1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2017-2018 Christoph Hellwig. 4 */ 5 6 #include <linux/backing-dev.h> 7 #include <linux/moduleparam.h> 8 #include <linux/vmalloc.h> 9 #include <trace/events/block.h> 10 #include "nvme.h" 11 12 bool multipath = true; 13 module_param(multipath, bool, 0444); 14 MODULE_PARM_DESC(multipath, 15 "turn on native support for multiple controllers per subsystem"); 16 17 static const char *nvme_iopolicy_names[] = { 18 [NVME_IOPOLICY_NUMA] = "numa", 19 [NVME_IOPOLICY_RR] = "round-robin", 20 }; 21 22 static int iopolicy = NVME_IOPOLICY_NUMA; 23 24 static int nvme_set_iopolicy(const char *val, const struct kernel_param *kp) 25 { 26 if (!val) 27 return -EINVAL; 28 if (!strncmp(val, "numa", 4)) 29 iopolicy = NVME_IOPOLICY_NUMA; 30 else if (!strncmp(val, "round-robin", 11)) 31 iopolicy = NVME_IOPOLICY_RR; 32 else 33 return -EINVAL; 34 35 return 0; 36 } 37 38 static int nvme_get_iopolicy(char *buf, const struct kernel_param *kp) 39 { 40 return sprintf(buf, "%s\n", nvme_iopolicy_names[iopolicy]); 41 } 42 43 module_param_call(iopolicy, nvme_set_iopolicy, nvme_get_iopolicy, 44 &iopolicy, 0644); 45 MODULE_PARM_DESC(iopolicy, 46 "Default multipath I/O policy; 'numa' (default) or 'round-robin'"); 47 48 void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys) 49 { 50 subsys->iopolicy = iopolicy; 51 } 52 53 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys) 54 { 55 struct nvme_ns_head *h; 56 57 lockdep_assert_held(&subsys->lock); 58 list_for_each_entry(h, &subsys->nsheads, entry) 59 if (h->disk) 60 blk_mq_unfreeze_queue(h->disk->queue); 61 } 62 63 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys) 64 { 65 struct nvme_ns_head *h; 66 67 lockdep_assert_held(&subsys->lock); 68 list_for_each_entry(h, &subsys->nsheads, entry) 69 if (h->disk) 70 blk_mq_freeze_queue_wait(h->disk->queue); 71 } 72 73 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys) 74 { 75 struct nvme_ns_head *h; 76 77 lockdep_assert_held(&subsys->lock); 78 list_for_each_entry(h, &subsys->nsheads, entry) 79 if (h->disk) 80 blk_freeze_queue_start(h->disk->queue); 81 } 82 83 void nvme_failover_req(struct request *req) 84 { 85 struct nvme_ns *ns = req->q->queuedata; 86 u16 status = nvme_req(req)->status & 0x7ff; 87 unsigned long flags; 88 struct bio *bio; 89 90 nvme_mpath_clear_current_path(ns); 91 92 /* 93 * If we got back an ANA error, we know the controller is alive but not 94 * ready to serve this namespace. Kick of a re-read of the ANA 95 * information page, and just try any other available path for now. 96 */ 97 if (nvme_is_ana_error(status) && ns->ctrl->ana_log_buf) { 98 set_bit(NVME_NS_ANA_PENDING, &ns->flags); 99 queue_work(nvme_wq, &ns->ctrl->ana_work); 100 } 101 102 spin_lock_irqsave(&ns->head->requeue_lock, flags); 103 for (bio = req->bio; bio; bio = bio->bi_next) { 104 bio_set_dev(bio, ns->head->disk->part0); 105 if (bio->bi_opf & REQ_POLLED) { 106 bio->bi_opf &= ~REQ_POLLED; 107 bio->bi_cookie = BLK_QC_T_NONE; 108 } 109 } 110 blk_steal_bios(&ns->head->requeue_list, req); 111 spin_unlock_irqrestore(&ns->head->requeue_lock, flags); 112 113 blk_mq_end_request(req, 0); 114 kblockd_schedule_work(&ns->head->requeue_work); 115 } 116 117 void nvme_mpath_start_request(struct request *rq) 118 { 119 struct nvme_ns *ns = rq->q->queuedata; 120 struct gendisk *disk = ns->head->disk; 121 122 if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq)) 123 return; 124 125 nvme_req(rq)->flags |= NVME_MPATH_IO_STATS; 126 nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, 127 blk_rq_bytes(rq) >> SECTOR_SHIFT, 128 req_op(rq), jiffies); 129 } 130 EXPORT_SYMBOL_GPL(nvme_mpath_start_request); 131 132 void nvme_mpath_end_request(struct request *rq) 133 { 134 struct nvme_ns *ns = rq->q->queuedata; 135 136 if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS)) 137 return; 138 bdev_end_io_acct(ns->head->disk->part0, req_op(rq), 139 nvme_req(rq)->start_time); 140 } 141 142 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl) 143 { 144 struct nvme_ns *ns; 145 146 down_read(&ctrl->namespaces_rwsem); 147 list_for_each_entry(ns, &ctrl->namespaces, list) { 148 if (!ns->head->disk) 149 continue; 150 kblockd_schedule_work(&ns->head->requeue_work); 151 if (ctrl->state == NVME_CTRL_LIVE) 152 disk_uevent(ns->head->disk, KOBJ_CHANGE); 153 } 154 up_read(&ctrl->namespaces_rwsem); 155 } 156 157 static const char *nvme_ana_state_names[] = { 158 [0] = "invalid state", 159 [NVME_ANA_OPTIMIZED] = "optimized", 160 [NVME_ANA_NONOPTIMIZED] = "non-optimized", 161 [NVME_ANA_INACCESSIBLE] = "inaccessible", 162 [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss", 163 [NVME_ANA_CHANGE] = "change", 164 }; 165 166 bool nvme_mpath_clear_current_path(struct nvme_ns *ns) 167 { 168 struct nvme_ns_head *head = ns->head; 169 bool changed = false; 170 int node; 171 172 if (!head) 173 goto out; 174 175 for_each_node(node) { 176 if (ns == rcu_access_pointer(head->current_path[node])) { 177 rcu_assign_pointer(head->current_path[node], NULL); 178 changed = true; 179 } 180 } 181 out: 182 return changed; 183 } 184 185 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl) 186 { 187 struct nvme_ns *ns; 188 189 down_read(&ctrl->namespaces_rwsem); 190 list_for_each_entry(ns, &ctrl->namespaces, list) { 191 nvme_mpath_clear_current_path(ns); 192 kblockd_schedule_work(&ns->head->requeue_work); 193 } 194 up_read(&ctrl->namespaces_rwsem); 195 } 196 197 void nvme_mpath_revalidate_paths(struct nvme_ns *ns) 198 { 199 struct nvme_ns_head *head = ns->head; 200 sector_t capacity = get_capacity(head->disk); 201 int node; 202 int srcu_idx; 203 204 srcu_idx = srcu_read_lock(&head->srcu); 205 list_for_each_entry_rcu(ns, &head->list, siblings) { 206 if (capacity != get_capacity(ns->disk)) 207 clear_bit(NVME_NS_READY, &ns->flags); 208 } 209 srcu_read_unlock(&head->srcu, srcu_idx); 210 211 for_each_node(node) 212 rcu_assign_pointer(head->current_path[node], NULL); 213 kblockd_schedule_work(&head->requeue_work); 214 } 215 216 static bool nvme_path_is_disabled(struct nvme_ns *ns) 217 { 218 /* 219 * We don't treat NVME_CTRL_DELETING as a disabled path as I/O should 220 * still be able to complete assuming that the controller is connected. 221 * Otherwise it will fail immediately and return to the requeue list. 222 */ 223 if (ns->ctrl->state != NVME_CTRL_LIVE && 224 ns->ctrl->state != NVME_CTRL_DELETING) 225 return true; 226 if (test_bit(NVME_NS_ANA_PENDING, &ns->flags) || 227 !test_bit(NVME_NS_READY, &ns->flags)) 228 return true; 229 return false; 230 } 231 232 static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node) 233 { 234 int found_distance = INT_MAX, fallback_distance = INT_MAX, distance; 235 struct nvme_ns *found = NULL, *fallback = NULL, *ns; 236 237 list_for_each_entry_rcu(ns, &head->list, siblings) { 238 if (nvme_path_is_disabled(ns)) 239 continue; 240 241 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA) 242 distance = node_distance(node, ns->ctrl->numa_node); 243 else 244 distance = LOCAL_DISTANCE; 245 246 switch (ns->ana_state) { 247 case NVME_ANA_OPTIMIZED: 248 if (distance < found_distance) { 249 found_distance = distance; 250 found = ns; 251 } 252 break; 253 case NVME_ANA_NONOPTIMIZED: 254 if (distance < fallback_distance) { 255 fallback_distance = distance; 256 fallback = ns; 257 } 258 break; 259 default: 260 break; 261 } 262 } 263 264 if (!found) 265 found = fallback; 266 if (found) 267 rcu_assign_pointer(head->current_path[node], found); 268 return found; 