block/export/vhost-user-blk-server.c

   1 /*
   2  * Sharing QEMU block devices via vhost-user protocal
   3  *
   4  * Parts of the code based on nbd/server.c.
   5  *
   6  * Copyright (c) Coiby Xu <coiby.xu@gmail.com>.
   7  * Copyright (c) 2020 Red Hat, Inc.
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or
  10  * later.  See the COPYING file in the top-level directory.
  11  */
  12 #include "qemu/osdep.h"
  13 #include "block/block.h"
  14 #include "subprojects/libvhost-user/libvhost-user.h" /* only for the type definitions */
  15 #include "standard-headers/linux/virtio_blk.h"
  16 #include "qemu/vhost-user-server.h"
  17 #include "vhost-user-blk-server.h"
  18 #include "qapi/error.h"
  19 #include "qom/object_interfaces.h"
  20 #include "sysemu/block-backend.h"
  21 #include "util/block-helpers.h"
  22
  23 /*
  24  * Sector units are 512 bytes regardless of the
  25  * virtio_blk_config->blk_size value.
  26  */
  27 #define VIRTIO_BLK_SECTOR_BITS 9
  28 #define VIRTIO_BLK_SECTOR_SIZE (1ull << VIRTIO_BLK_SECTOR_BITS)
  29
  30 enum {
  31     VHOST_USER_BLK_NUM_QUEUES_DEFAULT = 1,
  32     VHOST_USER_BLK_MAX_DISCARD_SECTORS = 32768,
  33     VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS = 32768,
  34 };
  35 struct virtio_blk_inhdr {
  36     unsigned char status;
  37 };
  38
  39 typedef struct VuBlkReq {
  40     VuVirtqElement elem;
  41     int64_t sector_num;
  42     size_t size;
  43     struct virtio_blk_inhdr *in;
  44     struct virtio_blk_outhdr out;
  45     VuServer *server;
  46     struct VuVirtq *vq;
  47 } VuBlkReq;
  48
  49 /* vhost user block device */
  50 typedef struct {
  51     BlockExport export;
  52     VuServer vu_server;
  53     uint32_t blk_size;
  54     QIOChannelSocket *sioc;
  55     struct virtio_blk_config blkcfg;
  56     bool writable;
  57 } VuBlkExport;
  58
  59 static void vu_blk_req_complete(VuBlkReq *req)
  60 {
  61     VuDev *vu_dev = &req->server->vu_dev;
  62
  63     /* IO size with 1 extra status byte */
  64     vu_queue_push(vu_dev, req->vq, &req->elem, req->size + 1);
  65     vu_queue_notify(vu_dev, req->vq);
  66
  67     free(req);
  68 }
  69
  70 static bool vu_blk_sect_range_ok(VuBlkExport *vexp, uint64_t sector,
  71                                  size_t size)
  72 {
  73     uint64_t nb_sectors = size >> BDRV_SECTOR_BITS;
  74     uint64_t total_sectors;
  75
  76     if (nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
  77         return false;
  78     }
  79     if ((sector << VIRTIO_BLK_SECTOR_BITS) % vexp->blk_size) {
  80         return false;
  81     }
  82     blk_get_geometry(vexp->export.blk, &total_sectors);
  83     if (sector > total_sectors || nb_sectors > total_sectors - sector) {
  84         return false;
  85     }
  86     return true;
  87 }
  88
  89 static int coroutine_fn
  90 vu_blk_discard_write_zeroes(VuBlkExport *vexp, struct iovec *iov,
  91                             uint32_t iovcnt, uint32_t type)
  92 {
  93     BlockBackend *blk = vexp->export.blk;
  94     struct virtio_blk_discard_write_zeroes desc;
  95     ssize_t size;
  96     uint64_t sector;
  97     uint32_t num_sectors;
  98     uint32_t max_sectors;
  99     uint32_t flags;
 100     int bytes;
 101
 102     /* Only one desc is currently supported */
 103     if (unlikely(iov_size(iov, iovcnt) > sizeof(desc))) {
 104         return VIRTIO_BLK_S_UNSUPP;
 105     }
 106
 107     size = iov_to_buf(iov, iovcnt, 0, &desc, sizeof(desc));
 108     if (unlikely(size != sizeof(desc))) {
 109         error_report("Invalid size %zd, expected %zu", size, sizeof(desc));
 110         return VIRTIO_BLK_S_IOERR;
 111     }
 112
 113     sector = le64_to_cpu(desc.sector);
 114     num_sectors = le32_to_cpu(desc.num_sectors);
 115     flags = le32_to_cpu(desc.flags);
 116     max_sectors = (type == VIRTIO_BLK_T_WRITE_ZEROES) ?
