command cmnd;
 870         blk_status_t ret;
 871
 872         iod->aborted = 0;
 873         iod->npages = -1;
 874         iod->nents = 0;
 875
 876         /*
 877          * We should not need to do this, but we're still using this to
 878          * ensure we can drain requests on a dying queue.
 879          */
 880         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 881                 return BLK_STS_IOERR;
 882
 883         ret = nvme_setup_cmd(ns, req, &cmnd);
 884         if (ret)
 885                 return ret;
 886
 887         if (blk_rq_nr_phys_segments(req)) {
 888                 ret = nvme_map_data(dev, req, &cmnd);
 889                 if (ret)
 890                         goto out_free_cmd;
 891         }
 892
 893         if (blk_integrity_rq(req)) {
 894                 ret = nvme_map_metadata(dev, req, &cmnd);
 895                 if (ret)
 896                         goto out_unmap_data;
 897         }
 898
 899         blk_mq_start_request(req);
 900         nvme_submit_cmd(nvmeq, &cmnd, bd->last);
 901         return BLK_STS_OK;
 902 out_unmap_data:
 903         nvme_unmap_data(dev, req);
 904 out_free_cmd:
 905         nvme_cleanup_cmd(req);
 906         return ret;
 907 }
 908
 909 static void nvme_pci_complete_rq(struct request *req)
 910 {
 911         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 912         struct nvme_dev *dev = iod->nvmeq->dev;
 913
 914         if (blk_integrity_rq(req))
 915                 dma_unmap_page(dev->dev, iod->meta_dma,
 916                                rq_integrity_vec(req)->bv_len, rq_data_dir(req));
 917         if (blk_rq_nr_phys_segments(req))
 918                 nvme_unmap_data(dev, req);
 919         nvme_complete_rq(req);
 920 }
 921
 922 /* We read the CQE phase first to check if the rest of the entry is valid */
 923 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
 924 {
 925         return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
 926                         nvmeq->cq_phase;
 927 }
 928
 929 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
 930 {
 931         u16 head = nvmeq->cq_head;
 932
 933         if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
 934                                               nvmeq->dbbuf_cq_ei))
 935                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 936 }
 937
 938 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
 939 {
 940         if (!nvmeq->qid)
 941                 return nvmeq->dev->admin_tagset.tags[0];
 942         return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
 943 }
 944
 945 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
 946 {
 947         volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
 948         struct request *req;
 949
 950         if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
 951                 dev_warn(nvmeq->dev->ctrl.device,
 952                         "invalid id %d completed on queue %d\n",
 953                         cqe->command_id, le16_to_cpu(cqe->sq_id));
 954                 return;
 955         }
 956
 957         /*
 958          * AEN requests are special as they don't time out and can
 959          * survive any kind of queue freeze and often don't respond to
 960          * aborts.  We don't even bother to allocate a struct request
 961          * for them but rather special case them here.
 962          */
 963         if (unlikely(nvme_is_aen_req(nvmeq->qid, cqe->command_id))) {
 964                 nvme_complete_async_event(&nvmeq->dev->ctrl,
 965                                 cqe->status, &cqe->result);
 966                 return;
 967         }
 968
 969         req = blk_mq_tag_to_rq(nvme_queue_tagset(nvmeq), cqe->command_id);
 970         trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
 971         nvme_end_request(req, cqe->status, cqe->result);
 972 }
 973
 974 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
 975 {
 976         while (start != end) {
 977                 nvme_handle_cqe(nvmeq, start);
 978                 if (++start == nvmeq->q_depth)
 979                         start = 0;
 980         }
 981 }
 982
 983 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
 984 {
 985         if (nvmeq->cq_head == nvmeq->q_depth - 1) {
 986                 nvmeq->cq_head = 0;
 987                 nvmeq->cq_phase = !nvmeq->cq_phase;
 988         } else {
 989                 nvmeq->cq_head++;
 990         }
 991 }
 992
 993 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
 994                                   u16 *end, unsigned int tag)
 995 {
 996         int found = 0;
 997
 998         *start = nvmeq->cq_head;
 999         while (nvme_cqe_pending(nvmeq)) {
1000                 if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1001                         found++;
1002                 nvme_update_cq_head(nvmeq);
1003         }
1004         *end = nvmeq->cq_head;
1005
1006         if (*start != *end)
1007                 nvme_ring_cq_doorbell(nvmeq);
1008         return found;
1009 }
1010
1011 static irqreturn_t nvme_irq(int irq, void *data)
1012 {
1013         struct nvme_queue *nvmeq = data;
1014         irqreturn_t ret = IRQ_NONE;
1015         u16 start, end;
1016
1017         /*
1018          * The rmb/wmb pair ensures we see all updates from a previous run of
1019          * the irq handler, even if that was on another CPU.
1020          */
1021         rmb();
1022         nvme_process_cq(nvmeq, &start, &end, -1);
1023         wmb();
1024
1025         if (start != end) {
1026                 nvme_complete_cqes(nvmeq, start, end);
1027                 return IRQ_HANDLED;
1028         }
1029
1030         return ret;
1031 }
1032
1033 static irqreturn_t nvme_irq_check(int irq, void *data)
1034 {
1035         struct nvme_queue *nvmeq = data;
1036         if (nvme_cqe_pending(nvmeq))
1037                 return IRQ_WAKE_THREAD;
1038         return IRQ_NONE;
1039 }
1040
1041 /*
1042  * Poll for completions any queue, including those not dedicated to polling.
1043  * Can be called from any context.
1044  */
1045 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1046 {
1047         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1048         u16 start, end;
1049         int found;
1050
1051         /*
1052          * For a poll queue we need to protect against the polling thread
1053          * using the CQ lock.  For normal interrupt driven threads we have
1054          * to disable the interrupt to avoid racing with it.
1055          */
1056         if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1057                 spin_lock(&nvmeq->cq_poll_lock);
1058                 found = nvme_process_cq(nvmeq, &start, &end, tag);
1059                 spin_unlock(&nvmeq->cq_poll_lock);
1060         } else {
1061                 disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1062                 found = nvme_process_cq(nvmeq, &start, &end, tag);
1063                 enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1064         }
1065
1066         nvme_complete_cqes(nvmeq, start, end);
1067         return found;
1068 }
1069
1070 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1071 {
1072         struct nvme_queue *nvmeq = hctx->driver_data;
1073         u16 start, end;
1074         bool found;
1075
1076         if (!nvme_cqe_pending(nvmeq))
1077                 return 0;
1078
1079         spin_lock(&nvmeq->cq_poll_lock);
1080         found = nvme_process_cq(nvmeq, &start, &end, -1);
1081         nvme_complete_cqes(nvmeq, start, end);
1082         spin_unlock(&nvmeq->cq_poll_lock);
1083
1084         return found;
1085 }
1086
1087 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1088 {
1089         struct nvme_dev *dev = to_nvme_dev(ctrl);
1090         struct nvme_queue *nvmeq = &dev->queues[0];
1091         struct nvme_command c;
1092
1093         memset(&c, 0, sizeof(c));
1094         c.common.opcode = nvme_admin_async_event;
1095         c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1096         nvme_submit_cmd(nvmeq, &c, true);
1097 }
1098
1099 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1100 {
1101         struct nvme_command c;
1102
1103         memset(&c, 0, sizeof(c));
1104         c.delete_queue.opcode = opcode;
1105         c.delete_queue.qid = cpu_to_le16(id);
1106
1107         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1108 }
1109
1110 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1111                 struct nvme_queue *nvmeq, s16 vector)
1112 {
1113         struct nvme_command c;
1114         int flags = NVME_QUEUE_PHYS_CONTIG;
1115
1116         if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1117                 flags |= NVME_CQ_IRQ_ENABLED;
1118
1119         /*
1120          * Note: we (ab)use the fact that the prp fields survive if no data
1121          * is attached to the request.
1122          */
1123         memset(&c, 0, sizeof(c));
1124         c.create_cq.opcode = nvme_admin_create_cq;
1125         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1126         c.create_cq.cqid = cpu_to_le16(qid);
1127         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1128         c.create_cq.cq_flags = cpu_to_le16(flags);
1129         c.create_cq.irq_vector = cpu_to_le16(vector);
1130
1131         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1132 }
1133
1134 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1135                                                 struct nvme_queue *nvmeq)
1136 {
1137         struct nvme_ctrl *ctrl = &dev->ctrl;
1138         struct nvme_command c;
1139         int flags = NVME_QUEUE_PHYS_CONTIG;
1140
1141         /*
1142          * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1143          * set. Since URGENT priority is zeroes, it makes all queues
1144          * URGENT.
1145          */
1146         if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1147                 flags |= NVME_SQ_PRIO_MEDIUM;
1148
1149         /*
1150          * Note: we (ab)use the fact that the prp fields survive if no data
1151          * is attached to the request.
