drivers/block/virtio_blk.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 //#define DEBUG
   3 #include <linux/spinlock.h>
   4 #include <linux/slab.h>
   5 #include <linux/blkdev.h>
   6 #include <linux/hdreg.h>
   7 #include <linux/module.h>
   8 #include <linux/mutex.h>
   9 #include <linux/interrupt.h>
  10 #include <linux/virtio.h>
  11 #include <linux/virtio_blk.h>
  12 #include <linux/scatterlist.h>
  13 #include <linux/string_helpers.h>
  14 #include <linux/idr.h>
  15 #include <linux/blk-mq.h>
  16 #include <linux/blk-mq-virtio.h>
  17 #include <linux/numa.h>
  18 #include <uapi/linux/virtio_ring.h>
  19
  20 #define PART_BITS 4
  21 #define VQ_NAME_LEN 16
  22 #define MAX_DISCARD_SEGMENTS 256u
  23
  24 /* The maximum number of sg elements that fit into a virtqueue */
  25 #define VIRTIO_BLK_MAX_SG_ELEMS 32768
  26
  27 #ifdef CONFIG_ARCH_NO_SG_CHAIN
  28 #define VIRTIO_BLK_INLINE_SG_CNT        0
  29 #else
  30 #define VIRTIO_BLK_INLINE_SG_CNT        2
  31 #endif
  32
  33 static unsigned int num_request_queues;
  34 module_param(num_request_queues, uint, 0644);
  35 MODULE_PARM_DESC(num_request_queues,
  36                  "Limit the number of request queues to use for blk device. "
  37                  "0 for no limit. "
  38                  "Values > nr_cpu_ids truncated to nr_cpu_ids.");
  39
  40 static unsigned int poll_queues;
  41 module_param(poll_queues, uint, 0644);
  42 MODULE_PARM_DESC(poll_queues, "The number of dedicated virtqueues for polling I/O");
  43
  44 static int major;
  45 static DEFINE_IDA(vd_index_ida);
  46
  47 static struct workqueue_struct *virtblk_wq;
  48
  49 struct virtio_blk_vq {
  50         struct virtqueue *vq;
  51         spinlock_t lock;
  52         char name[VQ_NAME_LEN];
  53 } ____cacheline_aligned_in_smp;
  54
  55 struct virtio_blk {
  56         /*
  57          * This mutex must be held by anything that may run after
  58          * virtblk_remove() sets vblk->vdev to NULL.
  59          *
  60          * blk-mq, virtqueue processing, and sysfs attribute code paths are
  61          * shut down before vblk->vdev is set to NULL and therefore do not need
  62          * to hold this mutex.
  63          */
  64         struct mutex vdev_mutex;
  65         struct virtio_device *vdev;
  66
  67         /* The disk structure for the kernel. */
  68         struct gendisk *disk;
  69
  70         /* Block layer tags. */
  71         struct blk_mq_tag_set tag_set;
  72
  73         /* Process context for config space updates */
  74         struct work_struct config_work;
  75
  76         /* Ida index - used to track minor number allocations. */
  77         int index;
  78
  79         /* num of vqs */
  80         int num_vqs;
  81         int io_queues[HCTX_MAX_TYPES];
  82         struct virtio_blk_vq *vqs;
  83 };
  84
  85 struct virtblk_req {
  86         struct virtio_blk_outhdr out_hdr;
  87         u8 status;
  88         struct sg_table sg_table;
  89         struct scatterlist sg[];
  90 };
  91
  92 static inline blk_status_t virtblk_result(struct virtblk_req *vbr)
  93 {
  94         switch (vbr->status) {
  95         case VIRTIO_BLK_S_OK:
  96                 return BLK_STS_OK;
  97         case VIRTIO_BLK_S_UNSUPP:
  98                 return BLK_STS_NOTSUPP;
  99         default:
 100                 return BLK_STS_IOERR;
 101         }
 102 }
 103
 104 static int virtblk_add_req(struct virtqueue *vq, struct virtblk_req *vbr)
 105 {
 106         struct scatterlist hdr, status, *sgs[3];
 107         unsigned int num_out = 0, num_in = 0;
 108
 109         sg_init_one(&hdr, &vbr->out_hdr, sizeof(vbr->out_hdr));
 110         sgs[num_out++] = &hdr;
 111
 112         if (vbr->sg_table.nents) {
 113                 if (vbr->out_hdr.type & cpu_to_virtio32(vq->vdev, VIRTIO_BLK_T_OUT))
 114                         sgs[num_out++] = vbr->sg_table.sgl;
 115                 else
 116                         sgs[num_out + num_in++] = vbr->sg_table.sgl;
 117         }
 118
 119         sg_init_one(&status, &vbr->status, sizeof(vbr->status));
 120         sgs[num_out + num_in++] = &status;
 121
 122         return virtqueue_add_sgs(vq, sgs, num_out, num_in, vbr, GFP_ATOMIC);
 123 }
 124
 125 static int virtblk_setup_discard_write_zeroes(struct request *req, bool unmap)
 126 {
 127         unsigned short segments = blk_rq_nr_discard_segments(req);
 128         unsigned short n = 0;
 129         struct virtio_blk_discard_write_zeroes *range;
 130         struct bio *bio;
 131         u32 flags = 0;
 132
 133         if (unmap)
 134                 flags |= VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP;
 135
 136         range = kmalloc_array(segments, sizeof(*range), GFP_ATOMIC);
 137         if (!range)
 138                 return -ENOMEM;
 139
 140         /*
 141          * Single max discard segment means multi-range discard isn't
 142          * supported, and block layer only runs contiguity merge like
 143          * normal RW request. So we can't reply on bio for retrieving
 144          * each range info.
