drivers/nvme/host/pci.c

   1 /*
   2  * NVM Express device driver
   3  * Copyright (c) 2011-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14
  15 #include <linux/aer.h>
  16 #include <linux/bitops.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/blk-mq.h>
  19 #include <linux/blk-mq-pci.h>
  20 #include <linux/cpu.h>
  21 #include <linux/delay.h>
  22 #include <linux/errno.h>
  23 #include <linux/fs.h>
  24 #include <linux/genhd.h>
  25 #include <linux/hdreg.h>
  26 #include <linux/idr.h>
  27 #include <linux/init.h>
  28 #include <linux/interrupt.h>
  29 #include <linux/io.h>
  30 #include <linux/kdev_t.h>
  31 #include <linux/kernel.h>
  32 #include <linux/mm.h>
  33 #include <linux/module.h>
  34 #include <linux/moduleparam.h>
  35 #include <linux/mutex.h>
  36 #include <linux/pci.h>
  37 #include <linux/poison.h>
  38 #include <linux/ptrace.h>
  39 #include <linux/sched.h>
  40 #include <linux/slab.h>
  41 #include <linux/t10-pi.h>
  42 #include <linux/timer.h>
  43 #include <linux/types.h>
  44 #include <linux/io-64-nonatomic-lo-hi.h>
  45 #include <asm/unaligned.h>
  46
  47 #include "nvme.h"
  48
  49 #define NVME_Q_DEPTH            1024
  50 #define NVME_AQ_DEPTH           256
  51 #define SQ_SIZE(depth)          (depth * sizeof(struct nvme_command))
  52 #define CQ_SIZE(depth)          (depth * sizeof(struct nvme_completion))
  53
  54 /*
  55  * We handle AEN commands ourselves and don't even let the
  56  * block layer know about them.
  57  */
  58 #define NVME_AQ_BLKMQ_DEPTH     (NVME_AQ_DEPTH - NVME_NR_AERS)
  59
  60 static int use_threaded_interrupts;
  61 module_param(use_threaded_interrupts, int, 0);
  62
  63 static bool use_cmb_sqes = true;
  64 module_param(use_cmb_sqes, bool, 0644);
  65 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
  66
  67 static struct workqueue_struct *nvme_workq;
  68
  69 struct nvme_dev;
  70 struct nvme_queue;
  71
  72 static int nvme_reset(struct nvme_dev *dev);
  73 static void nvme_process_cq(struct nvme_queue *nvmeq);
  74 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
  75
  76 /*
  77  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
  78  */
  79 struct nvme_dev {
  80         struct nvme_queue **queues;
  81         struct blk_mq_tag_set tagset;
  82         struct blk_mq_tag_set admin_tagset;
  83         u32 __iomem *dbs;
  84         struct device *dev;
  85         struct dma_pool *prp_page_pool;
  86         struct dma_pool *prp_small_pool;
  87         unsigned queue_count;
  88         unsigned online_queues;
  89         unsigned max_qid;
  90         int q_depth;
  91         u32 db_stride;
  92         void __iomem *bar;
  93         struct work_struct reset_work;
  94         struct work_struct remove_work;
  95         struct timer_list watchdog_timer;
  96         struct mutex shutdown_lock;
  97         bool subsystem;
  98         void __iomem *cmb;
  99         dma_addr_t cmb_dma_addr;
 100         u64 cmb_size;
 101         u32 cmbsz;
 102         u32 cmbloc;
 103         struct nvme_ctrl ctrl;
 104         struct completion ioq_wait;
 105 };
 106
 107 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 108 {
 109         return container_of(ctrl, struct nvme_dev, ctrl);
 110 }
 111
 112 /*
 113  * An NVM Express queue.  Each device has at least two (one for admin
 114  * commands and one for I/O commands).
 115  */
 116 struct nvme_queue {
 117         struct device *q_dmadev;
 118         struct nvme_dev *dev;
 119         char irqname[24];       /* nvme4294967295-65535\0 */
 120         spinlock_t q_lock;
 121         struct nvme_command *sq_cmds;
 122         struct nvme_command __iomem *sq_cmds_io;
 123         volatile struct nvme_completion *cqes;
 124         struct blk_mq_tags **tags;
 125         dma_addr_t sq_dma_addr;
 126         dma_addr_t cq_dma_addr;
 127         u32 __iomem *q_db;
 128         u16 q_depth;
 129         s16 cq_vector;
 130         u16 sq_tail;
 131         u16 cq_head;
 132         u16 qid;
 133         u8 cq_phase;
 134         u8 cqe_seen;
 135 };
 136
 137 /*
 138  * The nvme_iod describes the data in an I/O, including the list of PRP
 139  * entries.  You can't see it in this data structure because C doesn't let
 140  * me express that.  Use nvme_init_iod to ensure there's enough space
 141  * allocated to store the PRP list.
 142  */
 143 struct nvme_iod {
 144         struct nvme_request req;
 145         struct nvme_queue *nvmeq;
 146         int aborted;
 147         int npages;             /* In the PRP list. 0 means small pool in use */
 148         int nents;              /* Used in scatterlist */
 149         int length;             /* Of data, in bytes */
 150         dma_addr_t first_dma;
 151         struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
 152         struct scatterlist *sg;
 153         struct scatterlist inline_sg[0];
 154 };
 155
 156 /*
 157  * Check we didin't inadvertently grow the command struct
 158  */
 159 static inline void _nvme_check_size(void)
 160 {
 161         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
 162         BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
 163         BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
 164         BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
 165         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
 166         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
 167         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
 168         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
 169         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
 170         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
 171         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
 172         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
 173 }
 174
 175 /*
 176  * Max size of iod being embedded in the request payload
 177  */
 178 #define NVME_INT_PAGES          2
 179 #define NVME_INT_BYTES(dev)     (NVME_INT_PAGES * (dev)->ctrl.page_size)
 180
 181 /*
 182  * Will slightly overestimate the number of pages needed.  This is OK
 183  * as it only leads to a small amount of wasted memory for the lifetime of
 184  * the I/O.
