drivers/nvme/host/pci.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVM Express device driver
   4  * Copyright (c) 2011-2014, Intel Corporation.
   5  */
   6
   7 #include <linux/acpi.h>
   8 #include <linux/aer.h>
   9 #include <linux/async.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/blk-mq.h>
  12 #include <linux/blk-mq-pci.h>
  13 #include <linux/dmi.h>
  14 #include <linux/init.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/io.h>
  17 #include <linux/mm.h>
  18 #include <linux/module.h>
  19 #include <linux/mutex.h>
  20 #include <linux/once.h>
  21 #include <linux/pci.h>
  22 #include <linux/suspend.h>
  23 #include <linux/t10-pi.h>
  24 #include <linux/types.h>
  25 #include <linux/io-64-nonatomic-lo-hi.h>
  26 #include <linux/io-64-nonatomic-hi-lo.h>
  27 #include <linux/sed-opal.h>
  28 #include <linux/pci-p2pdma.h>
  29
  30 #include "trace.h"
  31 #include "nvme.h"
  32
  33 #define SQ_SIZE(q)      ((q)->q_depth << (q)->sqes)
  34 #define CQ_SIZE(q)      ((q)->q_depth * sizeof(struct nvme_completion))
  35
  36 #define SGES_PER_PAGE   (PAGE_SIZE / sizeof(struct nvme_sgl_desc))
  37
  38 /*
  39  * These can be higher, but we need to ensure that any command doesn't
  40  * require an sg allocation that needs more than a page of data.
  41  */
  42 #define NVME_MAX_KB_SZ  4096
  43 #define NVME_MAX_SEGS   127
  44
  45 static int use_threaded_interrupts;
  46 module_param(use_threaded_interrupts, int, 0);
  47
  48 static bool use_cmb_sqes = true;
  49 module_param(use_cmb_sqes, bool, 0444);
  50 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
  51
  52 static unsigned int max_host_mem_size_mb = 128;
  53 module_param(max_host_mem_size_mb, uint, 0444);
  54 MODULE_PARM_DESC(max_host_mem_size_mb,
  55         "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
  56
  57 static unsigned int sgl_threshold = SZ_32K;
  58 module_param(sgl_threshold, uint, 0644);
  59 MODULE_PARM_DESC(sgl_threshold,
  60                 "Use SGLs when average request segment size is larger or equal to "
  61                 "this size. Use 0 to disable SGLs.");
  62
  63 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
  64 static const struct kernel_param_ops io_queue_depth_ops = {
  65         .set = io_queue_depth_set,
  66         .get = param_get_uint,
  67 };
  68
  69 static unsigned int io_queue_depth = 1024;
  70 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
  71 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
  72
  73 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
  74 {
  75         unsigned int n;
  76         int ret;
  77
  78         ret = kstrtouint(val, 10, &n);
  79         if (ret != 0 || n > num_possible_cpus())
  80                 return -EINVAL;
  81         return param_set_uint(val, kp);
  82 }
  83
  84 static const struct kernel_param_ops io_queue_count_ops = {
  85         .set = io_queue_count_set,
  86         .get = param_get_uint,
  87 };
  88
  89 static unsigned int write_queues;
  90 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
  91 MODULE_PARM_DESC(write_queues,
  92         "Number of queues to use for writes. If not set, reads and writes "
  93         "will share a queue set.");
  94
  95 static unsigned int poll_queues;
  96 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
  97 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
  98
  99 static bool noacpi;
 100 module_param(noacpi, bool, 0444);
 101 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
 102
 103 struct nvme_dev;
 104 struct nvme_queue;
 105
 106 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
 107 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
 108
 109 /*
 110  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
 111  */
 112 struct nvme_dev {
 113         struct nvme_queue *queues;
 114         struct blk_mq_tag_set tagset;
 115         struct blk_mq_tag_set admin_tagset;
 116         u32 __iomem *dbs;
 117         struct device *dev;
 118         struct dma_pool *prp_page_pool;
 119         struct dma_pool *prp_small_pool;
 120         unsigned online_queues;
 121         unsigned max_qid;
 122         unsigned io_queues[HCTX_MAX_TYPES];
 123         unsigned int num_vecs;
 124         u32 q_depth;
 125         int io_sqes;
 126         u32 db_stride;
 127         void __iomem *bar;
 128         unsigned long bar_mapped_size;
 129         struct work_struct remove_work;
 130         struct mutex shutdown_lock;
 131         bool subsystem;
 132         u64 cmb_size;
 133         bool cmb_use_sqes;
 134         u32 cmbsz;
 135         u32 cmbloc;
 136         struct nvme_ctrl ctrl;
 137         u32 last_ps;
 138
 139         mempool_t *iod_mempool;
 140
 141         /* shadow doorbell buffer support: */
 142         u32 *dbbuf_dbs;
 143         dma_addr_t dbbuf_dbs_dma_addr;
 144         u32 *dbbuf_eis;
 145         dma_addr_t dbbuf_eis_dma_addr;
 146
 147         /* host memory buffer support: */
 148         u64 host_mem_size;
 149         u32 nr_host_mem_descs;
 150         dma_addr_t host_mem_descs_dma;
 151         struct nvme_host_mem_buf_desc *host_mem_descs;
 152         void **host_mem_desc_bufs;
 153         unsigned int nr_allocated_queues;
 154         unsigned int nr_write_queues;
 155         unsigned int nr_poll_queues;
 156 };
 157
 158 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
 159 {
 160         int ret;
 161         u32 n;
 162
 163         ret = kstrtou32(val, 10, &n);
 164         if (ret != 0 || n < 2)
 165                 return -EINVAL;
 166
 167         return param_set_uint(val, kp);
 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         u32 *dbbuf_sq_db;
 213         u32 *dbbuf_cq_db;
 214         u32 *dbbuf_sq_ei;
 215         u32 *dbbuf_cq_ei;
 216         struct completion delete_done;
 217 };
 218
 219 /*
 220  * The nvme_iod describes the data in an I/O.
 221  *
 222  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
 223  * to the actual struct scatterlist.
 224  */
 225 struct nvme_iod {
 226         struct nvme_request req;
 227         struct nvme_command cmd;
 228         struct nvme_queue *nvmeq;
 229         bool use_sgl;
 230         int aborted;
 231         int npages;             /* In the PRP list. 0 means small pool in use */
 232         int nents;              /* Used in scatterlist */
 233         dma_addr_t first_dma;
 234         unsigned int dma_len;   /* length of single DMA segment mapping */
 235         dma_addr_t meta_dma;
 236         struct scatterlist *sg;
 237 };
 238
 239 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
 240 {
 241         return dev->nr_allocated_queues * 8 * dev->db_stride;
 242 }
 243
 244 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
 245 {
 246         unsigned int mem_size = nvme_dbbuf_size(dev);
 247
 248         if (dev->dbbuf_dbs)
 249                 return 0;
 250
 251         dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
 252                                             &dev->dbbuf_dbs_dma_addr,
 253                                             GFP_KERNEL);
 254         if (!dev->dbbuf_dbs)
 255                 return -ENOMEM;
 256         dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
 257                                             &dev->dbbuf_eis_dma_addr,
 258                                             GFP_KERNEL);
 259         if (!dev->dbbuf_eis) {
 260                 dma_free_coherent(dev->dev, mem_size,
 261                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 262                 dev->dbbuf_dbs = NULL;
 263                 return -ENOMEM;
 264         }
 265
 266         return 0;
 267 }
 268
 269 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
 270 {
 271         unsigned int mem_size = nvme_dbbuf_size(dev);
 272
 273         if (dev->dbbuf_dbs) {
 274                 dma_free_coherent(dev->dev, mem_size,
 275                                   dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
 276                 dev->dbbuf_dbs = NULL;
 277         }
 278         if (dev->dbbuf_eis) {
 279                 dma_free_coherent(dev->dev, mem_size,
 280                                   dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
 281                 dev->dbbuf_eis = NULL;
 282         }
 283 }
 284
 285 static void nvme_dbbuf_init(struct nvme_dev *dev,
 286                             struct nvme_queue *nvmeq, int qid)
 287 {
 288         if (!dev->dbbuf_dbs || !qid)
 289                 return;
 290
 291         nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
 292         nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
 293         nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
 294         nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
 295 }
 296
 297 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
 298 {
 299         if (!nvmeq->qid)
 300                 return;
 301
 302         nvmeq->dbbuf_sq_db = NULL;
 303         nvmeq->dbbuf_cq_db = NULL;
 304         nvmeq->dbbuf_sq_ei = NULL;
 305         nvmeq->dbbuf_cq_ei = NULL;
 306 }
 307
 308 static void nvme_dbbuf_set(struct nvme_dev *dev)
 309 {
 310         struct nvme_command c = { };
 311         unsigned int i;
 312
 313         if (!dev->dbbuf_dbs)
 314                 return;
 315
 316         c.dbbuf.opcode = nvme_admin_dbbuf;
 317         c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
 318         c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
 319
 320         if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
 321                 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
 322                 /* Free memory and continue on */
 323                 nvme_dbbuf_dma_free(dev);
 324
 325                 for (i = 1; i <= dev->online_queues; i++)
 326                         nvme_dbbuf_free(&dev->queues[i]);
 327         }
 328 }
 329
 330 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
 331 {
 332         return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
 333 }
 334
 335 /* Update dbbuf and return true if an MMIO is required */
 336 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
 337                                               volatile u32 *dbbuf_ei)
 338 {
 339         if (dbbuf_db) {
 340                 u16 old_value;
 341
 342                 /*
 343                  * Ensure that the queue is written before updating
 344                  * the doorbell in memory
 345                  */
 346                 wmb();
 347
 348                 old_value = *dbbuf_db;
 349                 *dbbuf_db = value;
 350
 351                 /*
 352                  * Ensure that the doorbell is updated before reading the event
 353                  * index from memory.  The controller needs to provide similar
 354                  * ordering to ensure the envent index is updated before reading
 355                  * the doorbell.
 356                  */
 357                 mb();
 358
 359                 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
 360                         return false;
 361         }
 362
 363         return true;
 364 }
 365
 366 /*
 367  * Will slightly overestimate the number of pages needed.  This is OK
 368  * as it only leads to a small amount of wasted memory for the lifetime of
 369  * the I/O.
 370  */
 371 static int nvme_pci_npages_prp(void)
 372 {
 373         unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
 374                                       NVME_CTRL_PAGE_SIZE);
 375         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 376 }
 377
 378 /*
 379  * Calculates the number of pages needed for the SGL segments. For example a 4k
 380  * page can accommodate 256 SGL descriptors.
 381  */
 382 static int nvme_pci_npages_sgl(void)
 383 {
 384         return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
 385                         PAGE_SIZE);
 386 }
 387
 388 static size_t nvme_pci_iod_alloc_size(void)
 389 {
 390         size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
 391
 392         return sizeof(__le64 *) * npages +
 393                 sizeof(struct scatterlist) * NVME_MAX_SEGS;
 394 }
 395
 396 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 397                                 unsigned int hctx_idx)
 398 {
 399         struct nvme_dev *dev = data;
 400         struct nvme_queue *nvmeq = &dev->queues[0];
 401
 402         WARN_ON(hctx_idx != 0);
 403         WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
 404
 405         hctx->driver_data = nvmeq;
 406         return 0;
 407 }
 408
 409 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 410                           unsigned int hctx_idx)
 411 {
 412         struct nvme_dev *dev = data;
 413         struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
 414
 415         WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
 416         hctx->driver_data = nvmeq;
 417         return 0;
 418 }
 419
 420 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
 421                 unsigned int hctx_idx, unsigned int numa_node)
 422 {
 423         struct nvme_dev *dev = set->driver_data;
 424         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 425         int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
 426         struct nvme_queue *nvmeq = &dev->queues[queue_idx];
 427
 428         BUG_ON(!nvmeq);
 429         iod->nvmeq = nvmeq;
 430
 431         nvme_req(req)->ctrl = &dev->ctrl;
 432         nvme_req(req)->cmd = &iod->cmd;
 433         return 0;
 434 }
 435
 436 static int queue_irq_offset(struct nvme_dev *dev)
 437 {
 438         /* if we have more than 1 vec, admin queue offsets us by 1 */
 439         if (dev->num_vecs > 1)
 440                 return 1;
 441
 442         return 0;
 443 }
 444
 445 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
 446 {
 447         struct nvme_dev *dev = set->driver_data;
 448         int i, qoff, offset;
 449
 450         offset = queue_irq_offset(dev);
 451         for (i = 0, qoff = 0; i < set->nr_maps; i++) {
 452                 struct blk_mq_queue_map *map = &set->map[i];
 453
 454                 map->nr_queues = dev->io_queues[i];
 455                 if (!map->nr_queues) {
 456                         BUG_ON(i == HCTX_TYPE_DEFAULT);
 457                         continue;
 458                 }
 459
 460                 /*
 461                  * The poll queue(s) doesn't have an IRQ (and hence IRQ
 462                  * affinity), so use the regular blk-mq cpu mapping
 463                  */
 464                 map->queue_offset = qoff;
 465                 if (i != HCTX_TYPE_POLL && offset)
 466                         blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
 467                 else
 468                         blk_mq_map_queues(map);
 469                 qoff += map->nr_queues;
 470                 offset += map->nr_queues;
 471         }
 472
 473         return 0;
 474 }
 475
 476 /*
 477  * Write sq tail if we are asked to, or if the next command would wrap.
