drivers/nvme/host/pci.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVM Express device driver
   4  * Copyright (c) 2011-2014, Intel Corporation.
   5  */
   6
   7 #include <linux/acpi.h>
   8 #include <linux/async.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/blk-mq.h>
  11 #include <linux/blk-mq-pci.h>
  12 #include <linux/blk-integrity.h>
  13 #include <linux/dmi.h>
  14 #include <linux/init.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/io.h>
  17 #include <linux/kstrtox.h>
  18 #include <linux/memremap.h>
  19 #include <linux/mm.h>
  20 #include <linux/module.h>
  21 #include <linux/mutex.h>
  22 #include <linux/once.h>
  23 #include <linux/pci.h>
  24 #include <linux/suspend.h>
  25 #include <linux/t10-pi.h>
  26 #include <linux/types.h>
  27 #include <linux/io-64-nonatomic-lo-hi.h>
  28 #include <linux/io-64-nonatomic-hi-lo.h>
  29 #include <linux/sed-opal.h>
  30 #include <linux/pci-p2pdma.h>
  31
  32 #include "trace.h"
  33 #include "nvme.h"
  34
  35 #define SQ_SIZE(q)      ((q)->q_depth << (q)->sqes)
  36 #define CQ_SIZE(q)      ((q)->q_depth * sizeof(struct nvme_completion))
  37
  38 #define SGES_PER_PAGE   (NVME_CTRL_PAGE_SIZE / sizeof(struct nvme_sgl_desc))
  39
  40 /*
  41  * These can be higher, but we need to ensure that any command doesn't
  42  * require an sg allocation that needs more than a page of data.
  43  */
  44 #define NVME_MAX_KB_SZ  8192
  45 #define NVME_MAX_SEGS   128
  46 #define NVME_MAX_NR_ALLOCATIONS 5
  47
  48 static int use_threaded_interrupts;
  49 module_param(use_threaded_interrupts, int, 0444);
  50
  51 static bool use_cmb_sqes = true;
  52 module_param(use_cmb_sqes, bool, 0444);
  53 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
  54
  55 static unsigned int max_host_mem_size_mb = 128;
  56 module_param(max_host_mem_size_mb, uint, 0444);
  57 MODULE_PARM_DESC(max_host_mem_size_mb,
  58         "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
  59
  60 static unsigned int sgl_threshold = SZ_32K;
  61 module_param(sgl_threshold, uint, 0644);
  62 MODULE_PARM_DESC(sgl_threshold,
  63                 "Use SGLs when average request segment size is larger or equal to "
  64                 "this size. Use 0 to disable SGLs.");
  65
  66 #define NVME_PCI_MIN_QUEUE_SIZE 2
  67 #define NVME_PCI_MAX_QUEUE_SIZE 4095
  68 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
  69 static const struct kernel_param_ops io_queue_depth_ops = {
  70         .set = io_queue_depth_set,
  71         .get = param_get_uint,
  72 };
  73
  74 static unsigned int io_queue_depth = 1024;
  75 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
  76 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2 and < 4096");
  77
  78 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
  79 {
  80         unsigned int n;
  81         int ret;
  82
  83         ret = kstrtouint(val, 10, &n);
  84         if (ret != 0 || n > num_possible_cpus())
  85                 return -EINVAL;
  86         return param_set_uint(val, kp);
  87 }
  88
  89 static const struct kernel_param_ops io_queue_count_ops = {
  90         .set = io_queue_count_set,
  91         .get = param_get_uint,
  92 };
  93
  94 static unsigned int write_queues;
  95 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
  96 MODULE_PARM_DESC(write_queues,
  97         "Number of queues to use for writes. If not set, reads and writes "
  98         "will share a queue set.");
  99
 100 static unsigned int poll_queues;
 101 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
 102 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
 103
 104 static bool noacpi;
 105 module_param(noacpi, bool, 0444);
 106 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
 107
 108 struct nvme_dev;
 109 struct nvme_queue;
 110
 111 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
 112 static void nvme_delete_io_queues(struct nvme_dev *dev);
 113 static void nvme_update_attrs(struct nvme_dev *dev);
 114
 115 /*
 116  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
 117  */
 118 struct nvme_dev {
 119         struct nvme_queue *queues;
 120         struct blk_mq_tag_set tagset;
 121         struct blk_mq_tag_set admin_tagset;
 122         u32 __iomem *dbs;
 123         struct device *dev;
 124         struct dma_pool *prp_page_pool;
 125         struct dma_pool *prp_small_pool;
 126         unsigned online_queues;
 127         unsigned max_qid;
 128         unsigned io_queues[HCTX_MAX_TYPES];
 129         unsigned int num_vecs;
 130         u32 q_depth;
 131         int io_sqes;
 132         u32 db_stride;
 133         void __iomem *bar;
 134         unsigned long bar_mapped_size;
 135         struct mutex shutdown_lock;
 136         bool subsystem;
 137         u64 cmb_size;
 138         bool cmb_use_sqes;
 139         u32 cmbsz;
 140         u32 cmbloc;
 141         struct nvme_ctrl ctrl;
 142         u32 last_ps;
 143         bool hmb;
 144
 145         mempool_t *iod_mempool;
 146
 147         /* shadow doorbell buffer support: */
 148         __le32 *dbbuf_dbs;
 149         dma_addr_t dbbuf_dbs_dma_addr;
 150         __le32 *dbbuf_eis;
 151         dma_addr_t dbbuf_eis_dma_addr;
 152
 153         /* host memory buffer support: */
 154         u64 host_mem_size;
 155         u32 nr_host_mem_descs;
 156         dma_addr_t host_mem_descs_dma;
 157         struct nvme_host_mem_buf_desc *host_mem_descs;
 158         void **host_mem_desc_bufs;
 159         unsigned int nr_allocated_queues;
 160         unsigned int nr_write_queues;
 161         unsigned int nr_poll_queues;
 162 };
 163
 164 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
 165 {
 166         return param_set_uint_minmax(val, kp, NVME_PCI_MIN_QUEUE_SIZE,
 167                         NVME_PCI_MAX_QUEUE_SIZE);
 168 }
 169
 170 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
 171 {
 172         return qid * 2 * stride;
 173 }
 174
 175 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
 176 {
 177         return (qid * 2 + 1) * stride;
 178 }
 179
 180 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 181 {
 182         return container_of(ctrl, struct nvme_dev, ctrl);
 183 }
 184
 185 /*
 186  * An NVM Express queue.  Each device has at least two (one for admin
 187  * commands and one for I/O commands).
 188  */
 189 struct nvme_queue {
 190         struct nvme_dev *dev;
 191         spinlock_t sq_lock;
 192         void *sq_cmds;
 193          /* only used for poll queues: */
 194         spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
 195         struct nvme_completion *cqes;
 196         dma_addr_t sq_dma_addr;
 197         dma_addr_t cq_dma_addr;
 198         u32 __iomem *q_db;
 199         u32 q_depth;
 200         u16 cq_vector;
 201         u16 sq_tail;
 202         u16 last_sq_tail;
 203         u16 cq_head;
 204         u16 qid;
 205         u8 cq_phase;
 206         u8 sqes;
 207         unsigned long flags;
 208 #define NVMEQ_ENABLED           0
 209 #define NVMEQ_SQ_CMB            1
 210 #define NVMEQ_DELETE_ERROR      2
 211 #define NVMEQ_POLLED            3
 212         __le32 *dbbuf_sq_db;
 213         __le32 *dbbuf_cq_db;
 214         __le32 *dbbuf_sq_ei;
 215         __le32 *dbbuf_cq_ei;
 216         struct completion delete_done;
 217 };
 218
 219 union nvme_descriptor {
 220         struct nvme_sgl_desc    *sg_list;
 221         __le64                  *prp_list;
 222 };
 223
 224 /*
 225  * The nvme_iod describes the data in an I/O.
 226  *
 227  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
 228  * to the actual struct scatterlist.
 229  */
 230 struct nvme_iod {
 231         struct nvme_request req;
 232         struct nvme_command cmd;
 233         bool aborted;
 234         s8 nr_allocations;      /* PRP list pool allocations. 0 means small
 235                                    pool in use */
 236         unsigned int dma_len;   /* length of single DMA segment mapping */
 237         dma_addr_t first_dma;
 238         dma_addr_t meta_dma;
 239         struct sg_table sgt;
 240         union nvme_descriptor list[NVME_MAX_NR_ALLOCATIONS];
 241 };
 242
 243 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
 244 {
 245         return dev->nr_allocated_queues * 8 * dev->db_stride;
 246 }
 247
 248 static void nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
 249 {
 250         unsigned int mem_size = nvme_dbbuf_size(dev);
 251
 252         if (!(dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP))
 253                 return;
 254
 255         if (dev->dbbuf_dbs) {
 256                 /*
 257                  * Clear the dbbuf memory so the driver doesn't observe stale
 258                  * values from the previous instantiation.
 259                  */
 260                 memset(dev->dbbuf_dbs, 0, mem_size);
 261                 memset(dev->dbbuf_eis, 0, mem_size);
 262                 return;
 263         }
 264
 265         dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
 266                                             &dev->dbbuf_dbs_dma_addr,
 267                                             GFP_KERNEL);
 268         if (!dev->dbbuf_dbs)
 269                 goto fail;
 270         dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
 271                                             &dev->dbbuf_eis_dma_addr,
 272                                             GFP_KERNEL);
 273         if (!dev->dbbuf_eis)
 274                 goto fail_free_dbbuf_dbs;
 275         return;
 276
 277 fail_free_dbbuf_dbs:
 278         dma_free_coherent(dev->dev, mem_size, dev->dbbuf_dbs,
 279                           dev->dbbuf_dbs_dma_addr);
 280         dev->dbbuf_dbs = NULL;
 281 fail:
 282         dev_warn(dev->dev, "unable to allocate dma for dbbuf\n");
 283 }
 284
 285 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
 286 {
 287         unsigned int mem_size = nvme_dbbuf_size(dev);
 288
 289         if (dev->dbbuf_dbs) {
 290                 dma_free_coherent(dev->dev, mem_size,
 291                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 292                 dev->dbbuf_dbs = NULL;
 293         }
 294         if (dev->dbbuf_eis) {
 295                 dma_free_coherent(dev->dev, mem_size,
 296                                   dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
 297                 dev->dbbuf_eis = NULL;
 298         }
 299 }
 300
 301 static void nvme_dbbuf_init(struct nvme_dev *dev,
 302                             struct nvme_queue *nvmeq, int qid)
 303 {
 304         if (!dev->dbbuf_dbs || !qid)
 305                 return;
 306
 307         nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
 308         nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
 309         nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
 310         nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
 311 }
 312
 313 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
 314 {
 315         if (!nvmeq->qid)
 316                 return;
 317
 318         nvmeq->dbbuf_sq_db = NULL;
 319         nvmeq->dbbuf_cq_db = NULL;
 320         nvmeq->dbbuf_sq_ei = NULL;
 321         nvmeq->dbbuf_cq_ei = NULL;
 322 }
 323
 324 static void nvme_dbbuf_set(struct nvme_dev *dev)
 325 {
 326         struct nvme_command c = { };
 327         unsigned int i;
 328
 329         if (!dev->dbbuf_dbs)
 330                 return;
 331
 332         c.dbbuf.opcode = nvme_admin_dbbuf;
 333         c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
 334         c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
 335
 336         if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
 337                 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
 338                 /* Free memory and continue on */
 339                 nvme_dbbuf_dma_free(dev);
 340
 341                 for (i = 1; i <= dev->online_queues; i++)
 342                         nvme_dbbuf_free(&dev->queues[i]);
 343         }
 344 }
 345
 346 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
 347 {
 348         return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
 349 }
 350
 351 /* Update dbbuf and return true if an MMIO is required */
 352 static bool nvme_dbbuf_update_and_check_event(u16 value, __le32 *dbbuf_db,
 353                                               volatile __le32 *dbbuf_ei)
 354 {
 355         if (dbbuf_db) {
 356                 u16 old_value, event_idx;
 357
 358                 /*
 359                  * Ensure that the queue is written before updating
 360                  * the doorbell in memory
 361                  */
 362                 wmb();
 363
 364                 old_value = le32_to_cpu(*dbbuf_db);
 365                 *dbbuf_db = cpu_to_le32(value);
 366
 367                 /*
 368                  * Ensure that the doorbell is updated before reading the event
 369                  * index from memory.  The controller needs to provide similar
 370                  * ordering to ensure the envent index is updated before reading
 371                  * the doorbell.
 372                  */
 373                 mb();
 374
 375                 event_idx = le32_to_cpu(*dbbuf_ei);
 376                 if (!nvme_dbbuf_need_event(event_idx, value, old_value))
 377                         return false;
 378         }
 379
 380         return true;
 381 }
 382
 383 /*
 384  * Will slightly overestimate the number of pages needed.  This is OK
 385  * as it only leads to a small amount of wasted memory for the lifetime of
 386  * the I/O.
 387  */
 388 static int nvme_pci_npages_prp(void)
 389 {
 390         unsigned max_bytes = (NVME_MAX_KB_SZ * 1024) + NVME_CTRL_PAGE_SIZE;
 391         unsigned nprps = DIV_ROUND_UP(max_bytes, NVME_CTRL_PAGE_SIZE);
 392         return DIV_ROUND_UP(8 * nprps, NVME_CTRL_PAGE_SIZE - 8);
 393 }
 394
 395 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 396                                 unsigned int hctx_idx)
 397 {
 398         struct nvme_dev *dev = to_nvme_dev(data);
 399         struct nvme_queue *nvmeq = &dev->queues[0];
 400
 401         WARN_ON(hctx_idx != 0);
 402         WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
 403
 404         hctx->driver_data = nvmeq;
 405         return 0;
 406 }
 407
 408 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 409                           unsigned int hctx_idx)
 410 {
 411         struct nvme_dev *dev = to_nvme_dev(data);
 412         struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
 413
 414         WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
 415         hctx->driver_data = nvmeq;
 416         return 0;
 417 }
 418
 419 static int nvme_pci_init_request(struct blk_mq_tag_set *set,
 420                 struct request *req, unsigned int hctx_idx,
 421                 unsigned int numa_node)
 422 {
 423         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 424
 425         nvme_req(req)->ctrl = set->driver_data;
 426         nvme_req(req)->cmd = &iod->cmd;
 427         return 0;
 428 }
 429
 430 static int queue_irq_offset(struct nvme_dev *dev)
 431 {
 432         /* if we have more than 1 vec, admin queue offsets us by 1 */
 433         if (dev->num_vecs > 1)
 434                 return 1;
 435
 436         return 0;
 437 }
 438
 439 static void nvme_pci_map_queues(struct blk_mq_tag_set *set)
 440 {
 441         struct nvme_dev *dev = to_nvme_dev(set->driver_data);
 442         int i, qoff, offset;
 443
 444         offset = queue_irq_offset(dev);
 445         for (i = 0, qoff = 0; i < set->nr_maps; i++) {
 446                 struct blk_mq_queue_map *map = &set->map[i];
 447
 448                 map->nr_queues = dev->io_queues[i];
 449                 if (!map->nr_queues) {
 450                         BUG_ON(i == HCTX_TYPE_DEFAULT);
 451                         continue;
 452                 }
 453
 454                 /*
 455                  * The poll queue(s) doesn't have an IRQ (and hence IRQ
 456                  * affinity), so use the regular blk-mq cpu mapping
 457                  */
 458                 map->queue_offset = qoff;
 459                 if (i != HCTX_TYPE_POLL && offset)
 460                         blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
 461                 else
 462                         blk_mq_map_queues(map);
 463                 qoff += map->nr_queues;
 464                 offset += map->nr_queues;
 465         }
 466 }
 467
 468 /*
 469  * Write sq tail if we are asked to, or if the next command would wrap.