269 } 270 271 static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head, 272 struct nvme_ns *ns) 273 { 274 ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns, 275 siblings); 276 if (ns) 277 return ns; 278 return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings); 279 } 280 281 static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head, 282 int node, struct nvme_ns *old) 283 { 284 struct nvme_ns *ns, *found = NULL; 285 286 if (list_is_singular(&head->list)) { 287 if (nvme_path_is_disabled(old)) 288 return NULL; 289 return old; 290 } 291 292 for (ns = nvme_next_ns(head, old); 293 ns && ns != old; 294 ns = nvme_next_ns(head, ns)) { 295 if (nvme_path_is_disabled(ns)) 296 continue; 297 298 if (ns->ana_state == NVME_ANA_OPTIMIZED) { 299 found = ns; 300 goto out; 301 } 302 if (ns->ana_state == NVME_ANA_NONOPTIMIZED) 303 found = ns; 304 } 305 306 /* 307 * The loop above skips the current path for round-robin semantics. 308 * Fall back to the current path if either: 309 * - no other optimized path found and current is optimized, 310 * - no other usable path found and current is usable. 311 */ 312 if (!nvme_path_is_disabled(old) && 313 (old->ana_state == NVME_ANA_OPTIMIZED || 314 (!found && old->ana_state == NVME_ANA_NONOPTIMIZED))) 315 return old; 316 317 if (!found) 318 return NULL; 319 out: 320 rcu_assign_pointer(head->current_path[node], found); 321 return found; 322 } 323 324 static inline bool nvme_path_is_optimized(struct nvme_ns *ns) 325 { 326 return ns->ctrl->state == NVME_CTRL_LIVE && 327 ns->ana_state == NVME_ANA_OPTIMIZED; 328 } 329 330 inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head) 331 { 332 int node = numa_node_id(); 333 struct nvme_ns *ns; 334 335 ns = srcu_dereference(head->current_path[node], &head->srcu); 336 if (unlikely(!ns)) 337 return __nvme_find_path(head, node); 338 339 if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR) 340 return nvme_round_robin_path(head, node, ns); 341 if (unlikely(!nvme_path_is_optimized(ns))) 342 return __nvme_find_path(head, node); 343 return ns; 344 } 345 346 static bool nvme_available_path(struct nvme_ns_head *head) 347 { 348 struct nvme_ns *ns; 349 350 list_for_each_entry_rcu(ns, &head->list, siblings) { 351 if (test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ns->ctrl->flags)) 352 continue; 353 switch (ns->ctrl->state) { 354 case NVME_CTRL_LIVE: 355 case NVME_CTRL_RESETTING: 356 case NVME_CTRL_CONNECTING: 357 /* fallthru */ 358 return true; 359 default: 360 break; 361 } 362 } 363 return false; 364 } 365 366 static void nvme_ns_head_submit_bio(struct bio *bio) 367 { 368 struct nvme_ns_head *head = bio->bi_bdev->bd_disk->private_data; 369 struct device *dev = disk_to_dev(head->disk); 370 struct nvme_ns *ns; 371 int srcu_idx; 372 373 /* 374 * The namespace might be going away and the bio might be moved to a 375 * different queue via blk_steal_bios(), so we need to use the bio_split 376 * pool from the original queue to allocate the bvecs from. 377 */ 378 bio = bio_split_to_limits(bio); 379 380 srcu_idx = srcu_read_lock(&head->srcu); 381 ns = nvme_find_path(head); 382 if (likely(ns)) { 383 bio_set_dev(bio, ns->disk->part0); 384 bio->bi_opf |= REQ_NVME_MPATH; 385 trace_block_bio_remap(bio, disk_devt(ns->head->disk), 386 bio->bi_iter.bi_sector); 387 submit_bio_noacct(bio); 388 } else if (nvme_available_path(head)) { 389 dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n"); 390 391 spin_lock_irq(&head->requeue_lock); 392 bio_list_add(&head->requeue_list, bio); 393 spin_unlock_irq(&head->requeue_lock); 394 } else { 395 dev_warn_ratelimited(dev, "no available path - failing I/O\n"); 396 397 bio_io_error(bio); 398 } 399 400 srcu_read_unlock(&head->srcu, srcu_idx); 401 } 402 403 