 478  */
 479 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
 480 {
 481         if (!write_sq) {
 482                 u16 next_tail = nvmeq->sq_tail + 1;
 483
 484                 if (next_tail == nvmeq->q_depth)
 485                         next_tail = 0;
 486                 if (next_tail != nvmeq->last_sq_tail)
 487                         return;
 488         }
 489
 490         if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
 491                         nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
 492                 writel(nvmeq->sq_tail, nvmeq->q_db);
 493         nvmeq->last_sq_tail = nvmeq->sq_tail;
 494 }
 495
 496 /**
 497  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 498  * @nvmeq: The queue to use
 499  * @cmd: The command to send
 500  * @write_sq: whether to write to the SQ doorbell
 501  */
 502 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
 503                             bool write_sq)
 504 {
 505         spin_lock(&nvmeq->sq_lock);
 506         memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 507                cmd, sizeof(*cmd));
 508         if (++nvmeq->sq_tail == nvmeq->q_depth)
 509                 nvmeq->sq_tail = 0;
 510         nvme_write_sq_db(nvmeq, write_sq);
 511         spin_unlock(&nvmeq->sq_lock);
 512 }
 513
 514 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
 515 {
 516         struct nvme_queue *nvmeq = hctx->driver_data;
 517
 518         spin_lock(&nvmeq->sq_lock);
 519         if (nvmeq->sq_tail != nvmeq->last_sq_tail)
 520                 nvme_write_sq_db(nvmeq, true);
 521         spin_unlock(&nvmeq->sq_lock);
 522 }
 523
 524 static void **nvme_pci_iod_list(struct request *req)
 525 {
 526         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 527         return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
 528 }
 529
 530 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
 531 {
 532         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 533         int nseg = blk_rq_nr_phys_segments(req);
 534         unsigned int avg_seg_size;
 535
 536         avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
 537
 538         if (!nvme_ctrl_sgl_supported(&dev->ctrl))
 539                 return false;
 540         if (!iod->nvmeq->qid)
 541                 return false;
 542         if (!sgl_threshold || avg_seg_size < sgl_threshold)
 543                 return false;
 544         return true;
 545 }
 546
 547 static void nvme_free_prps(struct nvme_dev *dev, struct request *req)
 548 {
 549         const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
 550         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 551         dma_addr_t dma_addr = iod->first_dma;
 552         int i;
 553
 554         for (i = 0; i < iod->npages; i++) {
 555                 __le64 *prp_list = nvme_pci_iod_list(req)[i];
 556                 dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
 557
 558                 dma_pool_free(dev->prp_page_pool, prp_list, dma_addr);
 559                 dma_addr = next_dma_addr;
 560         }
 561 }
 562
 563 static void nvme_free_sgls(struct nvme_dev *dev, struct request *req)
 564 {
 565         const int last_sg = SGES_PER_PAGE - 1;
 566         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 567         dma_addr_t dma_addr = iod->first_dma;
 568         int i;
 569
 570         for (i = 0; i < iod->npages; i++) {
 571                 struct nvme_sgl_desc *sg_list = nvme_pci_iod_list(req)[i];
 572                 dma_addr_t next_dma_addr = le64_to_cpu((sg_list[last_sg]).addr);
 573
 574                 dma_pool_free(dev->prp_page_pool, sg_list, dma_addr);
 575                 dma_addr = next_dma_addr;
 576         }
 577 }
 578
 579 static void nvme_unmap_sg(struct nvme_dev *dev, struct request *req)
 580 {
 581         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 582
 583         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 584                 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
 585                                     rq_dma_dir(req));
 586         else
 587                 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
 588 }
 589
 590 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 591 {
 592         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 593
 594         if (iod->dma_len) {
 595                 dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len,
 596                                rq_dma_dir(req));
 597                 return;
 598         }
 599
 600         WARN_ON_ONCE(!iod->nents);
 601
 602         nvme_unmap_sg(dev, req);
 603         if (iod->npages == 0)
 604                 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
 605                               iod->first_dma);
 606         else if (iod->use_sgl)
 607                 nvme_free_sgls(dev, req);
 608         else
 609                 nvme_free_prps(dev, req);
 610         mempool_free(iod->sg, dev->iod_mempool);
 611 }
 612
 613 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
 614 {
 615         int i;
 616         struct scatterlist *sg;
 617
 618         for_each_sg(sgl, sg, nents, i) {
 619                 dma_addr_t phys = sg_phys(sg);
 620                 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
 621                         "dma_address:%pad dma_length:%d\n",
 622                         i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
 623                         sg_dma_len(sg));
 624         }
 625 }
 626
 627 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
 628                 struct request *req, struct nvme_rw_command *cmnd)
 629 {
 630         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 631         struct dma_pool *pool;
 632         int length = blk_rq_payload_bytes(req);
 633         struct scatterlist *sg = iod->sg;
 634         int dma_len = sg_dma_len(sg);
 635         u64 dma_addr = sg_dma_address(sg);
 636         int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
 637         __le64 *prp_list;
 638         void **list = nvme_pci_iod_list(req);
 639         dma_addr_t prp_dma;
 640         int nprps, i;
 641
 642         length -= (NVME_CTRL_PAGE_SIZE - offset);
 643         if (length <= 0) {
 644                 iod->first_dma = 0;
 645                 goto done;
 646         }
 647
 648         dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
 649         if (dma_len) {
 650                 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
 651         } else {
 652                 sg = sg_next(sg);
 653                 dma_addr = sg_dma_address(sg);
 654                 dma_len = sg_dma_len(sg);
 655         }
 656
 657         if (length <= NVME_CTRL_PAGE_SIZE) {
 658                 iod->first_dma = dma_addr;
 659                 goto done;
 660         }
 661
 662         nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
 663         if (nprps <= (256 / 8)) {
 664                 pool = dev->prp_small_pool;
 665                 iod->npages = 0;
 666         } else {
 667                 pool = dev->prp_page_pool;
 668                 iod->npages = 1;
 669         }
 670
 671         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 672         if (!prp_list) {
 673                 iod->first_dma = dma_addr;
 674                 iod->npages = -1;
 675                 return BLK_STS_RESOURCE;
 676         }
 677         list[0] = prp_list;
 678         iod->first_dma = prp_dma;
 679         i = 0;
 680         for (;;) {
 681                 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
 682                         __le64 *old_prp_list = prp_list;
 683                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 684                         if (!prp_list)
 685                                 goto free_prps;
 686                         list[iod->npages++] = prp_list;
 687                         prp_list[0] = old_prp_list[i - 1];
 688                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 689                         i = 1;
 690                 }
 691                 prp_list[i++] = cpu_to_le64(dma_addr);
 692                 dma_len -= NVME_CTRL_PAGE_SIZE;
 693                 dma_addr += NVME_CTRL_PAGE_SIZE;
 694                 length -= NVME_CTRL_PAGE_SIZE;
 695                 if (length <= 0)
 696                         break;
 697                 if (dma_len > 0)
 698                         continue;
 699                 if (unlikely(dma_len < 0))
 700                         goto bad_sgl;
 701                 sg = sg_next(sg);
 702                 dma_addr = sg_dma_address(sg);
 703                 dma_len = sg_dma_len(sg);
 704         }
 705 done:
 706         cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 707         cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
 708         return BLK_STS_OK;
 709 free_prps:
 710         nvme_free_prps(dev, req);
 711         return BLK_STS_RESOURCE;
 712 bad_sgl:
 713         WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
 714                         "Invalid SGL for payload:%d nents:%d\n",
 715                         blk_rq_payload_bytes(req), iod->nents);
 716         return BLK_STS_IOERR;
 717 }
 718
 719 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
 720                 struct scatterlist *sg)
 721 {
 722         sge->addr = cpu_to_le64(sg_dma_address(sg));
 723         sge->length = cpu_to_le32(sg_dma_len(sg));
 724         sge->type = NVME_SGL_FMT_DATA_DESC << 4;
 725 }
 726
 727 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
 728                 dma_addr_t dma_addr, int entries)
 729 {
 730         sge->addr = cpu_to_le64(dma_addr);
 731         if (entries < SGES_PER_PAGE) {
 732                 sge->length = cpu_to_le32(entries * sizeof(*sge));
 733                 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
 734         } else {
 735                 sge->length = cpu_to_le32(PAGE_SIZE);
 736                 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
 737         }
 738 }
 739
 740 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
 741                 struct request *req, struct nvme_rw_command *cmd, int entries)
 742 {
 743         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 744         struct dma_pool *pool;
 745         struct nvme_sgl_desc *sg_list;
 746         struct scatterlist *sg = iod->sg;
 747         dma_addr_t sgl_dma;
 748         int i = 0;
 749
 750         /* setting the transfer type as SGL */
 751         cmd->flags = NVME_CMD_SGL_METABUF;
 752
 753         if (entries == 1) {
 754                 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
 755                 return BLK_STS_OK;
 756         }
 757
 758         if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
 759                 pool = dev->prp_small_pool;
 760                 iod->npages = 0;
 761         } else {
 762                 pool = dev->prp_page_pool;
 763                 iod->npages = 1;
 764         }
 765
 766         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 767         if (!sg_list) {
 768                 iod->npages = -1;
 769                 return BLK_STS_RESOURCE;
 770         }
 771
 772         nvme_pci_iod_list(req)[0] = sg_list;
 773         iod->first_dma = sgl_dma;
 774
 775         nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
 776
 777         do {
 778                 if (i == SGES_PER_PAGE) {
 779                         struct nvme_sgl_desc *old_sg_desc = sg_list;
 780                         struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
 781
 782                         sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
 783                         if (!sg_list)
 784                                 goto free_sgls;
 785
 786                         i = 0;
 787                         nvme_pci_iod_list(req)[iod->npages++] = sg_list;
 788                         sg_list[i++] = *link;
 789                         nvme_pci_sgl_set_seg(link, sgl_dma, entries);
 790                 }
 791
 792                 nvme_pci_sgl_set_data(&sg_list[i++], sg);
 793                 sg = sg_next(sg);
 794         } while (--entries > 0);
 795
 796         return BLK_STS_OK;
 797 free_sgls:
 798         nvme_free_sgls(dev, req);
 799         return BLK_STS_RESOURCE;
 800 }
 801
 802 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
 803                 struct request *req, struct nvme_rw_command *cmnd,
 804                 struct bio_vec *bv)
 805 {
 806         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 807         unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
 808         unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
 809
 810         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 811         if (dma_mapping_error(dev->dev, iod->first_dma))
 812                 return BLK_STS_RESOURCE;
 813         iod->dma_len = bv->bv_len;
 814
 815         cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
 816         if (bv->bv_len > first_prp_len)
 817                 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
 818         return BLK_STS_OK;
 819 }
 820
 821 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
 822                 struct request *req, struct nvme_rw_command *cmnd,
 823                 struct bio_vec *bv)
 824 {
 825         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 826
 827         iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
 828         if (dma_mapping_error(dev->dev, iod->first_dma))
 829                 return BLK_STS_RESOURCE;
 830         iod->dma_len = bv->bv_len;
 831
 832         cmnd->flags = NVME_CMD_SGL_METABUF;
 833         cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
 834         cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
 835         cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
 836         return BLK_STS_OK;
 837 }
 838
 839 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 840                 struct nvme_command *cmnd)
 841 {
 842         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 843         blk_status_t ret = BLK_STS_RESOURCE;
 844         int nr_mapped;
 845
 846         if (blk_rq_nr_phys_segments(req) == 1) {
 847                 struct bio_vec bv = req_bvec(req);
 848
 849                 if (!is_pci_p2pdma_page(bv.bv_page)) {
 850                         if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
 851                                 return nvme_setup_prp_simple(dev, req,
 852                                                              &cmnd->rw, &bv);
 853
 854                         if (iod->nvmeq->qid && sgl_threshold &&
 855                             nvme_ctrl_sgl_supported(&dev->ctrl))
 856                                 return nvme_setup_sgl_simple(dev, req,
 857                                                              &cmnd->rw, &bv);
 858                 }
 859         }
 860
 861         iod->dma_len = 0;
 862         iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
 863         if (!iod->sg)
 864                 return BLK_STS_RESOURCE;
 865         sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
 866         iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 867         if (!iod->nents)
 868                 goto out_free_sg;
 869
 870         if (is_pci_p2pdma_page(sg_page(iod->sg)))
 871                 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
 872                                 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
 873         else
 874                 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
 875                                              rq_dma_dir(req), DMA_ATTR_NO_WARN);
 876         if (!nr_mapped)
 877                 goto out_free_sg;
 878
 879         iod->use_sgl = nvme_pci_use_sgls(dev, req);
 880         if (iod->use_sgl)
 881                 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
 882         else
 883                 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
 884         if (ret != BLK_STS_OK)
 885                 goto out_unmap_sg;
 886         return BLK_STS_OK;
 887
 888 out_unmap_sg:
 889         nvme_unmap_sg(dev, req);
 890 out_free_sg:
 891         mempool_free(iod->sg, dev->iod_mempool);
 892         return ret;
 893 }
 894
 895 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
 896                 struct nvme_command *cmnd)
 897 {
 898         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 899
 900         iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
 901                         rq_dma_dir(req), 0);
 902         if (dma_mapping_error(dev->dev, iod->meta_dma))
 903                 return BLK_STS_IOERR;
 904         cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
 905         return BLK_STS_OK;
 906 }
 907
 908 /*
 909  * NOTE: ns is NULL when called on the admin queue.