 470  */
 471 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
 472 {
 473         if (!write_sq) {
 474                 u16 next_tail = nvmeq->sq_tail + 1;
 475
 476                 if (next_tail == nvmeq->q_depth)
 477                         next_tail = 0;
 478                 if (next_tail != nvmeq->last_sq_tail)
 479                         return;
 480         }
 481
 482         if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
 483                         nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
 484                 writel(nvmeq->sq_tail, nvmeq->q_db);
 485         nvmeq->last_sq_tail = nvmeq->sq_tail;
 486 }
 487
 488 static inline void nvme_sq_copy_cmd(struct nvme_queue *nvmeq,
 489                                     struct nvme_command *cmd)
 490 {
 491         memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 492                 absolute_pointer(cmd), sizeof(*cmd));
 493         if (++nvmeq->sq_tail == nvmeq->q_depth)
 494                 nvmeq->sq_tail = 0;
 495 }
 496
 497 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
 498 {
 499         struct nvme_queue *nvmeq = hctx->driver_data;
 500
 501         spin_lock(&nvmeq->sq_lock);
 502         if (nvmeq->sq_tail != nvmeq->last_sq_tail)
 503                 nvme_write_sq_db(nvmeq, true);
 504         spin_unlock(&nvmeq->sq_lock);
 505 }
 506
 507 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req,
 508                                      int nseg)
 509 {
 510         struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
 511         unsigned int avg_seg_size;
 512
 513         avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
 514
 515         if (!nvme_ctrl_sgl_supported(&dev->ctrl))
 516                 return false;
 517         if (!nvmeq->qid)
 518                 return false;
 519         if (!sgl_threshold || avg_seg_size < sgl_threshold)
 520                 return false;
 521         return true;
 522 }
 523
 524 static void nvme_free_prps(struct nvme_dev *dev, struct request *req)
 525 {
 526         const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
 527         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 528         dma_addr_t dma_addr = iod->first_dma;
 529         int i;
 530
 531         for (i = 0; i < iod->nr_allocations; i++) {
 532                 __le64 *prp_list = iod->list[i].prp_list;
 533                 dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
 534
 535                 dma_pool_free(dev->prp_page_pool, prp_list, dma_addr);
 536                 dma_addr = next_dma_addr;
 537         }
 538 }
 539
 540 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 541 {
 542         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 543
 544         if (iod->dma_len) {
 545                 dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len,
 546                                rq_dma_dir(req));
 547                 return;
 548         }
 549
 550         WARN_ON_ONCE(!iod->sgt.nents);
 551
 552         dma_unmap_sgtable(dev->dev, &iod->sgt, rq_dma_dir(req), 0);
 553
 554         if (iod->nr_allocations == 0)
 555                 dma_pool_free(dev->prp_small_pool, iod->list[0].sg_list,
 556                               iod->first_dma);
 557         else if (iod->nr_allocations == 1)
 558                 dma_pool_free(dev->prp_page_pool, iod->list[0].sg_list,
 559                               iod->first_dma);
 560         else
 561                 nvme_free_prps(dev, req);
 562         mempool_free(iod->sgt.sgl, dev->iod_mempool);
 563 }
 564
 565 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
 566 {
 567         int i;
 568         struct scatterlist *sg;
 569
 570         for_each_sg(sgl, sg, nents, i) {
 571                 dma_addr_t phys = sg_phys(sg);
 572                 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
 573                         "dma_address:%pad dma_length:%d\n",
 574                         i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
 575                         sg_dma_len(sg));
 576         }
 577 }
 578
 579 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
 580                 struct request *req, struct nvme_rw_command *cmnd)
 581 {
 582         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 583         struct dma_pool *pool;
 584         int length = blk_rq_payload_bytes(req);
 585         struct scatterlist *sg = iod->sgt.sgl;
 586         int dma_len = sg_dma_len(sg);
 587         u64 dma_addr = sg_dma_address(sg);
 588         int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
 589         __le64 *prp_list;
 590         dma_addr_t prp_dma;
 591         int nprps, i;
 592
 593         length -= (NVME_CTRL_PAGE_SIZE - offset);
 594         if (length <= 0) {
 595                 iod->first_dma = 0;
 596                 goto done;
 597         }
 598
 599         dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
 600         if (dma_len) {
 601                 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
 602         } else {
 603                 sg = sg_next(sg);
 604                 dma_addr = sg_dma_address(sg);
 605                 dma_len = sg_dma_len(sg);
 606         }
 607
 608         if (length <= NVME_CTRL_PAGE_SIZE) {
 609                 iod->first_dma = dma_addr;
 610                 goto done;
 611         }
 612
 613         nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
 614         if (nprps <= (256 / 8)) {
 615                 pool = dev->prp_small_pool;
 616                 iod->nr_allocations = 0;
 617         } else {
 618                 pool = dev->prp_page_pool;
 619                 iod->nr_allocations = 1;
 620         }
 621
 622         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 623         if (!prp_list) {
 624                 iod->nr_allocations = -1;
 625                 return BLK_STS_RESOURCE;
 626         }
 627         iod->list[0].prp_list = prp_list;
 628         iod->first_dma = prp_dma;
 629         i = 0;
 630         for (;;) {
 631                 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
 632                         __le64 *old_prp_list = prp_list;
 633                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 634                         if (!prp_list)
 635                                 goto free_prps;
 636                         iod->list[iod->nr_allocations++].prp_list = prp_list;
 637                         prp_list[0] = old_prp_list[i - 1];
 638                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 639                         i = 1;
 640                 }
 641                 prp_list[i++] = cpu_to_le64(dma_addr);
 642                 dma_len -= NVME_CTRL_PAGE_SIZE;
 643                 dma_addr += NVME_CTRL_PAGE_SIZE;
 644                 length -= NVME_CTRL_PAGE_SIZE;
 645                 if (length <= 0)
 646                         break;
 647                 if (dma_len > 0)
 648                         continue;
 649                 if (unlikely(dma_len < 0))
 650                         goto bad_sgl;
 651                 sg = sg_next(sg);
 652                 dma_addr = sg_dma_address(sg);
 653                 dma_len = sg_dma_len(sg);
 654         }
 655 done:
 656         cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sgt.sgl));
 657         cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
 658         return BLK_STS_OK;
 659 free_prps:
 660         nvme_free_prps(dev, req);
 661         return BLK_STS_RESOURCE;
 662 bad_sgl:
 663         WARN(DO_ONCE(nvme_print_sgl, iod->sgt.sgl, iod->sgt.nents),
 664                         "Invalid SGL for payload:%d nents:%d\n",
 665                         blk_rq_payload_bytes(req), iod->sgt.nents);
 666         return BLK_STS_IOERR;
 667 }
 668
 669 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
 670                 struct scatterlist *sg)
 671 {
 672         sge->addr = cpu_to_le64(sg_dma_address(sg));
 673         sge->length = cpu_to_le32(sg_dma_len(sg));
 674         sge->type = NVME_SGL_FMT_DATA_DESC << 4;
 675 }
 676
 677 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
 678                 dma_addr_t dma_addr, int entries)
 679 {
 680         sge->addr = cpu_to_le64(dma_addr);
 681         sge->length = cpu_to_le32(entries * sizeof(*sge));
 682         sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
 683 }
 684
 685 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
 686                 struct request *req, struct nvme_rw_command *cmd)
 687 {
 688         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 689         struct dma_pool *pool;
 690         struct nvme_sgl_desc *sg_list;
 691         struct scatterlist *sg = iod->sgt.sgl;
 692         unsigned int entries = iod->sgt.nents;
 693         dma_addr_t sgl_dma;
 694         int i = 0;
 695
 696         /* setting the transfer type as SGL */
 697         cmd->flags = NVME_CMD_SGL_METABUF;
 698
 699         if (entries == 1) {
 700                 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
 701                 return BLK_STS_OK;
 702         }
 703
 704         if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
 705                 pool = dev->prp_small_pool;
 706                 iod->nr_allocations = 0;
 707         } else {
 708                 pool = dev->prp_page_pool;
 709                 iod->nr_allocations = 1;
 710         }
 711
 712         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 713         if (!sg_list) {
 714                 iod->nr_allocations = -1;
 715                 return BLK_STS_RESOURCE;
 716         }
 717
 718         iod->list[0].sg_list = sg_list;
 719         iod->first_dma = sgl_dma;
 720
 721         nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
 722         do {
 723                 nvme_pci_sgl_set_data(&sg_list[i++], sg);
 724                 sg = sg_next(sg);
 725         } while (--entries > 0);
 726
 727         return BLK_STS_OK;
 728 }
 729
 730 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
 731                 struct request *req, struct nvme_rw_command *cmnd,
 732                 struct bio_vec *bv)
 733 {
 734         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 735         unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
 736         unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
 737
 738         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 739         if (dma_mapping_error(dev->dev, iod->first_dma))
 740                 return BLK_STS_RESOURCE;
 741         iod->dma_len = bv->bv_len;
 742
 743         cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
 744         if (bv->bv_len > first_prp_len)
 745                 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
 746         else
 747                 cmnd->dptr.prp2 = 0;
 748         return BLK_STS_OK;
 749 }
 750
 751 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
 752                 struct request *req, struct nvme_rw_command *cmnd,
 753                 struct bio_vec *bv)
 754 {
 755         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 756
 757         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 758         if (dma_mapping_error(dev->dev, iod->first_dma))
 759                 return BLK_STS_RESOURCE;
 760         iod->dma_len = bv->bv_len;
 761
 762         cmnd->flags = NVME_CMD_SGL_METABUF;
 763         cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
 764         cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
 765         cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
 766         return BLK_STS_OK;
 767 }
 768
 769 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 770                 struct nvme_command *cmnd)
 771 {
 772         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 773         blk_status_t ret = BLK_STS_RESOURCE;
 774         int rc;
 775
 776         if (blk_rq_nr_phys_segments(req) == 1) {
 777                 struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
 778                 struct bio_vec bv = req_bvec(req);
 779
 780                 if (!is_pci_p2pdma_page(bv.bv_page)) {
 781                         if ((bv.bv_offset & (NVME_CTRL_PAGE_SIZE - 1)) +
 782                              bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
 783                                 return nvme_setup_prp_simple(dev, req,
 784                                                              &cmnd->rw, &bv);
 785
 786                         if (nvmeq->qid && sgl_threshold &&
 787                             nvme_ctrl_sgl_supported(&dev->ctrl))
 788                                 return nvme_setup_sgl_simple(dev, req,
 789                                                              &cmnd->rw, &bv);
 790                 }
 791         }
 792
 793         iod->dma_len = 0;
 794         iod->sgt.sgl = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
 795         if (!iod->sgt.sgl)
 796                 return BLK_STS_RESOURCE;
 797         sg_init_table(iod->sgt.sgl, blk_rq_nr_phys_segments(req));
 798         iod->sgt.orig_nents = blk_rq_map_sg(req->q, req, iod->sgt.sgl);
 799         if (!iod->sgt.orig_nents)
 800                 goto out_free_sg;
 801
 802         rc = dma_map_sgtable(dev->dev, &iod->sgt, rq_dma_dir(req),
 803                              DMA_ATTR_NO_WARN);
 804         if (rc) {
 805                 if (rc == -EREMOTEIO)
 806                         ret = BLK_STS_TARGET;
 807                 goto out_free_sg;
 808         }
 809
 810         if (nvme_pci_use_sgls(dev, req, iod->sgt.nents))
 811                 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw);
 812         else
 813                 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
 814         if (ret != BLK_STS_OK)
 815                 goto out_unmap_sg;
 816         return BLK_STS_OK;
 817
 818 out_unmap_sg:
 819         dma_unmap_sgtable(dev->dev, &iod->sgt, rq_dma_dir(req), 0);
 820 out_free_sg:
 821         mempool_free(iod->sgt.sgl, dev->iod_mempool);
 822         return ret;
 823 }
 824
 825 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
 826                 struct nvme_command *cmnd)
 827 {
 828         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 829         struct bio_vec bv = rq_integrity_vec(req);
 830
 831         iod->meta_dma = dma_map_bvec(dev->dev, &bv, rq_dma_dir(req), 0);
 832         if (dma_mapping_error(dev->dev, iod->meta_dma))
 833                 return BLK_STS_IOERR;
 834         cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
 835         return BLK_STS_OK;
 836 }
 837
 838 static blk_status_t nvme_prep_rq(struct nvme_dev *dev, struct request *req)
 839 {
 840         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 841         blk_status_t ret;
 842
 843         iod->aborted = false;
 844         iod->nr_allocations = -1;
 845         iod->sgt.nents = 0;
 846
 847         ret = nvme_setup_cmd(req->q->queuedata, req);
 848         if (ret)
 849                 return ret;
 850
 851         if (blk_rq_nr_phys_segments(req)) {
 852                 ret = nvme_map_data(dev, req, &iod->cmd);
 853                 if (ret)
 854                         goto out_free_cmd;
 855         }
 856
 857         if (blk_integrity_rq(req)) {
 858                 ret = nvme_map_metadata(dev, req, &iod->cmd);
 859                 if (ret)
 860                         goto out_unmap_data;
 861         }
 862
 863         nvme_start_request(req);
 864         return BLK_STS_OK;
 865 out_unmap_data:
 866         if (blk_rq_nr_phys_segments(req))
 867                 nvme_unmap_data(dev, req);
 868 out_free_cmd:
 869         nvme_cleanup_cmd(req);
 870         return ret;
 871 }
 872
 873 /*
 874  * NOTE: ns is NULL when called on the admin queue.
 875  */
 876 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 877                          const struct blk_mq_queue_data *bd)
 878 {
 879         struct nvme_queue *nvmeq = hctx->driver_data;
 880         struct nvme_dev *dev = nvmeq->dev;
 881         struct request *req = bd->rq;
 882         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 883         blk_status_t ret;
 884
 885         /*
 886          * We should not need to do this, but we're still using this to
 887          * ensure we can drain requests on a dying queue.
 888          */
 889         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 890                 return BLK_STS_IOERR;
 891
 892         if (unlikely(!nvme_check_ready(&dev->ctrl, req, true)))
 893                 return nvme_fail_nonready_command(&dev->ctrl, req);
 894
 895         ret = nvme_prep_rq(dev, req);
 896         if (unlikely(ret))
 897                 return ret;
 898         spin_lock(&nvmeq->sq_lock);
 899         nvme_sq_copy_cmd(nvmeq, &iod->cmd);
 900         nvme_write_sq_db(nvmeq, bd->last);
 901         spin_unlock(&nvmeq->sq_lock);
 902         return BLK_STS_OK;
 903 }
 904
 905 static void nvme_submit_cmds(struct nvme_queue *nvmeq, struct request **rqlist)
 906 {
 907         spin_lock(&nvmeq->sq_lock);
 908         while (!rq_list_empty(*rqlist)) {
 909                 struct request *req = rq_list_pop(rqlist);
 910                 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 911
 912                 nvme_sq_copy_cmd(nvmeq, &iod->cmd);
 913         }
 914         nvme_write_sq_db(nvmeq, true);
 915         spin_unlock(&nvmeq->sq_lock);
 916 }
 917
 918 static bool nvme_prep_rq_batch(struct nvme_queue *nvmeq, struct request *req)
 919 {
 920         /*
 921          * We should not need to do this, but we're still using this to
 922          * ensure we can drain requests on a dying queue.