 145          */
 146         if (queue_max_discard_segments(req->q) == 1) {
 147                 range[0].flags = cpu_to_le32(flags);
 148                 range[0].num_sectors = cpu_to_le32(blk_rq_sectors(req));
 149                 range[0].sector = cpu_to_le64(blk_rq_pos(req));
 150                 n = 1;
 151         } else {
 152                 __rq_for_each_bio(bio, req) {
 153                         u64 sector = bio->bi_iter.bi_sector;
 154                         u32 num_sectors = bio->bi_iter.bi_size >> SECTOR_SHIFT;
 155
 156                         range[n].flags = cpu_to_le32(flags);
 157                         range[n].num_sectors = cpu_to_le32(num_sectors);
 158                         range[n].sector = cpu_to_le64(sector);
 159                         n++;
 160                 }
 161         }
 162
 163         WARN_ON_ONCE(n != segments);
 164
 165         req->special_vec.bv_page = virt_to_page(range);
 166         req->special_vec.bv_offset = offset_in_page(range);
 167         req->special_vec.bv_len = sizeof(*range) * segments;
 168         req->rq_flags |= RQF_SPECIAL_PAYLOAD;
 169
 170         return 0;
 171 }
 172
 173 static void virtblk_unmap_data(struct request *req, struct virtblk_req *vbr)
 174 {
 175         if (blk_rq_nr_phys_segments(req))
 176                 sg_free_table_chained(&vbr->sg_table,
 177                                       VIRTIO_BLK_INLINE_SG_CNT);
 178 }
 179
 180 static int virtblk_map_data(struct blk_mq_hw_ctx *hctx, struct request *req,
 181                 struct virtblk_req *vbr)
 182 {
 183         int err;
 184
 185         if (!blk_rq_nr_phys_segments(req))
 186                 return 0;
 187
 188         vbr->sg_table.sgl = vbr->sg;
 189         err = sg_alloc_table_chained(&vbr->sg_table,
 190                                      blk_rq_nr_phys_segments(req),
 191                                      vbr->sg_table.sgl,
 192                                      VIRTIO_BLK_INLINE_SG_CNT);
 193         if (unlikely(err))
 194                 return -ENOMEM;
 195
 196         return blk_rq_map_sg(hctx->queue, req, vbr->sg_table.sgl);
 197 }
 198
 199 static void virtblk_cleanup_cmd(struct request *req)
 200 {
 201         if (req->rq_flags & RQF_SPECIAL_PAYLOAD)
 202                 kfree(bvec_virt(&req->special_vec));
 203 }
 204
 205 static blk_status_t virtblk_setup_cmd(struct virtio_device *vdev,
 206                                       struct request *req,
 207                                       struct virtblk_req *vbr)
 208 {
 209         bool unmap = false;
 210         u32 type;
 211
 212         vbr->out_hdr.sector = 0;
 213
 214         switch (req_op(req)) {
 215         case REQ_OP_READ:
 216                 type = VIRTIO_BLK_T_IN;
 217                 vbr->out_hdr.sector = cpu_to_virtio64(vdev,
 218                                                       blk_rq_pos(req));
 219                 break;
 220         case REQ_OP_WRITE:
 221                 type = VIRTIO_BLK_T_OUT;
 222                 vbr->out_hdr.sector = cpu_to_virtio64(vdev,
 223                                                       blk_rq_pos(req));
 224                 break;
 225         case REQ_OP_FLUSH:
 226                 type = VIRTIO_BLK_T_FLUSH;
 227                 break;
 228         case REQ_OP_DISCARD:
 229                 type = VIRTIO_BLK_T_DISCARD;
 230                 break;
 231         case REQ_OP_WRITE_ZEROES:
 232                 type = VIRTIO_BLK_T_WRITE_ZEROES;
 233                 unmap = !(req->cmd_flags & REQ_NOUNMAP);
 234                 break;
 235         case REQ_OP_DRV_IN:
 236                 type = VIRTIO_BLK_T_GET_ID;
 237                 break;
 238         default:
 239                 WARN_ON_ONCE(1);
 240                 return BLK_STS_IOERR;
 241         }
 242
 243         vbr->out_hdr.type = cpu_to_virtio32(vdev, type);
 244         vbr->out_hdr.ioprio = cpu_to_virtio32(vdev, req_get_ioprio(req));
 245
 246         if (type == VIRTIO_BLK_T_DISCARD || type == VIRTIO_BLK_T_WRITE_ZEROES) {
 247                 if (virtblk_setup_discard_write_zeroes(req, unmap))
 248                         return BLK_STS_RESOURCE;
 249         }
 250
 251         return 0;
 252 }
 253
 254 static inline void virtblk_request_done(struct request *req)
 255 {
 256         struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
 257
 258         virtblk_unmap_data(req, vbr);
 259         virtblk_cleanup_cmd(req);
 260         blk_mq_end_request(req, virtblk_result(vbr));
 261 }
 262
 263 static void virtblk_done(struct virtqueue *vq)
 264 {
 265         struct virtio_blk *vblk = vq->vdev->priv;
 266         bool req_done = false;
 267         int qid = vq->index;
 268         struct virtblk_req *vbr;
 269         unsigned long flags;
 270         unsigned int len;
 271
 272         spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
 273         do {
 274                 virtqueue_disable_cb(vq);
 275                 while ((vbr = virtqueue_get_buf(vblk->vqs[qid].vq, &len)) != NULL) {
 276                         struct request *req = blk_mq_rq_from_pdu(vbr);
 277
 278                         if (likely(!blk_should_fake_timeout(req->q)))
 279                                 blk_mq_complete_request(req);
 280                         req_done = true;
 281                 }
 282                 if (unlikely(virtqueue_is_broken(vq)))
 283                         break;
 284         } while (!virtqueue_enable_cb(vq));
 285
 286         /* In case queue is stopped waiting for more buffers. */
 287         if (req_done)
 288                 blk_mq_start_stopped_hw_queues(vblk->disk->queue, true);
 289         spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
 290 }
 291
 292 static void virtio_commit_rqs(struct blk_mq_hw_ctx *hctx)
 293 {
 294         struct virtio_blk *vblk = hctx->queue->queuedata;
 295         struct virtio_blk_vq *vq = &vblk->vqs[hctx->queue_num];
 296         bool kick;
 297
 298         spin_lock_irq(&vq->lock);
 299         kick = virtqueue_kick_prepare(vq->vq);
 300         spin_unlock_irq(&vq->lock);
 301
 302         if (kick)
 303                 virtqueue_notify(vq->vq);
 304 }
 305
 306 static blk_status_t virtblk_prep_rq(struct blk_mq_hw_ctx *hctx,