 910  */
 911 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 912                          const struct blk_mq_queue_data *bd)
 913 {
 914         struct nvme_ns *ns = hctx->queue->queuedata;
 915         struct nvme_queue *nvmeq = hctx->driver_data;
 916         struct nvme_dev *dev = nvmeq->dev;
 917         struct request *req = bd->rq;
 918         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 919         struct nvme_command *cmnd = &iod->cmd;
 920         blk_status_t ret;
 921
 922         iod->aborted = 0;
 923         iod->npages = -1;
 924         iod->nents = 0;
 925
 926         /*
 927          * We should not need to do this, but we're still using this to
 928          * ensure we can drain requests on a dying queue.
 929          */
 930         if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
 931                 return BLK_STS_IOERR;
 932
 933         if (!nvme_check_ready(&dev->ctrl, req, true))
 934                 return nvme_fail_nonready_command(&dev->ctrl, req);
 935
 936         ret = nvme_setup_cmd(ns, req);
 937         if (ret)
 938                 return ret;
 939
 940         if (blk_rq_nr_phys_segments(req)) {
 941                 ret = nvme_map_data(dev, req, cmnd);
 942                 if (ret)
 943                         goto out_free_cmd;
 944         }
 945
 946         if (blk_integrity_rq(req)) {
 947                 ret = nvme_map_metadata(dev, req, cmnd);
 948                 if (ret)
 949                         goto out_unmap_data;
 950         }
 951
 952         blk_mq_start_request(req);
 953         nvme_submit_cmd(nvmeq, cmnd, bd->last);
 954         return BLK_STS_OK;
 955 out_unmap_data:
 956         nvme_unmap_data(dev, req);
 957 out_free_cmd:
 958         nvme_cleanup_cmd(req);
 959         return ret;
 960 }
 961
 962 static void nvme_pci_complete_rq(struct request *req)
 963 {
 964         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 965         struct nvme_dev *dev = iod->nvmeq->dev;
 966
 967         if (blk_integrity_rq(req))
 968                 dma_unmap_page(dev->dev, iod->meta_dma,
 969                                rq_integrity_vec(req)->bv_len, rq_data_dir(req));
 970         if (blk_rq_nr_phys_segments(req))
 971                 nvme_unmap_data(dev, req);
 972         nvme_complete_rq(req);
 973 }
 974
 975 /* We read the CQE phase first to check if the rest of the entry is valid */
 976 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
 977 {
 978         struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
 979
 980         return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
 981 }
 982
 983 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
 984 {
 985         u16 head = nvmeq->cq_head;
 986
 987         if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
 988                                               nvmeq->dbbuf_cq_ei))
 989                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 990 }
 991
 992 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
 993 {
 994         if (!nvmeq->qid)
 995                 return nvmeq->dev->admin_tagset.tags[0];
 996         return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
 997 }
 998
 999 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
1000 {
1001         struct nvme_completion *cqe = &nvmeq->cqes[idx];
1002         __u16 command_id = READ_ONCE(cqe->command_id);
1003         struct request *req;
1004
1005         /*
1006          * AEN requests are special as they don't time out and can
1007          * survive any kind of queue freeze and often don't respond to
1008          * aborts.  We don't even bother to allocate a struct request
1009          * for them but rather special case them here.
1010          */
1011         if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1012                 nvme_complete_async_event(&nvmeq->dev->ctrl,
1013                                 cqe->status, &cqe->result);
1014                 return;
1015         }
1016
1017         req = blk_mq_tag_to_rq(nvme_queue_tagset(nvmeq), command_id);
1018         if (unlikely(!req)) {
1019                 dev_warn(nvmeq->dev->ctrl.device,
1020                         "invalid id %d completed on queue %d\n",
1021                         command_id, le16_to_cpu(cqe->sq_id));
1022                 return;
1023         }
1024
1025         trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
1026         if (!nvme_try_complete_req(req, cqe->status, cqe->result))
1027                 nvme_pci_complete_rq(req);
1028 }
1029
1030 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1031 {
1032         u32 tmp = nvmeq->cq_head + 1;
1033
1034         if (tmp == nvmeq->q_depth) {
1035                 nvmeq->cq_head = 0;
1036                 nvmeq->cq_phase ^= 1;
1037         } else {
1038                 nvmeq->cq_head = tmp;
1039         }
1040 }
1041
1042 static inline int nvme_process_cq(struct nvme_queue *nvmeq)
1043 {
1044         int found = 0;
1045
1046         while (nvme_cqe_pending(nvmeq)) {
1047                 found++;
1048                 /*
1049                  * load-load control dependency between phase and the rest of
1050                  * the cqe requires a full read memory barrier
1051                  */
1052                 dma_rmb();
1053                 nvme_handle_cqe(nvmeq, nvmeq->cq_head);
1054                 nvme_update_cq_head(nvmeq);
1055         }
1056
1057         if (found)
1058                 nvme_ring_cq_doorbell(nvmeq);
1059         return found;
1060 }
1061
1062 static irqreturn_t nvme_irq(int irq, void *data)
1063 {
1064         struct nvme_queue *nvmeq = data;
1065
1066         if (nvme_process_cq(nvmeq))
1067                 return IRQ_HANDLED;
1068         return IRQ_NONE;
1069 }
1070
1071 static irqreturn_t nvme_irq_check(int irq, void *data)
1072 {
1073         struct nvme_queue *nvmeq = data;
1074
1075         if (nvme_cqe_pending(nvmeq))
1076                 return IRQ_WAKE_THREAD;
1077         return IRQ_NONE;
1078 }
1079
1080 /*
1081  * Poll for completions for any interrupt driven queue
1082  * Can be called from any context.
1083  */
1084 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1085 {
1086         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1087
1088         WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1089
1090         disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1091         nvme_process_cq(nvmeq);
1092         enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1093 }
1094
1095 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1096 {
1097         struct nvme_queue *nvmeq = hctx->driver_data;
1098         bool found;
1099
1100         if (!nvme_cqe_pending(nvmeq))
1101                 return 0;
1102
1103         spin_lock(&nvmeq->cq_poll_lock);
1104         found = nvme_process_cq(nvmeq);
1105         spin_unlock(&nvmeq->cq_poll_lock);
1106
1107         return found;
1108 }
1109
1110 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1111 {
1112         struct nvme_dev *dev = to_nvme_dev(ctrl);
1113         struct nvme_queue *nvmeq = &dev->queues[0];
1114         struct nvme_command c = { };
1115
1116         c.common.opcode = nvme_admin_async_event;
1117         c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1118         nvme_submit_cmd(nvmeq, &c, true);
1119 }
1120
1121 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1122 {
1123         struct nvme_command c = { };
1124
1125         c.delete_queue.opcode = opcode;
1126         c.delete_queue.qid = cpu_to_le16(id);
1127
1128         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1129 }
1130
1131 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1132                 struct nvme_queue *nvmeq, s16 vector)
1133 {
1134         struct nvme_command c = { };
1135         int flags = NVME_QUEUE_PHYS_CONTIG;
1136
1137         if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1138                 flags |= NVME_CQ_IRQ_ENABLED;
1139
1140         /*
1141          * Note: we (ab)use the fact that the prp fields survive if no data
1142          * is attached to the request.
1143          */
1144         c.create_cq.opcode = nvme_admin_create_cq;
1145         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1146         c.create_cq.cqid = cpu_to_le16(qid);
1147         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1148         c.create_cq.cq_flags = cpu_to_le16(flags);
1149         c.create_cq.irq_vector = cpu_to_le16(vector);
1150
1151         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1152 }
1153
1154 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1155                                                 struct nvme_queue *nvmeq)
1156 {
1157         struct nvme_ctrl *ctrl = &dev->ctrl;
1158         struct nvme_command c = { };
1159         int flags = NVME_QUEUE_PHYS_CONTIG;
1160
1161         /*
1162          * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1163          * set. Since URGENT priority is zeroes, it makes all queues
1164          * URGENT.
1165          */
1166         if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1167                 flags |= NVME_SQ_PRIO_MEDIUM;
1168
1169         /*
1170          * Note: we (ab)use the fact that the prp fields survive if no data
1171          * is attached to the request.
1172          */
1173         c.create_sq.opcode = nvme_admin_create_sq;
1174         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1175         c.create_sq.sqid = cpu_to_le16(qid);
1176         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1177         c.create_sq.sq_flags = cpu_to_le16(flags);
1178         c.create_sq.cqid = cpu_to_le16(qid);
1179
1180         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1181 }
1182
1183 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1184 {
1185         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1186 }
1187
1188 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1189 {
1190         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1191 }
1192
1193 static void abort_endio(struct request *req, blk_status_t error)
1194 {
1195         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1196         struct nvme_queue *nvmeq = iod->nvmeq;
1197
1198         dev_warn(nvmeq->dev->ctrl.device,
1199                  "Abort status: 0x%x", nvme_req(req)->status);
1200         atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1201         blk_mq_free_request(req);
1202 }
1203
1204 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1205 {
1206         /* If true, indicates loss of adapter communication, possibly by a
1207          * NVMe Subsystem reset.
1208          */
1209         bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1210
1211         /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1212         switch (dev->ctrl.state) {
1213         case NVME_CTRL_RESETTING:
1214         case NVME_CTRL_CONNECTING:
1215                 return false;
1216         default:
1217                 break;
1218         }
1219
1220         /* We shouldn't reset unless the controller is on fatal error state
1221          * _or_ if we lost the communication with it.
1222          */
1223         if (!(csts & NVME_CSTS_CFS) && !nssro)
1224                 return false;
1225
1226         return true;
1227 }
1228
1229 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1230 {
1231         /* Read a config register to help see what died. */
1232         u16 pci_status;
1233         int result;
1234
1235         result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1236                                       &pci_status);
1237         if (result == PCIBIOS_SUCCESSFUL)
1238                 dev_warn(dev->ctrl.device,
1239                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1240                          csts, pci_status);
1241         else
1242                 dev_warn(dev->ctrl.device,
1243                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1244                          csts, result);
1245 }
1246
1247 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1248 {
1249         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1250         struct nvme_queue *nvmeq = iod->nvmeq;
1251         struct nvme_dev *dev = nvmeq->dev;
1252         struct request *abort_req;
1253         struct nvme_command cmd = { };
1254         u32 csts = readl(dev->bar + NVME_REG_CSTS);
1255
1256         /* If PCI error recovery process is happening, we cannot reset or
1257          * the recovery mechanism will surely fail.
1258          */
1259         mb();
1260         if (pci_channel_offline(to_pci_dev(dev->dev)))
1261                 return BLK_EH_RESET_TIMER;
1262
1263         /*
1264          * Reset immediately if the controller is failed
1265          */
1266         if (nvme_should_reset(dev, csts)) {
1267                 nvme_warn_reset(dev, csts);
1268                 nvme_dev_disable(dev, false);
1269                 nvme_reset_ctrl(&dev->ctrl);
1270                 return BLK_EH_DONE;
1271         }
1272
1273         /*
1274          * Did we miss an interrupt?