 923          */
 924         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 925                 return false;
 926         if (unlikely(!nvme_check_ready(&nvmeq->dev->ctrl, req, true)))
 927                 return false;
 928
 929         return nvme_prep_rq(nvmeq->dev, req) == BLK_STS_OK;
 930 }
 931
 932 static void nvme_queue_rqs(struct request **rqlist)
 933 {
 934         struct request *req, *next, *prev = NULL;
 935         struct request *requeue_list = NULL;
 936
 937         rq_list_for_each_safe(rqlist, req, next) {
 938                 struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
 939
 940                 if (!nvme_prep_rq_batch(nvmeq, req)) {
 941                         /* detach 'req' and add to remainder list */
 942                         rq_list_move(rqlist, &requeue_list, req, prev);
 943
 944                         req = prev;
 945                         if (!req)
 946                                 continue;
 947                 }
 948
 949                 if (!next || req->mq_hctx != next->mq_hctx) {
 950                         /* detach rest of list, and submit */
 951                         req->rq_next = NULL;
 952                         nvme_submit_cmds(nvmeq, rqlist);
 953                         *rqlist = next;
 954                         prev = NULL;
 955                 } else
 956                         prev = req;
 957         }
 958
 959         *rqlist = requeue_list;
 960 }
 961
 962 static __always_inline void nvme_pci_unmap_rq(struct request *req)
 963 {
 964         struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
 965         struct nvme_dev *dev = nvmeq->dev;
 966
 967         if (blk_integrity_rq(req)) {
 968                 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 969
 970                 dma_unmap_page(dev->dev, iod->meta_dma,
 971                                rq_integrity_vec(req).bv_len, rq_dma_dir(req));
 972         }
 973
 974         if (blk_rq_nr_phys_segments(req))
 975                 nvme_unmap_data(dev, req);
 976 }
 977
 978 static void nvme_pci_complete_rq(struct request *req)
 979 {
 980         nvme_pci_unmap_rq(req);
 981         nvme_complete_rq(req);
 982 }
 983
 984 static void nvme_pci_complete_batch(struct io_comp_batch *iob)
 985 {
 986         nvme_complete_batch(iob, nvme_pci_unmap_rq);
 987 }
 988
 989 /* We read the CQE phase first to check if the rest of the entry is valid */
 990 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
 991 {
 992         struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
 993
 994         return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
 995 }
 996
 997 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
 998 {
 999         u16 head = nvmeq->cq_head;
1000
1001         if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
1002                                               nvmeq->dbbuf_cq_ei))
1003                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
1004 }
1005
1006 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
1007 {
1008         if (!nvmeq->qid)
1009                 return nvmeq->dev->admin_tagset.tags[0];
1010         return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
1011 }
1012
1013 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
1014                                    struct io_comp_batch *iob, u16 idx)
1015 {
1016         struct nvme_completion *cqe = &nvmeq->cqes[idx];
1017         __u16 command_id = READ_ONCE(cqe->command_id);
1018         struct request *req;
1019
1020         /*
1021          * AEN requests are special as they don't time out and can
1022          * survive any kind of queue freeze and often don't respond to
1023          * aborts.  We don't even bother to allocate a struct request
1024          * for them but rather special case them here.
1025          */
1026         if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1027                 nvme_complete_async_event(&nvmeq->dev->ctrl,
1028                                 cqe->status, &cqe->result);
1029                 return;
1030         }
1031
1032         req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id);
1033         if (unlikely(!req)) {
1034                 dev_warn(nvmeq->dev->ctrl.device,
1035                         "invalid id %d completed on queue %d\n",
1036                         command_id, le16_to_cpu(cqe->sq_id));
1037                 return;
1038         }
1039
1040         trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
1041         if (!nvme_try_complete_req(req, cqe->status, cqe->result) &&
1042             !blk_mq_add_to_batch(req, iob, nvme_req(req)->status,
1043                                         nvme_pci_complete_batch))
1044                 nvme_pci_complete_rq(req);
1045 }
1046
1047 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1048 {
1049         u32 tmp = nvmeq->cq_head + 1;
1050
1051         if (tmp == nvmeq->q_depth) {
1052                 nvmeq->cq_head = 0;
1053                 nvmeq->cq_phase ^= 1;
1054         } else {
1055                 nvmeq->cq_head = tmp;
1056         }
1057 }
1058
1059 static inline int nvme_poll_cq(struct nvme_queue *nvmeq,
1060                                struct io_comp_batch *iob)
1061 {
1062         int found = 0;
1063
1064         while (nvme_cqe_pending(nvmeq)) {
1065                 found++;
1066                 /*
1067                  * load-load control dependency between phase and the rest of
1068                  * the cqe requires a full read memory barrier
1069                  */
1070                 dma_rmb();
1071                 nvme_handle_cqe(nvmeq, iob, nvmeq->cq_head);
1072                 nvme_update_cq_head(nvmeq);
1073         }
1074
1075         if (found)
1076                 nvme_ring_cq_doorbell(nvmeq);
1077         return found;
1078 }
1079
1080 static irqreturn_t nvme_irq(int irq, void *data)
1081 {
1082         struct nvme_queue *nvmeq = data;
1083         DEFINE_IO_COMP_BATCH(iob);
1084
1085         if (nvme_poll_cq(nvmeq, &iob)) {
1086                 if (!rq_list_empty(iob.req_list))
1087                         nvme_pci_complete_batch(&iob);
1088                 return IRQ_HANDLED;
1089         }
1090         return IRQ_NONE;
1091 }
1092
1093 static irqreturn_t nvme_irq_check(int irq, void *data)
1094 {
1095         struct nvme_queue *nvmeq = data;
1096
1097         if (nvme_cqe_pending(nvmeq))
1098                 return IRQ_WAKE_THREAD;
1099         return IRQ_NONE;
1100 }
1101
1102 /*
1103  * Poll for completions for any interrupt driven queue
1104  * Can be called from any context.
1105  */
1106 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1107 {
1108         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1109
1110         WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1111
1112         disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1113         nvme_poll_cq(nvmeq, NULL);
1114         enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1115 }
1116
1117 static int nvme_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1118 {
1119         struct nvme_queue *nvmeq = hctx->driver_data;
1120         bool found;
1121
1122         if (!nvme_cqe_pending(nvmeq))
1123                 return 0;
1124
1125         spin_lock(&nvmeq->cq_poll_lock);
1126         found = nvme_poll_cq(nvmeq, iob);
1127         spin_unlock(&nvmeq->cq_poll_lock);
1128
1129         return found;
1130 }
1131
1132 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1133 {
1134         struct nvme_dev *dev = to_nvme_dev(ctrl);
1135         struct nvme_queue *nvmeq = &dev->queues[0];
1136         struct nvme_command c = { };
1137
1138         c.common.opcode = nvme_admin_async_event;
1139         c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1140
1141         spin_lock(&nvmeq->sq_lock);
1142         nvme_sq_copy_cmd(nvmeq, &c);
1143         nvme_write_sq_db(nvmeq, true);
1144         spin_unlock(&nvmeq->sq_lock);
1145 }
1146
1147 static int nvme_pci_subsystem_reset(struct nvme_ctrl *ctrl)
1148 {
1149         struct nvme_dev *dev = to_nvme_dev(ctrl);
1150         int ret = 0;
1151
1152         /*
1153          * Taking the shutdown_lock ensures the BAR mapping is not being
1154          * altered by reset_work. Holding this lock before the RESETTING state
1155          * change, if successful, also ensures nvme_remove won't be able to
1156          * proceed to iounmap until we're done.
1157          */
1158         mutex_lock(&dev->shutdown_lock);
1159         if (!dev->bar_mapped_size) {
1160                 ret = -ENODEV;
1161                 goto unlock;
1162         }
1163
1164         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) {
1165                 ret = -EBUSY;
1166                 goto unlock;
1167         }
1168
1169         writel(NVME_SUBSYS_RESET, dev->bar + NVME_REG_NSSR);
1170         nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE);
1171
1172         /*
1173          * Read controller status to flush the previous write and trigger a
1174          * pcie read error.
1175          */
1176         readl(dev->bar + NVME_REG_CSTS);
1177 unlock:
1178         mutex_unlock(&dev->shutdown_lock);
1179         return ret;
1180 }
1181
1182 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1183 {
1184         struct nvme_command c = { };
1185
1186         c.delete_queue.opcode = opcode;
1187         c.delete_queue.qid = cpu_to_le16(id);
1188
1189         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1190 }
1191
1192 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1193                 struct nvme_queue *nvmeq, s16 vector)
1194 {
1195         struct nvme_command c = { };
1196         int flags = NVME_QUEUE_PHYS_CONTIG;
1197
1198         if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1199                 flags |= NVME_CQ_IRQ_ENABLED;
1200
1201         /*
1202          * Note: we (ab)use the fact that the prp fields survive if no data
1203          * is attached to the request.
1204          */
1205         c.create_cq.opcode = nvme_admin_create_cq;
1206         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1207         c.create_cq.cqid = cpu_to_le16(qid);
1208         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1209         c.create_cq.cq_flags = cpu_to_le16(flags);
1210         c.create_cq.irq_vector = cpu_to_le16(vector);
1211
1212         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1213 }
1214
1215 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1216                                                 struct nvme_queue *nvmeq)
1217 {
1218         struct nvme_ctrl *ctrl = &dev->ctrl;
1219         struct nvme_command c = { };
1220         int flags = NVME_QUEUE_PHYS_CONTIG;
1221
1222         /*
1223          * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1224          * set. Since URGENT priority is zeroes, it makes all queues
1225          * URGENT.
1226          */
1227         if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1228                 flags |= NVME_SQ_PRIO_MEDIUM;
1229
1230         /*
1231          * Note: we (ab)use the fact that the prp fields survive if no data
1232          * is attached to the request.
1233          */
1234         c.create_sq.opcode = nvme_admin_create_sq;
1235         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1236         c.create_sq.sqid = cpu_to_le16(qid);
1237         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1238         c.create_sq.sq_flags = cpu_to_le16(flags);
1239         c.create_sq.cqid = cpu_to_le16(qid);
1240
1241         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1242 }
1243
1244 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1245 {
1246         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1247 }
1248
1249 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1250 {
1251         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1252 }
1253
1254 static enum rq_end_io_ret abort_endio(struct request *req, blk_status_t error)
1255 {
1256         struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1257
1258         dev_warn(nvmeq->dev->ctrl.device,
1259                  "Abort status: 0x%x", nvme_req(req)->status);
1260         atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1261         blk_mq_free_request(req);
1262         return RQ_END_IO_NONE;
1263 }
1264
1265 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1266 {
1267         /* If true, indicates loss of adapter communication, possibly by a
1268          * NVMe Subsystem reset.
1269          */
1270         bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1271
1272         /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1273         switch (nvme_ctrl_state(&dev->ctrl)) {
1274         case NVME_CTRL_RESETTING:
1275         case NVME_CTRL_CONNECTING:
1276                 return false;
1277         default:
1278                 break;
1279         }
1280
1281         /* We shouldn't reset unless the controller is on fatal error state
1282          * _or_ if we lost the communication with it.
1283          */
1284         if (!(csts & NVME_CSTS_CFS) && !nssro)
1285                 return false;
1286
1287         return true;
1288 }
1289
1290 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1291 {
1292         /* Read a config register to help see what died. */
1293         u16 pci_status;
1294         int result;
1295
1296         result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1297                                       &pci_status);
1298         if (result == PCIBIOS_SUCCESSFUL)
1299                 dev_warn(dev->ctrl.device,
1300                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1301                          csts, pci_status);
1302         else
1303                 dev_warn(dev->ctrl.device,
1304                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1305                          csts, result);
1306
1307         if (csts != ~0)
1308                 return;
1309
1310         dev_warn(dev->ctrl.device,
1311                  "Does your device have a faulty power saving mode enabled?\n");
1312         dev_warn(dev->ctrl.device,
1313                  "Try \"nvme_core.default_ps_max_latency_us=0 pcie_aspm=off pcie_port_pm=off\" and report a bug\n");
1314 }
1315
1316 static enum blk_eh_timer_return nvme_timeout(struct request *req)
1317 {
1318         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1319         struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
1320         struct nvme_dev *dev = nvmeq->dev;
1321         struct request *abort_req;
1322         struct nvme_command cmd = { };
1323         u32 csts = readl(dev->bar + NVME_REG_CSTS);
1324         u8 opcode;
1325
1326         if (nvme_state_terminal(&dev->ctrl))
1327                 goto disable;
1328
1329         /* If PCI error recovery process is happening, we cannot reset or
1330          * the recovery mechanism will surely fail.
1331          */
1332         mb();
1333         if (pci_channel_offline(to_pci_dev(dev->dev)))
1334                 return BLK_EH_RESET_TIMER;
1335
1336         /*
1337          * Reset immediately if the controller is failed
1338          */
1339         if (nvme_should_reset(dev, csts)) {
1340                 nvme_warn_reset(dev, csts);
1341                 goto disable;
1342         }
1343
1344         /*
1345          * Did we miss an interrupt?
1346          */
1347         if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1348                 nvme_poll(req->mq_hctx, NULL);
1349         else
1350                 nvme_poll_irqdisable(nvmeq);
1351
1352         if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT) {
1353                 dev_warn(dev->ctrl.device,
1354                          "I/O tag %d (%04x) QID %d timeout, completion polled\n",
1355                          req->tag, nvme_cid(req), nvmeq->qid);
1356                 return BLK_EH_DONE;
1357         }
1358
1359         /*
1360          * Shutdown immediately if controller times out while starting. The
1361          * reset work will see the pci device disabled when it gets the forced
1362          * cancellation error. All outstanding requests are completed on
1363          * shutdown, so we return BLK_EH_DONE.
1364          */
1365         switch (nvme_ctrl_state(&dev->ctrl)) {
1366         case NVME_CTRL_CONNECTING:
1367                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1368                 fallthrough;
1369         case NVME_CTRL_DELETING:
1370                 dev_warn_ratelimited(dev->ctrl.device,
1371                          "I/O tag %d (%04x) QID %d timeout, disable controller\n",
1372                          req->tag, nvme_cid(req), nvmeq->qid);
1373                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1374                 nvme_dev_disable(dev, true);
1375                 return BLK_EH_DONE;
1376         case NVME_CTRL_RESETTING:
1377                 return BLK_EH_RESET_TIMER;
1378         default:
1379                 break;
1380         }
1381
1382         /*
1383          * Shutdown the controller immediately and schedule a reset if the
1384          * command was already aborted once before and still hasn't been
1385          * returned to the driver, or if this is the admin queue.