 307                                         struct virtio_blk *vblk,
 308                                         struct request *req,
 309                                         struct virtblk_req *vbr)
 310 {
 311         blk_status_t status;
 312
 313         status = virtblk_setup_cmd(vblk->vdev, req, vbr);
 314         if (unlikely(status))
 315                 return status;
 316
 317         blk_mq_start_request(req);
 318
 319         vbr->sg_table.nents = virtblk_map_data(hctx, req, vbr);
 320         if (unlikely(vbr->sg_table.nents < 0)) {
 321                 virtblk_cleanup_cmd(req);
 322                 return BLK_STS_RESOURCE;
 323         }
 324
 325         return BLK_STS_OK;
 326 }
 327
 328 static blk_status_t virtio_queue_rq(struct blk_mq_hw_ctx *hctx,
 329                            const struct blk_mq_queue_data *bd)
 330 {
 331         struct virtio_blk *vblk = hctx->queue->queuedata;
 332         struct request *req = bd->rq;
 333         struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
 334         unsigned long flags;
 335         int qid = hctx->queue_num;
 336         bool notify = false;
 337         blk_status_t status;
 338         int err;
 339
 340         status = virtblk_prep_rq(hctx, vblk, req, vbr);
 341         if (unlikely(status))
 342                 return status;
 343
 344         spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
 345         err = virtblk_add_req(vblk->vqs[qid].vq, vbr);
 346         if (err) {
 347                 virtqueue_kick(vblk->vqs[qid].vq);
 348                 /* Don't stop the queue if -ENOMEM: we may have failed to
 349                  * bounce the buffer due to global resource outage.
 350                  */
 351                 if (err == -ENOSPC)
 352                         blk_mq_stop_hw_queue(hctx);
 353                 spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
 354                 virtblk_unmap_data(req, vbr);
 355                 virtblk_cleanup_cmd(req);
 356                 switch (err) {
 357                 case -ENOSPC:
 358                         return BLK_STS_DEV_RESOURCE;
 359                 case -ENOMEM:
 360                         return BLK_STS_RESOURCE;
 361                 default:
 362                         return BLK_STS_IOERR;
 363                 }
 364         }
 365
 366         if (bd->last && virtqueue_kick_prepare(vblk->vqs[qid].vq))
 367                 notify = true;
 368         spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
 369
 370         if (notify)
 371                 virtqueue_notify(vblk->vqs[qid].vq);
 372         return BLK_STS_OK;
 373 }
 374
 375 static bool virtblk_prep_rq_batch(struct request *req)
 376 {
 377         struct virtio_blk *vblk = req->mq_hctx->queue->queuedata;
 378         struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
 379
 380         req->mq_hctx->tags->rqs[req->tag] = req;
 381
 382         return virtblk_prep_rq(req->mq_hctx, vblk, req, vbr) == BLK_STS_OK;
 383 }
 384
 385 static bool virtblk_add_req_batch(struct virtio_blk_vq *vq,
 386                                         struct request **rqlist,
 387                                         struct request **requeue_list)
 388 {
 389         unsigned long flags;
 390         int err;
 391         bool kick;
 392
 393         spin_lock_irqsave(&vq->lock, flags);
 394
 395         while (!rq_list_empty(*rqlist)) {
 396                 struct request *req = rq_list_pop(rqlist);
 397                 struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
 398
 399                 err = virtblk_add_req(vq->vq, vbr);
 400                 if (err) {
 401                         virtblk_unmap_data(req, vbr);
 402                         virtblk_cleanup_cmd(req);
 403                         rq_list_add(requeue_list, req);
 404                 }
 405         }
 406
 407         kick = virtqueue_kick_prepare(vq->vq);
 408         spin_unlock_irqrestore(&vq->lock, flags);
 409
 410         return kick;
 411 }
 412
 413 static void virtio_queue_rqs(struct request **rqlist)
 414 {
 415         struct request *req, *next, *prev = NULL;
 416         struct request *requeue_list = NULL;
 417
 418         rq_list_for_each_safe(rqlist, req, next) {
 419                 struct virtio_blk_vq *vq = req->mq_hctx->driver_data;
 420                 bool kick;
 421
 422                 if (!virtblk_prep_rq_batch(req)) {
 423                         rq_list_move(rqlist, &requeue_list, req, prev);
 424                         req = prev;
 425                         if (!req)
 426                                 continue;
 427                 }
 428
 429                 if (!next || req->mq_hctx != next->mq_hctx) {
 430                         req->rq_next = NULL;
 431                         kick = virtblk_add_req_batch(vq, rqlist, &requeue_list);
 432                         if (kick)
 433                                 virtqueue_notify(vq->vq);
 434
 435                         *rqlist = next;
 436                         prev = NULL;
 437                 } else
 438                         prev = req;
 439         }
 440
 441         *rqlist = requeue_list;
 442 }
 443
 444 /* return id (s/n) string for *disk to *id_str
 445  */
 446 static int virtblk_get_id(struct gendisk *disk, char *id_str)
 447 {
 448         struct virtio_blk *vblk = disk->private_data;
 449         struct request_queue *q = vblk->disk->queue;
 450         struct request *req;
 451         int err;
 452
 453         req = blk_mq_alloc_request(q, REQ_OP_DRV_IN, 0);
 454         if (IS_ERR(req))
 455                 return PTR_ERR(req);
 456
 457         err = blk_rq_map_kern(q, req, id_str, VIRTIO_BLK_ID_BYTES, GFP_KERNEL);
 458         if (err)
 459                 goto out;
 460
 461         blk_execute_rq(req, false);
 462         err = blk_status_to_errno(virtblk_result(blk_mq_rq_to_pdu(req)));
 463 out:
 464         blk_mq_free_request(req);
 465         return err;
 466 }
 467
 468 /* We provide getgeo only to please some old bootloader/partitioning tools */
 469 static int virtblk_getgeo(struct block_device *bd, struct hd_geometry *geo)
 470 {