1152          */
1153         memset(&c, 0, sizeof(c));
1154         c.create_sq.opcode = nvme_admin_create_sq;
1155         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1156         c.create_sq.sqid = cpu_to_le16(qid);
1157         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1158         c.create_sq.sq_flags = cpu_to_le16(flags);
1159         c.create_sq.cqid = cpu_to_le16(qid);
1160
1161         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1162 }
1163
1164 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1165 {
1166         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1167 }
1168
1169 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1170 {
1171         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1172 }
1173
1174 static void abort_endio(struct request *req, blk_status_t error)
1175 {
1176         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1177         struct nvme_queue *nvmeq = iod->nvmeq;
1178
1179         dev_warn(nvmeq->dev->ctrl.device,
1180                  "Abort status: 0x%x", nvme_req(req)->status);
1181         atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1182         blk_mq_free_request(req);
1183 }
1184
1185 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1186 {
1187
1188         /* If true, indicates loss of adapter communication, possibly by a
1189          * NVMe Subsystem reset.
1190          */
1191         bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1192
1193         /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1194         switch (dev->ctrl.state) {
1195         case NVME_CTRL_RESETTING:
1196         case NVME_CTRL_CONNECTING:
1197                 return false;
1198         default:
1199                 break;
1200         }
1201
1202         /* We shouldn't reset unless the controller is on fatal error state
1203          * _or_ if we lost the communication with it.
1204          */
1205         if (!(csts & NVME_CSTS_CFS) && !nssro)
1206                 return false;
1207
1208         return true;
1209 }
1210
1211 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1212 {
1213         /* Read a config register to help see what died. */
1214         u16 pci_status;
1215         int result;
1216
1217         result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1218                                       &pci_status);
1219         if (result == PCIBIOS_SUCCESSFUL)
1220                 dev_warn(dev->ctrl.device,
1221                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1222                          csts, pci_status);
1223         else
1224                 dev_warn(dev->ctrl.device,
1225                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1226                          csts, result);
1227 }
1228
1229 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1230 {
1231         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1232         struct nvme_queue *nvmeq = iod->nvmeq;
1233         struct nvme_dev *dev = nvmeq->dev;
1234         struct request *abort_req;
1235         struct nvme_command cmd;
1236         u32 csts = readl(dev->bar + NVME_REG_CSTS);
1237
1238         /* If PCI error recovery process is happening, we cannot reset or
1239          * the recovery mechanism will surely fail.
1240          */
1241         mb();
1242         if (pci_channel_offline(to_pci_dev(dev->dev)))
1243                 return BLK_EH_RESET_TIMER;
1244
1245         /*
1246          * Reset immediately if the controller is failed
1247          */
1248         if (nvme_should_reset(dev, csts)) {
1249                 nvme_warn_reset(dev, csts);
1250                 nvme_dev_disable(dev, false);
1251                 nvme_reset_ctrl(&dev->ctrl);
1252                 return BLK_EH_DONE;
1253         }
1254
1255         /*
1256          * Did we miss an interrupt?
1257          */
1258         if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1259                 dev_warn(dev->ctrl.device,
1260                          "I/O %d QID %d timeout, completion polled\n",
1261                          req->tag, nvmeq->qid);
1262                 return BLK_EH_DONE;
1263         }
1264
1265         /*
1266          * Shutdown immediately if controller times out while starting. The
1267          * reset work will see the pci device disabled when it gets the forced
1268          * cancellation error. All outstanding requests are completed on
1269          * shutdown, so we return BLK_EH_DONE.
1270          */
1271         switch (dev->ctrl.state) {
1272         case NVME_CTRL_CONNECTING:
1273                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1274                 /* fall through */
1275         case NVME_CTRL_DELETING:
1276                 dev_warn_ratelimited(dev->ctrl.device,
1277                          "I/O %d QID %d timeout, disable controller\n",
1278                          req->tag, nvmeq->qid);
1279                 nvme_dev_disable(dev, true);
1280                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1281                 return BLK_EH_DONE;
1282         case NVME_CTRL_RESETTING:
1283                 return BLK_EH_RESET_TIMER;
1284         default:
1285                 break;
1286         }
1287
1288         /*
1289          * Shutdown the controller immediately and schedule a reset if the
1290          * command was already aborted once before and still hasn't been
1291          * returned to the driver, or if this is the admin queue.
1292          */
1293         if (!nvmeq->qid || iod->aborted) {
1294                 dev_warn(dev->ctrl.device,
1295                          "I/O %d QID %d timeout, reset controller\n",
1296                          req->tag, nvmeq->qid);
1297                 nvme_dev_disable(dev, false);
1298                 nvme_reset_ctrl(&dev->ctrl);
1299
1300                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1301                 return BLK_EH_DONE;
1302         }
1303
1304         if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1305                 atomic_inc(&dev->ctrl.abort_limit);
1306                 return BLK_EH_RESET_TIMER;
1307         }
1308         iod->aborted = 1;
1309
1310         memset(&cmd, 0, sizeof(cmd));
1311         cmd.abort.opcode = nvme_admin_abort_cmd;
1312         cmd.abort.cid = req->tag;
1313         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1314
1315         dev_warn(nvmeq->dev->ctrl.device,
1316                 "I/O %d QID %d timeout, aborting\n",
1317                  req->tag, nvmeq->qid);
1318
1319         abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1320                         BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1321         if (IS_ERR(abort_req)) {
1322                 atomic_inc(&dev->ctrl.abort_limit);
1323                 return BLK_EH_RESET_TIMER;
1324         }
1325
1326         abort_req->timeout = ADMIN_TIMEOUT;
1327         abort_req->end_io_data = NULL;
1328         blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1329
1330         /*
1331          * The aborted req will be completed on receiving the abort req.
1332          * We enable the timer again. If hit twice, it'll cause a device reset,
1333          * as the device then is in a faulty state.
1334          */
1335         return BLK_EH_RESET_TIMER;
1336 }
1337
1338 static void nvme_free_queue(struct nvme_queue *nvmeq)
1339 {
1340         dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1341                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1342         if (!nvmeq->sq_cmds)
1343                 return;
1344
1345         if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1346                 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1347                                 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1348         } else {
1349                 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1350                                 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1351         }
1352 }
1353
1354 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1355 {
1356         int i;
1357
1358         for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1359                 dev->ctrl.queue_count--;
1360                 nvme_free_queue(&dev->queues[i]);
1361         }
1362 }
1363
1364 /**
1365  * nvme_suspend_queue - put queue into suspended state
1366  * @nvmeq: queue to suspend
1367  */
1368 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1369 {
1370         if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1371                 return 1;
1372
1373         /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1374         mb();
1375
1376         nvmeq->dev->online_queues--;
1377         if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1378                 blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1379         if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1380                 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1381         return 0;
1382 }
1383
1384 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1385 {
1386         int i;
1387
1388         for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1389                 nvme_suspend_queue(&dev->queues[i]);
1390 }
1391
1392 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1393 {
1394         struct nvme_queue *nvmeq = &dev->queues[0];
1395
1396         if (shutdown)
1397                 nvme_shutdown_ctrl(&dev->ctrl);
1398         else
1399                 nvme_disable_ctrl(&dev->ctrl);
1400
1401         nvme_poll_irqdisable(nvmeq, -1);
1402 }
1403
1404 /*
1405  * Called only on a device that has been disabled and after all other threads
1406  * that can check this device's completion queues have synced. This is the
1407  * last chance for the driver to see a natural completion before
1408  * nvme_cancel_request() terminates all incomplete requests.