1386          */
1387         opcode = nvme_req(req)->cmd->common.opcode;
1388         if (!nvmeq->qid || iod->aborted) {
1389                 dev_warn(dev->ctrl.device,
1390                          "I/O tag %d (%04x) opcode %#x (%s) QID %d timeout, reset controller\n",
1391                          req->tag, nvme_cid(req), opcode,
1392                          nvme_opcode_str(nvmeq->qid, opcode), nvmeq->qid);
1393                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1394                 goto disable;
1395         }
1396
1397         if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1398                 atomic_inc(&dev->ctrl.abort_limit);
1399                 return BLK_EH_RESET_TIMER;
1400         }
1401         iod->aborted = true;
1402
1403         cmd.abort.opcode = nvme_admin_abort_cmd;
1404         cmd.abort.cid = nvme_cid(req);
1405         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1406
1407         dev_warn(nvmeq->dev->ctrl.device,
1408                  "I/O tag %d (%04x) opcode %#x (%s) QID %d timeout, aborting req_op:%s(%u) size:%u\n",
1409                  req->tag, nvme_cid(req), opcode, nvme_get_opcode_str(opcode),
1410                  nvmeq->qid, blk_op_str(req_op(req)), req_op(req),
1411                  blk_rq_bytes(req));
1412
1413         abort_req = blk_mq_alloc_request(dev->ctrl.admin_q, nvme_req_op(&cmd),
1414                                          BLK_MQ_REQ_NOWAIT);
1415         if (IS_ERR(abort_req)) {
1416                 atomic_inc(&dev->ctrl.abort_limit);
1417                 return BLK_EH_RESET_TIMER;
1418         }
1419         nvme_init_request(abort_req, &cmd);
1420
1421         abort_req->end_io = abort_endio;
1422         abort_req->end_io_data = NULL;
1423         blk_execute_rq_nowait(abort_req, false);
1424
1425         /*
1426          * The aborted req will be completed on receiving the abort req.
1427          * We enable the timer again. If hit twice, it'll cause a device reset,
1428          * as the device then is in a faulty state.
1429          */
1430         return BLK_EH_RESET_TIMER;
1431
1432 disable:
1433         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING)) {
1434                 if (nvme_state_terminal(&dev->ctrl))
1435                         nvme_dev_disable(dev, true);
1436                 return BLK_EH_DONE;
1437         }
1438
1439         nvme_dev_disable(dev, false);
1440         if (nvme_try_sched_reset(&dev->ctrl))
1441                 nvme_unquiesce_io_queues(&dev->ctrl);
1442         return BLK_EH_DONE;
1443 }
1444
1445 static void nvme_free_queue(struct nvme_queue *nvmeq)
1446 {
1447         dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1448                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1449         if (!nvmeq->sq_cmds)
1450                 return;
1451
1452         if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1453                 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1454                                 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1455         } else {
1456                 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1457                                 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1458         }
1459 }
1460
1461 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1462 {
1463         int i;
1464
1465         for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1466                 dev->ctrl.queue_count--;
1467                 nvme_free_queue(&dev->queues[i]);
1468         }
1469 }
1470
1471 static void nvme_suspend_queue(struct nvme_dev *dev, unsigned int qid)
1472 {
1473         struct nvme_queue *nvmeq = &dev->queues[qid];
1474
1475         if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1476                 return;
1477
1478         /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1479         mb();
1480
1481         nvmeq->dev->online_queues--;
1482         if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1483                 nvme_quiesce_admin_queue(&nvmeq->dev->ctrl);
1484         if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1485                 pci_free_irq(to_pci_dev(dev->dev), nvmeq->cq_vector, nvmeq);
1486 }
1487
1488 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1489 {
1490         int i;
1491
1492         for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1493                 nvme_suspend_queue(dev, i);
1494 }
1495
1496 /*
1497  * Called only on a device that has been disabled and after all other threads
1498  * that can check this device's completion queues have synced, except
1499  * nvme_poll(). This is the last chance for the driver to see a natural
1500  * completion before nvme_cancel_request() terminates all incomplete requests.
1501  */
1502 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1503 {
1504         int i;
1505
1506         for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1507                 spin_lock(&dev->queues[i].cq_poll_lock);
1508                 nvme_poll_cq(&dev->queues[i], NULL);
1509                 spin_unlock(&dev->queues[i].cq_poll_lock);
1510         }
1511 }
1512
1513 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1514                                 int entry_size)
1515 {
1516         int q_depth = dev->q_depth;
1517         unsigned q_size_aligned = roundup(q_depth * entry_size,
1518                                           NVME_CTRL_PAGE_SIZE);
1519
1520         if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1521                 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1522
1523                 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1524                 q_depth = div_u64(mem_per_q, entry_size);
1525
1526                 /*
1527                  * Ensure the reduced q_depth is above some threshold where it
1528                  * would be better to map queues in system memory with the
1529                  * original depth
1530                  */
1531                 if (q_depth < 64)
1532                         return -ENOMEM;
1533         }
1534
1535         return q_depth;
1536 }
1537
1538 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1539                                 int qid)
1540 {
1541         struct pci_dev *pdev = to_pci_dev(dev->dev);
1542
1543         if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1544                 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1545                 if (nvmeq->sq_cmds) {
1546                         nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1547                                                         nvmeq->sq_cmds);
1548                         if (nvmeq->sq_dma_addr) {
1549                                 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1550                                 return 0;
1551                         }
1552
1553                         pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1554                 }
1555         }
1556
1557         nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1558                                 &nvmeq->sq_dma_addr, GFP_KERNEL);
1559         if (!nvmeq->sq_cmds)
1560                 return -ENOMEM;
1561         return 0;
1562 }
1563
1564 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1565 {
1566         struct nvme_queue *nvmeq = &dev->queues[qid];
1567
1568         if (dev->ctrl.queue_count > qid)
1569                 return 0;
1570
1571         nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1572         nvmeq->q_depth = depth;
1573         nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1574                                          &nvmeq->cq_dma_addr, GFP_KERNEL);
1575         if (!nvmeq->cqes)
1576                 goto free_nvmeq;
1577
1578         if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1579                 goto free_cqdma;
1580
1581         nvmeq->dev = dev;
1582         spin_lock_init(&nvmeq->sq_lock);
1583         spin_lock_init(&nvmeq->cq_poll_lock);
1584         nvmeq->cq_head = 0;
1585         nvmeq->cq_phase = 1;
1586         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1587         nvmeq->qid = qid;
1588         dev->ctrl.queue_count++;
1589
1590         return 0;
1591
1592  free_cqdma:
1593         dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1594                           nvmeq->cq_dma_addr);
1595  free_nvmeq:
1596         return -ENOMEM;
1597 }
1598
1599 static int queue_request_irq(struct nvme_queue *nvmeq)
1600 {
1601         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1602         int nr = nvmeq->dev->ctrl.instance;
1603
1604         if (use_threaded_interrupts) {
1605                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1606                                 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1607         } else {
1608                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1609                                 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1610         }
1611 }
1612
1613 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1614 {
1615         struct nvme_dev *dev = nvmeq->dev;
1616
1617         nvmeq->sq_tail = 0;
1618         nvmeq->last_sq_tail = 0;
1619         nvmeq->cq_head = 0;
1620         nvmeq->cq_phase = 1;
1621         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1622         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1623         nvme_dbbuf_init(dev, nvmeq, qid);
1624         dev->online_queues++;
1625         wmb(); /* ensure the first interrupt sees the initialization */
1626 }
1627
1628 /*
1629  * Try getting shutdown_lock while setting up IO queues.
1630  */
1631 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev)
1632 {
1633         /*
1634          * Give up if the lock is being held by nvme_dev_disable.
1635          */
1636         if (!mutex_trylock(&dev->shutdown_lock))
1637                 return -ENODEV;
1638
1639         /*
1640          * Controller is in wrong state, fail early.
1641          */
1642         if (nvme_ctrl_state(&dev->ctrl) != NVME_CTRL_CONNECTING) {
1643                 mutex_unlock(&dev->shutdown_lock);
1644                 return -ENODEV;
1645         }
1646
1647         return 0;
1648 }
1649
1650 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1651 {
1652         struct nvme_dev *dev = nvmeq->dev;
1653         int result;
1654         u16 vector = 0;
1655
1656         clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1657
1658         /*
1659          * A queue's vector matches the queue identifier unless the controller
1660          * has only one vector available.
1661          */
1662         if (!polled)
1663                 vector = dev->num_vecs == 1 ? 0 : qid;
1664         else
1665                 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1666
1667         result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1668         if (result)
1669                 return result;
1670
1671         result = adapter_alloc_sq(dev, qid, nvmeq);
1672         if (result < 0)
1673                 return result;
1674         if (result)
1675                 goto release_cq;
1676
1677         nvmeq->cq_vector = vector;
1678
1679         result = nvme_setup_io_queues_trylock(dev);
1680         if (result)
1681                 return result;
1682         nvme_init_queue(nvmeq, qid);
1683         if (!polled) {
1684                 result = queue_request_irq(nvmeq);
1685                 if (result < 0)
1686                         goto release_sq;
1687         }
1688
1689         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1690         mutex_unlock(&dev->shutdown_lock);
1691         return result;
1692
1693 release_sq:
1694         dev->online_queues--;
1695         mutex_unlock(&dev->shutdown_lock);
1696         adapter_delete_sq(dev, qid);
1697 release_cq:
1698         adapter_delete_cq(dev, qid);
1699         return result;
1700 }
1701
1702 static const struct blk_mq_ops nvme_mq_admin_ops = {
1703         .queue_rq       = nvme_queue_rq,
1704         .complete       = nvme_pci_complete_rq,
1705         .init_hctx      = nvme_admin_init_hctx,
1706         .init_request   = nvme_pci_init_request,
1707         .timeout        = nvme_timeout,
1708 };
1709
1710 static const struct blk_mq_ops nvme_mq_ops = {
1711         .queue_rq       = nvme_queue_rq,
1712         .queue_rqs      = nvme_queue_rqs,
1713         .complete       = nvme_pci_complete_rq,
1714         .commit_rqs     = nvme_commit_rqs,
1715         .init_hctx      = nvme_init_hctx,
1716         .init_request   = nvme_pci_init_request,
1717         .map_queues     = nvme_pci_map_queues,
1718         .timeout        = nvme_timeout,
1719         .poll           = nvme_poll,
1720 };
1721
1722 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1723 {
1724         if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1725                 /*
1726                  * If the controller was reset during removal, it's possible
1727                  * user requests may be waiting on a stopped queue. Start the
1728                  * queue to flush these to completion.
1729                  */
1730                 nvme_unquiesce_admin_queue(&dev->ctrl);
1731                 nvme_remove_admin_tag_set(&dev->ctrl);
1732         }
1733 }
1734
1735 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1736 {
1737         return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1738 }
1739
1740 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1741 {
1742         struct pci_dev *pdev = to_pci_dev(dev->dev);
1743
1744         if (size <= dev->bar_mapped_size)
1745                 return 0;
1746         if (size > pci_resource_len(pdev, 0))
1747                 return -ENOMEM;
1748         if (dev->bar)
1749                 iounmap(dev->bar);
1750         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1751         if (!dev->bar) {
1752                 dev->bar_mapped_size = 0;
1753                 return -ENOMEM;
1754         }
1755         dev->bar_mapped_size = size;
1756         dev->dbs = dev->bar + NVME_REG_DBS;
1757
1758         return 0;
1759 }
1760
1761 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1762 {
1763         int result;
1764         u32 aqa;
1765         struct nvme_queue *nvmeq;
1766
1767         result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1768         if (result < 0)
1769                 return result;
1770
1771         dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1772                                 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1773
1774         if (dev->subsystem &&
1775             (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1776                 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1777
1778         /*
1779          * If the device has been passed off to us in an enabled state, just
1780          * clear the enabled bit.  The spec says we should set the 'shutdown
1781          * notification bits', but doing so may cause the device to complete
1782          * commands to the admin queue ... and we don't know what memory that
1783          * might be pointing at!
1784          */
1785         result = nvme_disable_ctrl(&dev->ctrl, false);
1786         if (result < 0)
1787                 return result;
1788
1789         result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1790         if (result)
1791                 return result;
1792
1793         dev->ctrl.numa_node = dev_to_node(dev->dev);
1794
1795         nvmeq = &dev->queues[0];
1796         aqa = nvmeq->q_depth - 1;
1797         aqa |= aqa << 16;
1798
1799         writel(aqa, dev->bar + NVME_REG_AQA);
1800         lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1801         lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1802
1803         result = nvme_enable_ctrl(&dev->ctrl);
1804         if (result)
1805                 return result;
1806
1807         nvmeq->cq_vector = 0;
1808         nvme_init_queue(nvmeq, 0);
1809         result = queue_request_irq(nvmeq);
1810         if (result) {
1811                 dev->online_queues--;
1812                 return result;
1813         }
1814
1815         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1816         return result;
1817 }
1818
1819 static int nvme_create_io_queues(struct nvme_dev *dev)
1820 {
1821         unsigned i, max, rw_queues;
1822         int ret = 0;
1823
1824         for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1825                 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1826                         ret = -ENOMEM;
1827                         break;
1828                 }
1829         }
1830
1831         max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1832         if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1833                 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1834                                 dev->io_queues[HCTX_TYPE_READ];
1835         } else {
1836                 rw_queues = max;
1837         }
1838
1839         for (i = dev->online_queues; i <= max; i++) {
1840                 bool polled = i > rw_queues;
1841
1842                 ret = nvme_create_queue(&dev->queues[i], i, polled);
1843                 if (ret)
1844                         break;
1845         }
1846
1847         /*
1848          * Ignore failing Create SQ/CQ commands, we can continue with less
1849          * than the desired amount of queues, and even a controller without
1850          * I/O queues can still be used to issue admin commands.  This might
1851          * be useful to upgrade a buggy firmware for example.