 471         struct virtio_blk *vblk = bd->bd_disk->private_data;
 472         int ret = 0;
 473
 474         mutex_lock(&vblk->vdev_mutex);
 475
 476         if (!vblk->vdev) {
 477                 ret = -ENXIO;
 478                 goto out;
 479         }
 480
 481         /* see if the host passed in geometry config */
 482         if (virtio_has_feature(vblk->vdev, VIRTIO_BLK_F_GEOMETRY)) {
 483                 virtio_cread(vblk->vdev, struct virtio_blk_config,
 484                              geometry.cylinders, &geo->cylinders);
 485                 virtio_cread(vblk->vdev, struct virtio_blk_config,
 486                              geometry.heads, &geo->heads);
 487                 virtio_cread(vblk->vdev, struct virtio_blk_config,
 488                              geometry.sectors, &geo->sectors);
 489         } else {
 490                 /* some standard values, similar to sd */
 491                 geo->heads = 1 << 6;
 492                 geo->sectors = 1 << 5;
 493                 geo->cylinders = get_capacity(bd->bd_disk) >> 11;
 494         }
 495 out:
 496         mutex_unlock(&vblk->vdev_mutex);
 497         return ret;
 498 }
 499
 500 static void virtblk_free_disk(struct gendisk *disk)
 501 {
 502         struct virtio_blk *vblk = disk->private_data;
 503
 504         ida_simple_remove(&vd_index_ida, vblk->index);
 505         mutex_destroy(&vblk->vdev_mutex);
 506         kfree(vblk);
 507 }
 508
 509 static const struct block_device_operations virtblk_fops = {
 510         .owner          = THIS_MODULE,
 511         .getgeo         = virtblk_getgeo,
 512         .free_disk      = virtblk_free_disk,
 513 };
 514
 515 static int index_to_minor(int index)
 516 {
 517         return index << PART_BITS;
 518 }
 519
 520 static int minor_to_index(int minor)
 521 {
 522         return minor >> PART_BITS;
 523 }
 524
 525 static ssize_t serial_show(struct device *dev,
 526                            struct device_attribute *attr, char *buf)
 527 {
 528         struct gendisk *disk = dev_to_disk(dev);
 529         int err;
 530
 531         /* sysfs gives us a PAGE_SIZE buffer */
 532         BUILD_BUG_ON(PAGE_SIZE < VIRTIO_BLK_ID_BYTES);
 533
 534         buf[VIRTIO_BLK_ID_BYTES] = '\0';
 535         err = virtblk_get_id(disk, buf);
 536         if (!err)
 537                 return strlen(buf);
 538
 539         if (err == -EIO) /* Unsupported? Make it empty. */
 540                 return 0;
 541
 542         return err;
 543 }
 544
 545 static DEVICE_ATTR_RO(serial);
 546
 547 /* The queue's logical block size must be set before calling this */
 548 static void virtblk_update_capacity(struct virtio_blk *vblk, bool resize)
 549 {
 550         struct virtio_device *vdev = vblk->vdev;
 551         struct request_queue *q = vblk->disk->queue;
 552         char cap_str_2[10], cap_str_10[10];
 553         unsigned long long nblocks;
 554         u64 capacity;
 555
 556         /* Host must always specify the capacity. */
 557         virtio_cread(vdev, struct virtio_blk_config, capacity, &capacity);
 558
 559         nblocks = DIV_ROUND_UP_ULL(capacity, queue_logical_block_size(q) >> 9);
 560
 561         string_get_size(nblocks, queue_logical_block_size(q),
 562                         STRING_UNITS_2, cap_str_2, sizeof(cap_str_2));
 563         string_get_size(nblocks, queue_logical_block_size(q),
 564                         STRING_UNITS_10, cap_str_10, sizeof(cap_str_10));
 565
 566         dev_notice(&vdev->dev,
 567                    "[%s] %s%llu %d-byte logical blocks (%s/%s)\n",
 568                    vblk->disk->disk_name,
 569                    resize ? "new size: " : "",
 570                    nblocks,
 571                    queue_logical_block_size(q),
 572                    cap_str_10,
 573                    cap_str_2);
 574
 575         set_capacity_and_notify(vblk->disk, capacity);
 576 }
 577
 578 static void virtblk_config_changed_work(struct work_struct *work)
 579 {
 580         struct virtio_blk *vblk =
 581                 container_of(work, struct virtio_blk, config_work);
 582
 583         virtblk_update_capacity(vblk, true);
 584 }
 585
 586 static void virtblk_config_changed(struct virtio_device *vdev)
 587 {
 588         struct virtio_blk *vblk = vdev->priv;
 589
 590         queue_work(virtblk_wq, &vblk->config_work);
 591 }
 592
 593 static int init_vq(struct virtio_blk *vblk)
 594 {
 595         int err;
 596         int i;
 597         vq_callback_t **callbacks;
 598         const char **names;
 599         struct virtqueue **vqs;
 600         unsigned short num_vqs;
 601         unsigned int num_poll_vqs;
 602         struct virtio_device *vdev = vblk->vdev;
 603         struct irq_affinity desc = { 0, };
 604
 605         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_MQ,
 606                                    struct virtio_blk_config, num_queues,
 607                                    &num_vqs);
 608         if (err)
 609                 num_vqs = 1;
 610
 611         if (!err && !num_vqs) {
 612                 dev_err(&vdev->dev, "MQ advertised but zero queues reported\n");
 613                 return -EINVAL;
 614         }
 615
 616         num_vqs = min_t(unsigned int,
 617                         min_not_zero(num_request_queues, nr_cpu_ids),
 618                         num_vqs);
 619
 620         num_poll_vqs = min_t(unsigned int, poll_queues, num_vqs - 1);
 621
 622         vblk->io_queues[HCTX_TYPE_DEFAULT] = num_vqs - num_poll_vqs;
 623         vblk->io_queues[HCTX_TYPE_READ] = 0;
 624         vblk->io_queues[HCTX_TYPE_POLL] = num_poll_vqs;
 625
 626         dev_info(&vdev->dev, "%d/%d/%d default/read/poll queues\n",
 627                                 vblk->io_queues[HCTX_TYPE_DEFAULT],
 628                                 vblk->io_queues[HCTX_TYPE_READ],
 629                                 vblk->io_queues[HCTX_TYPE_POLL]);
 630
 631         vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL);
 632         if (!vblk->vqs)
 633                 return -ENOMEM;
 634
 635         names = kmalloc_array(num_vqs, sizeof(*names), GFP_KERNEL);
 636         callbacks = kmalloc_array(num_vqs, sizeof(*callbacks), GFP_KERNEL);
 637         vqs = kmalloc_array(num_vqs, sizeof(*vqs), GFP_KERNEL);