1275          */
1276         if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1277                 nvme_poll(req->mq_hctx);
1278         else
1279                 nvme_poll_irqdisable(nvmeq);
1280
1281         if (blk_mq_request_completed(req)) {
1282                 dev_warn(dev->ctrl.device,
1283                          "I/O %d QID %d timeout, completion polled\n",
1284                          req->tag, nvmeq->qid);
1285                 return BLK_EH_DONE;
1286         }
1287
1288         /*
1289          * Shutdown immediately if controller times out while starting. The
1290          * reset work will see the pci device disabled when it gets the forced
1291          * cancellation error. All outstanding requests are completed on
1292          * shutdown, so we return BLK_EH_DONE.
1293          */
1294         switch (dev->ctrl.state) {
1295         case NVME_CTRL_CONNECTING:
1296                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1297                 fallthrough;
1298         case NVME_CTRL_DELETING:
1299                 dev_warn_ratelimited(dev->ctrl.device,
1300                          "I/O %d QID %d timeout, disable controller\n",
1301                          req->tag, nvmeq->qid);
1302                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1303                 nvme_dev_disable(dev, true);
1304                 return BLK_EH_DONE;
1305         case NVME_CTRL_RESETTING:
1306                 return BLK_EH_RESET_TIMER;
1307         default:
1308                 break;
1309         }
1310
1311         /*
1312          * Shutdown the controller immediately and schedule a reset if the
1313          * command was already aborted once before and still hasn't been
1314          * returned to the driver, or if this is the admin queue.
1315          */
1316         if (!nvmeq->qid || iod->aborted) {
1317                 dev_warn(dev->ctrl.device,
1318                          "I/O %d QID %d timeout, reset controller\n",
1319                          req->tag, nvmeq->qid);
1320                 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1321                 nvme_dev_disable(dev, false);
1322                 nvme_reset_ctrl(&dev->ctrl);
1323
1324                 return BLK_EH_DONE;
1325         }
1326
1327         if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1328                 atomic_inc(&dev->ctrl.abort_limit);
1329                 return BLK_EH_RESET_TIMER;
1330         }
1331         iod->aborted = 1;
1332
1333         cmd.abort.opcode = nvme_admin_abort_cmd;
1334         cmd.abort.cid = req->tag;
1335         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1336
1337         dev_warn(nvmeq->dev->ctrl.device,
1338                 "I/O %d QID %d timeout, aborting\n",
1339                  req->tag, nvmeq->qid);
1340
1341         abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1342                         BLK_MQ_REQ_NOWAIT);
1343         if (IS_ERR(abort_req)) {
1344                 atomic_inc(&dev->ctrl.abort_limit);
1345                 return BLK_EH_RESET_TIMER;
1346         }
1347
1348         abort_req->end_io_data = NULL;
1349         blk_execute_rq_nowait(NULL, abort_req, 0, abort_endio);
1350
1351         /*
1352          * The aborted req will be completed on receiving the abort req.
1353          * We enable the timer again. If hit twice, it'll cause a device reset,
1354          * as the device then is in a faulty state.
1355          */
1356         return BLK_EH_RESET_TIMER;
1357 }
1358
1359 static void nvme_free_queue(struct nvme_queue *nvmeq)
1360 {
1361         dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1362                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1363         if (!nvmeq->sq_cmds)
1364                 return;
1365
1366         if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1367                 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1368                                 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1369         } else {
1370                 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1371                                 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1372         }
1373 }
1374
1375 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1376 {
1377         int i;
1378
1379         for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1380                 dev->ctrl.queue_count--;
1381                 nvme_free_queue(&dev->queues[i]);
1382         }
1383 }
1384
1385 /**
1386  * nvme_suspend_queue - put queue into suspended state
1387  * @nvmeq: queue to suspend
1388  */
1389 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1390 {
1391         if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1392                 return 1;
1393
1394         /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1395         mb();
1396
1397         nvmeq->dev->online_queues--;
1398         if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1399                 blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1400         if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1401                 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1402         return 0;
1403 }
1404
1405 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1406 {
1407         int i;
1408
1409         for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1410                 nvme_suspend_queue(&dev->queues[i]);
1411 }
1412
1413 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1414 {
1415         struct nvme_queue *nvmeq = &dev->queues[0];
1416
1417         if (shutdown)
1418                 nvme_shutdown_ctrl(&dev->ctrl);
1419         else
1420                 nvme_disable_ctrl(&dev->ctrl);
1421
1422         nvme_poll_irqdisable(nvmeq);
1423 }
1424
1425 /*
1426  * Called only on a device that has been disabled and after all other threads
1427  * that can check this device's completion queues have synced, except
1428  * nvme_poll(). This is the last chance for the driver to see a natural
1429  * completion before nvme_cancel_request() terminates all incomplete requests.
1430  */
1431 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1432 {
1433         int i;
1434
1435         for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1436                 spin_lock(&dev->queues[i].cq_poll_lock);
1437                 nvme_process_cq(&dev->queues[i]);
1438                 spin_unlock(&dev->queues[i].cq_poll_lock);
1439         }
1440 }
1441
1442 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1443                                 int entry_size)
1444 {
1445         int q_depth = dev->q_depth;
1446         unsigned q_size_aligned = roundup(q_depth * entry_size,
1447                                           NVME_CTRL_PAGE_SIZE);
1448
1449         if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1450                 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1451
1452                 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1453                 q_depth = div_u64(mem_per_q, entry_size);
1454
1455                 /*
1456                  * Ensure the reduced q_depth is above some threshold where it
1457                  * would be better to map queues in system memory with the
1458                  * original depth
1459                  */
1460                 if (q_depth < 64)
1461                         return -ENOMEM;
1462         }
1463
1464         return q_depth;
1465 }
1466
1467 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1468                                 int qid)
1469 {
1470         struct pci_dev *pdev = to_pci_dev(dev->dev);
1471
1472         if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1473                 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1474                 if (nvmeq->sq_cmds) {
1475                         nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1476                                                         nvmeq->sq_cmds);
1477                         if (nvmeq->sq_dma_addr) {
1478                                 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1479                                 return 0;
1480                         }
1481
1482                         pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1483                 }
1484         }
1485
1486         nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1487                                 &nvmeq->sq_dma_addr, GFP_KERNEL);
1488         if (!nvmeq->sq_cmds)
1489                 return -ENOMEM;
1490         return 0;
1491 }
1492
1493 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1494 {
1495         struct nvme_queue *nvmeq = &dev->queues[qid];
1496
1497         if (dev->ctrl.queue_count > qid)
1498                 return 0;
1499
1500         nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1501         nvmeq->q_depth = depth;
1502         nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1503                                          &nvmeq->cq_dma_addr, GFP_KERNEL);
1504         if (!nvmeq->cqes)
1505                 goto free_nvmeq;
1506
1507         if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1508                 goto free_cqdma;
1509
1510         nvmeq->dev = dev;
1511         spin_lock_init(&nvmeq->sq_lock);
1512         spin_lock_init(&nvmeq->cq_poll_lock);
1513         nvmeq->cq_head = 0;
1514         nvmeq->cq_phase = 1;
1515         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1516         nvmeq->qid = qid;
1517         dev->ctrl.queue_count++;
1518
1519         return 0;
1520
1521  free_cqdma:
1522         dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1523                           nvmeq->cq_dma_addr);
1524  free_nvmeq:
1525         return -ENOMEM;
1526 }
1527
1528 static int queue_request_irq(struct nvme_queue *nvmeq)
1529 {
1530         struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1531         int nr = nvmeq->dev->ctrl.instance;
1532
1533         if (use_threaded_interrupts) {
1534                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1535                                 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1536         } else {
1537                 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1538                                 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1539         }
1540 }
1541
1542 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1543 {
1544         struct nvme_dev *dev = nvmeq->dev;
1545
1546         nvmeq->sq_tail = 0;
1547         nvmeq->last_sq_tail = 0;
1548         nvmeq->cq_head = 0;
1549         nvmeq->cq_phase = 1;
1550         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1551         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1552         nvme_dbbuf_init(dev, nvmeq, qid);
1553         dev->online_queues++;
1554         wmb(); /* ensure the first interrupt sees the initialization */
1555 }
1556
1557 /*
1558  * Try getting shutdown_lock while setting up IO queues.
1559  */
1560 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev)
1561 {
1562         /*
1563          * Give up if the lock is being held by nvme_dev_disable.
1564          */
1565         if (!mutex_trylock(&dev->shutdown_lock))
1566                 return -ENODEV;
1567
1568         /*
1569          * Controller is in wrong state, fail early.
1570          */
1571         if (dev->ctrl.state != NVME_CTRL_CONNECTING) {
1572                 mutex_unlock(&dev->shutdown_lock);
1573                 return -ENODEV;
1574         }
1575
1576         return 0;
1577 }
1578
1579 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1580 {
1581         struct nvme_dev *dev = nvmeq->dev;
1582         int result;
1583         u16 vector = 0;
1584
1585         clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1586
1587         /*
1588          * A queue's vector matches the queue identifier unless the controller
1589          * has only one vector available.
1590          */
1591         if (!polled)
1592                 vector = dev->num_vecs == 1 ? 0 : qid;
1593         else
1594                 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1595
1596         result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1597         if (result)
1598                 return result;
1599
1600         result = adapter_alloc_sq(dev, qid, nvmeq);
1601         if (result < 0)
1602                 return result;
1603         if (result)
1604                 goto release_cq;
1605
1606         nvmeq->cq_vector = vector;
1607
1608         result = nvme_setup_io_queues_trylock(dev);
1609         if (result)
1610                 return result;
1611         nvme_init_queue(nvmeq, qid);
1612         if (!polled) {
1613                 result = queue_request_irq(nvmeq);
1614                 if (result < 0)
1615                         goto release_sq;
1616         }
1617
1618         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1619         mutex_unlock(&dev->shutdown_lock);
1620         return result;
1621
1622 release_sq:
1623         dev->online_queues--;
1624         mutex_unlock(&dev->shutdown_lock);
1625         adapter_delete_sq(dev, qid);
1626 release_cq:
1627         adapter_delete_cq(dev, qid);
1628         return result;
1629 }
1630
1631 static const struct blk_mq_ops nvme_mq_admin_ops = {
1632         .queue_rq       = nvme_queue_rq,
1633         .complete       = nvme_pci_complete_rq,
1634         .init_hctx      = nvme_admin_init_hctx,
1635         .init_request   = nvme_init_request,
1636         .timeout        = nvme_timeout,
1637 };
1638
1639 static const struct blk_mq_ops nvme_mq_ops = {
1640         .queue_rq       = nvme_queue_rq,
1641         .complete       = nvme_pci_complete_rq,
1642         .commit_rqs     = nvme_commit_rqs,
1643         .init_hctx      = nvme_init_hctx,
1644         .init_request   = nvme_init_request,
1645         .map_queues     = nvme_pci_map_queues,
1646         .timeout        = nvme_timeout,
1647         .poll           = nvme_poll,
1648 };
1649
1650 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1651 {
1652         if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1653                 /*
1654                  * If the controller was reset during removal, it's possible
1655                  * user requests may be waiting on a stopped queue. Start the
1656                  * queue to flush these to completion.