1852          */
1853         return ret >= 0 ? 0 : ret;
1854 }
1855
1856 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1857 {
1858         u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1859
1860         return 1ULL << (12 + 4 * szu);
1861 }
1862
1863 static u32 nvme_cmb_size(struct nvme_dev *dev)
1864 {
1865         return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1866 }
1867
1868 static void nvme_map_cmb(struct nvme_dev *dev)
1869 {
1870         u64 size, offset;
1871         resource_size_t bar_size;
1872         struct pci_dev *pdev = to_pci_dev(dev->dev);
1873         int bar;
1874
1875         if (dev->cmb_size)
1876                 return;
1877
1878         if (NVME_CAP_CMBS(dev->ctrl.cap))
1879                 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1880
1881         dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1882         if (!dev->cmbsz)
1883                 return;
1884         dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1885
1886         size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1887         offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1888         bar = NVME_CMB_BIR(dev->cmbloc);
1889         bar_size = pci_resource_len(pdev, bar);
1890
1891         if (offset > bar_size)
1892                 return;
1893
1894         /*
1895          * Tell the controller about the host side address mapping the CMB,
1896          * and enable CMB decoding for the NVMe 1.4+ scheme:
1897          */
1898         if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1899                 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1900                              (pci_bus_address(pdev, bar) + offset),
1901                              dev->bar + NVME_REG_CMBMSC);
1902         }
1903
1904         /*
1905          * Controllers may support a CMB size larger than their BAR,
1906          * for example, due to being behind a bridge. Reduce the CMB to
1907          * the reported size of the BAR
1908          */
1909         if (size > bar_size - offset)
1910                 size = bar_size - offset;
1911
1912         if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1913                 dev_warn(dev->ctrl.device,
1914                          "failed to register the CMB\n");
1915                 return;
1916         }
1917
1918         dev->cmb_size = size;
1919         dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1920
1921         if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1922                         (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1923                 pci_p2pmem_publish(pdev, true);
1924
1925         nvme_update_attrs(dev);
1926 }
1927
1928 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1929 {
1930         u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1931         u64 dma_addr = dev->host_mem_descs_dma;
1932         struct nvme_command c = { };
1933         int ret;
1934
1935         c.features.opcode       = nvme_admin_set_features;
1936         c.features.fid          = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1937         c.features.dword11      = cpu_to_le32(bits);
1938         c.features.dword12      = cpu_to_le32(host_mem_size);
1939         c.features.dword13      = cpu_to_le32(lower_32_bits(dma_addr));
1940         c.features.dword14      = cpu_to_le32(upper_32_bits(dma_addr));
1941         c.features.dword15      = cpu_to_le32(dev->nr_host_mem_descs);
1942
1943         ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1944         if (ret) {
1945                 dev_warn(dev->ctrl.device,
1946                          "failed to set host mem (err %d, flags %#x).\n",
1947                          ret, bits);
1948         } else
1949                 dev->hmb = bits & NVME_HOST_MEM_ENABLE;
1950
1951         return ret;
1952 }
1953
1954 static void nvme_free_host_mem(struct nvme_dev *dev)
1955 {
1956         int i;
1957
1958         for (i = 0; i < dev->nr_host_mem_descs; i++) {
1959                 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1960                 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
1961
1962                 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1963                                le64_to_cpu(desc->addr),
1964                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1965         }
1966
1967         kfree(dev->host_mem_desc_bufs);
1968         dev->host_mem_desc_bufs = NULL;
1969         dma_free_coherent(dev->dev,
1970                         dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1971                         dev->host_mem_descs, dev->host_mem_descs_dma);
1972         dev->host_mem_descs = NULL;
1973         dev->nr_host_mem_descs = 0;
1974 }
1975
1976 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1977                 u32 chunk_size)
1978 {
1979         struct nvme_host_mem_buf_desc *descs;
1980         u32 max_entries, len;
1981         dma_addr_t descs_dma;
1982         int i = 0;
1983         void **bufs;
1984         u64 size, tmp;
1985
1986         tmp = (preferred + chunk_size - 1);
1987         do_div(tmp, chunk_size);
1988         max_entries = tmp;
1989
1990         if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1991                 max_entries = dev->ctrl.hmmaxd;
1992
1993         descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1994                                    &descs_dma, GFP_KERNEL);
1995         if (!descs)
1996                 goto out;
1997
1998         bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1999         if (!bufs)
2000                 goto out_free_descs;
2001
2002         for (size = 0; size < preferred && i < max_entries; size += len) {
2003                 dma_addr_t dma_addr;
2004
2005                 len = min_t(u64, chunk_size, preferred - size);
2006                 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
2007                                 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2008                 if (!bufs[i])
2009                         break;
2010
2011                 descs[i].addr = cpu_to_le64(dma_addr);
2012                 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
2013                 i++;
2014         }
2015
2016         if (!size)
2017                 goto out_free_bufs;
2018
2019         dev->nr_host_mem_descs = i;
2020         dev->host_mem_size = size;
2021         dev->host_mem_descs = descs;
2022         dev->host_mem_descs_dma = descs_dma;
2023         dev->host_mem_desc_bufs = bufs;
2024         return 0;
2025
2026 out_free_bufs:
2027         while (--i >= 0) {
2028                 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
2029
2030                 dma_free_attrs(dev->dev, size, bufs[i],
2031                                le64_to_cpu(descs[i].addr),
2032                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2033         }
2034
2035         kfree(bufs);
2036 out_free_descs:
2037         dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
2038                         descs_dma);
2039 out:
2040         dev->host_mem_descs = NULL;
2041         return -ENOMEM;
2042 }
2043
2044 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2045 {
2046         u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2047         u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2048         u64 chunk_size;
2049
2050         /* start big and work our way down */
2051         for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2052                 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
2053                         if (!min || dev->host_mem_size >= min)
2054                                 return 0;
2055                         nvme_free_host_mem(dev);
2056                 }
2057         }
2058
2059         return -ENOMEM;
2060 }
2061
2062 static int nvme_setup_host_mem(struct nvme_dev *dev)
2063 {
2064         u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2065         u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2066         u64 min = (u64)dev->ctrl.hmmin * 4096;
2067         u32 enable_bits = NVME_HOST_MEM_ENABLE;
2068         int ret;
2069
2070         if (!dev->ctrl.hmpre)
2071                 return 0;
2072
2073         preferred = min(preferred, max);
2074         if (min > max) {
2075                 dev_warn(dev->ctrl.device,
2076                         "min host memory (%lld MiB) above limit (%d MiB).\n",
2077                         min >> ilog2(SZ_1M), max_host_mem_size_mb);
2078                 nvme_free_host_mem(dev);
2079                 return 0;
2080         }
2081
2082         /*
2083          * If we already have a buffer allocated check if we can reuse it.
2084          */
2085         if (dev->host_mem_descs) {
2086                 if (dev->host_mem_size >= min)
2087                         enable_bits |= NVME_HOST_MEM_RETURN;
2088                 else
2089                         nvme_free_host_mem(dev);
2090         }
2091
2092         if (!dev->host_mem_descs) {
2093                 if (nvme_alloc_host_mem(dev, min, preferred)) {
2094                         dev_warn(dev->ctrl.device,
2095                                 "failed to allocate host memory buffer.\n");
2096                         return 0; /* controller must work without HMB */
2097                 }
2098
2099                 dev_info(dev->ctrl.device,
2100                         "allocated %lld MiB host memory buffer.\n",
2101                         dev->host_mem_size >> ilog2(SZ_1M));
2102         }
2103
2104         ret = nvme_set_host_mem(dev, enable_bits);
2105         if (ret)
2106                 nvme_free_host_mem(dev);
2107         return ret;
2108 }
2109
2110 static ssize_t cmb_show(struct device *dev, struct device_attribute *attr,
2111                 char *buf)
2112 {
2113         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2114
2115         return sysfs_emit(buf, "cmbloc : x%08x\ncmbsz  : x%08x\n",
2116                        ndev->cmbloc, ndev->cmbsz);
2117 }
2118 static DEVICE_ATTR_RO(cmb);
2119
2120 static ssize_t cmbloc_show(struct device *dev, struct device_attribute *attr,
2121                 char *buf)
2122 {
2123         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2124
2125         return sysfs_emit(buf, "%u\n", ndev->cmbloc);
2126 }
2127 static DEVICE_ATTR_RO(cmbloc);
2128
2129 static ssize_t cmbsz_show(struct device *dev, struct device_attribute *attr,
2130                 char *buf)
2131 {
2132         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2133
2134         return sysfs_emit(buf, "%u\n", ndev->cmbsz);
2135 }
2136 static DEVICE_ATTR_RO(cmbsz);
2137
2138 static ssize_t hmb_show(struct device *dev, struct device_attribute *attr,
2139                         char *buf)
2140 {
2141         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2142
2143         return sysfs_emit(buf, "%d\n", ndev->hmb);
2144 }
2145
2146 static ssize_t hmb_store(struct device *dev, struct device_attribute *attr,
2147                          const char *buf, size_t count)
2148 {
2149         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
2150         bool new;
2151         int ret;
2152
2153         if (kstrtobool(buf, &new) < 0)
2154                 return -EINVAL;
2155
2156         if (new == ndev->hmb)
2157                 return count;
2158
2159         if (new) {
2160                 ret = nvme_setup_host_mem(ndev);
2161         } else {
2162                 ret = nvme_set_host_mem(ndev, 0);
2163                 if (!ret)
2164                         nvme_free_host_mem(ndev);
2165         }
2166
2167         if (ret < 0)
2168                 return ret;
2169
2170         return count;
2171 }
2172 static DEVICE_ATTR_RW(hmb);
2173
2174 static umode_t nvme_pci_attrs_are_visible(struct kobject *kobj,
2175                 struct attribute *a, int n)
2176 {
2177         struct nvme_ctrl *ctrl =
2178                 dev_get_drvdata(container_of(kobj, struct device, kobj));
2179         struct nvme_dev *dev = to_nvme_dev(ctrl);
2180
2181         if (a == &dev_attr_cmb.attr ||
2182             a == &dev_attr_cmbloc.attr ||
2183             a == &dev_attr_cmbsz.attr) {
2184                 if (!dev->cmbsz)
2185                         return 0;
2186         }
2187         if (a == &dev_attr_hmb.attr && !ctrl->hmpre)
2188                 return 0;
2189
2190         return a->mode;
2191 }
2192
2193 static struct attribute *nvme_pci_attrs[] = {
2194         &dev_attr_cmb.attr,
2195         &dev_attr_cmbloc.attr,
2196         &dev_attr_cmbsz.attr,
2197         &dev_attr_hmb.attr,
2198         NULL,
2199 };
2200
2201 static const struct attribute_group nvme_pci_dev_attrs_group = {
2202         .attrs          = nvme_pci_attrs,
2203         .is_visible     = nvme_pci_attrs_are_visible,
2204 };
2205
2206 static const struct attribute_group *nvme_pci_dev_attr_groups[] = {
2207         &nvme_dev_attrs_group,
2208         &nvme_pci_dev_attrs_group,
2209         NULL,
2210 };
2211
2212 static void nvme_update_attrs(struct nvme_dev *dev)
2213 {
2214         sysfs_update_group(&dev->ctrl.device->kobj, &nvme_pci_dev_attrs_group);
2215 }
2216
2217 /*
2218  * nirqs is the number of interrupts available for write and read
2219  * queues. The core already reserved an interrupt for the admin queue.
2220  */
2221 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2222 {
2223         struct nvme_dev *dev = affd->priv;
2224         unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2225
2226         /*
2227          * If there is no interrupt available for queues, ensure that
2228          * the default queue is set to 1. The affinity set size is
2229          * also set to one, but the irq core ignores it for this case.
2230          *
2231          * If only one interrupt is available or 'write_queue' == 0, combine
2232          * write and read queues.
2233          *
2234          * If 'write_queues' > 0, ensure it leaves room for at least one read
2235          * queue.
2236          */
2237         if (!nrirqs) {
2238                 nrirqs = 1;
2239                 nr_read_queues = 0;
2240         } else if (nrirqs == 1 || !nr_write_queues) {
2241                 nr_read_queues = 0;
2242         } else if (nr_write_queues >= nrirqs) {
2243                 nr_read_queues = 1;
2244         } else {
2245                 nr_read_queues = nrirqs - nr_write_queues;
2246         }
2247
2248         dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2249         affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2250         dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2251         affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2252         affd->nr_sets = nr_read_queues ? 2 : 1;
2253 }
2254
2255 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2256 {
2257         struct pci_dev *pdev = to_pci_dev(dev->dev);
2258         struct irq_affinity affd = {
2259                 .pre_vectors    = 1,
2260                 .calc_sets      = nvme_calc_irq_sets,
2261                 .priv           = dev,
2262         };
2263         unsigned int irq_queues, poll_queues;
2264         unsigned int flags = PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY;
2265
2266         /*
2267          * Poll queues don't need interrupts, but we need at least one I/O queue
2268          * left over for non-polled I/O.
2269          */
2270         poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2271         dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2272
2273         /*
2274          * Initialize for the single interrupt case, will be updated in
2275          * nvme_calc_irq_sets().
2276          */
2277         dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2278         dev->io_queues[HCTX_TYPE_READ] = 0;
2279
2280         /*
2281          * We need interrupts for the admin queue and each non-polled I/O queue,
2282          * but some Apple controllers require all queues to use the first
2283          * vector.
2284          */
2285         irq_queues = 1;
2286         if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2287                 irq_queues += (nr_io_queues - poll_queues);
2288         if (dev->ctrl.quirks & NVME_QUIRK_BROKEN_MSI)
2289                 flags &= ~PCI_IRQ_MSI;
2290         return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues, flags,
2291                                               &affd);
2292 }
2293
2294 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2295 {
2296         /*
2297          * If tags are shared with admin queue (Apple bug), then
2298          * make sure we only use one IO queue.
2299          */
2300         if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2301                 return 1;
2302         return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2303 }
2304
2305 static int nvme_setup_io_queues(struct nvme_dev *dev)
2306 {
2307         struct nvme_queue *adminq = &dev->queues[0];
2308         struct pci_dev *pdev = to_pci_dev(dev->dev);
2309         unsigned int nr_io_queues;
2310         unsigned long size;
2311         int result;
2312
2313         /*
2314          * Sample the module parameters once at reset time so that we have
2315          * stable values to work with.
2316          */
2317         dev->nr_write_queues = write_queues;
2318         dev->nr_poll_queues = poll_queues;
2319
2320         nr_io_queues = dev->nr_allocated_queues - 1;
2321         result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2322         if (result < 0)
2323                 return result;
2324
2325         if (nr_io_queues == 0)
2326                 return 0;
2327
2328         /*
2329          * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions
2330          * from set to unset. If there is a window to it is truely freed,
2331          * pci_free_irq_vectors() jumping into this window will crash.
2332          * And take lock to avoid racing with pci_free_irq_vectors() in
2333          * nvme_dev_disable() path.
2334          */
2335         result = nvme_setup_io_queues_trylock(dev);
2336         if (result)
2337                 return result;
2338         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2339                 pci_free_irq(pdev, 0, adminq);
2340
2341         if (dev->cmb_use_sqes) {
2342                 result = nvme_cmb_qdepth(dev, nr_io_queues,
2343                                 sizeof(struct nvme_command));
2344                 if (result > 0) {
2345                         dev->q_depth = result;
2346                         dev->ctrl.sqsize = result - 1;
2347                 } else {
2348                         dev->cmb_use_sqes = false;
2349                 }
2350         }
2351
2352         do {
2353                 size = db_bar_size(dev, nr_io_queues);
2354                 result = nvme_remap_bar(dev, size);
2355                 if (!result)
2356                         break;
2357                 if (!--nr_io_queues) {
2358                         result = -ENOMEM;
2359                         goto out_unlock;
2360                 }
2361         } while (1);
2362         adminq->q_db = dev->dbs;
2363
2364  retry:
2365         /* Deregister the admin queue's interrupt */
2366         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2367                 pci_free_irq(pdev, 0, adminq);
2368
2369         /*
2370          * If we enable msix early due to not intx, disable it again before
2371          * setting up the full range we need.
2372          */
2373         pci_free_irq_vectors(pdev);
2374
2375         result = nvme_setup_irqs(dev, nr_io_queues);
2376         if (result <= 0) {
2377                 result = -EIO;
2378                 goto out_unlock;
2379         }
2380
2381         dev->num_vecs = result;
2382         result = max(result - 1, 1);
2383         dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2384
2385         /*
2386          * Should investigate if there's a performance win from allocating
2387          * more queues than interrupt vectors; it might allow the submission
2388          * path to scale better, even if the receive path is limited by the
2389          * number of interrupts.