 638         if (!names || !callbacks || !vqs) {
 639                 err = -ENOMEM;
 640                 goto out;
 641         }
 642
 643         for (i = 0; i < num_vqs - num_poll_vqs; i++) {
 644                 callbacks[i] = virtblk_done;
 645                 snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", i);
 646                 names[i] = vblk->vqs[i].name;
 647         }
 648
 649         for (; i < num_vqs; i++) {
 650                 callbacks[i] = NULL;
 651                 snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req_poll.%d", i);
 652                 names[i] = vblk->vqs[i].name;
 653         }
 654
 655         /* Discover virtqueues and write information to configuration.  */
 656         err = virtio_find_vqs(vdev, num_vqs, vqs, callbacks, names, &desc);
 657         if (err)
 658                 goto out;
 659
 660         for (i = 0; i < num_vqs; i++) {
 661                 spin_lock_init(&vblk->vqs[i].lock);
 662                 vblk->vqs[i].vq = vqs[i];
 663         }
 664         vblk->num_vqs = num_vqs;
 665
 666 out:
 667         kfree(vqs);
 668         kfree(callbacks);
 669         kfree(names);
 670         if (err)
 671                 kfree(vblk->vqs);
 672         return err;
 673 }
 674
 675 /*
 676  * Legacy naming scheme used for virtio devices.  We are stuck with it for
 677  * virtio blk but don't ever use it for any new driver.
 678  */
 679 static int virtblk_name_format(char *prefix, int index, char *buf, int buflen)
 680 {
 681         const int base = 'z' - 'a' + 1;
 682         char *begin = buf + strlen(prefix);
 683         char *end = buf + buflen;
 684         char *p;
 685         int unit;
 686
 687         p = end - 1;
 688         *p = '\0';
 689         unit = base;
 690         do {
 691                 if (p == begin)
 692                         return -EINVAL;
 693                 *--p = 'a' + (index % unit);
 694                 index = (index / unit) - 1;
 695         } while (index >= 0);
 696
 697         memmove(begin, p, end - p);
 698         memcpy(buf, prefix, strlen(prefix));
 699
 700         return 0;
 701 }
 702
 703 static int virtblk_get_cache_mode(struct virtio_device *vdev)
 704 {
 705         u8 writeback;
 706         int err;
 707
 708         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE,
 709                                    struct virtio_blk_config, wce,
 710                                    &writeback);
 711
 712         /*
 713          * If WCE is not configurable and flush is not available,
 714          * assume no writeback cache is in use.
 715          */
 716         if (err)
 717                 writeback = virtio_has_feature(vdev, VIRTIO_BLK_F_FLUSH);
 718
 719         return writeback;
 720 }
 721
 722 static void virtblk_update_cache_mode(struct virtio_device *vdev)
 723 {
 724         u8 writeback = virtblk_get_cache_mode(vdev);
 725         struct virtio_blk *vblk = vdev->priv;
 726
 727         blk_queue_write_cache(vblk->disk->queue, writeback, false);
 728 }
 729
 730 static const char *const virtblk_cache_types[] = {
 731         "write through", "write back"
 732 };
 733
 734 static ssize_t
 735 cache_type_store(struct device *dev, struct device_attribute *attr,
 736                  const char *buf, size_t count)
 737 {
 738         struct gendisk *disk = dev_to_disk(dev);
 739         struct virtio_blk *vblk = disk->private_data;
 740         struct virtio_device *vdev = vblk->vdev;
 741         int i;
 742
 743         BUG_ON(!virtio_has_feature(vblk->vdev, VIRTIO_BLK_F_CONFIG_WCE));
 744         i = sysfs_match_string(virtblk_cache_types, buf);
 745         if (i < 0)
 746                 return i;
 747
 748         virtio_cwrite8(vdev, offsetof(struct virtio_blk_config, wce), i);
 749         virtblk_update_cache_mode(vdev);
 750         return count;
 751 }
 752
 753 static ssize_t
 754 cache_type_show(struct device *dev, struct device_attribute *attr, char *buf)
 755 {
 756         struct gendisk *disk = dev_to_disk(dev);
 757         struct virtio_blk *vblk = disk->private_data;
 758         u8 writeback = virtblk_get_cache_mode(vblk->vdev);
 759
 760         BUG_ON(writeback >= ARRAY_SIZE(virtblk_cache_types));
 761         return sysfs_emit(buf, "%s\n", virtblk_cache_types[writeback]);
 762 }
 763
 764 static DEVICE_ATTR_RW(cache_type);
 765
 766 static struct attribute *virtblk_attrs[] = {
 767         &dev_attr_serial.attr,
 768         &dev_attr_cache_type.attr,
 769         NULL,
 770 };
 771
 772 static umode_t virtblk_attrs_are_visible(struct kobject *kobj,
 773                 struct attribute *a, int n)
 774 {
 775         struct device *dev = kobj_to_dev(kobj);
 776         struct gendisk *disk = dev_to_disk(dev);
 777         struct virtio_blk *vblk = disk->private_data;
 778         struct virtio_device *vdev = vblk->vdev;
 779
 780         if (a == &dev_attr_cache_type.attr &&
 781             !virtio_has_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE))
 782                 return S_IRUGO;
 783
 784         return a->mode;
 785 }
 786
 787 static const struct attribute_group virtblk_attr_group = {
 788         .attrs = virtblk_attrs,
 789         .is_visible = virtblk_attrs_are_visible,
 790 };
 791
 792 static const struct attribute_group *virtblk_attr_groups[] = {
 793         &virtblk_attr_group,
 794         NULL,
 795 };
 796
 797 static int virtblk_map_queues(struct blk_mq_tag_set *set)
 798 {
 799         struct virtio_blk *vblk = set->driver_data;
 800         int i, qoff;
 801
 802         for (i = 0, qoff = 0; i < set->nr_maps; i++) {
 803                 struct blk_mq_queue_map *map = &set->map[i];
 804
 805                 map->nr_queues = vblk->io_queues[i];
 806                 map->queue_offset = qoff;
 807                 qoff += map->nr_queues;
 808
 809                 if (map->nr_queues == 0)
 810                         continue;
 811
 812                 /*
 813                  * Regular queues have interrupts and hence CPU affinity is
 814                  * defined by the core virtio code, but polling queues have
 815                  * no interrupts so we let the block layer assign CPU affinity.