 117                   VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS :
 118                   VHOST_USER_BLK_MAX_DISCARD_SECTORS;
 119
 120     /* This check ensures that 'bytes' fits in an int */
 121     if (unlikely(num_sectors > max_sectors)) {
 122         return VIRTIO_BLK_S_IOERR;
 123     }
 124
 125     bytes = num_sectors << VIRTIO_BLK_SECTOR_BITS;
 126
 127     if (unlikely(!vu_blk_sect_range_ok(vexp, sector, bytes))) {
 128         return VIRTIO_BLK_S_IOERR;
 129     }
 130
 131     /*
 132      * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard
 133      * and write zeroes commands if any unknown flag is set.
 134      */
 135     if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) {
 136         return VIRTIO_BLK_S_UNSUPP;
 137     }
 138
 139     if (type == VIRTIO_BLK_T_WRITE_ZEROES) {
 140         int blk_flags = 0;
 141
 142         if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) {
 143             blk_flags |= BDRV_REQ_MAY_UNMAP;
 144         }
 145
 146         if (blk_co_pwrite_zeroes(blk, sector << VIRTIO_BLK_SECTOR_BITS,
 147                                  bytes, blk_flags) == 0) {
 148             return VIRTIO_BLK_S_OK;
 149         }
 150     } else if (type == VIRTIO_BLK_T_DISCARD) {
 151         /*
 152          * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for
 153          * discard commands if the unmap flag is set.
 154          */
 155         if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) {
 156             return VIRTIO_BLK_S_UNSUPP;
 157         }
 158
 159         if (blk_co_pdiscard(blk, sector << VIRTIO_BLK_SECTOR_BITS,
 160                             bytes) == 0) {
 161             return VIRTIO_BLK_S_OK;
 162         }
 163     }
 164
 165     return VIRTIO_BLK_S_IOERR;
 166 }
 167
 168 static void coroutine_fn vu_blk_virtio_process_req(void *opaque)
 169 {
 170     VuBlkReq *req = opaque;
 171     VuServer *server = req->server;
 172     VuVirtqElement *elem = &req->elem;
 173     uint32_t type;
 174
 175     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
 176     BlockBackend *blk = vexp->export.blk;
 177
 178     struct iovec *in_iov = elem->in_sg;
 179     struct iovec *out_iov = elem->out_sg;
 180     unsigned in_num = elem->in_num;
 181     unsigned out_num = elem->out_num;
 182
 183     /* refer to hw/block/virtio_blk.c */
 184     if (elem->out_num < 1 || elem->in_num < 1) {
 185         error_report("virtio-blk request missing headers");
 186         goto err;
 187     }
 188
 189     if (unlikely(iov_to_buf(out_iov, out_num, 0, &req->out,
 190                             sizeof(req->out)) != sizeof(req->out))) {
 191         error_report("virtio-blk request outhdr too short");
 192         goto err;
 193     }
 194
 195     iov_discard_front(&out_iov, &out_num, sizeof(req->out));
 196
 197     if (in_iov[in_num - 1].iov_len < sizeof(struct virtio_blk_inhdr)) {
 198         error_report("virtio-blk request inhdr too short");
 199         goto err;
 200     }
 201
 202     /* We always touch the last byte, so just see how big in_iov is.  */
 203     req->in = (void *)in_iov[in_num - 1].iov_base
 204               + in_iov[in_num - 1].iov_len