2390          */
2391         result = queue_request_irq(adminq);
2392         if (result)
2393                 goto out_unlock;
2394         set_bit(NVMEQ_ENABLED, &adminq->flags);
2395         mutex_unlock(&dev->shutdown_lock);
2396
2397         result = nvme_create_io_queues(dev);
2398         if (result || dev->online_queues < 2)
2399                 return result;
2400
2401         if (dev->online_queues - 1 < dev->max_qid) {
2402                 nr_io_queues = dev->online_queues - 1;
2403                 nvme_delete_io_queues(dev);
2404                 result = nvme_setup_io_queues_trylock(dev);
2405                 if (result)
2406                         return result;
2407                 nvme_suspend_io_queues(dev);
2408                 goto retry;
2409         }
2410         dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2411                                         dev->io_queues[HCTX_TYPE_DEFAULT],
2412                                         dev->io_queues[HCTX_TYPE_READ],
2413                                         dev->io_queues[HCTX_TYPE_POLL]);
2414         return 0;
2415 out_unlock:
2416         mutex_unlock(&dev->shutdown_lock);
2417         return result;
2418 }
2419
2420 static enum rq_end_io_ret nvme_del_queue_end(struct request *req,
2421                                              blk_status_t error)
2422 {
2423         struct nvme_queue *nvmeq = req->end_io_data;
2424
2425         blk_mq_free_request(req);
2426         complete(&nvmeq->delete_done);
2427         return RQ_END_IO_NONE;
2428 }
2429
2430 static enum rq_end_io_ret nvme_del_cq_end(struct request *req,
2431                                           blk_status_t error)
2432 {
2433         struct nvme_queue *nvmeq = req->end_io_data;
2434
2435         if (error)
2436                 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2437
2438         return nvme_del_queue_end(req, error);
2439 }
2440
2441 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2442 {
2443         struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2444         struct request *req;
2445         struct nvme_command cmd = { };
2446
2447         cmd.delete_queue.opcode = opcode;
2448         cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2449
2450         req = blk_mq_alloc_request(q, nvme_req_op(&cmd), BLK_MQ_REQ_NOWAIT);
2451         if (IS_ERR(req))
2452                 return PTR_ERR(req);
2453         nvme_init_request(req, &cmd);
2454
2455         if (opcode == nvme_admin_delete_cq)
2456                 req->end_io = nvme_del_cq_end;
2457         else
2458                 req->end_io = nvme_del_queue_end;
2459         req->end_io_data = nvmeq;
2460
2461         init_completion(&nvmeq->delete_done);
2462         blk_execute_rq_nowait(req, false);
2463         return 0;
2464 }
2465
2466 static bool __nvme_delete_io_queues(struct nvme_dev *dev, u8 opcode)
2467 {
2468         int nr_queues = dev->online_queues - 1, sent = 0;
2469         unsigned long timeout;
2470
2471  retry:
2472         timeout = NVME_ADMIN_TIMEOUT;
2473         while (nr_queues > 0) {
2474                 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2475                         break;
2476                 nr_queues--;
2477                 sent++;
2478         }
2479         while (sent) {
2480                 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2481
2482                 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2483                                 timeout);
2484                 if (timeout == 0)
2485                         return false;
2486
2487                 sent--;
2488                 if (nr_queues)
2489                         goto retry;
2490         }
2491         return true;
2492 }
2493
2494 static void nvme_delete_io_queues(struct nvme_dev *dev)
2495 {
2496         if (__nvme_delete_io_queues(dev, nvme_admin_delete_sq))
2497                 __nvme_delete_io_queues(dev, nvme_admin_delete_cq);
2498 }
2499
2500 static unsigned int nvme_pci_nr_maps(struct nvme_dev *dev)
2501 {
2502         if (dev->io_queues[HCTX_TYPE_POLL])
2503                 return 3;
2504         if (dev->io_queues[HCTX_TYPE_READ])
2505                 return 2;
2506         return 1;
2507 }
2508
2509 static void nvme_pci_update_nr_queues(struct nvme_dev *dev)
2510 {
2511         blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2512         /* free previously allocated queues that are no longer usable */
2513         nvme_free_queues(dev, dev->online_queues);
2514 }
2515
2516 static int nvme_pci_enable(struct nvme_dev *dev)
2517 {
2518         int result = -ENOMEM;
2519         struct pci_dev *pdev = to_pci_dev(dev->dev);
2520         unsigned int flags = PCI_IRQ_ALL_TYPES;
2521
2522         if (pci_enable_device_mem(pdev))
2523                 return result;
2524
2525         pci_set_master(pdev);
2526
2527         if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2528                 result = -ENODEV;
2529                 goto disable;
2530         }
2531
2532         /*
2533          * Some devices and/or platforms don't advertise or work with INTx
2534          * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2535          * adjust this later.
2536          */
2537         if (dev->ctrl.quirks & NVME_QUIRK_BROKEN_MSI)
2538                 flags &= ~PCI_IRQ_MSI;
2539         result = pci_alloc_irq_vectors(pdev, 1, 1, flags);
2540         if (result < 0)
2541                 goto disable;
2542
2543         dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2544
2545         dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2546                                 io_queue_depth);
2547         dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2548         dev->dbs = dev->bar + 4096;
2549
2550         /*
2551          * Some Apple controllers require a non-standard SQE size.
2552          * Interestingly they also seem to ignore the CC:IOSQES register
2553          * so we don't bother updating it here.
2554          */
2555         if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2556                 dev->io_sqes = 7;
2557         else
2558                 dev->io_sqes = NVME_NVM_IOSQES;
2559
2560         /*
2561          * Temporary fix for the Apple controller found in the MacBook8,1 and
2562          * some MacBook7,1 to avoid controller resets and data loss.
2563          */
2564         if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2565                 dev->q_depth = 2;
2566                 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2567                         "set queue depth=%u to work around controller resets\n",
2568                         dev->q_depth);
2569         } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2570                    (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2571                    NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2572                 dev->q_depth = 64;
2573                 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2574                         "set queue depth=%u\n", dev->q_depth);
2575         }
2576
2577         /*
2578          * Controllers with the shared tags quirk need the IO queue to be
2579          * big enough so that we get 32 tags for the admin queue
2580          */
2581         if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2582             (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2583                 dev->q_depth = NVME_AQ_DEPTH + 2;
2584                 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2585                          dev->q_depth);
2586         }
2587         dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2588
2589         nvme_map_cmb(dev);
2590
2591         pci_save_state(pdev);
2592
2593         result = nvme_pci_configure_admin_queue(dev);
2594         if (result)
2595                 goto free_irq;
2596         return result;
2597
2598  free_irq:
2599         pci_free_irq_vectors(pdev);
2600  disable:
2601         pci_disable_device(pdev);
2602         return result;
2603 }
2604
2605 static void nvme_dev_unmap(struct nvme_dev *dev)
2606 {
2607         if (dev->bar)
2608                 iounmap(dev->bar);
2609         pci_release_mem_regions(to_pci_dev(dev->dev));
2610 }
2611
2612 static bool nvme_pci_ctrl_is_dead(struct nvme_dev *dev)
2613 {
2614         struct pci_dev *pdev = to_pci_dev(dev->dev);
2615         u32 csts;
2616
2617         if (!pci_is_enabled(pdev) || !pci_device_is_present(pdev))
2618                 return true;
2619         if (pdev->error_state != pci_channel_io_normal)
2620                 return true;
2621
2622         csts = readl(dev->bar + NVME_REG_CSTS);
2623         return (csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY);
2624 }
2625
2626 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2627 {
2628         enum nvme_ctrl_state state = nvme_ctrl_state(&dev->ctrl);
2629         struct pci_dev *pdev = to_pci_dev(dev->dev);
2630         bool dead;
2631
2632         mutex_lock(&dev->shutdown_lock);
2633         dead = nvme_pci_ctrl_is_dead(dev);
2634         if (state == NVME_CTRL_LIVE || state == NVME_CTRL_RESETTING) {
2635                 if (pci_is_enabled(pdev))
2636                         nvme_start_freeze(&dev->ctrl);
2637                 /*
2638                  * Give the controller a chance to complete all entered requests
2639                  * if doing a safe shutdown.
2640                  */
2641                 if (!dead && shutdown)
2642                         nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2643         }
2644
2645         nvme_quiesce_io_queues(&dev->ctrl);
2646
2647         if (!dead && dev->ctrl.queue_count > 0) {
2648                 nvme_delete_io_queues(dev);
2649                 nvme_disable_ctrl(&dev->ctrl, shutdown);
2650                 nvme_poll_irqdisable(&dev->queues[0]);
2651         }
2652         nvme_suspend_io_queues(dev);
2653         nvme_suspend_queue(dev, 0);
2654         pci_free_irq_vectors(pdev);
2655         if (pci_is_enabled(pdev))
2656                 pci_disable_device(pdev);
2657         nvme_reap_pending_cqes(dev);
2658
2659         nvme_cancel_tagset(&dev->ctrl);
2660         nvme_cancel_admin_tagset(&dev->ctrl);
2661
2662         /*
2663          * The driver will not be starting up queues again if shutting down so
2664          * must flush all entered requests to their failed completion to avoid
2665          * deadlocking blk-mq hot-cpu notifier.
2666          */
2667         if (shutdown) {
2668                 nvme_unquiesce_io_queues(&dev->ctrl);
2669                 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2670                         nvme_unquiesce_admin_queue(&dev->ctrl);
2671         }
2672         mutex_unlock(&dev->shutdown_lock);
2673 }
2674
2675 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2676 {
2677         if (!nvme_wait_reset(&dev->ctrl))
2678                 return -EBUSY;
2679         nvme_dev_disable(dev, shutdown);
2680         return 0;
2681 }
2682
2683 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2684 {
2685         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2686                                                 NVME_CTRL_PAGE_SIZE,
2687                                                 NVME_CTRL_PAGE_SIZE, 0);
2688         if (!dev->prp_page_pool)
2689                 return -ENOMEM;
2690
2691         /* Optimisation for I/Os between 4k and 128k */
2692         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2693                                                 256, 256, 0);
2694         if (!dev->prp_small_pool) {
2695                 dma_pool_destroy(dev->prp_page_pool);
2696                 return -ENOMEM;
2697         }
2698         return 0;
2699 }
2700
2701 static void nvme_release_prp_pools(struct nvme_dev *dev)
2702 {
2703         dma_pool_destroy(dev->prp_page_pool);
2704         dma_pool_destroy(dev->prp_small_pool);
2705 }
2706
2707 static int nvme_pci_alloc_iod_mempool(struct nvme_dev *dev)
2708 {
2709         size_t alloc_size = sizeof(struct scatterlist) * NVME_MAX_SEGS;
2710
2711         dev->iod_mempool = mempool_create_node(1,
2712                         mempool_kmalloc, mempool_kfree,
2713                         (void *)alloc_size, GFP_KERNEL,
2714                         dev_to_node(dev->dev));
2715         if (!dev->iod_mempool)
2716                 return -ENOMEM;
2717         return 0;
2718 }
2719
2720 static void nvme_free_tagset(struct nvme_dev *dev)
2721 {
2722         if (dev->tagset.tags)
2723                 nvme_remove_io_tag_set(&dev->ctrl);
2724         dev->ctrl.tagset = NULL;
2725 }
2726
2727 /* pairs with nvme_pci_alloc_dev */
2728 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2729 {
2730         struct nvme_dev *dev = to_nvme_dev(ctrl);
2731
2732         nvme_free_tagset(dev);
2733         put_device(dev->dev);
2734         kfree(dev->queues);
2735         kfree(dev);
2736 }
2737
2738 static void nvme_reset_work(struct work_struct *work)
2739 {
2740         struct nvme_dev *dev =
2741                 container_of(work, struct nvme_dev, ctrl.reset_work);
2742         bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2743         int result;
2744
2745         if (nvme_ctrl_state(&dev->ctrl) != NVME_CTRL_RESETTING) {
2746                 dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
2747                          dev->ctrl.state);
2748                 result = -ENODEV;
2749                 goto out;
2750         }
2751
2752         /*
2753          * If we're called to reset a live controller first shut it down before
2754          * moving on.
2755          */
2756         if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2757                 nvme_dev_disable(dev, false);
2758         nvme_sync_queues(&dev->ctrl);
2759
2760         mutex_lock(&dev->shutdown_lock);
2761         result = nvme_pci_enable(dev);
2762         if (result)
2763                 goto out_unlock;
2764         nvme_unquiesce_admin_queue(&dev->ctrl);
2765         mutex_unlock(&dev->shutdown_lock);
2766
2767         /*
2768          * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2769          * initializing procedure here.
2770          */
2771         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2772                 dev_warn(dev->ctrl.device,
2773                         "failed to mark controller CONNECTING\n");
2774                 result = -EBUSY;
2775                 goto out;
2776         }
2777
2778         result = nvme_init_ctrl_finish(&dev->ctrl, was_suspend);
2779         if (result)
2780                 goto out;
2781
2782         nvme_dbbuf_dma_alloc(dev);
2783
2784         result = nvme_setup_host_mem(dev);
2785         if (result < 0)
2786                 goto out;
2787
2788         result = nvme_setup_io_queues(dev);
2789         if (result)
2790                 goto out;
2791
2792         /*
2793          * Freeze and update the number of I/O queues as thos might have
2794          * changed.  If there are no I/O queues left after this reset, keep the
2795          * controller around but remove all namespaces.
2796          */
2797         if (dev->online_queues > 1) {
2798                 nvme_dbbuf_set(dev);
2799                 nvme_unquiesce_io_queues(&dev->ctrl);
2800                 nvme_wait_freeze(&dev->ctrl);
2801                 nvme_pci_update_nr_queues(dev);
2802                 nvme_unfreeze(&dev->ctrl);
2803         } else {
2804                 dev_warn(dev->ctrl.device, "IO queues lost\n");
2805                 nvme_mark_namespaces_dead(&dev->ctrl);
2806                 nvme_unquiesce_io_queues(&dev->ctrl);
2807                 nvme_remove_namespaces(&dev->ctrl);
2808                 nvme_free_tagset(dev);
2809         }
2810
2811         /*
2812          * If only admin queue live, keep it to do further investigation or
2813          * recovery.
2814          */
2815         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2816                 dev_warn(dev->ctrl.device,
2817                         "failed to mark controller live state\n");
2818                 result = -ENODEV;
2819                 goto out;
2820         }
2821
2822         nvme_start_ctrl(&dev->ctrl);
2823         return;
2824
2825  out_unlock:
2826         mutex_unlock(&dev->shutdown_lock);
2827  out:
2828         /*
2829          * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2830          * may be holding this pci_dev's device lock.