 816                  */
 817                 if (i == HCTX_TYPE_POLL)
 818                         blk_mq_map_queues(&set->map[i]);
 819                 else
 820                         blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0);
 821         }
 822
 823         return 0;
 824 }
 825
 826 static void virtblk_complete_batch(struct io_comp_batch *iob)
 827 {
 828         struct request *req;
 829
 830         rq_list_for_each(&iob->req_list, req) {
 831                 virtblk_unmap_data(req, blk_mq_rq_to_pdu(req));
 832                 virtblk_cleanup_cmd(req);
 833         }
 834         blk_mq_end_request_batch(iob);
 835 }
 836
 837 static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
 838 {
 839         struct virtio_blk *vblk = hctx->queue->queuedata;
 840         struct virtio_blk_vq *vq = hctx->driver_data;
 841         struct virtblk_req *vbr;
 842         unsigned long flags;
 843         unsigned int len;
 844         int found = 0;
 845
 846         spin_lock_irqsave(&vq->lock, flags);
 847
 848         while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) {
 849                 struct request *req = blk_mq_rq_from_pdu(vbr);
 850
 851                 found++;
 852                 if (!blk_mq_add_to_batch(req, iob, vbr->status,
 853                                                 virtblk_complete_batch))
 854                         blk_mq_complete_request(req);
 855         }
 856
 857         if (found)
 858                 blk_mq_start_stopped_hw_queues(vblk->disk->queue, true);
 859
 860         spin_unlock_irqrestore(&vq->lock, flags);
 861
 862         return found;
 863 }
 864
 865 static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 866                           unsigned int hctx_idx)
 867 {
 868         struct virtio_blk *vblk = data;
 869         struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx];
 870
 871         WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags);
 872         hctx->driver_data = vq;
 873         return 0;
 874 }
 875
 876 static const struct blk_mq_ops virtio_mq_ops = {
 877         .queue_rq       = virtio_queue_rq,
 878         .queue_rqs      = virtio_queue_rqs,
 879         .commit_rqs     = virtio_commit_rqs,
 880         .init_hctx      = virtblk_init_hctx,
 881         .complete       = virtblk_request_done,
 882         .map_queues     = virtblk_map_queues,
 883         .poll           = virtblk_poll,
 884 };
 885
 886 static unsigned int virtblk_queue_depth;
 887 module_param_named(queue_depth, virtblk_queue_depth, uint, 0444);
 888
 889 static int virtblk_probe(struct virtio_device *vdev)
 890 {
 891         struct virtio_blk *vblk;
 892         struct request_queue *q;
 893         int err, index;
 894
 895         u32 v, blk_size, max_size, sg_elems, opt_io_size;
 896         u16 min_io_size;
 897         u8 physical_block_exp, alignment_offset;
 898         unsigned int queue_depth;
 899
 900         if (!vdev->config->get) {
 901                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
 902                         __func__);
 903                 return -EINVAL;
 904         }
 905
 906         err = ida_simple_get(&vd_index_ida, 0, minor_to_index(1 << MINORBITS),
 907                              GFP_KERNEL);
 908         if (err < 0)
 909                 goto out;
 910         index = err;
 911
 912         /* We need to know how many segments before we allocate. */
 913         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_SEG_MAX,
 914                                    struct virtio_blk_config, seg_max,
 915                                    &sg_elems);
 916
 917         /* We need at least one SG element, whatever they say. */
 918         if (err || !sg_elems)
 919                 sg_elems = 1;
 920
 921         /* Prevent integer overflows and honor max vq size */
 922         sg_elems = min_t(u32, sg_elems, VIRTIO_BLK_MAX_SG_ELEMS - 2);
 923
 924         vdev->priv = vblk = kmalloc(sizeof(*vblk), GFP_KERNEL);
 925         if (!vblk) {
 926                 err = -ENOMEM;
 927                 goto out_free_index;
 928         }
 929
 930         mutex_init(&vblk->vdev_mutex);
 931
 932         vblk->vdev = vdev;
 933
 934         INIT_WORK(&vblk->config_work, virtblk_config_changed_work);
 935
 936         err = init_vq(vblk);
 937         if (err)
 938                 goto out_free_vblk;
 939
 940         /* Default queue sizing is to fill the ring. */
 941         if (!virtblk_queue_depth) {
 942                 queue_depth = vblk->vqs[0].vq->num_free;
 943                 /* ... but without indirect descs, we use 2 descs per req */
 944                 if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
 945                         queue_depth /= 2;
 946         } else {
 947                 queue_depth = virtblk_queue_depth;
 948         }
 949
 950         memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
 951         vblk->tag_set.ops = &virtio_mq_ops;
 952         vblk->tag_set.queue_depth = queue_depth;
 953         vblk->tag_set.numa_node = NUMA_NO_NODE;
 954         vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 955         vblk->tag_set.cmd_size =