 185  */
 186 static int nvme_npages(unsigned size, struct nvme_dev *dev)
 187 {
 188         unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
 189                                       dev->ctrl.page_size);
 190         return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 191 }
 192
 193 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
 194                 unsigned int size, unsigned int nseg)
 195 {
 196         return sizeof(__le64 *) * nvme_npages(size, dev) +
 197                         sizeof(struct scatterlist) * nseg;
 198 }
 199
 200 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
 201 {
 202         return sizeof(struct nvme_iod) +
 203                 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
 204 }
 205
 206 static int nvmeq_irq(struct nvme_queue *nvmeq)
 207 {
 208         return pci_irq_vector(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector);
 209 }
 210
 211 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 212                                 unsigned int hctx_idx)
 213 {
 214         struct nvme_dev *dev = data;
 215         struct nvme_queue *nvmeq = dev->queues[0];
 216
 217         WARN_ON(hctx_idx != 0);
 218         WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
 219         WARN_ON(nvmeq->tags);
 220
 221         hctx->driver_data = nvmeq;
 222         nvmeq->tags = &dev->admin_tagset.tags[0];
 223         return 0;
 224 }
 225
 226 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 227 {
 228         struct nvme_queue *nvmeq = hctx->driver_data;
 229
 230         nvmeq->tags = NULL;
 231 }
 232
 233 static int nvme_admin_init_request(void *data, struct request *req,
 234                                 unsigned int hctx_idx, unsigned int rq_idx,
 235                                 unsigned int numa_node)
 236 {
 237         struct nvme_dev *dev = data;
 238         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 239         struct nvme_queue *nvmeq = dev->queues[0];
 240
 241         BUG_ON(!nvmeq);
 242         iod->nvmeq = nvmeq;
 243         return 0;
 244 }
 245
 246 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 247                           unsigned int hctx_idx)
 248 {
 249         struct nvme_dev *dev = data;
 250         struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
 251
 252         if (!nvmeq->tags)
 253                 nvmeq->tags = &dev->tagset.tags[hctx_idx];
 254
 255         WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
 256         hctx->driver_data = nvmeq;
 257         return 0;
 258 }
 259
 260 static int nvme_init_request(void *data, struct request *req,
 261                                 unsigned int hctx_idx, unsigned int rq_idx,
 262                                 unsigned int numa_node)
 263 {
 264         struct nvme_dev *dev = data;
 265         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 266         struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
 267
 268         BUG_ON(!nvmeq);
 269         iod->nvmeq = nvmeq;
 270         return 0;
 271 }
 272
 273 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
 274 {
 275         struct nvme_dev *dev = set->driver_data;
 276
 277         return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
 278 }
 279
 280 /**
 281  * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
 282  * @nvmeq: The queue to use
 283  * @cmd: The command to send
 284  *
 285  * Safe to use from interrupt context
 286  */
 287 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
 288                                                 struct nvme_command *cmd)
 289 {
 290         u16 tail = nvmeq->sq_tail;
 291
 292         if (nvmeq->sq_cmds_io)
 293                 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
 294         else
 295                 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
 296
 297         if (++tail == nvmeq->q_depth)
 298                 tail = 0;
 299         writel(tail, nvmeq->q_db);
 300         nvmeq->sq_tail = tail;
 301 }
 302
 303 static __le64 **iod_list(struct request *req)
 304 {
 305         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 306         return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
 307 }
 308
 309 static int nvme_init_iod(struct request *rq, struct nvme_dev *dev)
 310 {
 311         struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
 312         int nseg = blk_rq_nr_phys_segments(rq);
 313         unsigned int size = blk_rq_payload_bytes(rq);
 314
 315         if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
 316                 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
 317                 if (!iod->sg)
 318                         return BLK_MQ_RQ_QUEUE_BUSY;
 319         } else {
 320                 iod->sg = iod->inline_sg;
 321         }
 322
 323         iod->aborted = 0;
 324         iod->npages = -1;
 325         iod->nents = 0;
 326         iod->length = size;
 327
 328         if (!(rq->rq_flags & RQF_DONTPREP)) {
 329                 rq->retries = 0;
 330                 rq->rq_flags |= RQF_DONTPREP;
 331         }
 332         return BLK_MQ_RQ_QUEUE_OK;
 333 }
 334
 335 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
 336 {
 337         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 338         const int last_prp = dev->ctrl.page_size / 8 - 1;
 339         int i;
 340         __le64 **list = iod_list(req);
 341         dma_addr_t prp_dma = iod->first_dma;
 342
 343         if (iod->npages == 0)
 344                 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
 345         for (i = 0; i < iod->npages; i++) {
 346                 __le64 *prp_list = list[i];
 347                 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
 348                 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
 349                 prp_dma = next_prp_dma;
 350         }
 351
 352         if (iod->sg != iod->inline_sg)
 353                 kfree(iod->sg);
 354 }
 355
 356 #ifdef CONFIG_BLK_DEV_INTEGRITY
 357 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
 358 {
 359         if (be32_to_cpu(pi->ref_tag) == v)
 360                 pi->ref_tag = cpu_to_be32(p);
 361 }
 362
 363 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
 364 {
 365         if (be32_to_cpu(pi->ref_tag) == p)
 366                 pi->ref_tag = cpu_to_be32(v);
 367 }
 368
 369 /**
 370  * nvme_dif_remap - remaps ref tags to bip seed and physical lba
 371  *
 372  * The virtual start sector is the one that was originally submitted by the
 373  * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
 374  * start sector may be different. Remap protection information to match the
 375  * physical LBA on writes, and back to the original seed on reads.
 376  *
 377  * Type 0 and 3 do not have a ref tag, so no remapping required.
 378  */
 379 static void nvme_dif_remap(struct request *req,
 380                         void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
 381 {
 382         struct nvme_ns *ns = req->rq_disk->private_data;
 383         struct bio_integrity_payload *bip;
 384         struct t10_pi_tuple *pi;
 385         void *p, *pmap;
 386         u32 i, nlb, ts, phys, virt;
 387
 388         if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
 389                 return;
 390
 391         bip = bio_integrity(req->bio);
 392         if (!bip)
 393                 return;
 394
 395         pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
 396
 397         p = pmap;
 398         virt = bip_get_seed(bip);
 399         phys = nvme_block_nr(ns, blk_rq_pos(req));
 400         nlb = (blk_rq_bytes(req) >> ns->lba_shift);
 401         ts = ns->disk->queue->integrity.tuple_size;
 402
 403         for (i = 0; i < nlb; i++, virt++, phys++) {
 404                 pi = (struct t10_pi_tuple *)p;
 405                 dif_swap(phys, virt, pi);
 406                 p += ts;
 407         }
 408         kunmap_atomic(pmap);
 409 }
 410 #else /* CONFIG_BLK_DEV_INTEGRITY */
 411 static void nvme_dif_remap(struct request *req,
 412                         void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
 413 {
 414 }
 415 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
 416 {
 417 }
 418 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
 419 {
 420 }
 421 #endif
 422
 423 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
 424 {
 425         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 426         struct dma_pool *pool;
 427         int length = blk_rq_payload_bytes(req);
 428         struct scatterlist *sg = iod->sg;
 429         int dma_len = sg_dma_len(sg);
 430         u64 dma_addr = sg_dma_address(sg);
 431         u32 page_size = dev->ctrl.page_size;
 432         int offset = dma_addr & (page_size - 1);
 433         __le64 *prp_list;
 434         __le64 **list = iod_list(req);
 435         dma_addr_t prp_dma;
 436         int nprps, i;
 437
 438         length -= (page_size - offset);
 439         if (length <= 0)
 440                 return true;
 441
 442         dma_len -= (page_size - offset);
 443         if (dma_len) {
 444                 dma_addr += (page_size - offset);
 445         } else {
 446                 sg = sg_next(sg);
 447                 dma_addr = sg_dma_address(sg);
 448                 dma_len = sg_dma_len(sg);
 449         }
 450
 451         if (length <= page_size) {
 452                 iod->first_dma = dma_addr;
 453                 return true;
 454         }
 455
 456         nprps = DIV_ROUND_UP(length, page_size);
 457         if (nprps <= (256 / 8)) {
 458                 pool = dev->prp_small_pool;
 459                 iod->npages = 0;
 460         } else {
 461                 pool = dev->prp_page_pool;
 462                 iod->npages = 1;
 463         }
 464
 465         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 466         if (!prp_list) {
 467                 iod->first_dma = dma_addr;
 468                 iod->npages = -1;
 469                 return false;
 470         }
 471         list[0] = prp_list;
 472         iod->first_dma = prp_dma;
 473         i = 0;
 474         for (;;) {
 475                 if (i == page_size >> 3) {
 476                         __le64 *old_prp_list = prp_list;
 477                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 478                         if (!prp_list)
 479                                 return false;
 480                         list[iod->npages++] = prp_list;
 481                         prp_list[0] = old_prp_list[i - 1];
 482                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 483                         i = 1;
 484                 }
 485                 prp_list[i++] = cpu_to_le64(dma_addr);
 486                 dma_len -= page_size;
 487                 dma_addr += page_size;
 488                 length -= page_size;
 489                 if (length <= 0)
 490                         break;
 491                 if (dma_len > 0)
 492                         continue;
 493                 BUG_ON(dma_len < 0);
 494                 sg = sg_next(sg);
 495                 dma_addr = sg_dma_address(sg);
 496                 dma_len = sg_dma_len(sg);
 497         }
 498
 499         return true;
 500 }
 501
 502 static int nvme_map_data(struct nvme_dev *dev, struct request *req,
 503                 struct nvme_command *cmnd)
 504 {
 505         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 506         struct request_queue *q = req->q;
 507         enum dma_data_direction dma_dir = rq_data_dir(req) ?
 508                         DMA_TO_DEVICE : DMA_FROM_DEVICE;
 509         int ret = BLK_MQ_RQ_QUEUE_ERROR;
 510
 511         sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
 512         iod->nents = blk_rq_map_sg(q, req, iod->sg);
 513         if (!iod->nents)
 514                 goto out;
 515
 516         ret = BLK_MQ_RQ_QUEUE_BUSY;
 517         if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
 518                                 DMA_ATTR_NO_WARN))
 519                 goto out;
 520
 521         if (!nvme_setup_prps(dev, req))
 522                 goto out_unmap;
 523
 524         ret = BLK_MQ_RQ_QUEUE_ERROR;
 525         if (blk_integrity_rq(req)) {
 526                 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
 527                         goto out_unmap;
 528
 529                 sg_init_table(&iod->meta_sg, 1);
 530                 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
 531                         goto out_unmap;
 532
 533                 if (rq_data_dir(req))
 534                         nvme_dif_remap(req, nvme_dif_prep);
 535
 536                 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
 537                         goto out_unmap;
 538         }
 539
 540         cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 541         cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
 542         if (blk_integrity_rq(req))
 543                 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
 544         return BLK_MQ_RQ_QUEUE_OK;
 545
 546 out_unmap:
 547         dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
 548 out:
 549         return ret;
 550 }
 551
 552 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 553 {
 554         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 555         enum dma_data_direction dma_dir = rq_data_dir(req) ?