1409  */
1410 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1411 {
1412         u16 start, end;
1413         int i;
1414
1415         for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1416                 nvme_process_cq(&dev->queues[i], &start, &end, -1);
1417                 nvme_complete_cqes(&dev->queues[i], start, end);
1418         }
1419 }
1420
1421 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1422                                 int entry_size)
1423 {
1424         int q_depth = dev->q_depth;
1425         unsigned q_size_aligned = roundup(q_depth * entry_size,
1426                                           dev->ctrl.page_size);
1427
1428         if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1429                 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1430                 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1431                 q_depth = div_u64(mem_per_q, entry_size);
1432
1433                 /*
1434                  * Ensure the reduced q_depth is above some threshold where it
1435                  * would be better to map queues in system memory with the
1436                  * original depth
1437                  */
1438                 if (q_depth < 64)
1439                         return -ENOMEM;
1440         }
1441
1442         return q_depth;
1443 }
1444
1445 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1446                                 int qid)
1447 {
1448         struct pci_dev *pdev = to_pci_dev(dev->dev);
1449
1450         if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1451                 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1452                 if (nvmeq->sq_cmds) {
1453                         nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1454                                                         nvmeq->sq_cmds);
1455                         if (nvmeq->sq_dma_addr) {
1456                                 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1457                                 return 0;
1458                         }
1459
1460                         pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1461                 }
1462         }
1463
1464         nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1465                                 &nvmeq->sq_dma_addr, GFP_KERNEL);
1466         if (!nvmeq->sq_cmds)
1467                 return -ENOMEM;
1468         return 0;
1469 }
1470
1471 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1472 {
1473         struct nvme_queue *nvmeq = &dev->queues[qid];
1474
1475         if (dev->ctrl.queue_count > qid)
1476                 return 0;
1477
1478         nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1479         nvmeq->q_depth = depth;
1480         nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1481                                          &nvmeq->cq_dma_addr, GFP_KERNEL);
1482         if (!nvmeq->cqes)
1483                 goto free_nvmeq;
1484
1485         if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1486                 goto free_cqdma;
1487
1488         nvmeq->dev = dev;
1489         spin_lock_init(&nvmeq->sq_lock);
1490         spin_lock_init(&nvmeq->cq_poll_lock);
1491         nvmeq->cq_head = 0;
1492         nvmeq->cq_phase = 1;
1493         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1494         nvmeq->qid = qid;
1495         dev->ctrl.queue_count++;
1496
1497         return 0;
1498
1499  free_cqdma:
1500         dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1501                           nvmeq->cq_dma_addr);
1502  free_nvmeq:
1503         return -ENOMEM;
1504 }
1505
1506 static int queue_request_irq(struct nvme_queue *nvmeq)
1507 {
1508         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1509         int nr = nvmeq->dev->ctrl.instance;
1510
1511         if (use_threaded_interrupts) {
1512                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1513                                 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1514         } else {
1515                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1516                                 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1517         }
1518 }
1519
1520 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1521 {
1522         struct nvme_dev *dev = nvmeq->dev;
1523
1524         nvmeq->sq_tail = 0;
1525         nvmeq->last_sq_tail = 0;
1526         nvmeq->cq_head = 0;
1527         nvmeq->cq_phase = 1;
1528         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1529         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1530         nvme_dbbuf_init(dev, nvmeq, qid);
1531         dev->online_queues++;
1532         wmb(); /* ensure the first interrupt sees the initialization */
1533 }
1534
1535 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1536 {
1537         struct nvme_dev *dev = nvmeq->dev;
1538         int result;
1539         u16 vector = 0;
1540
1541         clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1542
1543         /*
1544          * A queue's vector matches the queue identifier unless the controller
1545          * has only one vector available.
1546          */
1547         if (!polled)
1548                 vector = dev->num_vecs == 1 ? 0 : qid;
1549         else
1550                 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1551
1552         result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1553         if (result)
1554                 return result;
1555
1556         result = adapter_alloc_sq(dev, qid, nvmeq);
1557         if (result < 0)
1558                 return result;
1559         if (result)
1560                 goto release_cq;
1561
1562         nvmeq->cq_vector = vector;
1563         nvme_init_queue(nvmeq, qid);
1564
1565         if (!polled) {
1566                 result = queue_request_irq(nvmeq);
1567                 if (result < 0)
1568                         goto release_sq;
1569         }
1570
1571         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1572         return result;
1573
1574 release_sq:
1575         dev->online_queues--;
1576         adapter_delete_sq(dev, qid);
1577 release_cq:
1578         adapter_delete_cq(dev, qid);
1579         return result;
1580 }
1581
1582 static const struct blk_mq_ops nvme_mq_admin_ops = {
1583         .queue_rq       = nvme_queue_rq,
1584         .complete       = nvme_pci_complete_rq,
1585         .init_hctx      = nvme_admin_init_hctx,
1586         .init_request   = nvme_init_request,
1587         .timeout        = nvme_timeout,
1588 };
1589
1590 static const struct blk_mq_ops nvme_mq_ops = {
1591         .queue_rq       = nvme_queue_rq,
1592         .complete       = nvme_pci_complete_rq,
1593         .commit_rqs     = nvme_commit_rqs,
1594         .init_hctx      = nvme_init_hctx,
1595         .init_request   = nvme_init_request,
1596         .map_queues     = nvme_pci_map_queues,
1597         .timeout        = nvme_timeout,
1598         .poll           = nvme_poll,
1599 };
1600
1601 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1602 {
1603         if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1604                 /*
1605                  * If the controller was reset during removal, it's possible
1606                  * user requests may be waiting on a stopped queue. Start the
1607                  * queue to flush these to completion.
1608                  */
1609                 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1610                 blk_cleanup_queue(dev->ctrl.admin_q);
1611                 blk_mq_free_tag_set(&dev->admin_tagset);
1612         }
1613 }
1614
1615 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1616 {
1617         if (!dev->ctrl.admin_q) {
1618                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1619                 dev->admin_tagset.nr_hw_queues = 1;
1620
1621                 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1622                 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1623                 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1624                 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1625                 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1626                 dev->admin_tagset.driver_data = dev;
1627
1628                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1629                         return -ENOMEM;
1630                 dev->ctrl.admin_tagset = &dev->admin_tagset;
1631
1632                 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1633                 if (IS_ERR(dev->ctrl.admin_q)) {
1634                         blk_mq_free_tag_set(&dev->admin_tagset);
1635                         return -ENOMEM;
1636                 }
1637                 if (!blk_get_queue(dev->ctrl.admin_q)) {
1638                         nvme_dev_remove_admin(dev);
1639                         dev->ctrl.admin_q = NULL;
1640                         return -ENODEV;
1641                 }
1642         } else
1643                 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1644
1645         return 0;
1646 }
1647
1648 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1649 {
1650         return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1651 }
1652
1653 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1654 {
1655         struct pci_dev *pdev = to_pci_dev(dev->dev);
1656
1657         if (size <= dev->bar_mapped_size)
1658                 return 0;
1659         if (size > pci_resource_len(pdev, 0))
1660                 return -ENOMEM;
1661         if (dev->bar)
1662                 iounmap(dev->bar);
1663         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1664         if (!dev->bar) {
1665                 dev->bar_mapped_size = 0;
1666                 return -ENOMEM;
1667         }
1668         dev->bar_mapped_size = size;
1669         dev->dbs = dev->bar + NVME_REG_DBS;
1670
1671         return 0;
1672 }
1673
1674 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1675 {
1676         int result;
1677         u32 aqa;
1678         struct nvme_queue *nvmeq;
1679
1680         result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1681         if (result < 0)
1682                 return result;
1683
1684         dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1685                                 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1686
1687         if (dev->subsystem &&
1688             (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1689                 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1690
1691         result = nvme_disable_ctrl(&dev->ctrl);
1692         if (result < 0)
1693                 return result;
1694
1695         result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1696         if (result)
1697                 return result;
1698
1699         nvmeq = &dev->queues[0];
1700         aqa = nvmeq->q_depth - 1;
1701         aqa |= aqa << 16;
1702
1703         writel(aqa, dev->bar + NVME_REG_AQA);
1704         lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1705         lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1706
1707         result = nvme_enable_ctrl(&dev->ctrl);
1708         if (result)
1709                 return result;
1710
1711         nvmeq->cq_vector = 0;
1712         nvme_init_queue(nvmeq, 0);
1713         result = queue_request_irq(nvmeq);
1714         if (result) {
1715                 dev->online_queues--;
1716                 return result;
1717         }
1718
1719         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1720         return result;
1721 }
1722
1723 static int nvme_create_io_queues(struct nvme_dev *dev)
1724 {
1725         unsigned i, max, rw_queues;
1726         int ret = 0;
1727
1728         for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1729                 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1730                         ret = -ENOMEM;
1731                         break;
1732                 }
1733         }
1734
1735         max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1736         if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1737                 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1738                                 dev->io_queues[HCTX_TYPE_READ];
1739         } else {
1740                 rw_queues = max;
1741         }
1742
1743         for (i = dev->online_queues; i <= max; i++) {
1744                 bool polled = i > rw_queues;
1745
1746                 ret = nvme_create_queue(&dev->queues[i], i, polled);
1747                 if (ret)
1748                         break;
1749         }
1750
1751         /*
1752          * Ignore failing Create SQ/CQ commands, we can continue with less
1753          * than the desired amount of queues, and even a controller without
1754          * I/O queues can still be used to issue admin commands.  This might
1755          * be useful to upgrade a buggy firmware for example.