1657                  */
1658                 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1659                 blk_cleanup_queue(dev->ctrl.admin_q);
1660                 blk_mq_free_tag_set(&dev->admin_tagset);
1661         }
1662 }
1663
1664 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1665 {
1666         if (!dev->ctrl.admin_q) {
1667                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1668                 dev->admin_tagset.nr_hw_queues = 1;
1669
1670                 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1671                 dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1672                 dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1673                 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1674                 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1675                 dev->admin_tagset.driver_data = dev;
1676
1677                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1678                         return -ENOMEM;
1679                 dev->ctrl.admin_tagset = &dev->admin_tagset;
1680
1681                 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1682                 if (IS_ERR(dev->ctrl.admin_q)) {
1683                         blk_mq_free_tag_set(&dev->admin_tagset);
1684                         return -ENOMEM;
1685                 }
1686                 if (!blk_get_queue(dev->ctrl.admin_q)) {
1687                         nvme_dev_remove_admin(dev);
1688                         dev->ctrl.admin_q = NULL;
1689                         return -ENODEV;
1690                 }
1691         } else
1692                 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1693
1694         return 0;
1695 }
1696
1697 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1698 {
1699         return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1700 }
1701
1702 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1703 {
1704         struct pci_dev *pdev = to_pci_dev(dev->dev);
1705
1706         if (size <= dev->bar_mapped_size)
1707                 return 0;
1708         if (size > pci_resource_len(pdev, 0))
1709                 return -ENOMEM;
1710         if (dev->bar)
1711                 iounmap(dev->bar);
1712         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1713         if (!dev->bar) {
1714                 dev->bar_mapped_size = 0;
1715                 return -ENOMEM;
1716         }
1717         dev->bar_mapped_size = size;
1718         dev->dbs = dev->bar + NVME_REG_DBS;
1719
1720         return 0;
1721 }
1722
1723 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1724 {
1725         int result;
1726         u32 aqa;
1727         struct nvme_queue *nvmeq;
1728
1729         result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1730         if (result < 0)
1731                 return result;
1732
1733         dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1734                                 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1735
1736         if (dev->subsystem &&
1737             (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1738                 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1739
1740         result = nvme_disable_ctrl(&dev->ctrl);
1741         if (result < 0)
1742                 return result;
1743
1744         result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1745         if (result)
1746                 return result;
1747
1748         dev->ctrl.numa_node = dev_to_node(dev->dev);
1749
1750         nvmeq = &dev->queues[0];
1751         aqa = nvmeq->q_depth - 1;
1752         aqa |= aqa << 16;
1753
1754         writel(aqa, dev->bar + NVME_REG_AQA);
1755         lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1756         lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1757
1758         result = nvme_enable_ctrl(&dev->ctrl);
1759         if (result)
1760                 return result;
1761
1762         nvmeq->cq_vector = 0;
1763         nvme_init_queue(nvmeq, 0);
1764         result = queue_request_irq(nvmeq);
1765         if (result) {
1766                 dev->online_queues--;
1767                 return result;
1768         }
1769
1770         set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1771         return result;
1772 }
1773
1774 static int nvme_create_io_queues(struct nvme_dev *dev)
1775 {
1776         unsigned i, max, rw_queues;
1777         int ret = 0;
1778
1779         for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1780                 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1781                         ret = -ENOMEM;
1782                         break;
1783                 }
1784         }
1785
1786         max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1787         if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1788                 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1789                                 dev->io_queues[HCTX_TYPE_READ];
1790         } else {
1791                 rw_queues = max;
1792         }
1793
1794         for (i = dev->online_queues; i <= max; i++) {
1795                 bool polled = i > rw_queues;
1796
1797                 ret = nvme_create_queue(&dev->queues[i], i, polled);
1798                 if (ret)
1799                         break;
1800         }
1801
1802         /*
1803          * Ignore failing Create SQ/CQ commands, we can continue with less
1804          * than the desired amount of queues, and even a controller without
1805          * I/O queues can still be used to issue admin commands.  This might
1806          * be useful to upgrade a buggy firmware for example.
1807          */
1808         return ret >= 0 ? 0 : ret;
1809 }
1810
1811 static ssize_t nvme_cmb_show(struct device *dev,
1812                              struct device_attribute *attr,
1813                              char *buf)
1814 {
1815         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1816
1817         return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1818                        ndev->cmbloc, ndev->cmbsz);
1819 }
1820 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1821
1822 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1823 {
1824         u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1825
1826         return 1ULL << (12 + 4 * szu);
1827 }
1828
1829 static u32 nvme_cmb_size(struct nvme_dev *dev)
1830 {
1831         return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1832 }
1833
1834 static void nvme_map_cmb(struct nvme_dev *dev)
1835 {
1836         u64 size, offset;
1837         resource_size_t bar_size;
1838         struct pci_dev *pdev = to_pci_dev(dev->dev);
1839         int bar;
1840
1841         if (dev->cmb_size)
1842                 return;
1843
1844         if (NVME_CAP_CMBS(dev->ctrl.cap))
1845                 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1846
1847         dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1848         if (!dev->cmbsz)
1849                 return;
1850         dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1851
1852         size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1853         offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1854         bar = NVME_CMB_BIR(dev->cmbloc);
1855         bar_size = pci_resource_len(pdev, bar);
1856
1857         if (offset > bar_size)
1858                 return;
1859
1860         /*
1861          * Tell the controller about the host side address mapping the CMB,
1862          * and enable CMB decoding for the NVMe 1.4+ scheme:
1863          */
1864         if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1865                 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1866                              (pci_bus_address(pdev, bar) + offset),
1867                              dev->bar + NVME_REG_CMBMSC);
1868         }
1869
1870         /*
1871          * Controllers may support a CMB size larger than their BAR,
1872          * for example, due to being behind a bridge. Reduce the CMB to
1873          * the reported size of the BAR
1874          */
1875         if (size > bar_size - offset)
1876                 size = bar_size - offset;
1877
1878         if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1879                 dev_warn(dev->ctrl.device,
1880                          "failed to register the CMB\n");
1881                 return;
1882         }
1883
1884         dev->cmb_size = size;
1885         dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1886
1887         if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1888                         (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1889                 pci_p2pmem_publish(pdev, true);
1890
1891         if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1892                                     &dev_attr_cmb.attr, NULL))
1893                 dev_warn(dev->ctrl.device,
1894                          "failed to add sysfs attribute for CMB\n");
1895 }
1896
1897 static inline void nvme_release_cmb(struct nvme_dev *dev)
1898 {
1899         if (dev->cmb_size) {
1900                 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1901                                              &dev_attr_cmb.attr, NULL);
1902                 dev->cmb_size = 0;
1903         }
1904 }
1905
1906 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1907 {
1908         u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1909         u64 dma_addr = dev->host_mem_descs_dma;
1910         struct nvme_command c = { };
1911         int ret;
1912
1913         c.features.opcode       = nvme_admin_set_features;
1914         c.features.fid          = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1915         c.features.dword11      = cpu_to_le32(bits);
1916         c.features.dword12      = cpu_to_le32(host_mem_size);
1917         c.features.dword13      = cpu_to_le32(lower_32_bits(dma_addr));
1918         c.features.dword14      = cpu_to_le32(upper_32_bits(dma_addr));
1919         c.features.dword15      = cpu_to_le32(dev->nr_host_mem_descs);
1920
1921         ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1922         if (ret) {
1923                 dev_warn(dev->ctrl.device,
1924                          "failed to set host mem (err %d, flags %#x).\n",
1925                          ret, bits);
1926         }
1927         return ret;
1928 }
1929
1930 static void nvme_free_host_mem(struct nvme_dev *dev)
1931 {
1932         int i;
1933
1934         for (i = 0; i < dev->nr_host_mem_descs; i++) {
1935                 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1936                 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
1937
1938                 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1939                                le64_to_cpu(desc->addr),
1940                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1941         }
1942
1943         kfree(dev->host_mem_desc_bufs);
1944         dev->host_mem_desc_bufs = NULL;
1945         dma_free_coherent(dev->dev,
1946                         dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1947                         dev->host_mem_descs, dev->host_mem_descs_dma);
1948         dev->host_mem_descs = NULL;
1949         dev->nr_host_mem_descs = 0;
1950 }
1951
1952 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1953                 u32 chunk_size)
1954 {
1955         struct nvme_host_mem_buf_desc *descs;
1956         u32 max_entries, len;
1957         dma_addr_t descs_dma;
1958         int i = 0;
1959         void **bufs;
1960         u64 size, tmp;
1961
1962         tmp = (preferred + chunk_size - 1);
1963         do_div(tmp, chunk_size);
1964         max_entries = tmp;
1965
1966         if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1967                 max_entries = dev->ctrl.hmmaxd;
1968
1969         descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1970                                    &descs_dma, GFP_KERNEL);
1971         if (!descs)
1972                 goto out;
1973
1974         bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1975         if (!bufs)
1976                 goto out_free_descs;
1977
1978         for (size = 0; size < preferred && i < max_entries; size += len) {
1979                 dma_addr_t dma_addr;
1980
1981                 len = min_t(u64, chunk_size, preferred - size);
1982                 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1983                                 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1984                 if (!bufs[i])
1985                         break;
1986
1987                 descs[i].addr = cpu_to_le64(dma_addr);
1988                 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
1989                 i++;
1990         }
1991
1992         if (!size)
1993                 goto out_free_bufs;
1994
1995         dev->nr_host_mem_descs = i;
1996         dev->host_mem_size = size;
1997         dev->host_mem_descs = descs;
1998         dev->host_mem_descs_dma = descs_dma;
1999         dev->host_mem_desc_bufs = bufs;
2000         return 0;
2001
2002 out_free_bufs:
2003         while (--i >= 0) {
2004                 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
2005
2006                 dma_free_attrs(dev->dev, size, bufs[i],
2007                                le64_to_cpu(descs[i].addr),
2008                                DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2009         }
2010
2011         kfree(bufs);
2012 out_free_descs:
2013         dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
2014                         descs_dma);
2015 out:
2016         dev->host_mem_descs = NULL;
2017         return -ENOMEM;
2018 }
2019
2020 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2021 {
2022         u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2023         u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2024         u64 chunk_size;
2025
2026         /* start big and work our way down */
2027         for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2028                 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
2029                         if (!min || dev->host_mem_size >= min)
2030                                 return 0;
2031                         nvme_free_host_mem(dev);
2032                 }
2033         }
2034
2035         return -ENOMEM;
2036 }
2037
2038 static int nvme_setup_host_mem(struct nvme_dev *dev)
2039 {
2040         u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2041         u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2042         u64 min = (u64)dev->ctrl.hmmin * 4096;
2043         u32 enable_bits = NVME_HOST_MEM_ENABLE;
2044         int ret;
2045
2046         preferred = min(preferred, max);
2047         if (min > max) {
2048                 dev_warn(dev->ctrl.device,
2049                         "min host memory (%lld MiB) above limit (%d MiB).\n",
2050                         min >> ilog2(SZ_1M), max_host_mem_size_mb);
2051                 nvme_free_host_mem(dev);
2052                 return 0;
2053         }
2054
2055         /*
2056          * If we already have a buffer allocated check if we can reuse it.
2057          */
2058         if (dev->host_mem_descs) {
2059                 if (dev->host_mem_size >= min)
2060                         enable_bits |= NVME_HOST_MEM_RETURN;
2061                 else
2062                         nvme_free_host_mem(dev);
2063         }
2064
2065         if (!dev->host_mem_descs) {
2066                 if (nvme_alloc_host_mem(dev, min, preferred)) {
2067                         dev_warn(dev->ctrl.device,
2068                                 "failed to allocate host memory buffer.\n");
2069                         return 0; /* controller must work without HMB */
2070                 }
2071
2072                 dev_info(dev->ctrl.device,
2073                         "allocated %lld MiB host memory buffer.\n",
2074                         dev->host_mem_size >> ilog2(SZ_1M));
2075         }
2076
2077         ret = nvme_set_host_mem(dev, enable_bits);
2078         if (ret)
2079                 nvme_free_host_mem(dev);
2080         return ret;
2081 }
2082
2083 /*
2084  * nirqs is the number of interrupts available for write and read
2085  * queues. The core already reserved an interrupt for the admin queue.
2086  */
2087 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2088 {
2089         struct nvme_dev *dev = affd->priv;
2090         unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2091
2092         /*
2093          * If there is no interrupt available for queues, ensure that
2094          * the default queue is set to 1. The affinity set size is
2095          * also set to one, but the irq core ignores it for this case.
2096          *
2097          * If only one interrupt is available or 'write_queue' == 0, combine
2098          * write and read queues.
2099          *
2100          * If 'write_queues' > 0, ensure it leaves room for at least one read
2101          * queue.
2102          */
2103         if (!nrirqs) {
2104                 nrirqs = 1;
2105                 nr_read_queues = 0;
2106         } else if (nrirqs == 1 || !nr_write_queues) {
2107                 nr_read_queues = 0;
2108         } else if (nr_write_queues >= nrirqs) {
2109                 nr_read_queues = 1;
2110         } else {
2111                 nr_read_queues = nrirqs - nr_write_queues;
2112         }
2113
2114         dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2115         affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2116         dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2117         affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2118         affd->nr_sets = nr_read_queues ? 2 : 1;
2119 }
2120
2121 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2122 {
2123         struct pci_dev *pdev = to_pci_dev(dev->dev);
2124         struct irq_affinity affd = {
2125                 .pre_vectors    = 1,
2126                 .calc_sets      = nvme_calc_irq_sets,
2127                 .priv           = dev,
2128         };
2129         unsigned int irq_queues, poll_queues;
2130
2131         /*
2132          * Poll queues don't need interrupts, but we need at least one I/O queue
2133          * left over for non-polled I/O.
2134          */
2135         poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2136         dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2137
2138         /*
2139          * Initialize for the single interrupt case, will be updated in
2140          * nvme_calc_irq_sets().