 956                 sizeof(struct virtblk_req) +
 957                 sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT;
 958         vblk->tag_set.driver_data = vblk;
 959         vblk->tag_set.nr_hw_queues = vblk->num_vqs;
 960         vblk->tag_set.nr_maps = 1;
 961         if (vblk->io_queues[HCTX_TYPE_POLL])
 962                 vblk->tag_set.nr_maps = 3;
 963
 964         err = blk_mq_alloc_tag_set(&vblk->tag_set);
 965         if (err)
 966                 goto out_free_vq;
 967
 968         vblk->disk = blk_mq_alloc_disk(&vblk->tag_set, vblk);
 969         if (IS_ERR(vblk->disk)) {
 970                 err = PTR_ERR(vblk->disk);
 971                 goto out_free_tags;
 972         }
 973         q = vblk->disk->queue;
 974
 975         virtblk_name_format("vd", index, vblk->disk->disk_name, DISK_NAME_LEN);
 976
 977         vblk->disk->major = major;
 978         vblk->disk->first_minor = index_to_minor(index);
 979         vblk->disk->minors = 1 << PART_BITS;
 980         vblk->disk->private_data = vblk;
 981         vblk->disk->fops = &virtblk_fops;
 982         vblk->index = index;
 983
 984         /* configure queue flush support */
 985         virtblk_update_cache_mode(vdev);
 986
 987         /* If disk is read-only in the host, the guest should obey */
 988         if (virtio_has_feature(vdev, VIRTIO_BLK_F_RO))
 989                 set_disk_ro(vblk->disk, 1);
 990
 991         /* We can handle whatever the host told us to handle. */
 992         blk_queue_max_segments(q, sg_elems);
 993
 994         /* No real sector limit. */
 995         blk_queue_max_hw_sectors(q, -1U);
 996
 997         max_size = virtio_max_dma_size(vdev);
 998
 999         /* Host can optionally specify maximum segment size and number of
1000          * segments. */
1001         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_SIZE_MAX,
1002                                    struct virtio_blk_config, size_max, &v);
1003         if (!err)
1004                 max_size = min(max_size, v);
1005
1006         blk_queue_max_segment_size(q, max_size);
1007
1008         /* Host can optionally specify the block size of the device */
1009         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_BLK_SIZE,
1010                                    struct virtio_blk_config, blk_size,
1011                                    &blk_size);
1012         if (!err) {
1013                 err = blk_validate_block_size(blk_size);
1014                 if (err) {
1015                         dev_err(&vdev->dev,
1016                                 "virtio_blk: invalid block size: 0x%x\n",
1017                                 blk_size);
1018                         goto out_cleanup_disk;
1019                 }
1020
1021                 blk_queue_logical_block_size(q, blk_size);
1022         } else
1023                 blk_size = queue_logical_block_size(q);
1024
1025         /* Use topology information if available */
1026         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1027                                    struct virtio_blk_config, physical_block_exp,
1028                                    &physical_block_exp);
1029         if (!err && physical_block_exp)
1030                 blk_queue_physical_block_size(q,
1031                                 blk_size * (1 << physical_block_exp));
1032
1033         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1034                                    struct virtio_blk_config, alignment_offset,
1035                                    &alignment_offset);
1036         if (!err && alignment_offset)
1037                 blk_queue_alignment_offset(q, blk_size * alignment_offset);
1038
1039         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1040                                    struct virtio_blk_config, min_io_size,
1041                                    &min_io_size);
1042         if (!err && min_io_size)
1043                 blk_queue_io_min(q, blk_size * min_io_size);
1044
1045         err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1046                                    struct virtio_blk_config, opt_io_size,
1047                                    &opt_io_size);
1048         if (!err && opt_io_size)
1049                 blk_queue_io_opt(q, blk_size * opt_io_size);
1050
1051         if (virtio_has_feature(vdev, VIRTIO_BLK_F_DISCARD)) {
1052                 virtio_cread(vdev, struct virtio_blk_config,
1053                              discard_sector_alignment, &v);
1054                 if (v)
1055                         q->limits.discard_granularity = v << SECTOR_SHIFT;
1056                 else
1057                         q->limits.discard_granularity = blk_size;
1058
1059                 virtio_cread(vdev, struct virtio_blk_config,
1060                              max_discard_sectors, &v);
1061                 blk_queue_max_discard_sectors(q, v ? v : UINT_MAX);
1062
1063                 virtio_cread(vdev, struct virtio_blk_config, max_discard_seg,
1064                              &v);
1065
1066                 /*
1067                  * max_discard_seg == 0 is out of spec but we always
1068                  * handled it.