1756          */
1757         return ret >= 0 ? 0 : ret;
1758 }
1759
1760 static ssize_t nvme_cmb_show(struct device *dev,
1761                              struct device_attribute *attr,
1762                              char *buf)
1763 {
1764         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1765
1766         return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1767                        ndev->cmbloc, ndev->cmbsz);
1768 }
1769 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1770
1771 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1772 {
1773         u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1774
1775         return 1ULL << (12 + 4 * szu);
1776 }
1777
1778 static u32 nvme_cmb_size(struct nvme_dev *dev)
1779 {
1780         return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1781 }
1782
1783 static void nvme_map_cmb(struct nvme_dev *dev)
1784 {
1785         u64 size, offset;
1786         resource_size_t bar_size;
1787         struct pci_dev *pdev = to_pci_dev(dev->dev);
1788         int bar;
1789
1790         if (dev->cmb_size)
1791                 return;
1792
1793         dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1794         if (!dev->cmbsz)
1795                 return;
1796         dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1797
1798         size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1799         offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1800         bar = NVME_CMB_BIR(dev->cmbloc);
1801         bar_size = pci_resource_len(pdev, bar);
1802
1803         if (offset > bar_size)
1804                 return;
1805
1806         /*
1807          * Controllers may support a CMB size larger than their BAR,
1808          * for example, due to being behind a bridge. Reduce the CMB to
1809          * the reported size of the BAR
1810          */
1811         if (size > bar_size - offset)
1812                 size = bar_size - offset;
1813
1814         if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1815                 dev_warn(dev->ctrl.device,
1816                          "failed to register the CMB\n");
1817                 return;
1818         }
1819
1820         dev->cmb_size = size;
1821         dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1822
1823         if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1824                         (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1825                 pci_p2pmem_publish(pdev, true);
1826
1827         if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1828                                     &dev_attr_cmb.attr, NULL))
1829                 dev_warn(dev->ctrl.device,
1830                          "failed to add sysfs attribute for CMB\n");
1831 }
1832
1833 static inline void nvme_release_cmb(struct nvme_dev *dev)
1834 {
1835         if (dev->cmb_size) {
1836                 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1837                                              &dev_attr_cmb.attr, NULL);
1838                 dev->cmb_size = 0;
1839         }
1840 }
1841
1842 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1843 {
1844         u64 dma_addr = dev->host_mem_descs_dma;
1845         struct nvme_command c;
1846         int ret;
1847
1848         memset(&c, 0, sizeof(c));
1849         c.features.opcode       = nvme_admin_set_features;
1850         c.features.fid          = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1851         c.features.dword11      = cpu_to_le32(bits);
1852         c.features.dword12      = cpu_to_le32(dev->host_mem_size >>
1853                                               ilog2(dev->ctrl.page_size));
1854         c.features.dword13      = cpu_to_le32(lower_32_bits(dma_addr));
1855         c.features.dword14      = cpu_to_le32(upper_32_bits(dma_addr));
1856         c.features.dword15      = cpu_to_le32(dev->nr_host_mem_descs);
1857
1858         ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1859         if (ret) {
1860                 dev_warn(dev->ctrl.device,
1861                          "failed to set host mem (err %d, flags %#x).\n",
1862                          ret, bits);
1863         }
1864         return ret;
1865 }
1866
1867 static void nvme_free_host_mem(struct nvme_dev *dev)
1868 {
1869         int i;
1870
1871         for (i = 0; i < dev->nr_host_mem_descs; i++) {
1872                 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1873                 size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1874
1875                 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1876                                le64_to_cpu(desc->addr),
1877                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1878         }
1879
1880         kfree(dev->host_mem_desc_bufs);
1881         dev->host_mem_desc_bufs = NULL;
1882         dma_free_coherent(dev->dev,
1883                         dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1884                         dev->host_mem_descs, dev->host_mem_descs_dma);
1885         dev->host_mem_descs = NULL;
1886         dev->nr_host_mem_descs = 0;
1887 }
1888
1889 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1890                 u32 chunk_size)
1891 {
1892         struct nvme_host_mem_buf_desc *descs;
1893         u32 max_entries, len;
1894         dma_addr_t descs_dma;
1895         int i = 0;
1896         void **bufs;
1897         u64 size, tmp;
1898
1899         tmp = (preferred + chunk_size - 1);
1900         do_div(tmp, chunk_size);
1901         max_entries = tmp;
1902
1903         if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1904                 max_entries = dev->ctrl.hmmaxd;
1905
1906         descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1907                                    &descs_dma, GFP_KERNEL);
1908         if (!descs)
1909                 goto out;
1910
1911         bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1912         if (!bufs)
1913                 goto out_free_descs;
1914
1915         for (size = 0; size < preferred && i < max_entries; size += len) {
1916                 dma_addr_t dma_addr;
1917
1918                 len = min_t(u64, chunk_size, preferred - size);
1919                 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1920                                 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1921                 if (!bufs[i])
1922                         break;
1923
1924                 descs[i].addr = cpu_to_le64(dma_addr);
1925                 descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1926                 i++;
1927         }
1928
1929         if (!size)
1930                 goto out_free_bufs;
1931
1932         dev->nr_host_mem_descs = i;
1933         dev->host_mem_size = size;
1934         dev->host_mem_descs = descs;
1935         dev->host_mem_descs_dma = descs_dma;
1936         dev->host_mem_desc_bufs = bufs;
1937         return 0;
1938
1939 out_free_bufs:
1940         while (--i >= 0) {
1941                 size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1942
1943                 dma_free_attrs(dev->dev, size, bufs[i],
1944                                le64_to_cpu(descs[i].addr),
1945                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1946         }
1947
1948         kfree(bufs);
1949 out_free_descs:
1950         dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1951                         descs_dma);
1952 out:
1953         dev->host_mem_descs = NULL;
1954         return -ENOMEM;
1955 }
1956
1957 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1958 {
1959         u32 chunk_size;
1960
1961         /* start big and work our way down */
1962         for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1963              chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1964              chunk_size /= 2) {
1965                 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1966                         if (!min || dev->host_mem_size >= min)
1967                                 return 0;
1968                         nvme_free_host_mem(dev);
1969                 }
1970         }
1971
1972         return -ENOMEM;
1973 }
1974
1975 static int nvme_setup_host_mem(struct nvme_dev *dev)
1976 {
1977         u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1978         u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1979         u64 min = (u64)dev->ctrl.hmmin * 4096;
1980         u32 enable_bits = NVME_HOST_MEM_ENABLE;
1981         int ret;
1982
1983         preferred = min(preferred, max);
1984         if (min > max) {
1985                 dev_warn(dev->ctrl.device,
1986                         "min host memory (%lld MiB) above limit (%d MiB).\n",
1987                         min >> ilog2(SZ_1M), max_host_mem_size_mb);
1988                 nvme_free_host_mem(dev);
1989                 return 0;
1990         }
1991
1992         /*
1993          * If we already have a buffer allocated check if we can reuse it.
1994          */
1995         if (dev->host_mem_descs) {
1996                 if (dev->host_mem_size >= min)
1997                         enable_bits |= NVME_HOST_MEM_RETURN;
1998                 else
1999                         nvme_free_host_mem(dev);
2000         }
2001
2002         if (!dev->host_mem_descs) {
2003                 if (nvme_alloc_host_mem(dev, min, preferred)) {
2004                         dev_warn(dev->ctrl.device,
2005                                 "failed to allocate host memory buffer.\n");
2006                         return 0; /* controller must work without HMB */
2007                 }
2008
2009                 dev_info(dev->ctrl.device,
2010                         "allocated %lld MiB host memory buffer.\n",
2011                         dev->host_mem_size >> ilog2(SZ_1M));
2012         }
2013
2014         ret = nvme_set_host_mem(dev, enable_bits);
2015         if (ret)
2016                 nvme_free_host_mem(dev);
2017         return ret;
2018 }
2019
2020 /*
2021  * nirqs is the number of interrupts available for write and read
2022  * queues. The core already reserved an interrupt for the admin queue.
2023  */
2024 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2025 {
2026         struct nvme_dev *dev = affd->priv;
2027         unsigned int nr_read_queues;
2028
2029         /*
2030          * If there is no interupt available for queues, ensure that
2031          * the default queue is set to 1. The affinity set size is
2032          * also set to one, but the irq core ignores it for this case.
2033          *
2034          * If only one interrupt is available or 'write_queue' == 0, combine
2035          * write and read queues.
2036          *
2037          * If 'write_queues' > 0, ensure it leaves room for at least one read
2038          * queue.
2039          */
2040         if (!nrirqs) {
2041                 nrirqs = 1;
2042                 nr_read_queues = 0;
2043         } else if (nrirqs == 1 || !write_queues) {
2044                 nr_read_queues = 0;
2045         } else if (write_queues >= nrirqs) {
2046                 nr_read_queues = 1;
2047         } else {
2048                 nr_read_queues = nrirqs - write_queues;
2049         }
2050
2051         dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2052         affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2053         dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2054         affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2055         affd->nr_sets = nr_read_queues ? 2 : 1;
2056 }
2057
2058 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2059 {
2060         struct pci_dev *pdev = to_pci_dev(dev->dev);
2061         struct irq_affinity affd = {
2062                 .pre_vectors    = 1,
2063                 .calc_sets      = nvme_calc_irq_sets,
2064                 .priv           = dev,
2065         };
2066         unsigned int irq_queues, this_p_queues;
2067
2068         /*
2069          * Poll queues don't need interrupts, but we need at least one IO
2070          * queue left over for non-polled IO.