2141          */
2142         dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2143         dev->io_queues[HCTX_TYPE_READ] = 0;
2144
2145         /*
2146          * We need interrupts for the admin queue and each non-polled I/O queue,
2147          * but some Apple controllers require all queues to use the first
2148          * vector.
2149          */
2150         irq_queues = 1;
2151         if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2152                 irq_queues += (nr_io_queues - poll_queues);
2153         return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2154                               PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2155 }
2156
2157 static void nvme_disable_io_queues(struct nvme_dev *dev)
2158 {
2159         if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2160                 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2161 }
2162
2163 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2164 {
2165         /*
2166          * If tags are shared with admin queue (Apple bug), then
2167          * make sure we only use one IO queue.
2168          */
2169         if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2170                 return 1;
2171         return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2172 }
2173
2174 static int nvme_setup_io_queues(struct nvme_dev *dev)
2175 {
2176         struct nvme_queue *adminq = &dev->queues[0];
2177         struct pci_dev *pdev = to_pci_dev(dev->dev);
2178         unsigned int nr_io_queues;
2179         unsigned long size;
2180         int result;
2181
2182         /*
2183          * Sample the module parameters once at reset time so that we have
2184          * stable values to work with.
2185          */
2186         dev->nr_write_queues = write_queues;
2187         dev->nr_poll_queues = poll_queues;
2188
2189         nr_io_queues = dev->nr_allocated_queues - 1;
2190         result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2191         if (result < 0)
2192                 return result;
2193
2194         if (nr_io_queues == 0)
2195                 return 0;
2196
2197         /*
2198          * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions
2199          * from set to unset. If there is a window to it is truely freed,
2200          * pci_free_irq_vectors() jumping into this window will crash.
2201          * And take lock to avoid racing with pci_free_irq_vectors() in
2202          * nvme_dev_disable() path.
2203          */
2204         result = nvme_setup_io_queues_trylock(dev);
2205         if (result)
2206                 return result;
2207         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2208                 pci_free_irq(pdev, 0, adminq);
2209
2210         if (dev->cmb_use_sqes) {
2211                 result = nvme_cmb_qdepth(dev, nr_io_queues,
2212                                 sizeof(struct nvme_command));
2213                 if (result > 0)
2214                         dev->q_depth = result;
2215                 else
2216                         dev->cmb_use_sqes = false;
2217         }
2218
2219         do {
2220                 size = db_bar_size(dev, nr_io_queues);
2221                 result = nvme_remap_bar(dev, size);
2222                 if (!result)
2223                         break;
2224                 if (!--nr_io_queues) {
2225                         result = -ENOMEM;
2226                         goto out_unlock;
2227                 }
2228         } while (1);
2229         adminq->q_db = dev->dbs;
2230
2231  retry:
2232         /* Deregister the admin queue's interrupt */
2233         if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags))
2234                 pci_free_irq(pdev, 0, adminq);
2235
2236         /*
2237          * If we enable msix early due to not intx, disable it again before
2238          * setting up the full range we need.
2239          */
2240         pci_free_irq_vectors(pdev);
2241
2242         result = nvme_setup_irqs(dev, nr_io_queues);
2243         if (result <= 0) {
2244                 result = -EIO;
2245                 goto out_unlock;
2246         }
2247
2248         dev->num_vecs = result;
2249         result = max(result - 1, 1);
2250         dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2251
2252         /*
2253          * Should investigate if there's a performance win from allocating
2254          * more queues than interrupt vectors; it might allow the submission
2255          * path to scale better, even if the receive path is limited by the
2256          * number of interrupts.
2257          */
2258         result = queue_request_irq(adminq);
2259         if (result)
2260                 goto out_unlock;
2261         set_bit(NVMEQ_ENABLED, &adminq->flags);
2262         mutex_unlock(&dev->shutdown_lock);
2263
2264         result = nvme_create_io_queues(dev);
2265         if (result || dev->online_queues < 2)
2266                 return result;
2267
2268         if (dev->online_queues - 1 < dev->max_qid) {
2269                 nr_io_queues = dev->online_queues - 1;
2270                 nvme_disable_io_queues(dev);
2271                 result = nvme_setup_io_queues_trylock(dev);
2272                 if (result)
2273                         return result;
2274                 nvme_suspend_io_queues(dev);
2275                 goto retry;
2276         }
2277         dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2278                                         dev->io_queues[HCTX_TYPE_DEFAULT],
2279                                         dev->io_queues[HCTX_TYPE_READ],
2280                                         dev->io_queues[HCTX_TYPE_POLL]);
2281         return 0;
2282 out_unlock:
2283         mutex_unlock(&dev->shutdown_lock);
2284         return result;
2285 }
2286
2287 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2288 {
2289         struct nvme_queue *nvmeq = req->end_io_data;
2290
2291         blk_mq_free_request(req);
2292         complete(&nvmeq->delete_done);
2293 }
2294
2295 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2296 {
2297         struct nvme_queue *nvmeq = req->end_io_data;
2298
2299         if (error)
2300                 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2301
2302         nvme_del_queue_end(req, error);
2303 }
2304
2305 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2306 {
2307         struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2308         struct request *req;
2309         struct nvme_command cmd = { };
2310
2311         cmd.delete_queue.opcode = opcode;
2312         cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2313
2314         req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
2315         if (IS_ERR(req))
2316                 return PTR_ERR(req);
2317
2318         req->end_io_data = nvmeq;
2319
2320         init_completion(&nvmeq->delete_done);
2321         blk_execute_rq_nowait(NULL, req, false,
2322                         opcode == nvme_admin_delete_cq ?
2323                                 nvme_del_cq_end : nvme_del_queue_end);
2324         return 0;
2325 }
2326
2327 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2328 {
2329         int nr_queues = dev->online_queues - 1, sent = 0;
2330         unsigned long timeout;
2331
2332  retry:
2333         timeout = NVME_ADMIN_TIMEOUT;
2334         while (nr_queues > 0) {
2335                 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2336                         break;
2337                 nr_queues--;
2338                 sent++;
2339         }
2340         while (sent) {
2341                 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2342
2343                 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2344                                 timeout);
2345                 if (timeout == 0)
2346                         return false;
2347
2348                 sent--;
2349                 if (nr_queues)
2350                         goto retry;
2351         }
2352         return true;
2353 }
2354
2355 static void nvme_dev_add(struct nvme_dev *dev)
2356 {
2357         int ret;
2358
2359         if (!dev->ctrl.tagset) {
2360                 dev->tagset.ops = &nvme_mq_ops;
2361                 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2362                 dev->tagset.nr_maps = 2; /* default + read */
2363                 if (dev->io_queues[HCTX_TYPE_POLL])
2364                         dev->tagset.nr_maps++;
2365                 dev->tagset.timeout = NVME_IO_TIMEOUT;
2366                 dev->tagset.numa_node = dev->ctrl.numa_node;
2367                 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2368                                                 BLK_MQ_MAX_DEPTH) - 1;
2369                 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2370                 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2371                 dev->tagset.driver_data = dev;
2372
2373                 /*
2374                  * Some Apple controllers requires tags to be unique
2375                  * across admin and IO queue, so reserve the first 32
2376                  * tags of the IO queue.
2377                  */
2378                 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2379                         dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2380
2381                 ret = blk_mq_alloc_tag_set(&dev->tagset);
2382                 if (ret) {
2383                         dev_warn(dev->ctrl.device,
2384                                 "IO queues tagset allocation failed %d\n", ret);
2385                         return;
2386                 }
2387                 dev->ctrl.tagset = &dev->tagset;
2388         } else {
2389                 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2390
2391                 /* Free previously allocated queues that are no longer usable */
2392                 nvme_free_queues(dev, dev->online_queues);
2393         }
2394
2395         nvme_dbbuf_set(dev);
2396 }
2397
2398 static int nvme_pci_enable(struct nvme_dev *dev)
2399 {
2400         int result = -ENOMEM;
2401         struct pci_dev *pdev = to_pci_dev(dev->dev);
2402         int dma_address_bits = 64;
2403
2404         if (pci_enable_device_mem(pdev))
2405                 return result;
2406
2407         pci_set_master(pdev);
2408
2409         if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48)
2410                 dma_address_bits = 48;
2411         if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(dma_address_bits)))
2412                 goto disable;
2413
2414         if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2415                 result = -ENODEV;
2416                 goto disable;
2417         }
2418
2419         /*
2420          * Some devices and/or platforms don't advertise or work with INTx
2421          * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2422          * adjust this later.
2423          */
2424         result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2425         if (result < 0)
2426                 return result;
2427
2428         dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2429
2430         dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2431                                 io_queue_depth);
2432         dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2433         dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2434         dev->dbs = dev->bar + 4096;
2435
2436         /*
2437          * Some Apple controllers require a non-standard SQE size.
2438          * Interestingly they also seem to ignore the CC:IOSQES register
2439          * so we don't bother updating it here.
2440          */
2441         if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2442                 dev->io_sqes = 7;
2443         else
2444                 dev->io_sqes = NVME_NVM_IOSQES;
2445
2446         /*
2447          * Temporary fix for the Apple controller found in the MacBook8,1 and
2448          * some MacBook7,1 to avoid controller resets and data loss.
2449          */
2450         if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2451                 dev->q_depth = 2;
2452                 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2453                         "set queue depth=%u to work around controller resets\n",
2454                         dev->q_depth);
2455         } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2456                    (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2457                    NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2458                 dev->q_depth = 64;
2459                 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2460                         "set queue depth=%u\n", dev->q_depth);
2461         }
2462
2463         /*
2464          * Controllers with the shared tags quirk need the IO queue to be
2465          * big enough so that we get 32 tags for the admin queue
2466          */
2467         if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2468             (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2469                 dev->q_depth = NVME_AQ_DEPTH + 2;
2470                 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2471                          dev->q_depth);
2472         }
2473
2474
2475         nvme_map_cmb(dev);
2476
2477         pci_enable_pcie_error_reporting(pdev);
2478         pci_save_state(pdev);
2479         return 0;
2480
2481  disable:
2482         pci_disable_device(pdev);
2483         return result;
2484 }
2485
2486 static void nvme_dev_unmap(struct nvme_dev *dev)
2487 {
2488         if (dev->bar)
2489                 iounmap(dev->bar);
2490         pci_release_mem_regions(to_pci_dev(dev->dev));
2491 }
2492
2493 static void nvme_pci_disable(struct nvme_dev *dev)
2494 {
2495         struct pci_dev *pdev = to_pci_dev(dev->dev);
2496
2497         pci_free_irq_vectors(pdev);
2498
2499         if (pci_is_enabled(pdev)) {
2500                 pci_disable_pcie_error_reporting(pdev);
2501                 pci_disable_device(pdev);
2502         }
2503 }
2504
2505 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2506 {
2507         bool dead = true, freeze = false;
2508         struct pci_dev *pdev = to_pci_dev(dev->dev);
2509
2510         mutex_lock(&dev->shutdown_lock);
2511         if (pci_is_enabled(pdev)) {
2512                 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2513
2514                 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2515                     dev->ctrl.state == NVME_CTRL_RESETTING) {
2516                         freeze = true;
2517                         nvme_start_freeze(&dev->ctrl);
2518                 }
2519                 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2520                         pdev->error_state  != pci_channel_io_normal);
2521         }
2522
2523         /*
2524          * Give the controller a chance to complete all entered requests if
2525          * doing a safe shutdown.
2526          */
2527         if (!dead && shutdown && freeze)
2528                 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2529
2530         nvme_stop_queues(&dev->ctrl);
2531
2532         if (!dead && dev->ctrl.queue_count > 0) {
2533                 nvme_disable_io_queues(dev);
2534                 nvme_disable_admin_queue(dev, shutdown);
2535         }
2536         nvme_suspend_io_queues(dev);
2537         nvme_suspend_queue(&dev->queues[0]);
2538         nvme_pci_disable(dev);
2539         nvme_reap_pending_cqes(dev);
2540
2541         blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2542         blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2543         blk_mq_tagset_wait_completed_request(&dev->tagset);
2544         blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2545
2546         /*
2547          * The driver will not be starting up queues again if shutting down so
2548          * must flush all entered requests to their failed completion to avoid
2549          * deadlocking blk-mq hot-cpu notifier.