 205               - sizeof(struct virtio_blk_inhdr);
 206     iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr));
 207
 208     type = le32_to_cpu(req->out.type);
 209     switch (type & ~VIRTIO_BLK_T_BARRIER) {
 210     case VIRTIO_BLK_T_IN:
 211     case VIRTIO_BLK_T_OUT: {
 212         QEMUIOVector qiov;
 213         int64_t offset;
 214         ssize_t ret = 0;
 215         bool is_write = type & VIRTIO_BLK_T_OUT;
 216         req->sector_num = le64_to_cpu(req->out.sector);
 217
 218         if (is_write && !vexp->writable) {
 219             req->in->status = VIRTIO_BLK_S_IOERR;
 220             break;
 221         }
 222
 223         if (is_write) {
 224             qemu_iovec_init_external(&qiov, out_iov, out_num);
 225         } else {
 226             qemu_iovec_init_external(&qiov, in_iov, in_num);
 227         }
 228
 229         if (unlikely(!vu_blk_sect_range_ok(vexp,
 230                                            req->sector_num,
 231                                            qiov.size))) {
 232             req->in->status = VIRTIO_BLK_S_IOERR;
 233             break;
 234         }
 235
 236         offset = req->sector_num << VIRTIO_BLK_SECTOR_BITS;
 237
 238         if (is_write) {
 239             ret = blk_co_pwritev(blk, offset, qiov.size, &qiov, 0);
 240         } else {
 241             ret = blk_co_preadv(blk, offset, qiov.size, &qiov, 0);
 242         }
 243         if (ret >= 0) {
 244             req->in->status = VIRTIO_BLK_S_OK;
 245         } else {
 246             req->in->status = VIRTIO_BLK_S_IOERR;
 247         }
 248         break;
 249     }
 250     case VIRTIO_BLK_T_FLUSH:
 251         if (blk_co_flush(blk) == 0) {
 252             req->in->status = VIRTIO_BLK_S_OK;
 253         } else {
 254             req->in->status = VIRTIO_BLK_S_IOERR;
 255         }
 256         break;
 257     case VIRTIO_BLK_T_GET_ID: {
 258         size_t size = MIN(iov_size(&elem->in_sg[0], in_num),
 259                           VIRTIO_BLK_ID_BYTES);
 260         snprintf(elem->in_sg[0].iov_base, size, "%s", "vhost_user_blk");
 261         req->in->status = VIRTIO_BLK_S_OK;
 262         req->size = elem->in_sg[0].iov_len;
 263         break;
 264     }
 265     case VIRTIO_BLK_T_DISCARD:
 266     case VIRTIO_BLK_T_WRITE_ZEROES: {
 267         if (!vexp->writable) {
 268             req->in->status = VIRTIO_BLK_S_IOERR;
 269             break;
 270         }
 271
 272         req->in->status = vu_blk_discard_write_zeroes(vexp, out_iov, out_num,
 273                                                       type);
 274         break;
 275     }
 276     default:
 277         req->in->status = VIRTIO_BLK_S_UNSUPP;
 278         break;
 279     }
 280
 281     vu_blk_req_complete(req);
 282     return;
 283
 284 err:
 285     free(req);
 286 }
 287
 288 static void vu_blk_process_vq(VuDev *vu_dev, int idx)
 289 {
 290     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
 291     VuVirtq *vq = vu_get_queue(vu_dev, idx);
 292
 293     while (1) {
 294         VuBlkReq *req;
 295
 296         req = vu_queue_pop(vu_dev, vq, sizeof(VuBlkReq));
 297         if (!req) {
 298             break;