 556                         DMA_TO_DEVICE : DMA_FROM_DEVICE;
 557
 558         if (iod->nents) {
 559                 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
 560                 if (blk_integrity_rq(req)) {
 561                         if (!rq_data_dir(req))
 562                                 nvme_dif_remap(req, nvme_dif_complete);
 563                         dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
 564                 }
 565         }
 566
 567         nvme_cleanup_cmd(req);
 568         nvme_free_iod(dev, req);
 569 }
 570
 571 /*
 572  * NOTE: ns is NULL when called on the admin queue.
 573  */
 574 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 575                          const struct blk_mq_queue_data *bd)
 576 {
 577         struct nvme_ns *ns = hctx->queue->queuedata;
 578         struct nvme_queue *nvmeq = hctx->driver_data;
 579         struct nvme_dev *dev = nvmeq->dev;
 580         struct request *req = bd->rq;
 581         struct nvme_command cmnd;
 582         int ret = BLK_MQ_RQ_QUEUE_OK;
 583
 584         /*
 585          * If formated with metadata, require the block layer provide a buffer
 586          * unless this namespace is formated such that the metadata can be
 587          * stripped/generated by the controller with PRACT=1.
 588          */
 589         if (ns && ns->ms && !blk_integrity_rq(req)) {
 590                 if (!(ns->pi_type && ns->ms == 8) &&
 591                                         req->cmd_type != REQ_TYPE_DRV_PRIV) {
 592                         blk_mq_end_request(req, -EFAULT);
 593                         return BLK_MQ_RQ_QUEUE_OK;
 594                 }
 595         }
 596
 597         ret = nvme_setup_cmd(ns, req, &cmnd);
 598         if (ret != BLK_MQ_RQ_QUEUE_OK)
 599                 return ret;
 600
 601         ret = nvme_init_iod(req, dev);
 602         if (ret != BLK_MQ_RQ_QUEUE_OK)
 603                 goto out_free_cmd;
 604
 605         if (blk_rq_nr_phys_segments(req))
 606                 ret = nvme_map_data(dev, req, &cmnd);
 607
 608         if (ret != BLK_MQ_RQ_QUEUE_OK)
 609                 goto out_cleanup_iod;
 610
 611         blk_mq_start_request(req);
 612
 613         spin_lock_irq(&nvmeq->q_lock);
 614         if (unlikely(nvmeq->cq_vector < 0)) {
 615                 if (ns && !test_bit(NVME_NS_DEAD, &ns->flags))
 616                         ret = BLK_MQ_RQ_QUEUE_BUSY;
 617                 else
 618                         ret = BLK_MQ_RQ_QUEUE_ERROR;
 619                 spin_unlock_irq(&nvmeq->q_lock);
 620                 goto out_cleanup_iod;
 621         }
 622         __nvme_submit_cmd(nvmeq, &cmnd);
 623         nvme_process_cq(nvmeq);
 624         spin_unlock_irq(&nvmeq->q_lock);
 625         return BLK_MQ_RQ_QUEUE_OK;
 626 out_cleanup_iod:
 627         nvme_free_iod(dev, req);
 628 out_free_cmd:
 629         nvme_cleanup_cmd(req);
 630         return ret;
 631 }
 632
 633 static void nvme_complete_rq(struct request *req)
 634 {
 635         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 636         struct nvme_dev *dev = iod->nvmeq->dev;
 637         int error = 0;
 638
 639         nvme_unmap_data(dev, req);
 640
 641         if (unlikely(req->errors)) {
 642                 if (nvme_req_needs_retry(req, req->errors)) {
 643                         req->retries++;
 644                         nvme_requeue_req(req);
 645                         return;
 646                 }
 647
 648                 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 649                         error = req->errors;
 650                 else
 651                         error = nvme_error_status(req->errors);
 652         }
 653
 654         if (unlikely(iod->aborted)) {
 655                 dev_warn(dev->ctrl.device,
 656                         "completing aborted command with status: %04x\n",
 657                         req->errors);
 658         }
 659
 660         blk_mq_end_request(req, error);
 661 }
 662
 663 /* We read the CQE phase first to check if the rest of the entry is valid */
 664 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
 665                 u16 phase)
 666 {
 667         return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
 668 }
 669
 670 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
 671 {
 672         u16 head, phase;
 673
 674         head = nvmeq->cq_head;
 675         phase = nvmeq->cq_phase;
 676
 677         while (nvme_cqe_valid(nvmeq, head, phase)) {
 678                 struct nvme_completion cqe = nvmeq->cqes[head];
 679                 struct request *req;
 680
 681                 if (++head == nvmeq->q_depth) {
 682                         head = 0;
 683                         phase = !phase;
 684                 }
 685
 686                 if (tag && *tag == cqe.command_id)
 687                         *tag = -1;
 688
 689                 if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
 690                         dev_warn(nvmeq->dev->ctrl.device,
 691                                 "invalid id %d completed on queue %d\n",
 692                                 cqe.command_id, le16_to_cpu(cqe.sq_id));
 693                         continue;
 694                 }
 695
 696                 /*
 697                  * AEN requests are special as they don't time out and can
 698                  * survive any kind of queue freeze and often don't respond to
 699                  * aborts.  We don't even bother to allocate a struct request
 700                  * for them but rather special case them here.
 701                  */
 702                 if (unlikely(nvmeq->qid == 0 &&
 703                                 cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
 704                         nvme_complete_async_event(&nvmeq->dev->ctrl,
 705                                         cqe.status, &cqe.result);
 706                         continue;
 707                 }
 708
 709                 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
 710                 nvme_req(req)->result = cqe.result;
 711                 blk_mq_complete_request(req, le16_to_cpu(cqe.status) >> 1);
 712         }
 713
 714         if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
 715                 return;
 716
 717         if (likely(nvmeq->cq_vector >= 0))
 718                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 719         nvmeq->cq_head = head;
 720         nvmeq->cq_phase = phase;
 721
 722         nvmeq->cqe_seen = 1;
 723 }
 724
 725 static void nvme_process_cq(struct nvme_queue *nvmeq)
 726 {
 727         __nvme_process_cq(nvmeq, NULL);
 728 }
 729
 730 static irqreturn_t nvme_irq(int irq, void *data)
 731 {
 732         irqreturn_t result;
 733         struct nvme_queue *nvmeq = data;
 734         spin_lock(&nvmeq->q_lock);
 735         nvme_process_cq(nvmeq);
 736         result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
 737         nvmeq->cqe_seen = 0;
 738         spin_unlock(&nvmeq->q_lock);
 739         return result;
 740 }
 741
 742 static irqreturn_t nvme_irq_check(int irq, void *data)
 743 {
 744         struct nvme_queue *nvmeq = data;
 745         if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
 746                 return IRQ_WAKE_THREAD;
 747         return IRQ_NONE;
 748 }
 749
 750 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
 751 {
 752         struct nvme_queue *nvmeq = hctx->driver_data;
 753
 754         if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
 755                 spin_lock_irq(&nvmeq->q_lock);
 756                 __nvme_process_cq(nvmeq, &tag);
 757                 spin_unlock_irq(&nvmeq->q_lock);
 758
 759                 if (tag == -1)
 760                         return 1;
 761         }
 762
 763         return 0;
 764 }
 765
 766 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
 767 {
 768         struct nvme_dev *dev = to_nvme_dev(ctrl);
 769         struct nvme_queue *nvmeq = dev->queues[0];
 770         struct nvme_command c;
 771
 772         memset(&c, 0, sizeof(c));
 773         c.common.opcode = nvme_admin_async_event;
 774         c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
 775
 776         spin_lock_irq(&nvmeq->q_lock);
 777         __nvme_submit_cmd(nvmeq, &c);
 778         spin_unlock_irq(&nvmeq->q_lock);
 779 }
 780
 781 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
 782 {
 783         struct nvme_command c;
 784
 785         memset(&c, 0, sizeof(c));
 786         c.delete_queue.opcode = opcode;
 787         c.delete_queue.qid = cpu_to_le16(id);
 788
 789         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
 790 }
 791
 792 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
 793                                                 struct nvme_queue *nvmeq)
 794 {
 795         struct nvme_command c;
 796         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
 797
 798         /*
 799          * Note: we (ab)use the fact the the prp fields survive if no data
 800          * is attached to the request.