2550          */
2551         if (shutdown) {
2552                 nvme_start_queues(&dev->ctrl);
2553                 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2554                         blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2555         }
2556         mutex_unlock(&dev->shutdown_lock);
2557 }
2558
2559 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2560 {
2561         if (!nvme_wait_reset(&dev->ctrl))
2562                 return -EBUSY;
2563         nvme_dev_disable(dev, shutdown);
2564         return 0;
2565 }
2566
2567 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2568 {
2569         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2570                                                 NVME_CTRL_PAGE_SIZE,
2571                                                 NVME_CTRL_PAGE_SIZE, 0);
2572         if (!dev->prp_page_pool)
2573                 return -ENOMEM;
2574
2575         /* Optimisation for I/Os between 4k and 128k */
2576         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2577                                                 256, 256, 0);
2578         if (!dev->prp_small_pool) {
2579                 dma_pool_destroy(dev->prp_page_pool);
2580                 return -ENOMEM;
2581         }
2582         return 0;
2583 }
2584
2585 static void nvme_release_prp_pools(struct nvme_dev *dev)
2586 {
2587         dma_pool_destroy(dev->prp_page_pool);
2588         dma_pool_destroy(dev->prp_small_pool);
2589 }
2590
2591 static void nvme_free_tagset(struct nvme_dev *dev)
2592 {
2593         if (dev->tagset.tags)
2594                 blk_mq_free_tag_set(&dev->tagset);
2595         dev->ctrl.tagset = NULL;
2596 }
2597
2598 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2599 {
2600         struct nvme_dev *dev = to_nvme_dev(ctrl);
2601
2602         nvme_dbbuf_dma_free(dev);
2603         nvme_free_tagset(dev);
2604         if (dev->ctrl.admin_q)
2605                 blk_put_queue(dev->ctrl.admin_q);
2606         free_opal_dev(dev->ctrl.opal_dev);
2607         mempool_destroy(dev->iod_mempool);
2608         put_device(dev->dev);
2609         kfree(dev->queues);
2610         kfree(dev);
2611 }
2612
2613 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2614 {
2615         /*
2616          * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2617          * may be holding this pci_dev's device lock.
2618          */
2619         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2620         nvme_get_ctrl(&dev->ctrl);
2621         nvme_dev_disable(dev, false);
2622         nvme_kill_queues(&dev->ctrl);
2623         if (!queue_work(nvme_wq, &dev->remove_work))
2624                 nvme_put_ctrl(&dev->ctrl);
2625 }
2626
2627 static void nvme_reset_work(struct work_struct *work)
2628 {
2629         struct nvme_dev *dev =
2630                 container_of(work, struct nvme_dev, ctrl.reset_work);
2631         bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2632         int result;
2633
2634         if (dev->ctrl.state != NVME_CTRL_RESETTING) {
2635                 dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
2636                          dev->ctrl.state);
2637                 result = -ENODEV;
2638                 goto out;
2639         }
2640
2641         /*
2642          * If we're called to reset a live controller first shut it down before
2643          * moving on.
2644          */
2645         if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2646                 nvme_dev_disable(dev, false);
2647         nvme_sync_queues(&dev->ctrl);
2648
2649         mutex_lock(&dev->shutdown_lock);
2650         result = nvme_pci_enable(dev);
2651         if (result)
2652                 goto out_unlock;
2653
2654         result = nvme_pci_configure_admin_queue(dev);
2655         if (result)
2656                 goto out_unlock;
2657
2658         result = nvme_alloc_admin_tags(dev);
2659         if (result)
2660                 goto out_unlock;
2661
2662         /*
2663          * Limit the max command size to prevent iod->sg allocations going
2664          * over a single page.
2665          */
2666         dev->ctrl.max_hw_sectors = min_t(u32,
2667                 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2668         dev->ctrl.max_segments = NVME_MAX_SEGS;
2669
2670         /*
2671          * Don't limit the IOMMU merged segment size.
2672          */
2673         dma_set_max_seg_size(dev->dev, 0xffffffff);
2674         dma_set_min_align_mask(dev->dev, NVME_CTRL_PAGE_SIZE - 1);
2675
2676         mutex_unlock(&dev->shutdown_lock);
2677
2678         /*
2679          * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2680          * initializing procedure here.
2681          */
2682         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2683                 dev_warn(dev->ctrl.device,
2684                         "failed to mark controller CONNECTING\n");
2685                 result = -EBUSY;
2686                 goto out;
2687         }
2688
2689         /*
2690          * We do not support an SGL for metadata (yet), so we are limited to a
2691          * single integrity segment for the separate metadata pointer.
2692          */
2693         dev->ctrl.max_integrity_segments = 1;
2694
2695         result = nvme_init_ctrl_finish(&dev->ctrl);
2696         if (result)
2697                 goto out;
2698
2699         if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2700                 if (!dev->ctrl.opal_dev)
2701                         dev->ctrl.opal_dev =
2702                                 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2703                 else if (was_suspend)
2704                         opal_unlock_from_suspend(dev->ctrl.opal_dev);
2705         } else {
2706                 free_opal_dev(dev->ctrl.opal_dev);
2707                 dev->ctrl.opal_dev = NULL;
2708         }
2709
2710         if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2711                 result = nvme_dbbuf_dma_alloc(dev);
2712                 if (result)
2713                         dev_warn(dev->dev,
2714                                  "unable to allocate dma for dbbuf\n");
2715         }
2716
2717         if (dev->ctrl.hmpre) {
2718                 result = nvme_setup_host_mem(dev);
2719                 if (result < 0)
2720                         goto out;
2721         }
2722
2723         result = nvme_setup_io_queues(dev);
2724         if (result)
2725                 goto out;
2726
2727         /*
2728          * Keep the controller around but remove all namespaces if we don't have
2729          * any working I/O queue.
2730          */
2731         if (dev->online_queues < 2) {
2732                 dev_warn(dev->ctrl.device, "IO queues not created\n");
2733                 nvme_kill_queues(&dev->ctrl);
2734                 nvme_remove_namespaces(&dev->ctrl);
2735                 nvme_free_tagset(dev);
2736         } else {
2737                 nvme_start_queues(&dev->ctrl);
2738                 nvme_wait_freeze(&dev->ctrl);
2739                 nvme_dev_add(dev);
2740                 nvme_unfreeze(&dev->ctrl);
2741         }
2742
2743         /*
2744          * If only admin queue live, keep it to do further investigation or
2745          * recovery.
2746          */
2747         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2748                 dev_warn(dev->ctrl.device,
2749                         "failed to mark controller live state\n");
2750                 result = -ENODEV;
2751                 goto out;
2752         }
2753
2754         nvme_start_ctrl(&dev->ctrl);
2755         return;
2756
2757  out_unlock:
2758         mutex_unlock(&dev->shutdown_lock);
2759  out:
2760         if (result)
2761                 dev_warn(dev->ctrl.device,
2762                          "Removing after probe failure status: %d\n", result);
2763         nvme_remove_dead_ctrl(dev);
2764 }
2765
2766 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2767 {
2768         struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2769         struct pci_dev *pdev = to_pci_dev(dev->dev);
2770
2771         if (pci_get_drvdata(pdev))
2772                 device_release_driver(&pdev->dev);
2773         nvme_put_ctrl(&dev->ctrl);
2774 }
2775
2776 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2777 {
2778         *val = readl(to_nvme_dev(ctrl)->bar + off);
2779         return 0;
2780 }
2781
2782 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2783 {
2784         writel(val, to_nvme_dev(ctrl)->bar + off);
2785         return 0;
2786 }
2787
2788 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2789 {
2790         *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2791         return 0;
2792 }
2793
2794 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2795 {
2796         struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2797
2798         return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2799 }
2800
2801 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2802         .name                   = "pcie",
2803         .module                 = THIS_MODULE,
2804         .flags                  = NVME_F_METADATA_SUPPORTED |
2805                                   NVME_F_PCI_P2PDMA,
2806         .reg_read32             = nvme_pci_reg_read32,
2807         .reg_write32            = nvme_pci_reg_write32,
2808         .reg_read64             = nvme_pci_reg_read64,
2809         .free_ctrl              = nvme_pci_free_ctrl,
2810         .submit_async_event     = nvme_pci_submit_async_event,
2811         .get_address            = nvme_pci_get_address,
2812 };
2813
2814 static int nvme_dev_map(struct nvme_dev *dev)
2815 {
2816         struct pci_dev *pdev = to_pci_dev(dev->dev);
2817
2818         if (pci_request_mem_regions(pdev, "nvme"))
2819                 return -ENODEV;
2820
2821         if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2822                 goto release;
2823
2824         return 0;
2825   release:
2826         pci_release_mem_regions(pdev);
2827         return -ENODEV;
2828 }
2829
2830 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2831 {
2832         if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2833                 /*
2834                  * Several Samsung devices seem to drop off the PCIe bus
2835                  * randomly when APST is on and uses the deepest sleep state.
2836                  * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2837                  * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2838                  * 950 PRO 256GB", but it seems to be restricted to two Dell
2839                  * laptops.
2840                  */
2841                 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2842                     (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2843                      dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2844                         return NVME_QUIRK_NO_DEEPEST_PS;
2845         } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2846                 /*
2847                  * Samsung SSD 960 EVO drops off the PCIe bus after system
2848                  * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2849                  * within few minutes after bootup on a Coffee Lake board -
2850                  * ASUS PRIME Z370-A
2851                  */
2852                 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2853                     (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2854                      dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2855                         return NVME_QUIRK_NO_APST;
2856         } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2857                     pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2858                    (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2859                 /*
2860                  * Forcing to use host managed nvme power settings for
2861                  * lowest idle power with quick resume latency on
2862                  * Samsung and Toshiba SSDs based on suspend behavior
2863                  * on Coffee Lake board for LENOVO C640
2864                  */
2865                 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2866                      dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2867                         return NVME_QUIRK_SIMPLE_SUSPEND;
2868         }
2869
2870         return 0;
2871 }
2872
2873 static void nvme_async_probe(void *data, async_cookie_t cookie)
2874 {
2875         struct nvme_dev *dev = data;
2876
2877         flush_work(&dev->ctrl.reset_work);
2878         flush_work(&dev->ctrl.scan_work);
2879         nvme_put_ctrl(&dev->ctrl);
2880 }
2881
2882 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2883 {
2884         int node, result = -ENOMEM;
2885         struct nvme_dev *dev;
2886         unsigned long quirks = id->driver_data;
2887         size_t alloc_size;
2888
2889         node = dev_to_node(&pdev->dev);
2890         if (node == NUMA_NO_NODE)
2891                 set_dev_node(&pdev->dev, first_memory_node);
2892
2893         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2894         if (!dev)
2895                 return -ENOMEM;
2896
2897         dev->nr_write_queues = write_queues;
2898         dev->nr_poll_queues = poll_queues;
2899         dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2900         dev->queues = kcalloc_node(dev->nr_allocated_queues,
2901                         sizeof(struct nvme_queue), GFP_KERNEL, node);
2902         if (!dev->queues)
2903                 goto free;
2904
2905         dev->dev = get_device(&pdev->dev);
2906         pci_set_drvdata(pdev, dev);
2907
2908         result = nvme_dev_map(dev);
2909         if (result)
2910                 goto put_pci;
2911
2912         INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2913         INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2914         mutex_init(&dev->shutdown_lock);
2915
2916         result = nvme_setup_prp_pools(dev);
2917         if (result)
2918                 goto unmap;
2919
2920         quirks |= check_vendor_combination_bug(pdev);
2921
2922         if (!noacpi && acpi_storage_d3(&pdev->dev)) {
2923                 /*
2924                  * Some systems use a bios work around to ask for D3 on
2925                  * platforms that support kernel managed suspend.
2926                  */
2927                 dev_info(&pdev->dev,
2928                          "platform quirk: setting simple suspend\n");
2929                 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
2930         }
2931
2932         /*
2933          * Double check that our mempool alloc size will cover the biggest
2934          * command we support.
2935          */
2936         alloc_size = nvme_pci_iod_alloc_size();
2937         WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2938
2939         dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2940                                                 mempool_kfree,
2941                                                 (void *) alloc_size,
2942                                                 GFP_KERNEL, node);
2943         if (!dev->iod_mempool) {
2944                 result = -ENOMEM;
2945                 goto release_pools;
2946         }
2947
2948         result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2949                         quirks);
2950         if (result)
2951                 goto release_mempool;
2952
2953         dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2954
2955         nvme_reset_ctrl(&dev->ctrl);
2956         async_schedule(nvme_async_probe, dev);
2957
2958         return 0;
2959
2960  release_mempool:
2961         mempool_destroy(dev->iod_mempool);
2962  release_pools:
2963         nvme_release_prp_pools(dev);
2964  unmap:
2965         nvme_dev_unmap(dev);
2966  put_pci:
2967         put_device(dev->dev);
2968  free:
2969         kfree(dev->queues);
2970         kfree(dev);
2971         return result;
2972 }
2973
2974 static void nvme_reset_prepare(struct pci_dev *pdev)
2975 {
2976         struct nvme_dev *dev = pci_get_drvdata(pdev);
2977
2978         /*
2979          * We don't need to check the return value from waiting for the reset
2980          * state as pci_dev device lock is held, making it impossible to race
2981          * with ->remove().