 299         }
 300
 301         req->server = server;
 302         req->vq = vq;
 303
 304         Coroutine *co =
 305             qemu_coroutine_create(vu_blk_virtio_process_req, req);
 306         qemu_coroutine_enter(co);
 307     }
 308 }
 309
 310 static void vu_blk_queue_set_started(VuDev *vu_dev, int idx, bool started)
 311 {
 312     VuVirtq *vq;
 313
 314     assert(vu_dev);
 315
 316     vq = vu_get_queue(vu_dev, idx);
 317     vu_set_queue_handler(vu_dev, vq, started ? vu_blk_process_vq : NULL);
 318 }
 319
 320 static uint64_t vu_blk_get_features(VuDev *dev)
 321 {
 322     uint64_t features;
 323     VuServer *server = container_of(dev, VuServer, vu_dev);
 324     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
 325     features = 1ull << VIRTIO_BLK_F_SIZE_MAX |
 326                1ull << VIRTIO_BLK_F_SEG_MAX |
 327                1ull << VIRTIO_BLK_F_TOPOLOGY |
 328                1ull << VIRTIO_BLK_F_BLK_SIZE |
 329                1ull << VIRTIO_BLK_F_FLUSH |
 330                1ull << VIRTIO_BLK_F_DISCARD |
 331                1ull << VIRTIO_BLK_F_WRITE_ZEROES |
 332                1ull << VIRTIO_BLK_F_CONFIG_WCE |
 333                1ull << VIRTIO_BLK_F_MQ |
 334                1ull << VIRTIO_F_VERSION_1 |
 335                1ull << VIRTIO_RING_F_INDIRECT_DESC |
 336                1ull << VIRTIO_RING_F_EVENT_IDX |
 337                1ull << VHOST_USER_F_PROTOCOL_FEATURES;
 338
 339     if (!vexp->writable) {
 340         features |= 1ull << VIRTIO_BLK_F_RO;
 341     }
 342
 343     return features;
 344 }
 345
 346 static uint64_t vu_blk_get_protocol_features(VuDev *dev)
 347 {
 348     return 1ull << VHOST_USER_PROTOCOL_F_CONFIG |
 349            1ull << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD;
 350 }
 351
 352 static int
 353 vu_blk_get_config(VuDev *vu_dev, uint8_t *config, uint32_t len)
 354 {
 355     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
 356     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
 357
 358     if (len > sizeof(struct virtio_blk_config)) {
 359         return -1;
 360     }
 361
 362     memcpy(config, &vexp->blkcfg, len);
 363     return 0;
 364 }
 365
 366 static int
 367 vu_blk_set_config(VuDev *vu_dev, const uint8_t *data,
 368                     uint32_t offset, uint32_t size, uint32_t flags)
 369 {
 370     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
 371     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
 372     uint8_t wce;
 373
 374     /* don't support live migration */
 375     if (flags != VHOST_SET_CONFIG_TYPE_MASTER) {
 376         return -EINVAL;
 377     }
 378
 379     if (offset != offsetof(struct virtio_blk_config, wce) ||
 380         size != 1) {
 381         return -EINVAL;
 382     }
 383
 384     wce = *data;
 385     vexp->blkcfg.wce = wce;
 386     blk_set_enable_write_cache(vexp->export.blk, wce);
 387     return 0;
 388 }
 389
 390 /*
 391  * When the client disconnects, it sends a VHOST_USER_NONE request
 392  * and vu_process_message will simple call exit which cause the VM
 393  * to exit abruptly.
 394  * To avoid this issue,  process VHOST_USER_NONE request ahead
 395  * of vu_process_message.