 801          */
 802         memset(&c, 0, sizeof(c));
 803         c.create_cq.opcode = nvme_admin_create_cq;
 804         c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
 805         c.create_cq.cqid = cpu_to_le16(qid);
 806         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 807         c.create_cq.cq_flags = cpu_to_le16(flags);
 808         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
 809
 810         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
 811 }
 812
 813 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
 814                                                 struct nvme_queue *nvmeq)
 815 {
 816         struct nvme_command c;
 817         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
 818
 819         /*
 820          * Note: we (ab)use the fact the the prp fields survive if no data
 821          * is attached to the request.
 822          */
 823         memset(&c, 0, sizeof(c));
 824         c.create_sq.opcode = nvme_admin_create_sq;
 825         c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
 826         c.create_sq.sqid = cpu_to_le16(qid);
 827         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 828         c.create_sq.sq_flags = cpu_to_le16(flags);
 829         c.create_sq.cqid = cpu_to_le16(qid);
 830
 831         return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
 832 }
 833
 834 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
 835 {
 836         return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
 837 }
 838
 839 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
 840 {
 841         return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
 842 }
 843
 844 static void abort_endio(struct request *req, int error)
 845 {
 846         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 847         struct nvme_queue *nvmeq = iod->nvmeq;
 848         u16 status = req->errors;
 849
 850         dev_warn(nvmeq->dev->ctrl.device, "Abort status: 0x%x", status);
 851         atomic_inc(&nvmeq->dev->ctrl.abort_limit);
 852         blk_mq_free_request(req);
 853 }
 854
 855 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
 856 {
 857         struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
 858         struct nvme_queue *nvmeq = iod->nvmeq;
 859         struct nvme_dev *dev = nvmeq->dev;
 860         struct request *abort_req;
 861         struct nvme_command cmd;
 862
 863         /*
 864          * Shutdown immediately if controller times out while starting. The
 865          * reset work will see the pci device disabled when it gets the forced
 866          * cancellation error. All outstanding requests are completed on
 867          * shutdown, so we return BLK_EH_HANDLED.
 868          */
 869         if (dev->ctrl.state == NVME_CTRL_RESETTING) {
 870                 dev_warn(dev->ctrl.device,
 871                          "I/O %d QID %d timeout, disable controller\n",
 872                          req->tag, nvmeq->qid);
 873                 nvme_dev_disable(dev, false);
 874                 req->errors = NVME_SC_CANCELLED;
 875                 return BLK_EH_HANDLED;
 876         }
 877
 878         /*
 879          * Shutdown the controller immediately and schedule a reset if the
 880          * command was already aborted once before and still hasn't been
 881          * returned to the driver, or if this is the admin queue.
 882          */
 883         if (!nvmeq->qid || iod->aborted) {
 884                 dev_warn(dev->ctrl.device,
 885                          "I/O %d QID %d timeout, reset controller\n",
 886                          req->tag, nvmeq->qid);
 887                 nvme_dev_disable(dev, false);
 888                 nvme_reset(dev);
 889
 890                 /*
 891                  * Mark the request as handled, since the inline shutdown
 892                  * forces all outstanding requests to complete.
 893                  */
 894                 req->errors = NVME_SC_CANCELLED;
 895                 return BLK_EH_HANDLED;
 896         }
 897
 898         iod->aborted = 1;
 899
 900         if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
 901                 atomic_inc(&dev->ctrl.abort_limit);
 902                 return BLK_EH_RESET_TIMER;
 903         }
 904
 905         memset(&cmd, 0, sizeof(cmd));
 906         cmd.abort.opcode = nvme_admin_abort_cmd;
 907         cmd.abort.cid = req->tag;
 908         cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
 909
 910         dev_warn(nvmeq->dev->ctrl.device,
 911                 "I/O %d QID %d timeout, aborting\n",
 912                  req->tag, nvmeq->qid);
 913
 914         abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
 915                         BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
 916         if (IS_ERR(abort_req)) {
 917                 atomic_inc(&dev->ctrl.abort_limit);
 918                 return BLK_EH_RESET_TIMER;
 919         }
 920
 921         abort_req->timeout = ADMIN_TIMEOUT;
 922         abort_req->end_io_data = NULL;
 923         blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
 924
 925         /*
 926          * The aborted req will be completed on receiving the abort req.
 927          * We enable the timer again. If hit twice, it'll cause a device reset,
 928          * as the device then is in a faulty state.
 929          */
 930         return BLK_EH_RESET_TIMER;
 931 }
 932
 933 static void nvme_free_queue(struct nvme_queue *nvmeq)
 934 {
 935         dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
 936                                 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
 937         if (nvmeq->sq_cmds)
 938                 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
 939                                         nvmeq->sq_cmds, nvmeq->sq_dma_addr);
 940         kfree(nvmeq);
 941 }
 942
 943 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
 944 {
 945         int i;
 946
 947         for (i = dev->queue_count - 1; i >= lowest; i--) {
 948                 struct nvme_queue *nvmeq = dev->queues[i];
 949                 dev->queue_count--;
 950                 dev->queues[i] = NULL;
 951                 nvme_free_queue(nvmeq);
 952         }
 953 }
 954
 955 /**
 956  * nvme_suspend_queue - put queue into suspended state
 957  * @nvmeq - queue to suspend
 958  */
 959 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
 960 {
 961         int vector;
 962
 963         spin_lock_irq(&nvmeq->q_lock);
 964         if (nvmeq->cq_vector == -1) {
 965                 spin_unlock_irq(&nvmeq->q_lock);
 966                 return 1;
 967         }
 968         vector = nvmeq_irq(nvmeq);
 969         nvmeq->dev->online_queues--;
 970         nvmeq->cq_vector = -1;
 971         spin_unlock_irq(&nvmeq->q_lock);
 972
 973         if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
 974                 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
 975
 976         free_irq(vector, nvmeq);
 977
 978         return 0;
 979 }
 980
 981 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
 982 {
 983         struct nvme_queue *nvmeq = dev->queues[0];
 984
 985         if (!nvmeq)
 986                 return;
 987         if (nvme_suspend_queue(nvmeq))
 988                 return;
 989
 990         if (shutdown)
 991                 nvme_shutdown_ctrl(&dev->ctrl);
 992         else
 993                 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
 994                                                 dev->bar + NVME_REG_CAP));
 995
 996         spin_lock_irq(&nvmeq->q_lock);
 997         nvme_process_cq(nvmeq);
 998         spin_unlock_irq(&nvmeq->q_lock);
 999 }
1000
1001 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1002                                 int entry_size)
1003 {
1004         int q_depth = dev->q_depth;
1005         unsigned q_size_aligned = roundup(q_depth * entry_size,
1006                                           dev->ctrl.page_size);
1007
1008         if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1009                 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1010                 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1011                 q_depth = div_u64(mem_per_q, entry_size);
1012
1013                 /*
1014                  * Ensure the reduced q_depth is above some threshold where it
1015                  * would be better to map queues in system memory with the
1016                  * original depth
1017                  */
1018                 if (q_depth < 64)
1019                         return -ENOMEM;
1020         }
1021
1022         return q_depth;
1023 }
1024
1025 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1026                                 int qid, int depth)
1027 {
1028         if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1029                 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1030                                                       dev->ctrl.page_size);
1031                 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1032                 nvmeq->sq_cmds_io = dev->cmb + offset;
1033         } else {
1034                 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1035                                         &nvmeq->sq_dma_addr, GFP_KERNEL);
1036                 if (!nvmeq->sq_cmds)
1037                         return -ENOMEM;
1038         }
1039
1040         return 0;
1041 }
1042
1043 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1044                                                         int depth)
1045 {
1046         struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1047         if (!nvmeq)
1048                 return NULL;
1049
1050         nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1051                                           &nvmeq->cq_dma_addr, GFP_KERNEL);
1052         if (!nvmeq->cqes)
1053                 goto free_nvmeq;
1054
1055         if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1056                 goto free_cqdma;
1057
1058         nvmeq->q_dmadev = dev->dev;
1059         nvmeq->dev = dev;