2982          */
2983         nvme_disable_prepare_reset(dev, false);
2984         nvme_sync_queues(&dev->ctrl);
2985 }
2986
2987 static void nvme_reset_done(struct pci_dev *pdev)
2988 {
2989         struct nvme_dev *dev = pci_get_drvdata(pdev);
2990
2991         if (!nvme_try_sched_reset(&dev->ctrl))
2992                 flush_work(&dev->ctrl.reset_work);
2993 }
2994
2995 static void nvme_shutdown(struct pci_dev *pdev)
2996 {
2997         struct nvme_dev *dev = pci_get_drvdata(pdev);
2998
2999         nvme_disable_prepare_reset(dev, true);
3000 }
3001
3002 /*
3003  * The driver's remove may be called on a device in a partially initialized
3004  * state. This function must not have any dependencies on the device state in
3005  * order to proceed.
3006  */
3007 static void nvme_remove(struct pci_dev *pdev)
3008 {
3009         struct nvme_dev *dev = pci_get_drvdata(pdev);
3010
3011         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3012         pci_set_drvdata(pdev, NULL);
3013
3014         if (!pci_device_is_present(pdev)) {
3015                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3016                 nvme_dev_disable(dev, true);
3017         }
3018
3019         flush_work(&dev->ctrl.reset_work);
3020         nvme_stop_ctrl(&dev->ctrl);
3021         nvme_remove_namespaces(&dev->ctrl);
3022         nvme_dev_disable(dev, true);
3023         nvme_release_cmb(dev);
3024         nvme_free_host_mem(dev);
3025         nvme_dev_remove_admin(dev);
3026         nvme_free_queues(dev, 0);
3027         nvme_release_prp_pools(dev);
3028         nvme_dev_unmap(dev);
3029         nvme_uninit_ctrl(&dev->ctrl);
3030 }
3031
3032 #ifdef CONFIG_PM_SLEEP
3033 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3034 {
3035         return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3036 }
3037
3038 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3039 {
3040         return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3041 }
3042
3043 static int nvme_resume(struct device *dev)
3044 {
3045         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3046         struct nvme_ctrl *ctrl = &ndev->ctrl;
3047
3048         if (ndev->last_ps == U32_MAX ||
3049             nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3050                 return nvme_try_sched_reset(&ndev->ctrl);
3051         return 0;
3052 }
3053
3054 static int nvme_suspend(struct device *dev)
3055 {
3056         struct pci_dev *pdev = to_pci_dev(dev);
3057         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3058         struct nvme_ctrl *ctrl = &ndev->ctrl;
3059         int ret = -EBUSY;
3060
3061         ndev->last_ps = U32_MAX;
3062
3063         /*
3064          * The platform does not remove power for a kernel managed suspend so
3065          * use host managed nvme power settings for lowest idle power if
3066          * possible. This should have quicker resume latency than a full device
3067          * shutdown.  But if the firmware is involved after the suspend or the
3068          * device does not support any non-default power states, shut down the
3069          * device fully.
3070          *
3071          * If ASPM is not enabled for the device, shut down the device and allow
3072          * the PCI bus layer to put it into D3 in order to take the PCIe link
3073          * down, so as to allow the platform to achieve its minimum low-power
3074          * state (which may not be possible if the link is up).
3075          *
3076          * If a host memory buffer is enabled, shut down the device as the NVMe
3077          * specification allows the device to access the host memory buffer in
3078          * host DRAM from all power states, but hosts will fail access to DRAM
3079          * during S3.
3080          */
3081         if (pm_suspend_via_firmware() || !ctrl->npss ||
3082             !pcie_aspm_enabled(pdev) ||
3083             ndev->nr_host_mem_descs ||
3084             (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3085                 return nvme_disable_prepare_reset(ndev, true);
3086
3087         nvme_start_freeze(ctrl);
3088         nvme_wait_freeze(ctrl);
3089         nvme_sync_queues(ctrl);
3090
3091         if (ctrl->state != NVME_CTRL_LIVE)
3092                 goto unfreeze;
3093
3094         ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3095         if (ret < 0)
3096                 goto unfreeze;
3097
3098         /*
3099          * A saved state prevents pci pm from generically controlling the
3100          * device's power. If we're using protocol specific settings, we don't
3101          * want pci interfering.
3102          */
3103         pci_save_state(pdev);
3104
3105         ret = nvme_set_power_state(ctrl, ctrl->npss);
3106         if (ret < 0)
3107                 goto unfreeze;
3108
3109         if (ret) {
3110                 /* discard the saved state */
3111                 pci_load_saved_state(pdev, NULL);
3112
3113                 /*
3114                  * Clearing npss forces a controller reset on resume. The
3115                  * correct value will be rediscovered then.
3116                  */
3117                 ret = nvme_disable_prepare_reset(ndev, true);
3118                 ctrl->npss = 0;
3119         }
3120 unfreeze:
3121         nvme_unfreeze(ctrl);
3122         return ret;
3123 }
3124
3125 static int nvme_simple_suspend(struct device *dev)
3126 {
3127         struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3128
3129         return nvme_disable_prepare_reset(ndev, true);
3130 }
3131
3132 static int nvme_simple_resume(struct device *dev)
3133 {
3134         struct pci_dev *pdev = to_pci_dev(dev);
3135         struct nvme_dev *ndev = pci_get_drvdata(pdev);
3136
3137         return nvme_try_sched_reset(&ndev->ctrl);
3138 }
3139
3140 static const struct dev_pm_ops nvme_dev_pm_ops = {
3141         .suspend        = nvme_suspend,
3142         .resume         = nvme_resume,
3143         .freeze         = nvme_simple_suspend,
3144         .thaw           = nvme_simple_resume,
3145         .poweroff       = nvme_simple_suspend,
3146         .restore        = nvme_simple_resume,
3147 };
3148 #endif /* CONFIG_PM_SLEEP */
3149
3150 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3151                                                 pci_channel_state_t state)
3152 {
3153         struct nvme_dev *dev = pci_get_drvdata(pdev);
3154
3155         /*
3156          * A frozen channel requires a reset. When detected, this method will
3157          * shutdown the controller to quiesce. The controller will be restarted
3158          * after the slot reset through driver's slot_reset callback.
3159          */
3160         switch (state) {
3161         case pci_channel_io_normal:
3162                 return PCI_ERS_RESULT_CAN_RECOVER;
3163         case pci_channel_io_frozen:
3164                 dev_warn(dev->ctrl.device,
3165                         "frozen state error detected, reset controller\n");
3166                 nvme_dev_disable(dev, false);
3167                 return PCI_ERS_RESULT_NEED_RESET;
3168         case pci_channel_io_perm_failure:
3169                 dev_warn(dev->ctrl.device,
3170                         "failure state error detected, request disconnect\n");
3171                 return PCI_ERS_RESULT_DISCONNECT;
3172         }
3173         return PCI_ERS_RESULT_NEED_RESET;
3174 }
3175
3176 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3177 {
3178         struct nvme_dev *dev = pci_get_drvdata(pdev);
3179
3180         dev_info(dev->ctrl.device, "restart after slot reset\n");
3181         pci_restore_state(pdev);
3182         nvme_reset_ctrl(&dev->ctrl);
3183         return PCI_ERS_RESULT_RECOVERED;
3184 }
3185
3186 static void nvme_error_resume(struct pci_dev *pdev)
3187 {
3188         struct nvme_dev *dev = pci_get_drvdata(pdev);
3189
3190         flush_work(&dev->ctrl.reset_work);
3191 }
3192
3193 static const struct pci_error_handlers nvme_err_handler = {
3194         .error_detected = nvme_error_detected,
3195         .slot_reset     = nvme_slot_reset,
3196         .resume         = nvme_error_resume,
3197         .reset_prepare  = nvme_reset_prepare,
3198         .reset_done     = nvme_reset_done,
3199 };
3200
3201 static const struct pci_device_id nvme_id_table[] = {
3202         { PCI_VDEVICE(INTEL, 0x0953),   /* Intel 750/P3500/P3600/P3700 */
3203                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3204                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3205         { PCI_VDEVICE(INTEL, 0x0a53),   /* Intel P3520 */
3206                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3207                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3208         { PCI_VDEVICE(INTEL, 0x0a54),   /* Intel P4500/P4600 */
3209                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3210                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3211         { PCI_VDEVICE(INTEL, 0x0a55),   /* Dell Express Flash P4600 */
3212                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3213                                 NVME_QUIRK_DEALLOCATE_ZEROES, },
3214         { PCI_VDEVICE(INTEL, 0xf1a5),   /* Intel 600P/P3100 */
3215                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3216                                 NVME_QUIRK_MEDIUM_PRIO_SQ |
3217                                 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3218                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3219         { PCI_VDEVICE(INTEL, 0xf1a6),   /* Intel 760p/Pro 7600p */
3220                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3221         { PCI_VDEVICE(INTEL, 0x5845),   /* Qemu emulated controller */
3222                 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3223                                 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3224         { PCI_DEVICE(0x126f, 0x2263),   /* Silicon Motion unidentified */
3225                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
3226         { PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3227                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3228                                 NVME_QUIRK_NO_NS_DESC_LIST, },
3229         { PCI_DEVICE(0x1c58, 0x0003),   /* HGST adapter */
3230                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3231         { PCI_DEVICE(0x1c58, 0x0023),   /* WDC SN200 adapter */
3232                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3233         { PCI_DEVICE(0x1c5f, 0x0540),   /* Memblaze Pblaze4 adapter */
3234                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3235         { PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3236                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3237         { PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3238                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3239                                 NVME_QUIRK_DISABLE_WRITE_ZEROES|
3240                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3241         { PCI_DEVICE(0x1987, 0x5016),   /* Phison E16 */
3242                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3243         { PCI_DEVICE(0x1b4b, 0x1092),   /* Lexar 256 GB SSD */
3244                 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3245                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3246         { PCI_DEVICE(0x1d1d, 0x1f1f),   /* LighNVM qemu device */
3247                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3248         { PCI_DEVICE(0x1d1d, 0x2807),   /* CNEX WL */
3249                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3250         { PCI_DEVICE(0x1d1d, 0x2601),   /* CNEX Granby */
3251                 .driver_data = NVME_QUIRK_LIGHTNVM, },
3252         { PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3253                 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3254         { PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3255                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3256                                 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3257         { PCI_DEVICE(0x1c5c, 0x1504),   /* SK Hynix PC400 */
3258                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3259         { PCI_DEVICE(0x15b7, 0x2001),   /*  Sandisk Skyhawk */
3260                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3261         { PCI_DEVICE(0x1d97, 0x2263),   /* SPCC */
3262                 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3263         { PCI_DEVICE(0x2646, 0x2262),   /* KINGSTON SKC2000 NVMe SSD */
3264                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3265         { PCI_DEVICE(0x2646, 0x2263),   /* KINGSTON A2000 NVMe SSD  */
3266                 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3267         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061),
3268                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3269         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065),
3270                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3271         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061),
3272                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3273         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00),
3274                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3275         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01),
3276                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3277         { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02),
3278                 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, },
3279         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3280                 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3281         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3282         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3283                 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3284                                 NVME_QUIRK_128_BYTES_SQES |
3285                                 NVME_QUIRK_SHARED_TAGS },
3286
3287         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3288         { 0, }
3289 };
3290 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3291
3292 static struct pci_driver nvme_driver = {
3293         .name           = "nvme",
3294         .id_table       = nvme_id_table,
3295         .probe          = nvme_probe,
3296         .remove         = nvme_remove,
3297         .shutdown       = nvme_shutdown,
3298 #ifdef CONFIG_PM_SLEEP
3299         .driver         = {
3300                 .pm     = &nvme_dev_pm_ops,
3301         },
3302 #endif
3303         .sriov_configure = pci_sriov_configure_simple,
3304         .err_handler    = &nvme_err_handler,
3305 };
3306
3307 static int __init nvme_init(void)
3308 {
3309         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3310         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3311         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3312         BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3313
3314         return pci_register_driver(&nvme_driver);
3315 }
3316
3317 static void __exit nvme_exit(void)
3318 {
3319         pci_unregister_driver(&nvme_driver);
3320         flush_workqueue(nvme_wq);
3321 }
3322
3323 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3324 MODULE_LICENSE("GPL");
3325 MODULE_VERSION("1.0");
3326 module_init(nvme_init);
3327 module_exit(nvme_exit);