2071          */
2072         this_p_queues = poll_queues;
2073         if (this_p_queues >= nr_io_queues) {
2074                 this_p_queues = nr_io_queues - 1;
2075                 irq_queues = 1;
2076         } else {
2077                 irq_queues = nr_io_queues - this_p_queues + 1;
2078         }
2079         dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2080
2081         /* Initialize for the single interrupt case */
2082         dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2083         dev->io_queues[HCTX_TYPE_READ] = 0;
2084
2085         /*
2086          * Some Apple controllers require all queues to use the
2087          * first vector.
2088          */
2089         if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2090                 irq_queues = 1;
2091
2092         return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2093                               PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2094 }
2095
2096 static void nvme_disable_io_queues(struct nvme_dev *dev)
2097 {
2098         if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2099                 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2100 }
2101
2102 static int nvme_setup_io_queues(struct nvme_dev *dev)
2103 {
2104         struct nvme_queue *adminq = &dev->queues[0];
2105         struct pci_dev *pdev = to_pci_dev(dev->dev);
2106         int result, nr_io_queues;
2107         unsigned long size;
2108
2109         nr_io_queues = max_io_queues();
2110
2111         /*
2112          * If tags are shared with admin queue (Apple bug), then
2113          * make sure we only use one IO queue.
2114          */
2115         if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2116                 nr_io_queues = 1;
2117
2118         result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2119         if (result < 0)
2120                 return result;
2121
2122         if (nr_io_queues == 0)
2123                 return 0;
2124
2125         clear_bit(NVMEQ_ENABLED, &adminq->flags);
2126
2127         if (dev->cmb_use_sqes) {
2128                 result = nvme_cmb_qdepth(dev, nr_io_queues,
2129                                 sizeof(struct nvme_command));
2130                 if (result > 0)
2131                         dev->q_depth = result;
2132                 else
2133                         dev->cmb_use_sqes = false;
2134         }
2135
2136         do {
2137                 size = db_bar_size(dev, nr_io_queues);
2138                 result = nvme_remap_bar(dev, size);
2139                 if (!result)
2140                         break;
2141                 if (!--nr_io_queues)
2142                         return -ENOMEM;
2143         } while (1);
2144         adminq->q_db = dev->dbs;
2145
2146  retry:
2147         /* Deregister the admin queue's interrupt */
2148         pci_free_irq(pdev, 0, adminq);
2149
2150         /*
2151          * If we enable msix early due to not intx, disable it again before
2152          * setting up the full range we need.
2153          */
2154         pci_free_irq_vectors(pdev);
2155
2156         result = nvme_setup_irqs(dev, nr_io_queues);
2157         if (result <= 0)
2158                 return -EIO;
2159
2160         dev->num_vecs = result;
2161         result = max(result - 1, 1);
2162         dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2163
2164         /*
2165          * Should investigate if there's a performance win from allocating
2166          * more queues than interrupt vectors; it might allow the submission
2167          * path to scale better, even if the receive path is limited by the
2168          * number of interrupts.
2169          */
2170         result = queue_request_irq(adminq);
2171         if (result)
2172                 return result;
2173         set_bit(NVMEQ_ENABLED, &adminq->flags);
2174
2175         result = nvme_create_io_queues(dev);
2176         if (result || dev->online_queues < 2)
2177                 return result;
2178
2179         if (dev->online_queues - 1 < dev->max_qid) {
2180                 nr_io_queues = dev->online_queues - 1;
2181                 nvme_disable_io_queues(dev);
2182                 nvme_suspend_io_queues(dev);
2183                 goto retry;
2184         }
2185         dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2186                                         dev->io_queues[HCTX_TYPE_DEFAULT],
2187                                         dev->io_queues[HCTX_TYPE_READ],
2188                                         dev->io_queues[HCTX_TYPE_POLL]);
2189         return 0;
2190 }
2191
2192 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2193 {
2194         struct nvme_queue *nvmeq = req->end_io_data;
2195
2196         blk_mq_free_request(req);
2197         complete(&nvmeq->delete_done);
2198 }
2199
2200 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2201 {
2202         struct nvme_queue *nvmeq = req->end_io_data;
2203
2204         if (error)
2205                 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2206
2207         nvme_del_queue_end(req, error);
2208 }
2209
2210 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2211 {
2212         struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2213         struct request *req;
2214         struct nvme_command cmd;
2215
2216         memset(&cmd, 0, sizeof(cmd));
2217         cmd.delete_queue.opcode = opcode;
2218         cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2219
2220         req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2221         if (IS_ERR(req))
2222                 return PTR_ERR(req);
2223
2224         req->timeout = ADMIN_TIMEOUT;
2225         req->end_io_data = nvmeq;
2226
2227         init_completion(&nvmeq->delete_done);
2228         blk_execute_rq_nowait(q, NULL, req, false,
2229                         opcode == nvme_admin_delete_cq ?
2230                                 nvme_del_cq_end : nvme_del_queue_end);
2231         return 0;
2232 }
2233
2234 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2235 {
2236         int nr_queues = dev->online_queues - 1, sent = 0;
2237         unsigned long timeout;
2238
2239  retry:
2240         timeout = ADMIN_TIMEOUT;
2241         while (nr_queues > 0) {
2242                 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2243                         break;
2244                 nr_queues--;
2245                 sent++;
2246         }
2247         while (sent) {
2248                 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2249
2250                 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2251                                 timeout);
2252                 if (timeout == 0)
2253                         return false;
2254
2255                 sent--;
2256                 if (nr_queues)
2257                         goto retry;
2258         }
2259         return true;
2260 }
2261
2262 static void nvme_dev_add(struct nvme_dev *dev)
2263 {
2264         int ret;
2265
2266         if (!dev->ctrl.tagset) {
2267                 dev->tagset.ops = &nvme_mq_ops;
2268                 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2269                 dev->tagset.nr_maps = 2; /* default + read */
2270                 if (dev->io_queues[HCTX_TYPE_POLL])
2271                         dev->tagset.nr_maps++;
2272                 dev->tagset.timeout = NVME_IO_TIMEOUT;
2273                 dev->tagset.numa_node = dev_to_node(dev->dev);
2274                 dev->tagset.queue_depth =
2275                                 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2276                 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2277                 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2278                 dev->tagset.driver_data = dev;
2279
2280                 /*
2281                  * Some Apple controllers requires tags to be unique
2282                  * across admin and IO queue, so reserve the first 32
2283                  * tags of the IO queue.
2284                  */
2285                 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2286                         dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2287
2288                 ret = blk_mq_alloc_tag_set(&dev->tagset);
2289                 if (ret) {
2290                         dev_warn(dev->ctrl.device,
2291                                 "IO queues tagset allocation failed %d\n", ret);
2292                         return;
2293                 }
2294                 dev->ctrl.tagset = &dev->tagset;
2295         } else {
2296                 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2297
2298                 /* Free previously allocated queues that are no longer usable */
2299                 nvme_free_queues(dev, dev->online_queues);
2300         }
2301
2302         nvme_dbbuf_set(dev);
2303 }
2304
2305 static int nvme_pci_enable(struct nvme_dev *dev)
2306 {
2307         int result = -ENOMEM;
2308         struct pci_dev *pdev = to_pci_dev(dev->dev);
2309
2310         if (pci_enable_device_mem(pdev))
2311                 return result;
2312
2313         pci_set_master(pdev);
2314
2315         if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2316                 goto disable;
2317
2318         if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2319                 result = -ENODEV;
2320                 goto disable;
2321         }
2322
2323         /*
2324          * Some devices and/or platforms don't advertise or work with INTx
2325          * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2326          * adjust this later.
2327          */
2328         result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2329         if (result < 0)
2330                 return result;
2331
2332         dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2333
2334         dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2335                                 io_queue_depth);
2336         dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2337         dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2338         dev->dbs = dev->bar + 4096;
2339
2340         /*
2341          * Some Apple controllers require a non-standard SQE size.
2342          * Interestingly they also seem to ignore the CC:IOSQES register
2343          * so we don't bother updating it here.
2344          */
2345         if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2346                 dev->io_sqes = 7;
2347         else
2348                 dev->io_sqes = NVME_NVM_IOSQES;
2349
2350         /*
2351          * Temporary fix for the Apple controller found in the MacBook8,1 and
2352          * some MacBook7,1 to avoid controller resets and data loss.