2831          */
2832         dev_warn(dev->ctrl.device, "Disabling device after reset failure: %d\n",
2833                  result);
2834         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2835         nvme_dev_disable(dev, true);
2836         nvme_sync_queues(&dev->ctrl);
2837         nvme_mark_namespaces_dead(&dev->ctrl);
2838         nvme_unquiesce_io_queues(&dev->ctrl);
2839         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2840 }
2841
2842 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2843 {
2844         *val = readl(to_nvme_dev(ctrl)->bar + off);
2845         return 0;
2846 }
2847
2848 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2849 {
2850         writel(val, to_nvme_dev(ctrl)->bar + off);
2851         return 0;
2852 }
2853
2854 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2855 {
2856         *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2857         return 0;
2858 }
2859
2860 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2861 {
2862         struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2863
2864         return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2865 }
2866
2867 static void nvme_pci_print_device_info(struct nvme_ctrl *ctrl)
2868 {
2869         struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2870         struct nvme_subsystem *subsys = ctrl->subsys;
2871
2872         dev_err(ctrl->device,
2873                 "VID:DID %04x:%04x model:%.*s firmware:%.*s\n",
2874                 pdev->vendor, pdev->device,
2875                 nvme_strlen(subsys->model, sizeof(subsys->model)),
2876                 subsys->model, nvme_strlen(subsys->firmware_rev,
2877                                            sizeof(subsys->firmware_rev)),
2878                 subsys->firmware_rev);
2879 }
2880
2881 static bool nvme_pci_supports_pci_p2pdma(struct nvme_ctrl *ctrl)
2882 {
2883         struct nvme_dev *dev = to_nvme_dev(ctrl);
2884
2885         return dma_pci_p2pdma_supported(dev->dev);
2886 }
2887
2888 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2889         .name                   = "pcie",
2890         .module                 = THIS_MODULE,
2891         .flags                  = NVME_F_METADATA_SUPPORTED,
2892         .dev_attr_groups        = nvme_pci_dev_attr_groups,
2893         .reg_read32             = nvme_pci_reg_read32,
2894         .reg_write32            = nvme_pci_reg_write32,
2895         .reg_read64             = nvme_pci_reg_read64,
2896         .free_ctrl              = nvme_pci_free_ctrl,
2897         .submit_async_event     = nvme_pci_submit_async_event,
2898         .subsystem_reset        = nvme_pci_subsystem_reset,
2899         .get_address            = nvme_pci_get_address,
2900         .print_device_info      = nvme_pci_print_device_info,
2901         .supports_pci_p2pdma    = nvme_pci_supports_pci_p2pdma,
2902 };
2903
2904 static int nvme_dev_map(struct nvme_dev *dev)
2905 {
2906         struct pci_dev *pdev = to_pci_dev(dev->dev);
2907
2908         if (pci_request_mem_regions(pdev, "nvme"))
2909                 return -ENODEV;
2910
2911         if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2912                 goto release;
2913
2914         return 0;
2915   release:
2916         pci_release_mem_regions(pdev);
2917         return -ENODEV;
2918 }
2919
2920 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2921 {
2922         if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2923                 /*
2924                  * Several Samsung devices seem to drop off the PCIe bus
2925                  * randomly when APST is on and uses the deepest sleep state.
2926                  * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2927                  * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2928                  * 950 PRO 256GB", but it seems to be restricted to two Dell
2929                  * laptops.
2930                  */
2931                 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2932                     (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2933                      dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2934                         return NVME_QUIRK_NO_DEEPEST_PS;
2935         } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2936                 /*
2937                  * Samsung SSD 960 EVO drops off the PCIe bus after system
2938                  * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2939                  * within few minutes after bootup on a Coffee Lake board -
2940                  * ASUS PRIME Z370-A
2941                  */
2942                 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2943                     (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2944                      dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2945                         return NVME_QUIRK_NO_APST;
2946         } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2947                     pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2948                    (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2949                 /*
2950                  * Forcing to use host managed nvme power settings for
2951                  * lowest idle power with quick resume latency on
2952                  * Samsung and Toshiba SSDs based on suspend behavior
2953                  * on Coffee Lake board for LENOVO C640
2954                  */
2955                 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2956                      dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2957                         return NVME_QUIRK_SIMPLE_SUSPEND;
2958         } else if (pdev->vendor == 0x2646 && (pdev->device == 0x2263 ||
2959                    pdev->device == 0x500f)) {
2960                 /*
2961                  * Exclude some Kingston NV1 and A2000 devices from
2962                  * NVME_QUIRK_SIMPLE_SUSPEND. Do a full suspend to save a
2963                  * lot fo energy with s2idle sleep on some TUXEDO platforms.
2964                  */
2965                 if (dmi_match(DMI_BOARD_NAME, "NS5X_NS7XAU") ||
2966                     dmi_match(DMI_BOARD_NAME, "NS5x_7xAU") ||
2967                     dmi_match(DMI_BOARD_NAME, "NS5x_7xPU") ||
2968                     dmi_match(DMI_BOARD_NAME, "PH4PRX1_PH6PRX1"))
2969                         return NVME_QUIRK_FORCE_NO_SIMPLE_SUSPEND;
2970         }
2971
2972         /*
2973          * NVMe SSD drops off the PCIe bus after system idle
2974          * for 10 hours on a Lenovo N60z board.
2975          */
2976         if (dmi_match(DMI_BOARD_NAME, "LXKT-ZXEG-N6"))
2977                 return NVME_QUIRK_NO_APST;
2978
2979         return 0;
2980 }
2981
2982 static struct nvme_dev *nvme_pci_alloc_dev(struct pci_dev *pdev,
2983                 const struct pci_device_id *id)
2984 {
2985         unsigned long quirks = id->driver_data;
2986         int node = dev_to_node(&pdev->dev);
2987         struct nvme_dev *dev;
2988         int ret = -ENOMEM;
2989
2990         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2991         if (!dev)
2992                 return ERR_PTR(-ENOMEM);
2993         INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2994         mutex_init(&dev->shutdown_lock);
2995
2996         dev->nr_write_queues = write_queues;
2997         dev->nr_poll_queues = poll_queues;
2998         dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2999         dev->queues = kcalloc_node(dev->nr_allocated_queues,
3000                         sizeof(struct nvme_queue), GFP_KERNEL, node);
3001         if (!dev->queues)
3002                 goto out_free_dev;
3003
3004         dev->dev = get_device(&pdev->dev);
3005
3006         quirks |= check_vendor_combination_bug(pdev);
3007         if (!noacpi &&
3008             !(quirks & NVME_QUIRK_FORCE_NO_SIMPLE_SUSPEND) &&
3009             acpi_storage_d3(&pdev->dev)) {
3010                 /*
3011                  * Some systems use a bios work around to ask for D3 on
3012                  * platforms that support kernel managed suspend.
3013                  */
3014                 dev_info(&pdev->dev,
3015                          "platform quirk: setting simple suspend\n");
3016                 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
3017         }
3018         ret = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
3019                              quirks);
3020         if (ret)
3021                 goto out_put_device;
3022
3023         if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48)
3024                 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(48));
3025         else
3026                 dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
3027         dma_set_min_align_mask(&pdev->dev, NVME_CTRL_PAGE_SIZE - 1);
3028         dma_set_max_seg_size(&pdev->dev, 0xffffffff);
3029
3030         /*
3031          * Limit the max command size to prevent iod->sg allocations going
3032          * over a single page.
3033          */
3034         dev->ctrl.max_hw_sectors = min_t(u32,
3035                 NVME_MAX_KB_SZ << 1, dma_opt_mapping_size(&pdev->dev) >> 9);
3036         dev->ctrl.max_segments = NVME_MAX_SEGS;
3037
3038         /*
3039          * There is no support for SGLs for metadata (yet), so we are limited to
3040          * a single integrity segment for the separate metadata pointer.
3041          */
3042         dev->ctrl.max_integrity_segments = 1;
3043         return dev;
3044
3045 out_put_device:
3046         put_device(dev->dev);
3047         kfree(dev->queues);
3048 out_free_dev:
3049         kfree(dev);
3050         return ERR_PTR(ret);
3051 }
3052
3053 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
3054 {
3055         struct nvme_dev *dev;
3056         int result = -ENOMEM;
3057
3058         dev = nvme_pci_alloc_dev(pdev, id);
3059         if (IS_ERR(dev))
3060                 return PTR_ERR(dev);
3061
3062         result = nvme_add_ctrl(&dev->ctrl);
3063         if (result)
3064                 goto out_put_ctrl;
3065
3066         result = nvme_dev_map(dev);
3067         if (result)
3068                 goto out_uninit_ctrl;
3069
3070         result = nvme_setup_prp_pools(dev);
3071         if (result)
3072                 goto out_dev_unmap;
3073
3074         result = nvme_pci_alloc_iod_mempool(dev);
3075         if (result)
3076                 goto out_release_prp_pools;
3077
3078         dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
3079
3080         result = nvme_pci_enable(dev);
3081         if (result)
3082                 goto out_release_iod_mempool;
3083
3084         result = nvme_alloc_admin_tag_set(&dev->ctrl, &dev->admin_tagset,
3085                                 &nvme_mq_admin_ops, sizeof(struct nvme_iod));
3086         if (result)
3087                 goto out_disable;
3088
3089         /*
3090          * Mark the controller as connecting before sending admin commands to
3091          * allow the timeout handler to do the right thing.
3092          */
3093         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
3094                 dev_warn(dev->ctrl.device,
3095                         "failed to mark controller CONNECTING\n");
3096                 result = -EBUSY;
3097                 goto out_disable;
3098         }
3099
3100         result = nvme_init_ctrl_finish(&dev->ctrl, false);
3101         if (result)
3102                 goto out_disable;
3103
3104         nvme_dbbuf_dma_alloc(dev);
3105
3106         result = nvme_setup_host_mem(dev);
3107         if (result < 0)
3108                 goto out_disable;
3109
3110         result = nvme_setup_io_queues(dev);
3111         if (result)
3112                 goto out_disable;
3113
3114         if (dev->online_queues > 1) {
3115                 nvme_alloc_io_tag_set(&dev->ctrl, &dev->tagset, &nvme_mq_ops,
3116                                 nvme_pci_nr_maps(dev), sizeof(struct nvme_iod));
3117                 nvme_dbbuf_set(dev);
3118         }
3119
3120         if (!dev->ctrl.tagset)
3121                 dev_warn(dev->ctrl.device, "IO queues not created\n");
3122
3123         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
3124                 dev_warn(dev->ctrl.device,
3125                         "failed to mark controller live state\n");
3126                 result = -ENODEV;
3127                 goto out_disable;
3128         }
3129
3130         pci_set_drvdata(pdev, dev);
3131
3132         nvme_start_ctrl(&dev->ctrl);
3133         nvme_put_ctrl(&dev->ctrl);
3134         flush_work(&dev->ctrl.scan_work);
3135         return 0;
3136
3137 out_disable:
3138         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3139         nvme_dev_disable(dev, true);
3140         nvme_free_host_mem(dev);
3141         nvme_dev_remove_admin(dev);
3142         nvme_dbbuf_dma_free(dev);
3143         nvme_free_queues(dev, 0);
3144 out_release_iod_mempool:
3145         mempool_destroy(dev->iod_mempool);
3146 out_release_prp_pools:
3147         nvme_release_prp_pools(dev);
3148 out_dev_unmap:
3149         nvme_dev_unmap(dev);
3150 out_uninit_ctrl:
3151         nvme_uninit_ctrl(&dev->ctrl);
3152 out_put_ctrl:
3153         nvme_put_ctrl(&dev->ctrl);
3154         return result;
3155 }
3156
3157 static void nvme_reset_prepare(struct pci_dev *pdev)
3158 {
3159         struct nvme_dev *dev = pci_get_drvdata(pdev);
3160
3161         /*
3162          * We don't need to check the return value from waiting for the reset
3163          * state as pci_dev device lock is held, making it impossible to race
3164          * with ->remove().
3165          */
3166         nvme_disable_prepare_reset(dev, false);
3167         nvme_sync_queues(&dev->ctrl);
3168 }
3169
3170 static void nvme_reset_done(struct pci_dev *pdev)
3171 {
3172         struct nvme_dev *dev = pci_get_drvdata(pdev);
3173
3174         if (!nvme_try_sched_reset(&dev->ctrl))
3175                 flush_work(&dev->ctrl.reset_work);
3176 }
3177
3178 static void nvme_shutdown(struct pci_dev *pdev)
3179 {
3180         struct nvme_dev *dev = pci_get_drvdata(pdev);
3181
3182         nvme_disable_prepare_reset(dev, true);
3183 }
3184
3185 /*
3186  * The driver's remove may be called on a device in a partially initialized
3187  * state. This function must not have any dependencies on the device state in
3188  * order to proceed.
3189  */
3190 static void nvme_remove(struct pci_dev *pdev)
3191 {
3192         struct nvme_dev *dev = pci_get_drvdata(pdev);
3193
3194         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3195         pci_set_drvdata(pdev, NULL);
3196
3197         if (!pci_device_is_present(pdev)) {
3198                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3199                 nvme_dev_disable(dev, true);
3200         }
3201
3202         flush_work(&dev->ctrl.reset_work);
3203         nvme_stop_ctrl(&dev->ctrl);
3204         nvme_remove_namespaces(&dev->ctrl);
3205         nvme_dev_disable(dev, true);
3206         nvme_free_host_mem(dev);
3207         nvme_dev_remove_admin(dev);
3208         nvme_dbbuf_dma_free(dev);
3209         nvme_free_queues(dev, 0);
3210         mempool_destroy(dev->iod_mempool);
3211         nvme_release_prp_pools(dev);
3212         nvme_dev_unmap(dev);
3213         nvme_uninit_ctrl(&dev->ctrl);
3214 }
3215
3216 #ifdef CONFIG_PM_SLEEP
3217 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3218 {
3219         return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3220 }
3221
3222 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3223 {
3224         return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3225 }
3226
3227 static int nvme_resume(struct device *dev)
3228 {
3229         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3230         struct nvme_ctrl *ctrl = &ndev->ctrl;
3231
3232         if (ndev->last_ps == U32_MAX ||
3233             nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3234                 goto reset;
3235         if (ctrl->hmpre && nvme_setup_host_mem(ndev))
3236                 goto reset;
3237
3238         return 0;
3239 reset:
3240         return nvme_try_sched_reset(ctrl);
3241 }
3242
3243 static int nvme_suspend(struct device *dev)
3244 {
3245         struct pci_dev *pdev = to_pci_dev(dev);
3246         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3247         struct nvme_ctrl *ctrl = &ndev->ctrl;
3248         int ret = -EBUSY;
3249
3250         ndev->last_ps = U32_MAX;
3251
3252         /*
3253          * The platform does not remove power for a kernel managed suspend so
3254          * use host managed nvme power settings for lowest idle power if
3255          * possible. This should have quicker resume latency than a full device
3256          * shutdown.  But if the firmware is involved after the suspend or the
3257          * device does not support any non-default power states, shut down the
3258          * device fully.
3259          *
3260          * If ASPM is not enabled for the device, shut down the device and allow
3261          * the PCI bus layer to put it into D3 in order to take the PCIe link
3262          * down, so as to allow the platform to achieve its minimum low-power
3263          * state (which may not be possible if the link is up).
3264          */
3265         if (pm_suspend_via_firmware() || !ctrl->npss ||
3266             !pcie_aspm_enabled(pdev) ||
3267             (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3268                 return nvme_disable_prepare_reset(ndev, true);
3269
3270         nvme_start_freeze(ctrl);
3271         nvme_wait_freeze(ctrl);
3272         nvme_sync_queues(ctrl);
3273
3274         if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
3275                 goto unfreeze;
3276
3277         /*
3278          * Host memory access may not be successful in a system suspend state,
3279          * but the specification allows the controller to access memory in a
3280          * non-operational power state.