1060         snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1061                         dev->ctrl.instance, qid);
1062         spin_lock_init(&nvmeq->q_lock);
1063         nvmeq->cq_head = 0;
1064         nvmeq->cq_phase = 1;
1065         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1066         nvmeq->q_depth = depth;
1067         nvmeq->qid = qid;
1068         nvmeq->cq_vector = -1;
1069         dev->queues[qid] = nvmeq;
1070         dev->queue_count++;
1071
1072         return nvmeq;
1073
1074  free_cqdma:
1075         dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1076                                                         nvmeq->cq_dma_addr);
1077  free_nvmeq:
1078         kfree(nvmeq);
1079         return NULL;
1080 }
1081
1082 static int queue_request_irq(struct nvme_queue *nvmeq)
1083 {
1084         if (use_threaded_interrupts)
1085                 return request_threaded_irq(nvmeq_irq(nvmeq), nvme_irq_check,
1086                                 nvme_irq, IRQF_SHARED, nvmeq->irqname, nvmeq);
1087         else
1088                 return request_irq(nvmeq_irq(nvmeq), nvme_irq, IRQF_SHARED,
1089                                 nvmeq->irqname, nvmeq);
1090 }
1091
1092 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1093 {
1094         struct nvme_dev *dev = nvmeq->dev;
1095
1096         spin_lock_irq(&nvmeq->q_lock);
1097         nvmeq->sq_tail = 0;
1098         nvmeq->cq_head = 0;
1099         nvmeq->cq_phase = 1;
1100         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1101         memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1102         dev->online_queues++;
1103         spin_unlock_irq(&nvmeq->q_lock);
1104 }
1105
1106 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1107 {
1108         struct nvme_dev *dev = nvmeq->dev;
1109         int result;
1110
1111         nvmeq->cq_vector = qid - 1;
1112         result = adapter_alloc_cq(dev, qid, nvmeq);
1113         if (result < 0)
1114                 return result;
1115
1116         result = adapter_alloc_sq(dev, qid, nvmeq);
1117         if (result < 0)
1118                 goto release_cq;
1119
1120         result = queue_request_irq(nvmeq);
1121         if (result < 0)
1122                 goto release_sq;
1123
1124         nvme_init_queue(nvmeq, qid);
1125         return result;
1126
1127  release_sq:
1128         adapter_delete_sq(dev, qid);
1129  release_cq:
1130         adapter_delete_cq(dev, qid);
1131         return result;
1132 }
1133
1134 static struct blk_mq_ops nvme_mq_admin_ops = {
1135         .queue_rq       = nvme_queue_rq,
1136         .complete       = nvme_complete_rq,
1137         .init_hctx      = nvme_admin_init_hctx,
1138         .exit_hctx      = nvme_admin_exit_hctx,
1139         .init_request   = nvme_admin_init_request,
1140         .timeout        = nvme_timeout,
1141 };
1142
1143 static struct blk_mq_ops nvme_mq_ops = {
1144         .queue_rq       = nvme_queue_rq,
1145         .complete       = nvme_complete_rq,
1146         .init_hctx      = nvme_init_hctx,
1147         .init_request   = nvme_init_request,
1148         .map_queues     = nvme_pci_map_queues,
1149         .timeout        = nvme_timeout,
1150         .poll           = nvme_poll,
1151 };
1152
1153 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1154 {
1155         if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1156                 /*
1157                  * If the controller was reset during removal, it's possible
1158                  * user requests may be waiting on a stopped queue. Start the
1159                  * queue to flush these to completion.
1160                  */
1161                 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1162                 blk_cleanup_queue(dev->ctrl.admin_q);
1163                 blk_mq_free_tag_set(&dev->admin_tagset);
1164         }
1165 }
1166
1167 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1168 {
1169         if (!dev->ctrl.admin_q) {
1170                 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1171                 dev->admin_tagset.nr_hw_queues = 1;
1172
1173                 /*
1174                  * Subtract one to leave an empty queue entry for 'Full Queue'
1175                  * condition. See NVM-Express 1.2 specification, section 4.1.2.
1176                  */
1177                 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1178                 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1179                 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1180                 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1181                 dev->admin_tagset.driver_data = dev;
1182
1183                 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1184                         return -ENOMEM;
1185
1186                 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1187                 if (IS_ERR(dev->ctrl.admin_q)) {
1188                         blk_mq_free_tag_set(&dev->admin_tagset);
1189                         return -ENOMEM;
1190                 }
1191                 if (!blk_get_queue(dev->ctrl.admin_q)) {
1192                         nvme_dev_remove_admin(dev);
1193                         dev->ctrl.admin_q = NULL;
1194                         return -ENODEV;
1195                 }
1196         } else
1197                 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1198
1199         return 0;
1200 }
1201
1202 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1203 {
1204         int result;
1205         u32 aqa;
1206         u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1207         struct nvme_queue *nvmeq;
1208
1209         dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1210                                                 NVME_CAP_NSSRC(cap) : 0;
1211
1212         if (dev->subsystem &&
1213             (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1214                 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1215
1216         result = nvme_disable_ctrl(&dev->ctrl, cap);
1217         if (result < 0)
1218                 return result;
1219
1220         nvmeq = dev->queues[0];
1221         if (!nvmeq) {
1222                 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1223                 if (!nvmeq)
1224                         return -ENOMEM;
1225         }
1226
1227         aqa = nvmeq->q_depth - 1;
1228         aqa |= aqa << 16;
1229
1230         writel(aqa, dev->bar + NVME_REG_AQA);
1231         lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1232         lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1233
1234         result = nvme_enable_ctrl(&dev->ctrl, cap);
1235         if (result)
1236                 return result;
1237
1238         nvmeq->cq_vector = 0;
1239         result = queue_request_irq(nvmeq);
1240         if (result) {
1241                 nvmeq->cq_vector = -1;
1242                 return result;
1243         }
1244
1245         return result;
1246 }
1247
1248 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1249 {
1250
1251         /* If true, indicates loss of adapter communication, possibly by a
1252          * NVMe Subsystem reset.
1253          */
1254         bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1255
1256         /* If there is a reset ongoing, we shouldn't reset again. */
1257         if (work_busy(&dev->reset_work))
1258                 return false;
1259
1260         /* We shouldn't reset unless the controller is on fatal error state
1261          * _or_ if we lost the communication with it.
1262          */
1263         if (!(csts & NVME_CSTS_CFS) && !nssro)
1264                 return false;
1265
1266         /* If PCI error recovery process is happening, we cannot reset or
1267          * the recovery mechanism will surely fail.
1268          */
1269         if (pci_channel_offline(to_pci_dev(dev->dev)))
1270                 return false;
1271
1272         return true;
1273 }
1274
1275 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1276 {
1277         /* Read a config register to help see what died. */
1278         u16 pci_status;
1279         int result;
1280
1281         result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1282                                       &pci_status);
1283         if (result == PCIBIOS_SUCCESSFUL)
1284                 dev_warn(dev->dev,
1285                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1286                          csts, pci_status);
1287         else
1288                 dev_warn(dev->dev,
1289                          "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1290                          csts, result);
1291 }
1292
1293 static void nvme_watchdog_timer(unsigned long data)
1294 {
1295         struct nvme_dev *dev = (struct nvme_dev *)data;
1296         u32 csts = readl(dev->bar + NVME_REG_CSTS);
1297
1298         /* Skip controllers under certain specific conditions. */
1299         if (nvme_should_reset(dev, csts)) {
1300                 if (!nvme_reset(dev))
1301                         nvme_warn_reset(dev, csts);
1302                 return;
1303         }
1304
1305         mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1306 }
1307
1308 static int nvme_create_io_queues(struct nvme_dev *dev)
1309 {
1310         unsigned i, max;
1311         int ret = 0;
1312
1313         for (i = dev->queue_count; i <= dev->max_qid; i++) {
1314                 if (!nvme_alloc_queue(dev, i, dev->q_depth)) {
1315                         ret = -ENOMEM;
1316                         break;
1317                 }
1318         }
1319
1320         max = min(dev->max_qid, dev->queue_count - 1);
1321         for (i = dev->online_queues; i <= max; i++) {
1322                 ret = nvme_create_queue(dev->queues[i], i);
1323                 if (ret)
1324                         break;
1325         }
1326
1327         /*
1328          * Ignore failing Create SQ/CQ commands, we can continue with less
1329          * than the desired aount of queues, and even a controller without
1330          * I/O queues an still be used to issue admin commands.  This might
1331          * be useful to upgrade a buggy firmware for example.