2353          */
2354         if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2355                 dev->q_depth = 2;
2356                 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2357                         "set queue depth=%u to work around controller resets\n",
2358                         dev->q_depth);
2359         } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2360                    (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2361                    NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2362                 dev->q_depth = 64;
2363                 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2364                         "set queue depth=%u\n", dev->q_depth);
2365         }
2366
2367         /*
2368          * Controllers with the shared tags quirk need the IO queue to be
2369          * big enough so that we get 32 tags for the admin queue
2370          */
2371         if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2372             (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2373                 dev->q_depth = NVME_AQ_DEPTH + 2;
2374                 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2375                          dev->q_depth);
2376         }
2377
2378
2379         nvme_map_cmb(dev);
2380
2381         pci_enable_pcie_error_reporting(pdev);
2382         pci_save_state(pdev);
2383         return 0;
2384
2385  disable:
2386         pci_disable_device(pdev);
2387         return result;
2388 }
2389
2390 static void nvme_dev_unmap(struct nvme_dev *dev)
2391 {
2392         if (dev->bar)
2393                 iounmap(dev->bar);
2394         pci_release_mem_regions(to_pci_dev(dev->dev));
2395 }
2396
2397 static void nvme_pci_disable(struct nvme_dev *dev)
2398 {
2399         struct pci_dev *pdev = to_pci_dev(dev->dev);
2400
2401         pci_free_irq_vectors(pdev);
2402
2403         if (pci_is_enabled(pdev)) {
2404                 pci_disable_pcie_error_reporting(pdev);
2405                 pci_disable_device(pdev);
2406         }
2407 }
2408
2409 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2410 {
2411         bool dead = true, freeze = false;
2412         struct pci_dev *pdev = to_pci_dev(dev->dev);
2413
2414         mutex_lock(&dev->shutdown_lock);
2415         if (pci_is_enabled(pdev)) {
2416                 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2417
2418                 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2419                     dev->ctrl.state == NVME_CTRL_RESETTING) {
2420                         freeze = true;
2421                         nvme_start_freeze(&dev->ctrl);
2422                 }
2423                 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2424                         pdev->error_state  != pci_channel_io_normal);
2425         }
2426
2427         /*
2428          * Give the controller a chance to complete all entered requests if
2429          * doing a safe shutdown.
2430          */
2431         if (!dead && shutdown && freeze)
2432                 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2433
2434         nvme_stop_queues(&dev->ctrl);
2435
2436         if (!dead && dev->ctrl.queue_count > 0) {
2437                 nvme_disable_io_queues(dev);
2438                 nvme_disable_admin_queue(dev, shutdown);
2439         }
2440         nvme_suspend_io_queues(dev);
2441         nvme_suspend_queue(&dev->queues[0]);
2442         nvme_pci_disable(dev);
2443         nvme_reap_pending_cqes(dev);
2444
2445         blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2446         blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2447         blk_mq_tagset_wait_completed_request(&dev->tagset);
2448         blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2449
2450         /*
2451          * The driver will not be starting up queues again if shutting down so
2452          * must flush all entered requests to their failed completion to avoid
2453          * deadlocking blk-mq hot-cpu notifier.
2454          */
2455         if (shutdown) {
2456                 nvme_start_queues(&dev->ctrl);
2457                 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2458                         blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2459         }
2460         mutex_unlock(&dev->shutdown_lock);
2461 }
2462
2463 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2464 {
2465         if (!nvme_wait_reset(&dev->ctrl))
2466                 return -EBUSY;
2467         nvme_dev_disable(dev, shutdown);
2468         return 0;
2469 }
2470
2471 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2472 {
2473         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2474                                                 PAGE_SIZE, PAGE_SIZE, 0);
2475         if (!dev->prp_page_pool)
2476                 return -ENOMEM;
2477
2478         /* Optimisation for I/Os between 4k and 128k */
2479         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2480                                                 256, 256, 0);
2481         if (!dev->prp_small_pool) {
2482                 dma_pool_destroy(dev->prp_page_pool);
2483                 return -ENOMEM;
2484         }
2485         return 0;
2486 }
2487
2488 static void nvme_release_prp_pools(struct nvme_dev *dev)
2489 {
2490         dma_pool_destroy(dev->prp_page_pool);
2491         dma_pool_destroy(dev->prp_small_pool);
2492 }
2493
2494 static void nvme_free_tagset(struct nvme_dev *dev)
2495 {
2496         if (dev->tagset.tags)
2497                 blk_mq_free_tag_set(&dev->tagset);
2498         dev->ctrl.tagset = NULL;
2499 }
2500
2501 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2502 {
2503         struct nvme_dev *dev = to_nvme_dev(ctrl);
2504
2505         nvme_dbbuf_dma_free(dev);
2506         put_device(dev->dev);
2507         nvme_free_tagset(dev);
2508         if (dev->ctrl.admin_q)
2509                 blk_put_queue(dev->ctrl.admin_q);
2510         kfree(dev->queues);
2511         free_opal_dev(dev->ctrl.opal_dev);
2512         mempool_destroy(dev->iod_mempool);
2513         kfree(dev);
2514 }
2515
2516 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2517 {
2518         /*
2519          * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2520          * may be holding this pci_dev's device lock.
2521          */
2522         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2523         nvme_get_ctrl(&dev->ctrl);
2524         nvme_dev_disable(dev, false);
2525         nvme_kill_queues(&dev->ctrl);
2526         if (!queue_work(nvme_wq, &dev->remove_work))
2527                 nvme_put_ctrl(&dev->ctrl);
2528 }
2529
2530 static void nvme_reset_work(struct work_struct *work)
2531 {
2532         struct nvme_dev *dev =
2533                 container_of(work, struct nvme_dev, ctrl.reset_work);
2534         bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2535         int result;
2536
2537         if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2538                 result = -ENODEV;
2539                 goto out;
2540         }
2541
2542         /*
2543          * If we're called to reset a live controller first shut it down before
2544          * moving on.
2545          */
2546         if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2547                 nvme_dev_disable(dev, false);
2548         nvme_sync_queues(&dev->ctrl);
2549
2550         mutex_lock(&dev->shutdown_lock);
2551         result = nvme_pci_enable(dev);
2552         if (result)
2553                 goto out_unlock;
2554
2555         result = nvme_pci_configure_admin_queue(dev);
2556         if (result)
2557                 goto out_unlock;
2558
2559         result = nvme_alloc_admin_tags(dev);
2560         if (result)
2561                 goto out_unlock;
2562
2563         /*
2564          * Limit the max command size to prevent iod->sg allocations going
2565          * over a single page.
2566          */
2567         dev->ctrl.max_hw_sectors = min_t(u32,
2568                 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2569         dev->ctrl.max_segments = NVME_MAX_SEGS;
2570
2571         /*
2572          * Don't limit the IOMMU merged segment size.
2573          */
2574         dma_set_max_seg_size(dev->dev, 0xffffffff);
2575
2576         mutex_unlock(&dev->shutdown_lock);
2577
2578         /*
2579          * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2580          * initializing procedure here.
2581          */
2582         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2583                 dev_warn(dev->ctrl.device,
2584                         "failed to mark controller CONNECTING\n");
2585                 result = -EBUSY;
2586                 goto out;
2587         }
2588
2589         result = nvme_init_identify(&dev->ctrl);
2590         if (result)
2591                 goto out;
2592
2593         if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2594                 if (!dev->ctrl.opal_dev)
2595                         dev->ctrl.opal_dev =
2596                                 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2597                 else if (was_suspend)
2598                         opal_unlock_from_suspend(dev->ctrl.opal_dev);
2599         } else {
2600                 free_opal_dev(dev->ctrl.opal_dev);
2601                 dev->ctrl.opal_dev = NULL;
2602         }
2603
2604         if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2605                 result = nvme_dbbuf_dma_alloc(dev);
2606                 if (result)
2607                         dev_warn(dev->dev,
2608                                  "unable to allocate dma for dbbuf\n");
2609         }
2610
2611         if (dev->ctrl.hmpre) {
2612                 result = nvme_setup_host_mem(dev);
2613                 if (result < 0)
2614                         goto out;
2615         }
2616
2617         result = nvme_setup_io_queues(dev);
2618         if (result)
2619                 goto out;
2620
2621         /*
2622          * Keep the controller around but remove all namespaces if we don't have
2623          * any working I/O queue.
2624          */
2625         if (dev->online_queues < 2) {
2626                 dev_warn(dev->ctrl.device, "IO queues not created\n");
2627                 nvme_kill_queues(&dev->ctrl);
2628                 nvme_remove_namespaces(&dev->ctrl);
2629                 nvme_free_tagset(dev);
2630         } else {
2631                 nvme_start_queues(&dev->ctrl);
2632                 nvme_wait_freeze(&dev->ctrl);
2633                 nvme_dev_add(dev);
2634                 nvme_unfreeze(&dev->ctrl);
2635         }
2636
2637         /*
2638          * If only admin queue live, keep it to do further investigation or
2639          * recovery.