3281          */
3282         if (ndev->hmb) {
3283                 ret = nvme_set_host_mem(ndev, 0);
3284                 if (ret < 0)
3285                         goto unfreeze;
3286         }
3287
3288         ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3289         if (ret < 0)
3290                 goto unfreeze;
3291
3292         /*
3293          * A saved state prevents pci pm from generically controlling the
3294          * device's power. If we're using protocol specific settings, we don't
3295          * want pci interfering.
3296          */
3297         pci_save_state(pdev);
3298
3299         ret = nvme_set_power_state(ctrl, ctrl->npss);
3300         if (ret < 0)
3301                 goto unfreeze;
3302
3303         if (ret) {
3304                 /* discard the saved state */
3305                 pci_load_saved_state(pdev, NULL);
3306
3307                 /*
3308                  * Clearing npss forces a controller reset on resume. The
3309                  * correct value will be rediscovered then.
3310                  */
3311                 ret = nvme_disable_prepare_reset(ndev, true);
3312                 ctrl->npss = 0;
3313         }
3314 unfreeze:
3315         nvme_unfreeze(ctrl);
3316         return ret;
3317 }
3318
3319 static int nvme_simple_suspend(struct device *dev)
3320 {
3321         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3322
3323         return nvme_disable_prepare_reset(ndev, true);
3324 }
3325
3326 static int nvme_simple_resume(struct device *dev)
3327 {
3328         struct pci_dev *pdev = to_pci_dev(dev);
3329         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3330
3331         return nvme_try_sched_reset(&ndev->ctrl);
3332 }
3333
3334 static const struct dev_pm_ops nvme_dev_pm_ops = {
3335         .suspend        = nvme_suspend,
3336         .resume         = nvme_resume,
3337         .freeze         = nvme_simple_suspend,
3338         .thaw           = nvme_simple_resume,
3339         .poweroff       = nvme_simple_suspend,
3340         .restore        = nvme_simple_resume,
3341 };
3342 #endif /* CONFIG_PM_SLEEP */
3343
3344 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3345                                                 pci_channel_state_t state)
3346 {
3347         struct nvme_dev *dev = pci_get_drvdata(pdev);
3348
3349         /*
3350          * A frozen channel requires a reset. When detected, this method will
3351          * shutdown the controller to quiesce. The controller will be restarted
3352          * after the slot reset through driver's slot_reset callback.
3353          */
3354         switch (state) {
3355         case pci_channel_io_normal:
3356                 return PCI_ERS_RESULT_CAN_RECOVER;
3357         case pci_channel_io_frozen:
3358                 dev_warn(dev->ctrl.device,
3359                         "frozen state error detected, reset controller\n");
3360                 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING)) {
3361                         nvme_dev_disable(dev, true);
3362                         return PCI_ERS_RESULT_DISCONNECT;
3363                 }
3364                 nvme_dev_disable(dev, false);
3365                 return PCI_ERS_RESULT_NEED_RESET;
3366         case pci_channel_io_perm_failure:
3367                 dev_warn(dev->ctrl.device,
3368                         "failure state error detected, request disconnect\n");
3369                 return PCI_ERS_RESULT_DISCONNECT;
3370         }
3371         return PCI_ERS_RESULT_NEED_RESET;
3372 }
3373
3374 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3375 {
3376         struct nvme_dev *dev = pci_get_drvdata(pdev);
3377
3378         dev_info(dev->ctrl.device, "restart after slot reset\n");
3379         pci_restore_state(pdev);
3380         if (!nvme_try_sched_reset(&dev->ctrl))
3381                 nvme_unquiesce_io_queues(&dev->ctrl);
3382         return PCI_ERS_RESULT_RECOVERED;
3383 }
3384
3385 static void nvme_error_resume(struct pci_dev *pdev)
3386 {
3387         struct nvme_dev *dev = pci_get_drvdata(pdev);
3388
3389         flush_work(&dev->ctrl.reset_work);
3390 }
3391
3392 static const struct pci_error_handlers nvme_err_handler = {
3393         .error_detected = nvme_error_detected,
3394         .slot_reset     = nvme_slot_reset,
3395         .resume         = nvme_error_resume,
3396         .reset_prepare  = nvme_reset_prepare,
3397         .reset_done     = nvme_reset_done,
3398 };
3399
3400 static const struct pci_device_id nvme_id_table[] = {
3401         { PCI_VDEVICE(INTEL, 0x0953),   /* Intel 750/P3500/P3600/P3700 */
3402                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3403                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3404         { PCI_VDEVICE(INTEL, 0x0a53),   /* Intel P3520 */
3405                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3406                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3407         { PCI_VDEVICE(INTEL, 0x0a54),   /* Intel P4500/P4600 */
3408                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3409                                 NVME_QUIRK_DEALLOCATE_ZEROES |
3410                                 NVME_QUIRK_IGNORE_DEV_SUBNQN |
3411                                 NVME_QUIRK_BOGUS_NID, },
3412         { PCI_VDEVICE(INTEL, 0x0a55),   /* Dell Express Flash P4600 */
3413                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3414                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3415         { PCI_VDEVICE(INTEL, 0xf1a5),   /* Intel 600P/P3100 */
3416                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3417                                 NVME_QUIRK_MEDIUM_PRIO_SQ |
3418                                 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3419                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3420         { PCI_VDEVICE(INTEL, 0xf1a6),   /* Intel 760p/Pro 7600p */
3421                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3422         { PCI_VDEVICE(INTEL, 0x5845),   /* Qemu emulated controller */
3423                 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3424                                 NVME_QUIRK_DISABLE_WRITE_ZEROES |
3425                                 NVME_QUIRK_BOGUS_NID, },
3426         { PCI_VDEVICE(REDHAT, 0x0010),  /* Qemu emulated controller */
3427                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3428         { PCI_DEVICE(0x126f, 0x2262),   /* Silicon Motion generic */
3429                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3430                                 NVME_QUIRK_BOGUS_NID, },
3431         { PCI_DEVICE(0x126f, 0x2263),   /* Silicon Motion unidentified */
3432                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3433                                 NVME_QUIRK_BOGUS_NID, },
3434         { PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3435                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3436                                 NVME_QUIRK_NO_NS_DESC_LIST, },
3437         { PCI_DEVICE(0x1c58, 0x0003),   /* HGST adapter */
3438                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3439         { PCI_DEVICE(0x1c58, 0x0023),   /* WDC SN200 adapter */
3440                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3441         { PCI_DEVICE(0x1c5f, 0x0540),   /* Memblaze Pblaze4 adapter */
3442                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3443         { PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3444                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3445         { PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3446                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3447                                 NVME_QUIRK_DISABLE_WRITE_ZEROES|
3448                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3449         { PCI_DEVICE(0x15b7, 0x5008),   /* Sandisk SN530 */
3450                 .driver_data = NVME_QUIRK_BROKEN_MSI },
3451         { PCI_DEVICE(0x1987, 0x5012),   /* Phison E12 */
3452                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3453         { PCI_DEVICE(0x1987, 0x5016),   /* Phison E16 */
3454                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
3455                                 NVME_QUIRK_BOGUS_NID, },
3456         { PCI_DEVICE(0x1987, 0x5019),  /* phison E19 */
3457                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3458         { PCI_DEVICE(0x1987, 0x5021),   /* Phison E21 */
3459                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3460         { PCI_DEVICE(0x1b4b, 0x1092),   /* Lexar 256 GB SSD */
3461                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3462                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3463         { PCI_DEVICE(0x1cc1, 0x33f8),   /* ADATA IM2P33F8ABR1 1 TB */
3464                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3465         { PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3466                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
3467                                 NVME_QUIRK_BOGUS_NID, },
3468         { PCI_DEVICE(0x10ec, 0x5763),  /* ADATA SX6000PNP */
3469                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3470         { PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3471                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3472                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3473          { PCI_DEVICE(0x1344, 0x5407), /* Micron Technology Inc NVMe SSD */
3474                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN },
3475          { PCI_DEVICE(0x1344, 0x6001),   /* Micron Nitro NVMe */
3476                  .driver_data = NVME_QUIRK_BOGUS_NID, },
3477         { PCI_DEVICE(0x1c5c, 0x1504),   /* SK Hynix PC400 */
3478                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3479         { PCI_DEVICE(0x1c5c, 0x174a),   /* SK Hynix P31 SSD */
3480                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3481         { PCI_DEVICE(0x1c5c, 0x1D59),   /* SK Hynix BC901 */
3482                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3483         { PCI_DEVICE(0x15b7, 0x2001),   /*  Sandisk Skyhawk */
3484                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3485         { PCI_DEVICE(0x1d97, 0x2263),   /* SPCC */
3486                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3487         { PCI_DEVICE(0x144d, 0xa80b),   /* Samsung PM9B1 256G and 512G */
3488                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES |
3489                                 NVME_QUIRK_BOGUS_NID, },
3490         { PCI_DEVICE(0x144d, 0xa809),   /* Samsung MZALQ256HBJD 256G */
3491                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3492         { PCI_DEVICE(0x144d, 0xa802),   /* Samsung SM953 */
3493                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3494         { PCI_DEVICE(0x1cc4, 0x6303),   /* UMIS RPJTJ512MGE1QDY 512G */
3495                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3496         { PCI_DEVICE(0x1cc4, 0x6302),   /* UMIS RPJTJ256MGE1QDY 256G */
3497                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3498         { PCI_DEVICE(0x2646, 0x2262),   /* KINGSTON SKC2000 NVMe SSD */
3499                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3500         { PCI_DEVICE(0x2646, 0x2263),   /* KINGSTON A2000 NVMe SSD  */
3501                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3502         { PCI_DEVICE(0x2646, 0x5013),   /* Kingston KC3000, Kingston FURY Renegade */
3503                 .driver_data = NVME_QUIRK_NO_SECONDARY_TEMP_THRESH, },
3504         { PCI_DEVICE(0x2646, 0x5018),   /* KINGSTON OM8SFP4xxxxP OS21012 NVMe SSD */
3505                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3506         { PCI_DEVICE(0x2646, 0x5016),   /* KINGSTON OM3PGP4xxxxP OS21011 NVMe SSD */
3507                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3508         { PCI_DEVICE(0x2646, 0x501A),   /* KINGSTON OM8PGP4xxxxP OS21005 NVMe SSD */
3509                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3510         { PCI_DEVICE(0x2646, 0x501B),   /* KINGSTON OM8PGP4xxxxQ OS21005 NVMe SSD */
3511                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3512         { PCI_DEVICE(0x2646, 0x501E),   /* KINGSTON OM3PGP4xxxxQ OS21011 NVMe SSD */
3513                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3514         { PCI_DEVICE(0x1f40, 0x1202),   /* Netac Technologies Co. NV3000 NVMe SSD */
3515                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3516         { PCI_DEVICE(0x1f40, 0x5236),   /* Netac Technologies Co. NV7000 NVMe SSD */
3517                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3518         { PCI_DEVICE(0x1e4B, 0x1001),   /* MAXIO MAP1001 */
3519                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3520         { PCI_DEVICE(0x1e4B, 0x1002),   /* MAXIO MAP1002 */
3521                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3522         { PCI_DEVICE(0x1e4B, 0x1202),   /* MAXIO MAP1202 */
3523                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3524         { PCI_DEVICE(0x1e4B, 0x1602),   /* MAXIO MAP1602 */
3525                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3526         { PCI_DEVICE(0x1cc1, 0x5350),   /* ADATA XPG GAMMIX S50 */
3527                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3528         { PCI_DEVICE(0x1dbe, 0x5236),   /* ADATA XPG GAMMIX S70 */
3529                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3530         { PCI_DEVICE(0x1e49, 0x0021),   /* ZHITAI TiPro5000 NVMe SSD */
3531                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3532         { PCI_DEVICE(0x1e49, 0x0041),   /* ZHITAI TiPro7000 NVMe SSD */
3533                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3534         { PCI_DEVICE(0xc0a9, 0x540a),   /* Crucial P2 */
3535                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3536         { PCI_DEVICE(0x1d97, 0x2263), /* Lexar NM610 */
3537                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3538         { PCI_DEVICE(0x1d97, 0x1d97), /* Lexar NM620 */
3539                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3540         { PCI_DEVICE(0x1d97, 0x2269), /* Lexar NM760 */
3541                 .driver_data = NVME_QUIRK_BOGUS_NID |
3542                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3543         { PCI_DEVICE(0x10ec, 0x5763), /* TEAMGROUP T-FORCE CARDEA ZERO Z330 SSD */
3544                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3545         { PCI_DEVICE(0x1e4b, 0x1602), /* HS-SSD-FUTURE 2048G  */
3546                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3547         { PCI_DEVICE(0x10ec, 0x5765), /* TEAMGROUP MP33 2TB SSD */
3548                 .driver_data = NVME_QUIRK_BOGUS_NID, },
3549         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061),
3550                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3551         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065),
3552                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3553         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061),
3554                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3555         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00),
3556                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3557         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01),
3558                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3559         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02),
3560                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3561         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3562                 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3563         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3564         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3565                 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3566                                 NVME_QUIRK_128_BYTES_SQES |
3567                                 NVME_QUIRK_SHARED_TAGS |
3568                                 NVME_QUIRK_SKIP_CID_GEN |
3569                                 NVME_QUIRK_IDENTIFY_CNS },
3570         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3571         { 0, }
3572 };
3573 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3574
3575 static struct pci_driver nvme_driver = {
3576         .name           = "nvme",
3577         .id_table       = nvme_id_table,
3578         .probe          = nvme_probe,
3579         .remove         = nvme_remove,
3580         .shutdown       = nvme_shutdown,
3581         .driver         = {
3582                 .probe_type     = PROBE_PREFER_ASYNCHRONOUS,
3583 #ifdef CONFIG_PM_SLEEP
3584                 .pm             = &nvme_dev_pm_ops,
3585 #endif
3586         },
3587         .sriov_configure = pci_sriov_configure_simple,
3588         .err_handler    = &nvme_err_handler,
3589 };
3590
3591 static int __init nvme_init(void)
3592 {
3593         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3594         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3595         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3596         BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3597         BUILD_BUG_ON(NVME_MAX_SEGS > SGES_PER_PAGE);
3598         BUILD_BUG_ON(sizeof(struct scatterlist) * NVME_MAX_SEGS > PAGE_SIZE);
3599         BUILD_BUG_ON(nvme_pci_npages_prp() > NVME_MAX_NR_ALLOCATIONS);
3600
3601         return pci_register_driver(&nvme_driver);
3602 }
3603
3604 static void __exit nvme_exit(void)
3605 {
3606         pci_unregister_driver(&nvme_driver);
3607         flush_workqueue(nvme_wq);
3608 }
3609
3610 MODULE_AUTHOR("Matthew Wilcox <[email protected]>");
3611 MODULE_LICENSE("GPL");
3612 MODULE_VERSION("1.0");
3613 MODULE_DESCRIPTION("NVMe host PCIe transport driver");
3614 module_init(nvme_init);
3615 module_exit(nvme_exit);