1332          */
1333         return ret >= 0 ? 0 : ret;
1334 }
1335
1336 static ssize_t nvme_cmb_show(struct device *dev,
1337                              struct device_attribute *attr,
1338                              char *buf)
1339 {
1340         struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1341
1342         return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1343                        ndev->cmbloc, ndev->cmbsz);
1344 }
1345 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1346
1347 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1348 {
1349         u64 szu, size, offset;
1350         resource_size_t bar_size;
1351         struct pci_dev *pdev = to_pci_dev(dev->dev);
1352         void __iomem *cmb;
1353         dma_addr_t dma_addr;
1354
1355         dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1356         if (!(NVME_CMB_SZ(dev->cmbsz)))
1357                 return NULL;
1358         dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1359
1360         if (!use_cmb_sqes)
1361                 return NULL;
1362
1363         szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1364         size = szu * NVME_CMB_SZ(dev->cmbsz);
1365         offset = szu * NVME_CMB_OFST(dev->cmbloc);
1366         bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1367
1368         if (offset > bar_size)
1369                 return NULL;
1370
1371         /*
1372          * Controllers may support a CMB size larger than their BAR,
1373          * for example, due to being behind a bridge. Reduce the CMB to
1374          * the reported size of the BAR
1375          */
1376         if (size > bar_size - offset)
1377                 size = bar_size - offset;
1378
1379         dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1380         cmb = ioremap_wc(dma_addr, size);
1381         if (!cmb)
1382                 return NULL;
1383
1384         dev->cmb_dma_addr = dma_addr;
1385         dev->cmb_size = size;
1386         return cmb;
1387 }
1388
1389 static inline void nvme_release_cmb(struct nvme_dev *dev)
1390 {
1391         if (dev->cmb) {
1392                 iounmap(dev->cmb);
1393                 dev->cmb = NULL;
1394         }
1395 }
1396
1397 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1398 {
1399         return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1400 }
1401
1402 static int nvme_setup_io_queues(struct nvme_dev *dev)
1403 {
1404         struct nvme_queue *adminq = dev->queues[0];
1405         struct pci_dev *pdev = to_pci_dev(dev->dev);
1406         int result, nr_io_queues, size;
1407
1408         nr_io_queues = num_online_cpus();
1409         result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1410         if (result < 0)
1411                 return result;
1412
1413         if (nr_io_queues == 0)
1414                 return 0;
1415
1416         if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1417                 result = nvme_cmb_qdepth(dev, nr_io_queues,
1418                                 sizeof(struct nvme_command));
1419                 if (result > 0)
1420                         dev->q_depth = result;
1421                 else
1422                         nvme_release_cmb(dev);
1423         }
1424
1425         size = db_bar_size(dev, nr_io_queues);
1426         if (size > 8192) {
1427                 iounmap(dev->bar);
1428                 do {
1429                         dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1430                         if (dev->bar)
1431                                 break;
1432                         if (!--nr_io_queues)
1433                                 return -ENOMEM;
1434                         size = db_bar_size(dev, nr_io_queues);
1435                 } while (1);
1436                 dev->dbs = dev->bar + 4096;
1437                 adminq->q_db = dev->dbs;
1438         }
1439
1440         /* Deregister the admin queue's interrupt */
1441         free_irq(pci_irq_vector(pdev, 0), adminq);
1442
1443         /*
1444          * If we enable msix early due to not intx, disable it again before
1445          * setting up the full range we need.
1446          */
1447         pci_free_irq_vectors(pdev);
1448         nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1449                         PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1450         if (nr_io_queues <= 0)
1451                 return -EIO;
1452         dev->max_qid = nr_io_queues;
1453
1454         /*
1455          * Should investigate if there's a performance win from allocating
1456          * more queues than interrupt vectors; it might allow the submission
1457          * path to scale better, even if the receive path is limited by the
1458          * number of interrupts.
1459          */
1460
1461         result = queue_request_irq(adminq);
1462         if (result) {
1463                 adminq->cq_vector = -1;
1464                 return result;
1465         }
1466         return nvme_create_io_queues(dev);
1467 }
1468
1469 static void nvme_del_queue_end(struct request *req, int error)
1470 {
1471         struct nvme_queue *nvmeq = req->end_io_data;
1472
1473         blk_mq_free_request(req);
1474         complete(&nvmeq->dev->ioq_wait);
1475 }
1476
1477 static void nvme_del_cq_end(struct request *req, int error)
1478 {
1479         struct nvme_queue *nvmeq = req->end_io_data;
1480
1481         if (!error) {
1482                 unsigned long flags;
1483
1484                 /*
1485                  * We might be called with the AQ q_lock held
1486                  * and the I/O queue q_lock should always
1487                  * nest inside the AQ one.
1488                  */
1489                 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1490                                         SINGLE_DEPTH_NESTING);
1491                 nvme_process_cq(nvmeq);
1492                 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1493         }
1494
1495         nvme_del_queue_end(req, error);
1496 }
1497
1498 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1499 {
1500         struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1501         struct request *req;
1502         struct nvme_command cmd;
1503
1504         memset(&cmd, 0, sizeof(cmd));
1505         cmd.delete_queue.opcode = opcode;
1506         cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1507
1508         req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1509         if (IS_ERR(req))
1510                 return PTR_ERR(req);
1511
1512         req->timeout = ADMIN_TIMEOUT;
1513         req->end_io_data = nvmeq;
1514
1515         blk_execute_rq_nowait(q, NULL, req, false,
1516                         opcode == nvme_admin_delete_cq ?
1517                                 nvme_del_cq_end : nvme_del_queue_end);
1518         return 0;
1519 }
1520
1521 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1522 {
1523         int pass;
1524         unsigned long timeout;
1525         u8 opcode = nvme_admin_delete_sq;
1526
1527         for (pass = 0; pass < 2; pass++) {
1528                 int sent = 0, i = queues;
1529
1530                 reinit_completion(&dev->ioq_wait);
1531  retry:
1532                 timeout = ADMIN_TIMEOUT;
1533                 for (; i > 0; i--, sent++)
1534                         if (nvme_delete_queue(dev->queues[i], opcode))
1535                                 break;
1536
1537                 while (sent--) {
1538                         timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1539                         if (timeout == 0)
1540                                 return;
1541                         if (i)
1542                                 goto retry;
1543                 }
1544                 opcode = nvme_admin_delete_cq;
1545         }
1546 }
1547
1548 /*
1549  * Return: error value if an error occurred setting up the queues or calling
1550  * Identify Device.  0 if these succeeded, even if adding some of the
1551  * namespaces failed.  At the moment, these failures are silent.  TBD which
1552  * failures should be reported.
1553  */
1554 static int nvme_dev_add(struct nvme_dev *dev)
1555 {
1556         if (!dev->ctrl.tagset) {
1557                 dev->tagset.ops = &nvme_mq_ops;
1558                 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1559                 dev->tagset.timeout = NVME_IO_TIMEOUT;
1560                 dev->tagset.numa_node = dev_to_node(dev->dev);
1561                 dev->tagset.queue_depth =
1562                                 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1563                 dev->tagset.cmd_size = nvme_cmd_size(dev);
1564                 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1565                 dev->tagset.driver_data = dev;
1566
1567                 if (blk_mq_alloc_tag_set(&dev->tagset))
1568                         return 0;
1569                 dev->ctrl.tagset = &dev->tagset;
1570         } else {
1571                 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1572
1573                 /* Free previously allocated queues that are no longer usable */
1574                 nvme_free_queues(dev, dev->online_queues);
1575         }
1576
1577         return 0;
1578 }
1579
1580 static int nvme_pci_enable(struct nvme_dev *dev)
1581 {
1582         u64 cap;
1583         int result = -ENOMEM;
1584         struct pci_dev *pdev = to_pci_dev(dev->dev);
1585
1586         if (pci_enable_device_mem(pdev))
1587                 return result;
1588
1589         pci_set_master(pdev);
1590
1591         if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1592             dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1593                 goto disable;
1594
1595         if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1596                 result = -ENODEV;
1597                 goto disable;
1598         }
1599
1600         /*
1601          * Some devices and/or platforms don't advertise or work with INTx
1602          * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1603          * adjust this later.
1604          */
1605         result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1606         if (result < 0)
1607                 return result;
1608
1609         cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1610
1611         dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1612         dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1613         dev->dbs = dev->bar + 4096;
1614
1615         /*
1616          * Temporary fix for the Apple controller found in the MacBook8,1 and
1617          * some MacBook7,1 to avoid controller resets and data loss.
1618          */
1619         if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1620                 dev->q_depth = 2;
1621                 dev_warn(dev->dev, "detected Apple NVMe controller, set "
1622                         "queue depth=%u to work around controller resets\n",
1623                         dev->q_depth);
1624         }
1625
1626         /*
1627          * CMBs can currently only exist on >=1.2 PCIe devices. We only
1628          * populate sysfs if a CMB is implemented. Note that we add the
1629          * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1630          * it on exit. Since nvme_dev_attrs_group has no name we can pass
1631          * NULL as final argument to sysfs_add_file_to_group.