2640          */
2641         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2642                 dev_warn(dev->ctrl.device,
2643                         "failed to mark controller live state\n");
2644                 result = -ENODEV;
2645                 goto out;
2646         }
2647
2648         nvme_start_ctrl(&dev->ctrl);
2649         return;
2650
2651  out_unlock:
2652         mutex_unlock(&dev->shutdown_lock);
2653  out:
2654         if (result)
2655                 dev_warn(dev->ctrl.device,
2656                          "Removing after probe failure status: %d\n", result);
2657         nvme_remove_dead_ctrl(dev);
2658 }
2659
2660 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2661 {
2662         struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2663         struct pci_dev *pdev = to_pci_dev(dev->dev);
2664
2665         if (pci_get_drvdata(pdev))
2666                 device_release_driver(&pdev->dev);
2667         nvme_put_ctrl(&dev->ctrl);
2668 }
2669
2670 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2671 {
2672         *val = readl(to_nvme_dev(ctrl)->bar + off);
2673         return 0;
2674 }
2675
2676 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2677 {
2678         writel(val, to_nvme_dev(ctrl)->bar + off);
2679         return 0;
2680 }
2681
2682 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2683 {
2684         *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2685         return 0;
2686 }
2687
2688 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2689 {
2690         struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2691
2692         return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2693 }
2694
2695 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2696         .name                   = "pcie",
2697         .module                 = THIS_MODULE,
2698         .flags                  = NVME_F_METADATA_SUPPORTED |
2699                                   NVME_F_PCI_P2PDMA,
2700         .reg_read32             = nvme_pci_reg_read32,
2701         .reg_write32            = nvme_pci_reg_write32,
2702         .reg_read64             = nvme_pci_reg_read64,
2703         .free_ctrl              = nvme_pci_free_ctrl,
2704         .submit_async_event     = nvme_pci_submit_async_event,
2705         .get_address            = nvme_pci_get_address,
2706 };
2707
2708 static int nvme_dev_map(struct nvme_dev *dev)
2709 {
2710         struct pci_dev *pdev = to_pci_dev(dev->dev);
2711
2712         if (pci_request_mem_regions(pdev, "nvme"))
2713                 return -ENODEV;
2714
2715         if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2716                 goto release;
2717
2718         return 0;
2719   release:
2720         pci_release_mem_regions(pdev);
2721         return -ENODEV;
2722 }
2723
2724 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2725 {
2726         if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2727                 /*
2728                  * Several Samsung devices seem to drop off the PCIe bus
2729                  * randomly when APST is on and uses the deepest sleep state.
2730                  * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2731                  * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2732                  * 950 PRO 256GB", but it seems to be restricted to two Dell
2733                  * laptops.
2734                  */
2735                 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2736                     (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2737                      dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2738                         return NVME_QUIRK_NO_DEEPEST_PS;
2739         } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2740                 /*
2741                  * Samsung SSD 960 EVO drops off the PCIe bus after system
2742                  * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2743                  * within few minutes after bootup on a Coffee Lake board -
2744                  * ASUS PRIME Z370-A
2745                  */
2746                 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2747                     (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2748                      dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2749                         return NVME_QUIRK_NO_APST;
2750         } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2751                     pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2752                    (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2753                 /*
2754                  * Forcing to use host managed nvme power settings for
2755                  * lowest idle power with quick resume latency on
2756                  * Samsung and Toshiba SSDs based on suspend behavior
2757                  * on Coffee Lake board for LENOVO C640
2758                  */
2759                 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2760                      dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2761                         return NVME_QUIRK_SIMPLE_SUSPEND;
2762         }
2763
2764         return 0;
2765 }
2766
2767 static void nvme_async_probe(void *data, async_cookie_t cookie)
2768 {
2769         struct nvme_dev *dev = data;
2770
2771         flush_work(&dev->ctrl.reset_work);
2772         flush_work(&dev->ctrl.scan_work);
2773         nvme_put_ctrl(&dev->ctrl);
2774 }
2775
2776 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2777 {
2778         int node, result = -ENOMEM;
2779         struct nvme_dev *dev;
2780         unsigned long quirks = id->driver_data;
2781         size_t alloc_size;
2782
2783         node = dev_to_node(&pdev->dev);
2784         if (node == NUMA_NO_NODE)
2785                 set_dev_node(&pdev->dev, first_memory_node);
2786
2787         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2788         if (!dev)
2789                 return -ENOMEM;
2790
2791         dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2792                                         GFP_KERNEL, node);
2793         if (!dev->queues)
2794                 goto free;
2795
2796         dev->dev = get_device(&pdev->dev);
2797         pci_set_drvdata(pdev, dev);
2798
2799         result = nvme_dev_map(dev);
2800         if (result)
2801                 goto put_pci;
2802
2803         INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2804         INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2805         mutex_init(&dev->shutdown_lock);
2806
2807         result = nvme_setup_prp_pools(dev);
2808         if (result)
2809                 goto unmap;
2810
2811         quirks |= check_vendor_combination_bug(pdev);
2812
2813         /*
2814          * Double check that our mempool alloc size will cover the biggest
2815          * command we support.
2816          */
2817         alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2818                                                 NVME_MAX_SEGS, true);
2819         WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2820
2821         dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2822                                                 mempool_kfree,
2823                                                 (void *) alloc_size,
2824                                                 GFP_KERNEL, node);
2825         if (!dev->iod_mempool) {
2826                 result = -ENOMEM;
2827                 goto release_pools;
2828         }
2829
2830         result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2831                         quirks);
2832         if (result)
2833                 goto release_mempool;
2834
2835         dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2836
2837         nvme_reset_ctrl(&dev->ctrl);
2838         nvme_get_ctrl(&dev->ctrl);
2839         async_schedule(nvme_async_probe, dev);
2840
2841         return 0;
2842
2843  release_mempool:
2844         mempool_destroy(dev->iod_mempool);
2845  release_pools:
2846         nvme_release_prp_pools(dev);
2847  unmap:
2848         nvme_dev_unmap(dev);
2849  put_pci:
2850         put_device(dev->dev);
2851  free:
2852         kfree(dev->queues);
2853         kfree(dev);
2854         return result;
2855 }
2856
2857 static void nvme_reset_prepare(struct pci_dev *pdev)
2858 {
2859         struct nvme_dev *dev = pci_get_drvdata(pdev);
2860
2861         /*
2862          * We don't need to check the return value from waiting for the reset
2863          * state as pci_dev device lock is held, making it impossible to race
2864          * with ->remove().
2865          */
2866         nvme_disable_prepare_reset(dev, false);
2867         nvme_sync_queues(&dev->ctrl);
2868 }
2869
2870 static void nvme_reset_done(struct pci_dev *pdev)
2871 {
2872         struct nvme_dev *dev = pci_get_drvdata(pdev);
2873
2874         if (!nvme_try_sched_reset(&dev->ctrl))
2875                 flush_work(&dev->ctrl.reset_work);
2876 }
2877
2878 static void nvme_shutdown(struct pci_dev *pdev)
2879 {
2880         struct nvme_dev *dev = pci_get_drvdata(pdev);
2881         nvme_disable_prepare_reset(dev, true);
2882 }
2883
2884 /*
2885  * The driver's remove may be called on a device in a partially initialized
2886  * state. This function must not have any dependencies on the device state in
2887  * order to proceed.
2888  */
2889 static void nvme_remove(struct pci_dev *pdev)
2890 {
2891         struct nvme_dev *dev = pci_get_drvdata(pdev);
2892
2893         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2894         pci_set_drvdata(pdev, NULL);
2895
2896         if (!pci_device_is_present(pdev)) {
2897                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2898                 nvme_dev_disable(dev, true);
2899                 nvme_dev_remove_admin(dev);
2900         }
2901
2902         flush_work(&dev->ctrl.reset_work);
2903         nvme_stop_ctrl(&dev->ctrl);
2904         nvme_remove_namespaces(&dev->ctrl);
2905         nvme_dev_disable(dev, true);
2906         nvme_release_cmb(dev);
2907         nvme_free_host_mem(dev);
2908         nvme_dev_remove_admin(dev);
2909         nvme_free_queues(dev, 0);
2910         nvme_uninit_ctrl(&dev->ctrl);
2911         nvme_release_prp_pools(dev);
2912         nvme_dev_unmap(dev);
2913         nvme_put_ctrl(&dev->ctrl);
2914 }
2915
2916 #ifdef CONFIG_PM_SLEEP
2917 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2918 {
2919         return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2920 }
2921
2922 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2923 {
2924         return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2925 }
2926
2927 static int nvme_resume(struct device *dev)
2928 {
2929         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2930         struct nvme_ctrl *ctrl = &ndev->ctrl;
2931
2932         if (ndev->last_ps == U32_MAX ||
2933             nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2934                 return nvme_try_sched_reset(&ndev->ctrl);
2935         return 0;
2936 }
2937
2938 static int nvme_suspend(struct device *dev)
2939 {
2940         struct pci_dev *pdev = to_pci_dev(dev);
2941         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2942         struct nvme_ctrl *ctrl = &ndev->ctrl;
2943         int ret = -EBUSY;
2944
2945         ndev->last_ps = U32_MAX;
2946
2947         /*
2948          * The platform does not remove power for a kernel managed suspend so
2949          * use host managed nvme power settings for lowest idle power if
2950          * possible. This should have quicker resume latency than a full device
2951          * shutdown.  But if the firmware is involved after the suspend or the
2952          * device does not support any non-default power states, shut down the
2953          * device fully.