1069                  */
1070                 if (!v)
1071                         v = sg_elems;
1072                 blk_queue_max_discard_segments(q,
1073                                                min(v, MAX_DISCARD_SEGMENTS));
1074         }
1075
1076         if (virtio_has_feature(vdev, VIRTIO_BLK_F_WRITE_ZEROES)) {
1077                 virtio_cread(vdev, struct virtio_blk_config,
1078                              max_write_zeroes_sectors, &v);
1079                 blk_queue_max_write_zeroes_sectors(q, v ? v : UINT_MAX);
1080         }
1081
1082         virtblk_update_capacity(vblk, false);
1083         virtio_device_ready(vdev);
1084
1085         err = device_add_disk(&vdev->dev, vblk->disk, virtblk_attr_groups);
1086         if (err)
1087                 goto out_cleanup_disk;
1088
1089         return 0;
1090
1091 out_cleanup_disk:
1092         put_disk(vblk->disk);
1093 out_free_tags:
1094         blk_mq_free_tag_set(&vblk->tag_set);
1095 out_free_vq:
1096         vdev->config->del_vqs(vdev);
1097         kfree(vblk->vqs);
1098 out_free_vblk:
1099         kfree(vblk);
1100 out_free_index:
1101         ida_simple_remove(&vd_index_ida, index);
1102 out:
1103         return err;
1104 }
1105
1106 static void virtblk_remove(struct virtio_device *vdev)
1107 {
1108         struct virtio_blk *vblk = vdev->priv;
1109
1110         /* Make sure no work handler is accessing the device. */
1111         flush_work(&vblk->config_work);
1112
1113         del_gendisk(vblk->disk);
1114         blk_mq_free_tag_set(&vblk->tag_set);
1115
1116         mutex_lock(&vblk->vdev_mutex);
1117
1118         /* Stop all the virtqueues. */
1119         virtio_reset_device(vdev);
1120
1121         /* Virtqueues are stopped, nothing can use vblk->vdev anymore. */
1122         vblk->vdev = NULL;
1123
1124         vdev->config->del_vqs(vdev);
1125         kfree(vblk->vqs);
1126
1127         mutex_unlock(&vblk->vdev_mutex);
1128
1129         put_disk(vblk->disk);
1130 }
1131
1132 #ifdef CONFIG_PM_SLEEP
1133 static int virtblk_freeze(struct virtio_device *vdev)
1134 {
1135         struct virtio_blk *vblk = vdev->priv;
1136
1137         /* Ensure we don't receive any more interrupts */
1138         virtio_reset_device(vdev);
1139
1140         /* Make sure no work handler is accessing the device. */
1141         flush_work(&vblk->config_work);
1142
1143         blk_mq_quiesce_queue(vblk->disk->queue);
1144
1145         vdev->config->del_vqs(vdev);
1146         kfree(vblk->vqs);
1147
1148         return 0;
1149 }
1150
1151 static int virtblk_restore(struct virtio_device *vdev)
1152 {
1153         struct virtio_blk *vblk = vdev->priv;
1154         int ret;
1155
1156         ret = init_vq(vdev->priv);
1157         if (ret)
1158                 return ret;
1159
1160         virtio_device_ready(vdev);
1161
1162         blk_mq_unquiesce_queue(vblk->disk->queue);
1163         return 0;
1164 }
1165 #endif
1166
1167 static const struct virtio_device_id id_table[] = {
1168         { VIRTIO_ID_BLOCK, VIRTIO_DEV_ANY_ID },
1169         { 0 },
1170 };
1171
1172 static unsigned int features_legacy[] = {
1173         VIRTIO_BLK_F_SEG_MAX, VIRTIO_BLK_F_SIZE_MAX, VIRTIO_BLK_F_GEOMETRY,
1174         VIRTIO_BLK_F_RO, VIRTIO_BLK_F_BLK_SIZE,
1175         VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY, VIRTIO_BLK_F_CONFIG_WCE,
1176         VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD, VIRTIO_BLK_F_WRITE_ZEROES,
1177 }
1178 ;
1179 static unsigned int features[] = {
1180         VIRTIO_BLK_F_SEG_MAX, VIRTIO_BLK_F_SIZE_MAX, VIRTIO_BLK_F_GEOMETRY,
1181         VIRTIO_BLK_F_RO, VIRTIO_BLK_F_BLK_SIZE,
1182         VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY, VIRTIO_BLK_F_CONFIG_WCE,
1183         VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD, VIRTIO_BLK_F_WRITE_ZEROES,
1184 };
1185
1186 static struct virtio_driver virtio_blk = {
1187         .feature_table                  = features,
1188         .feature_table_size             = ARRAY_SIZE(features),
1189         .feature_table_legacy           = features_legacy,
1190         .feature_table_size_legacy      = ARRAY_SIZE(features_legacy),
1191         .driver.name                    = KBUILD_MODNAME,
1192         .driver.owner                   = THIS_MODULE,
1193         .id_table                       = id_table,
1194         .probe                          = virtblk_probe,
1195         .remove                         = virtblk_remove,
1196         .config_changed                 = virtblk_config_changed,
1197 #ifdef CONFIG_PM_SLEEP
1198         .freeze                         = virtblk_freeze,
1199         .restore                        = virtblk_restore,
1200 #endif
1201 };
1202
1203 static int __init virtio_blk_init(void)
1204 {
1205         int error;
1206
1207         virtblk_wq = alloc_workqueue("virtio-blk", 0, 0);
1208         if (!virtblk_wq)
1209                 return -ENOMEM;
1210
1211         major = register_blkdev(0, "virtblk");
1212         if (major < 0) {
1213                 error = major;
1214                 goto out_destroy_workqueue;
1215         }
1216
1217         error = register_virtio_driver(&virtio_blk);
1218         if (error)
1219                 goto out_unregister_blkdev;
1220         return 0;
1221
1222 out_unregister_blkdev:
1223         unregister_blkdev(major, "virtblk");
1224 out_destroy_workqueue:
1225         destroy_workqueue(virtblk_wq);
1226         return error;
1227 }
1228
1229 static void __exit virtio_blk_fini(void)
1230 {
1231         unregister_virtio_driver(&virtio_blk);
1232         unregister_blkdev(major, "virtblk");
1233         destroy_workqueue(virtblk_wq);
1234 }
1235 module_init(virtio_blk_init);
1236 module_exit(virtio_blk_fini);
1237
1238 MODULE_DEVICE_TABLE(virtio, id_table);
1239 MODULE_DESCRIPTION("Virtio block driver");
1240 MODULE_LICENSE("GPL");