1632          */
1633
1634         if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1635                 dev->cmb = nvme_map_cmb(dev);
1636
1637                 if (dev->cmbsz) {
1638                         if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1639                                                     &dev_attr_cmb.attr, NULL))
1640                                 dev_warn(dev->dev,
1641                                          "failed to add sysfs attribute for CMB\n");
1642                 }
1643         }
1644
1645         pci_enable_pcie_error_reporting(pdev);
1646         pci_save_state(pdev);
1647         return 0;
1648
1649  disable:
1650         pci_disable_device(pdev);
1651         return result;
1652 }
1653
1654 static void nvme_dev_unmap(struct nvme_dev *dev)
1655 {
1656         if (dev->bar)
1657                 iounmap(dev->bar);
1658         pci_release_mem_regions(to_pci_dev(dev->dev));
1659 }
1660
1661 static void nvme_pci_disable(struct nvme_dev *dev)
1662 {
1663         struct pci_dev *pdev = to_pci_dev(dev->dev);
1664
1665         pci_free_irq_vectors(pdev);
1666
1667         if (pci_is_enabled(pdev)) {
1668                 pci_disable_pcie_error_reporting(pdev);
1669                 pci_disable_device(pdev);
1670         }
1671 }
1672
1673 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1674 {
1675         int i, queues;
1676         u32 csts = -1;
1677
1678         del_timer_sync(&dev->watchdog_timer);
1679
1680         mutex_lock(&dev->shutdown_lock);
1681         if (pci_is_enabled(to_pci_dev(dev->dev))) {
1682                 nvme_stop_queues(&dev->ctrl);
1683                 csts = readl(dev->bar + NVME_REG_CSTS);
1684         }
1685
1686         queues = dev->online_queues - 1;
1687         for (i = dev->queue_count - 1; i > 0; i--)
1688                 nvme_suspend_queue(dev->queues[i]);
1689
1690         if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
1691                 /* A device might become IO incapable very soon during
1692                  * probe, before the admin queue is configured. Thus,
1693                  * queue_count can be 0 here.
1694                  */
1695                 if (dev->queue_count)
1696                         nvme_suspend_queue(dev->queues[0]);
1697         } else {
1698                 nvme_disable_io_queues(dev, queues);
1699                 nvme_disable_admin_queue(dev, shutdown);
1700         }
1701         nvme_pci_disable(dev);
1702
1703         blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
1704         blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
1705         mutex_unlock(&dev->shutdown_lock);
1706 }
1707
1708 static int nvme_setup_prp_pools(struct nvme_dev *dev)
1709 {
1710         dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
1711                                                 PAGE_SIZE, PAGE_SIZE, 0);
1712         if (!dev->prp_page_pool)
1713                 return -ENOMEM;
1714
1715         /* Optimisation for I/Os between 4k and 128k */
1716         dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
1717                                                 256, 256, 0);
1718         if (!dev->prp_small_pool) {
1719                 dma_pool_destroy(dev->prp_page_pool);
1720                 return -ENOMEM;
1721         }
1722         return 0;
1723 }
1724
1725 static void nvme_release_prp_pools(struct nvme_dev *dev)
1726 {
1727         dma_pool_destroy(dev->prp_page_pool);
1728         dma_pool_destroy(dev->prp_small_pool);
1729 }
1730
1731 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
1732 {
1733         struct nvme_dev *dev = to_nvme_dev(ctrl);
1734
1735         put_device(dev->dev);
1736         if (dev->tagset.tags)
1737                 blk_mq_free_tag_set(&dev->tagset);
1738         if (dev->ctrl.admin_q)
1739                 blk_put_queue(dev->ctrl.admin_q);
1740         kfree(dev->queues);
1741         kfree(dev);
1742 }
1743
1744 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
1745 {
1746         dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
1747
1748         kref_get(&dev->ctrl.kref);
1749         nvme_dev_disable(dev, false);
1750         if (!schedule_work(&dev->remove_work))
1751                 nvme_put_ctrl(&dev->ctrl);
1752 }
1753
1754 static void nvme_reset_work(struct work_struct *work)
1755 {
1756         struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
1757         int result = -ENODEV;
1758
1759         if (WARN_ON(dev->ctrl.state == NVME_CTRL_RESETTING))
1760                 goto out;
1761
1762         /*
1763          * If we're called to reset a live controller first shut it down before
1764          * moving on.
1765          */
1766         if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
1767                 nvme_dev_disable(dev, false);
1768
1769         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
1770                 goto out;
1771
1772         result = nvme_pci_enable(dev);
1773         if (result)
1774                 goto out;
1775
1776         result = nvme_configure_admin_queue(dev);
1777         if (result)
1778                 goto out;
1779
1780         nvme_init_queue(dev->queues[0], 0);
1781         result = nvme_alloc_admin_tags(dev);
1782         if (result)
1783                 goto out;
1784
1785         result = nvme_init_identify(&dev->ctrl);
1786         if (result)
1787                 goto out;
1788
1789         result = nvme_setup_io_queues(dev);
1790         if (result)
1791                 goto out;
1792
1793         /*
1794          * A controller that can not execute IO typically requires user
1795          * intervention to correct. For such degraded controllers, the driver
1796          * should not submit commands the user did not request, so skip
1797          * registering for asynchronous event notification on this condition.
1798          */
1799         if (dev->online_queues > 1)
1800                 nvme_queue_async_events(&dev->ctrl);
1801
1802         mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1803
1804         /*
1805          * Keep the controller around but remove all namespaces if we don't have
1806          * any working I/O queue.
1807          */
1808         if (dev->online_queues < 2) {
1809                 dev_warn(dev->ctrl.device, "IO queues not created\n");
1810                 nvme_kill_queues(&dev->ctrl);
1811                 nvme_remove_namespaces(&dev->ctrl);
1812         } else {
1813                 nvme_start_queues(&dev->ctrl);
1814                 nvme_dev_add(dev);
1815         }
1816
1817         if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
1818                 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
1819                 goto out;
1820         }
1821
1822         if (dev->online_queues > 1)
1823                 nvme_queue_scan(&dev->ctrl);
1824         return;
1825
1826  out:
1827         nvme_remove_dead_ctrl(dev, result);
1828 }
1829
1830 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
1831 {
1832         struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
1833         struct pci_dev *pdev = to_pci_dev(dev->dev);
1834
1835         nvme_kill_queues(&dev->ctrl);
1836         if (pci_get_drvdata(pdev))
1837                 device_release_driver(&pdev->dev);
1838         nvme_put_ctrl(&dev->ctrl);
1839 }
1840
1841 static int nvme_reset(struct nvme_dev *dev)
1842 {
1843         if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
1844                 return -ENODEV;
1845         if (work_busy(&dev->reset_work))
1846                 return -ENODEV;
1847         if (!queue_work(nvme_workq, &dev->reset_work))
1848                 return -EBUSY;
1849         return 0;
1850 }
1851
1852 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
1853 {
1854         *val = readl(to_nvme_dev(ctrl)->bar + off);
1855         return 0;
1856 }
1857
1858 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
1859 {
1860         writel(val, to_nvme_dev(ctrl)->bar + off);
1861         return 0;
1862 }
1863
1864 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
1865 {
1866         *val = readq(to_nvme_dev(ctrl)->bar + off);
1867         return 0;
1868 }
1869
1870 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
1871 {
1872         struct nvme_dev *dev = to_nvme_dev(ctrl);
1873         int ret = nvme_reset(dev);
1874
1875         if (!ret)
1876                 flush_work(&dev->reset_work);
1877         return ret;
1878 }
1879
1880 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
1881         .name                   = "pcie",
1882         .module                 = THIS_MODULE,
1883         .reg_read32             = nvme_pci_reg_read32,
1884         .reg_write32            = nvme_pci_reg_write32,
1885         .reg_read64             = nvme_pci_reg_read64,
1886         .reset_ctrl             = nvme_pci_reset_ctrl,
1887         .free_ctrl              = nvme_pci_free_ctrl,
1888         .submit_async_event     = nvme_pci_submit_async_event,
1889 };
1890
1891 static int nvme_dev_map(struct nvme_dev *dev)
1892 {
1893         struct pci_dev *pdev = to_pci_dev(dev->dev);
1894
1895         if (pci_request_mem_regions(pdev, "nvme"))