2954          *
2955          * If ASPM is not enabled for the device, shut down the device and allow
2956          * the PCI bus layer to put it into D3 in order to take the PCIe link
2957          * down, so as to allow the platform to achieve its minimum low-power
2958          * state (which may not be possible if the link is up).
2959          */
2960         if (pm_suspend_via_firmware() || !ctrl->npss ||
2961             !pcie_aspm_enabled(pdev) ||
2962             (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
2963                 return nvme_disable_prepare_reset(ndev, true);
2964
2965         nvme_start_freeze(ctrl);
2966         nvme_wait_freeze(ctrl);
2967         nvme_sync_queues(ctrl);
2968
2969         if (ctrl->state != NVME_CTRL_LIVE)
2970                 goto unfreeze;
2971
2972         ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2973         if (ret < 0)
2974                 goto unfreeze;
2975
2976         /*
2977          * A saved state prevents pci pm from generically controlling the
2978          * device's power. If we're using protocol specific settings, we don't
2979          * want pci interfering.
2980          */
2981         pci_save_state(pdev);
2982
2983         ret = nvme_set_power_state(ctrl, ctrl->npss);
2984         if (ret < 0)
2985                 goto unfreeze;
2986
2987         if (ret) {
2988                 /* discard the saved state */
2989                 pci_load_saved_state(pdev, NULL);
2990
2991                 /*
2992                  * Clearing npss forces a controller reset on resume. The
2993                  * correct value will be rediscovered then.
2994                  */
2995                 ret = nvme_disable_prepare_reset(ndev, true);
2996                 ctrl->npss = 0;
2997         }
2998 unfreeze:
2999         nvme_unfreeze(ctrl);
3000         return ret;
3001 }
3002
3003 static int nvme_simple_suspend(struct device *dev)
3004 {
3005         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3006         return nvme_disable_prepare_reset(ndev, true);
3007 }
3008
3009 static int nvme_simple_resume(struct device *dev)
3010 {
3011         struct pci_dev *pdev = to_pci_dev(dev);
3012         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3013
3014         return nvme_try_sched_reset(&ndev->ctrl);
3015 }
3016
3017 static const struct dev_pm_ops nvme_dev_pm_ops = {
3018         .suspend        = nvme_suspend,
3019         .resume         = nvme_resume,
3020         .freeze         = nvme_simple_suspend,
3021         .thaw           = nvme_simple_resume,
3022         .poweroff       = nvme_simple_suspend,
3023         .restore        = nvme_simple_resume,
3024 };
3025 #endif /* CONFIG_PM_SLEEP */
3026
3027 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3028                                                 pci_channel_state_t state)
3029 {
3030         struct nvme_dev *dev = pci_get_drvdata(pdev);
3031
3032         /*
3033          * A frozen channel requires a reset. When detected, this method will
3034          * shutdown the controller to quiesce. The controller will be restarted
3035          * after the slot reset through driver's slot_reset callback.
3036          */
3037         switch (state) {
3038         case pci_channel_io_normal:
3039                 return PCI_ERS_RESULT_CAN_RECOVER;
3040         case pci_channel_io_frozen:
3041                 dev_warn(dev->ctrl.device,
3042                         "frozen state error detected, reset controller\n");
3043                 nvme_dev_disable(dev, false);
3044                 return PCI_ERS_RESULT_NEED_RESET;
3045         case pci_channel_io_perm_failure:
3046                 dev_warn(dev->ctrl.device,
3047                         "failure state error detected, request disconnect\n");
3048                 return PCI_ERS_RESULT_DISCONNECT;
3049         }
3050         return PCI_ERS_RESULT_NEED_RESET;
3051 }
3052
3053 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3054 {
3055         struct nvme_dev *dev = pci_get_drvdata(pdev);
3056
3057         dev_info(dev->ctrl.device, "restart after slot reset\n");
3058         pci_restore_state(pdev);
3059         nvme_reset_ctrl(&dev->ctrl);
3060         return PCI_ERS_RESULT_RECOVERED;
3061 }
3062
3063 static void nvme_error_resume(struct pci_dev *pdev)
3064 {
3065         struct nvme_dev *dev = pci_get_drvdata(pdev);
3066
3067         flush_work(&dev->ctrl.reset_work);
3068 }
3069
3070 static const struct pci_error_handlers nvme_err_handler = {
3071         .error_detected = nvme_error_detected,
3072         .slot_reset     = nvme_slot_reset,
3073         .resume         = nvme_error_resume,
3074         .reset_prepare  = nvme_reset_prepare,
3075         .reset_done     = nvme_reset_done,
3076 };
3077
3078 static const struct pci_device_id nvme_id_table[] = {
3079         { PCI_VDEVICE(INTEL, 0x0953),
3080                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3081                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3082         { PCI_VDEVICE(INTEL, 0x0a53),
3083                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3084                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3085         { PCI_VDEVICE(INTEL, 0x0a54),
3086                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3087                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3088         { PCI_VDEVICE(INTEL, 0x0a55),
3089                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3090                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3091         { PCI_VDEVICE(INTEL, 0xf1a5),   /* Intel 600P/P3100 */
3092                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3093                                 NVME_QUIRK_MEDIUM_PRIO_SQ |
3094                                 NVME_QUIRK_NO_TEMP_THRESH_CHANGE },
3095         { PCI_VDEVICE(INTEL, 0xf1a6),   /* Intel 760p/Pro 7600p */
3096                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3097         { PCI_VDEVICE(INTEL, 0x5845),   /* Qemu emulated controller */
3098                 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3099                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3100         { PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3101                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3102         { PCI_DEVICE(0x1c58, 0x0003),   /* HGST adapter */
3103                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3104         { PCI_DEVICE(0x1c58, 0x0023),   /* WDC SN200 adapter */
3105                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3106         { PCI_DEVICE(0x1c5f, 0x0540),   /* Memblaze Pblaze4 adapter */
3107                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3108         { PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3109                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3110         { PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3111                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3112         { PCI_DEVICE(0x1d1d, 0x1f1f),   /* LighNVM qemu device */
3113                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3114         { PCI_DEVICE(0x1d1d, 0x2807),   /* CNEX WL */
3115                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3116         { PCI_DEVICE(0x1d1d, 0x2601),   /* CNEX Granby */
3117                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3118         { PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3119                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3120         { PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3121                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3122                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3123         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3124         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3125                 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3126         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3127         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3128                 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3129                                 NVME_QUIRK_128_BYTES_SQES |
3130                                 NVME_QUIRK_SHARED_TAGS },
3131         { 0, }
3132 };
3133 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3134
3135 static struct pci_driver nvme_driver = {
3136         .name           = "nvme",
3137         .id_table       = nvme_id_table,
3138         .probe          = nvme_probe,
3139         .remove         = nvme_remove,
3140         .shutdown       = nvme_shutdown,
3141 #ifdef CONFIG_PM_SLEEP
3142         .driver         = {
3143                 .pm     = &nvme_dev_pm_ops,
3144         },
3145 #endif
3146         .sriov_configure = pci_sriov_configure_simple,
3147         .err_handler    = &nvme_err_handler,
3148 };
3149
3150 static int __init nvme_init(void)
3151 {
3152         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3153         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3154         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3155         BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3156
3157         write_queues = min(write_queues, num_possible_cpus());
3158         poll_queues = min(poll_queues, num_possible_cpus());
3159         return pci_register_driver(&nvme_driver);
3160 }
3161
3162 static void __exit nvme_exit(void)
3163 {
3164         pci_unregister_driver(&nvme_driver);
3165         flush_workqueue(nvme_wq);
3166 }
3167
3168 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3169 MODULE_LICENSE("GPL");
3170 MODULE_VERSION("1.0");
3171 module_init(nvme_init);
3172 module_exit(nvme_exit);