 396  *
 397  */
 398 static int vu_blk_process_msg(VuDev *dev, VhostUserMsg *vmsg, int *do_reply)
 399 {
 400     if (vmsg->request == VHOST_USER_NONE) {
 401         dev->panic(dev, "disconnect");
 402         return true;
 403     }
 404     return false;
 405 }
 406
 407 static const VuDevIface vu_blk_iface = {
 408     .get_features          = vu_blk_get_features,
 409     .queue_set_started     = vu_blk_queue_set_started,
 410     .get_protocol_features = vu_blk_get_protocol_features,
 411     .get_config            = vu_blk_get_config,
 412     .set_config            = vu_blk_set_config,
 413     .process_msg           = vu_blk_process_msg,
 414 };
 415
 416 static void blk_aio_attached(AioContext *ctx, void *opaque)
 417 {
 418     VuBlkExport *vexp = opaque;
 419
 420     vexp->export.ctx = ctx;
 421     vhost_user_server_attach_aio_context(&vexp->vu_server, ctx);
 422 }
 423
 424 static void blk_aio_detach(void *opaque)
 425 {
 426     VuBlkExport *vexp = opaque;
 427
 428     vhost_user_server_detach_aio_context(&vexp->vu_server);
 429     vexp->export.ctx = NULL;
 430 }
 431
 432 static void
 433 vu_blk_initialize_config(BlockDriverState *bs,
 434                          struct virtio_blk_config *config,
 435                          uint32_t blk_size,
 436                          uint16_t num_queues)
 437 {
 438     config->capacity =
 439         cpu_to_le64(bdrv_getlength(bs) >> VIRTIO_BLK_SECTOR_BITS);
 440     config->blk_size = cpu_to_le32(blk_size);
 441     config->size_max = cpu_to_le32(0);
 442     config->seg_max = cpu_to_le32(128 - 2);
 443     config->min_io_size = cpu_to_le16(1);
 444     config->opt_io_size = cpu_to_le32(1);
 445     config->num_queues = cpu_to_le16(num_queues);
 446     config->max_discard_sectors =
 447         cpu_to_le32(VHOST_USER_BLK_MAX_DISCARD_SECTORS);
 448     config->max_discard_seg = cpu_to_le32(1);
 449     config->discard_sector_alignment =
 450         cpu_to_le32(blk_size >> VIRTIO_BLK_SECTOR_BITS);
 451     config->max_write_zeroes_sectors
 452         = cpu_to_le32(VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS);
 453     config->max_write_zeroes_seg = cpu_to_le32(1);
 454 }
 455
 456 static void vu_blk_exp_request_shutdown(BlockExport *exp)
 457 {
 458     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
 459
 460     vhost_user_server_stop(&vexp->vu_server);
 461 }
 462
 463 static int vu_blk_exp_create(BlockExport *exp, BlockExportOptions *opts,
 464                              Error **errp)
 465 {
 466     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
 467     BlockExportOptionsVhostUserBlk *vu_opts = &opts->u.vhost_user_blk;
 468     Error *local_err = NULL;
 469     uint64_t logical_block_size;
 470     uint16_t num_queues = VHOST_USER_BLK_NUM_QUEUES_DEFAULT;
 471
 472     vexp->writable = opts->writable;
 473     vexp->blkcfg.wce = 0;
 474
 475     if (vu_opts->has_logical_block_size) {
 476         logical_block_size = vu_opts->logical_block_size;
 477     } else {
 478         logical_block_size = VIRTIO_BLK_SECTOR_SIZE;
 479     }
 480     check_block_size(exp->id, "logical-block-size", logical_block_size,
 481                      &local_err);
 482     if (local_err) {
 483         error_propagate(errp, local_err);
 484         return -EINVAL;
 485     }
 486     vexp->blk_size = logical_block_size;
 487     blk_set_guest_block_size(exp->blk, logical_block_size);
 488
 489     if (vu_opts->has_num_queues) {
 490         num_queues = vu_opts->num_queues;
 491     }
 492     if (num_queues == 0) {
 493         error_setg(errp, "num-queues must be greater than 0");
 494         return -EINVAL;
 495     }
 496
 497     vu_blk_initialize_config(blk_bs(exp->blk), &vexp->blkcfg,
 498                              logical_block_size, num_queues);
 499
 500     blk_add_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
 501                                  vexp);
 502
 503     if (!vhost_user_server_start(&vexp->vu_server, vu_opts->addr, exp->ctx,
 504                                  num_queues, &vu_blk_iface, errp)) {
 505         blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 506                                         blk_aio_detach, vexp);
 507         return -EADDRNOTAVAIL;
 508     }
 509
 510     return 0;
 511 }
 512
 513 static void vu_blk_exp_delete(BlockExport *exp)
 514 {
 515     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
 516
 517     blk_remove_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
 518                                     vexp);
 519 }
 520
 521 const BlockExportDriver blk_exp_vhost_user_blk = {
 522     .type               = BLOCK_EXPORT_TYPE_VHOST_USER_BLK,
 523     .instance_size      = sizeof(VuBlkExport),
 524     .create             = vu_blk_exp_create,
 525     .delete             = vu_blk_exp_delete,
 526     .request_shutdown   = vu_blk_exp_request_shutdown,
 527 };