1896                 return -ENODEV;
1897
1898         dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1899         if (!dev->bar)
1900                 goto release;
1901
1902         return 0;
1903   release:
1904         pci_release_mem_regions(pdev);
1905         return -ENODEV;
1906 }
1907
1908 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1909 {
1910         int node, result = -ENOMEM;
1911         struct nvme_dev *dev;
1912
1913         node = dev_to_node(&pdev->dev);
1914         if (node == NUMA_NO_NODE)
1915                 set_dev_node(&pdev->dev, first_memory_node);
1916
1917         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
1918         if (!dev)
1919                 return -ENOMEM;
1920         dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
1921                                                         GFP_KERNEL, node);
1922         if (!dev->queues)
1923                 goto free;
1924
1925         dev->dev = get_device(&pdev->dev);
1926         pci_set_drvdata(pdev, dev);
1927
1928         result = nvme_dev_map(dev);
1929         if (result)
1930                 goto free;
1931
1932         INIT_WORK(&dev->reset_work, nvme_reset_work);
1933         INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
1934         setup_timer(&dev->watchdog_timer, nvme_watchdog_timer,
1935                 (unsigned long)dev);
1936         mutex_init(&dev->shutdown_lock);
1937         init_completion(&dev->ioq_wait);
1938
1939         result = nvme_setup_prp_pools(dev);
1940         if (result)
1941                 goto put_pci;
1942
1943         result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
1944                         id->driver_data);
1945         if (result)
1946                 goto release_pools;
1947
1948         dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
1949
1950         queue_work(nvme_workq, &dev->reset_work);
1951         return 0;
1952
1953  release_pools:
1954         nvme_release_prp_pools(dev);
1955  put_pci:
1956         put_device(dev->dev);
1957         nvme_dev_unmap(dev);
1958  free:
1959         kfree(dev->queues);
1960         kfree(dev);
1961         return result;
1962 }
1963
1964 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
1965 {
1966         struct nvme_dev *dev = pci_get_drvdata(pdev);
1967
1968         if (prepare)
1969                 nvme_dev_disable(dev, false);
1970         else
1971                 nvme_reset(dev);
1972 }
1973
1974 static void nvme_shutdown(struct pci_dev *pdev)
1975 {
1976         struct nvme_dev *dev = pci_get_drvdata(pdev);
1977         nvme_dev_disable(dev, true);
1978 }
1979
1980 /*
1981  * The driver's remove may be called on a device in a partially initialized
1982  * state. This function must not have any dependencies on the device state in
1983  * order to proceed.
1984  */
1985 static void nvme_remove(struct pci_dev *pdev)
1986 {
1987         struct nvme_dev *dev = pci_get_drvdata(pdev);
1988
1989         nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1990
1991         pci_set_drvdata(pdev, NULL);
1992
1993         if (!pci_device_is_present(pdev))
1994                 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
1995
1996         flush_work(&dev->reset_work);
1997         nvme_uninit_ctrl(&dev->ctrl);
1998         nvme_dev_disable(dev, true);
1999         nvme_dev_remove_admin(dev);
2000         nvme_free_queues(dev, 0);
2001         nvme_release_cmb(dev);
2002         nvme_release_prp_pools(dev);
2003         nvme_dev_unmap(dev);
2004         nvme_put_ctrl(&dev->ctrl);
2005 }
2006
2007 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2008 {
2009         int ret = 0;
2010
2011         if (numvfs == 0) {
2012                 if (pci_vfs_assigned(pdev)) {
2013                         dev_warn(&pdev->dev,
2014                                 "Cannot disable SR-IOV VFs while assigned\n");
2015                         return -EPERM;
2016                 }
2017                 pci_disable_sriov(pdev);
2018                 return 0;
2019         }
2020
2021         ret = pci_enable_sriov(pdev, numvfs);
2022         return ret ? ret : numvfs;
2023 }
2024
2025 #ifdef CONFIG_PM_SLEEP
2026 static int nvme_suspend(struct device *dev)
2027 {
2028         struct pci_dev *pdev = to_pci_dev(dev);
2029         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2030
2031         nvme_dev_disable(ndev, true);
2032         return 0;
2033 }
2034
2035 static int nvme_resume(struct device *dev)
2036 {
2037         struct pci_dev *pdev = to_pci_dev(dev);
2038         struct nvme_dev *ndev = pci_get_drvdata(pdev);
2039
2040         nvme_reset(ndev);
2041         return 0;
2042 }
2043 #endif
2044
2045 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2046
2047 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2048                                                 pci_channel_state_t state)
2049 {
2050         struct nvme_dev *dev = pci_get_drvdata(pdev);
2051
2052         /*
2053          * A frozen channel requires a reset. When detected, this method will
2054          * shutdown the controller to quiesce. The controller will be restarted
2055          * after the slot reset through driver's slot_reset callback.
2056          */
2057         switch (state) {
2058         case pci_channel_io_normal:
2059                 return PCI_ERS_RESULT_CAN_RECOVER;
2060         case pci_channel_io_frozen:
2061                 dev_warn(dev->ctrl.device,
2062                         "frozen state error detected, reset controller\n");
2063                 nvme_dev_disable(dev, false);
2064                 return PCI_ERS_RESULT_NEED_RESET;
2065         case pci_channel_io_perm_failure:
2066                 dev_warn(dev->ctrl.device,
2067                         "failure state error detected, request disconnect\n");
2068                 return PCI_ERS_RESULT_DISCONNECT;
2069         }
2070         return PCI_ERS_RESULT_NEED_RESET;
2071 }
2072
2073 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2074 {
2075         struct nvme_dev *dev = pci_get_drvdata(pdev);
2076
2077         dev_info(dev->ctrl.device, "restart after slot reset\n");
2078         pci_restore_state(pdev);
2079         nvme_reset(dev);
2080         return PCI_ERS_RESULT_RECOVERED;
2081 }
2082
2083 static void nvme_error_resume(struct pci_dev *pdev)
2084 {
2085         pci_cleanup_aer_uncorrect_error_status(pdev);
2086 }
2087
2088 static const struct pci_error_handlers nvme_err_handler = {
2089         .error_detected = nvme_error_detected,
2090         .slot_reset     = nvme_slot_reset,
2091         .resume         = nvme_error_resume,
2092         .reset_notify   = nvme_reset_notify,
2093 };
2094
2095 static const struct pci_device_id nvme_id_table[] = {
2096         { PCI_VDEVICE(INTEL, 0x0953),
2097                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2098                                 NVME_QUIRK_DISCARD_ZEROES, },
2099         { PCI_VDEVICE(INTEL, 0x0a53),
2100                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2101                                 NVME_QUIRK_DISCARD_ZEROES, },
2102         { PCI_VDEVICE(INTEL, 0x0a54),
2103                 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2104                                 NVME_QUIRK_DISCARD_ZEROES, },
2105         { PCI_VDEVICE(INTEL, 0x5845),   /* Qemu emulated controller */
2106                 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2107         { PCI_DEVICE(0x1c58, 0x0003),   /* HGST adapter */
2108                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2109         { PCI_DEVICE(0x1c5f, 0x0540),   /* Memblaze Pblaze4 adapter */
2110                 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2111         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2112         { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2113         { 0, }
2114 };
2115 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2116
2117 static struct pci_driver nvme_driver = {
2118         .name           = "nvme",
2119         .id_table       = nvme_id_table,
2120         .probe          = nvme_probe,
2121         .remove         = nvme_remove,
2122         .shutdown       = nvme_shutdown,
2123         .driver         = {
2124                 .pm     = &nvme_dev_pm_ops,
2125         },
2126         .sriov_configure = nvme_pci_sriov_configure,
2127         .err_handler    = &nvme_err_handler,
2128 };
2129
2130 static int __init nvme_init(void)
2131 {
2132         int result;
2133
2134         nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
2135         if (!nvme_workq)
2136                 return -ENOMEM;
2137
2138         result = pci_register_driver(&nvme_driver);
2139         if (result)
2140                 destroy_workqueue(nvme_workq);
2141         return result;
2142 }
2143
2144 static void __exit nvme_exit(void)
2145 {
2146         pci_unregister_driver(&nvme_driver);
2147         destroy_workqueue(nvme_workq);
2148         _nvme_check_size();
2149 }
2150
2151 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2152 MODULE_LICENSE("GPL");
2153 MODULE_VERSION("1.0");
2154 module_init(nvme_init);
2155 module_exit(nvme_exit);