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode) 404 { 405 if (!nvme_tryget_ns_head(bdev->bd_disk->private_data)) 406 return -ENXIO; 407 return 0; 408 } 409 410 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode) 411 { 412 nvme_put_ns_head(disk->private_data); 413 } 414 415 #ifdef CONFIG_BLK_DEV_ZONED 416 static int nvme_ns_head_report_zones(struct gendisk *disk, sector_t sector, 417 unsigned int nr_zones, report_zones_cb cb, void *data) 418 { 419 struct nvme_ns_head *head = disk->private_data; 420 struct nvme_ns *ns; 421 int srcu_idx, ret = -EWOULDBLOCK; 422 423 srcu_idx = srcu_read_lock(&head->srcu); 424 ns = nvme_find_path(head); 425 if (ns) 426 ret = nvme_ns_report_zones(ns, sector, nr_zones, cb, data); 427 srcu_read_unlock(&head->srcu, srcu_idx); 428 return ret; 429 } 430 #else 431 #define nvme_ns_head_report_zones NULL 432 #endif /* CONFIG_BLK_DEV_ZONED */ 433 434 const struct block_device_operations nvme_ns_head_ops = { 435 .owner = THIS_MODULE, 436 .submit_bio = nvme_ns_head_submit_bio, 437 .open = nvme_ns_head_open, 438 .release = nvme_ns_head_release, 439 .ioctl = nvme_ns_head_ioctl, 440 .compat_ioctl = blkdev_compat_ptr_ioctl, 441 .getgeo = nvme_getgeo, 442 .report_zones = nvme_ns_head_report_zones, 443 .pr_ops = &nvme_pr_ops, 444 }; 445 446 static inline struct nvme_ns_head *cdev_to_ns_head(struct cdev *cdev) 447 { 448 return container_of(cdev, struct nvme_ns_head, cdev); 449 } 450 451 static int nvme_ns_head_chr_open(struct inode *inode, struct file *file) 452 { 453 if (!nvme_tryget_ns_head(cdev_to_ns_head(inode->i_cdev))) 454 return -ENXIO; 455 return 0; 456 } 457 458 static int nvme_ns_head_chr_release(struct inode *inode, struct file *file) 459 { 460 nvme_put_ns_head(cdev_to_ns_head(inode->i_cdev)); 461 return 0; 462 } 463 464 static const struct file_operations nvme_ns_head_chr_fops = { 465 .owner = THIS_MODULE, 466 .open = nvme_ns_head_chr_open, 467 .release = nvme_ns_head_chr_release, 468 .unlocked_ioctl = nvme_ns_head_chr_ioctl, 469 .compat_ioctl = compat_ptr_ioctl, 470 .uring_cmd = nvme_ns_head_chr_uring_cmd, 471 .uring_cmd_iopoll = nvme_ns_head_chr_uring_cmd_iopoll, 472 }; 473 474 static int nvme_add_ns_head_cdev(struct nvme_ns_head *head) 475 { 476 int ret; 477 478 head->cdev_device.parent = &head->subsys->dev; 479 ret = dev_set_name(&head->cdev_device, "ng%dn%d", 480 head->subsys->instance, head->instance); 481 if (ret) 482 return ret; 483 ret = nvme_cdev_add(&head->cdev, &head->cdev_device, 484 &nvme_ns_head_chr_fops, THIS_MODULE); 485 return ret; 486 } 487 488 static void nvme_requeue_work(struct work_struct *work) 489 { 490 struct nvme_ns_head *head = 491 container_of(work, struct nvme_ns_head, requeue_work); 492 struct bio *bio, *next; 493 494 spin_lock_irq(&head->requeue_lock); 495 next = bio_list_get(&head->requeue_list); 496 spin_unlock_irq(&head->requeue_lock); 497 498 while ((bio = next) != NULL) { 499 next = bio->bi_next; 500 bio->bi_next = NULL; 501 502 submit_bio_noacct(bio); 503 } 504 } 505 506 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head) 507 { 508 bool vwc = false; 509 510 mutex_init(&head->lock); 511 bio_list_init(&head->requeue_list); 512 spin_lock_init(&head->requeue_lock); 513 INIT_WORK(&head->requeue_work, nvme_requeue_work); 514 515 /* 516 * Add a multipath node if the subsystems supports multiple controllers. 517 * We also do this for private namespaces as the namespace sharing flag 518 * could change after a rescan. 519 */ 520 if (!(ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) || 521 !nvme_is_unique_nsid(ctrl, head) || !multipath) 522 return 0; 523 524 head->disk = blk_alloc_disk(ctrl->numa_node); 525 if (!head->disk) 526 return -ENOMEM; 527 head->disk->fops = &nvme_ns_head_ops; 528 head->disk->private_data = head; 529 sprintf(head->disk->disk_name, "nvme%dn%d", 530 ctrl->subsys->instance, head->instance); 531 532 blk_queue_flag_set(QUEUE_FLAG_NONROT, head->disk->queue); 533 blk_queue_flag_set(QUEUE_FLAG_NOWAIT, head->disk->queue); 534 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, head->disk->queue); 535 /* 536 * This assumes all controllers that refer to a namespace either 537 * support poll queues or not. That is not a strict guarantee, 538 * but if the assumption is wrong the effect is only suboptimal 539 * performance but not correctness problem. 540 */ 541 if (ctrl->tagset->nr_maps > HCTX_TYPE_POLL && 542 ctrl->tagset->map[HCTX_TYPE_POLL].nr_queues) 543 blk_queue_flag_set(QUEUE_FLAG_POLL, head->disk->queue); 544 545 /* set to a default value of 512 until the disk is validated */ 546 blk_queue_logical_block_size(head->disk->queue, 512); 547 blk_set_stacking_limits(&head->disk->queue->limits); 548 blk_queue_dma_alignment(head->disk->queue, 3); 549 550 /* we need to propagate up the VMC settings */ 551 if (ctrl->vwc & NVME_CTRL_VWC_PRESENT) 552 vwc = true; 553 blk_queue_write_cache(head->disk->queue, vwc, vwc); 554 return 0; 555 } 556 557 static void nvme_mpath_set_live(struct nvme_ns *ns) 558 { 559 struct nvme_ns_head *head = ns->head; 560 int rc; 561 562 if (!head->disk) 563 return; 564 565 /* 566 * test_and_set_bit() is used because it is protecting against two nvme 567 * paths simultaneously calling device_add_disk() on the same namespace 568 * head. 569 */ 570 if (!test_and_set_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) { 571 rc = device_add_disk(&head->subsys->dev, head->disk, 572 nvme_ns_id_attr_groups); 573 if (rc) { 574 clear_bit(NVME_NSHEAD_DISK_LIVE, &ns->flags); 575 return; 576 } 577 nvme_add_ns_head_cdev(head); 578 } 579 580 mutex_lock(&head->lock); 581 if (nvme_path_is_optimized(ns)) { 582 int node, srcu_idx; 583 584 srcu_idx = srcu_read_lock(&head->srcu); 585 for_each_node(node) 586 __nvme_find_path(head, node); 587 srcu_read_unlock(&head->srcu, srcu_idx); 588 } 589 mutex_unlock(&head->lock); 590 591 synchronize_srcu(&head->srcu); 592 kblockd_schedule_work(&head->requeue_work); 593 } 594 595 static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data, 596 int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *, 597 void *)) 598 { 599 void *base = ctrl->ana_log_buf; 600 size_t offset = sizeof(struct nvme_ana_rsp_hdr); 601 int error, i; 602 603 lockdep_assert_held(&ctrl->ana_lock); 604 605 for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) { 606 struct nvme_ana_group_desc *desc = base + offset; 607 u32 nr_nsids; 608 size_t nsid_buf_size; 609 610 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc))) 611 return -EINVAL; 612 613 nr_nsids = le32_to_cpu(desc->nnsids); 614 nsid_buf_size = flex_array_size(desc, nsids, nr_nsids); 615 616 if (WARN_ON_ONCE(desc->grpid == 0)) 617 return -EINVAL; 618 if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax)) 619 return -EINVAL; 620 if (WARN_ON_ONCE(desc->state == 0)) 621 return -EINVAL; 622 if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE)) 623 return -EINVAL; 624 625 offset += sizeof(*desc); 626 if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size)) 627 return -EINVAL; 628 629 error = cb(ctrl, desc, data); 630 if (error) 631 return error; 632 633 offset += nsid_buf_size; 634 } 635 636 return 0; 637 } 638 639 static inline bool nvme_state_is_live(enum nvme_ana_state state) 640 { 641 return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED; 642 } 643 644 static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc, 645 struct nvme_ns *ns) 646 { 647 ns->ana_grpid = le32_to_cpu(desc->grpid); 648 ns->ana_state = desc->state; 649 clear_bit(NVME_NS_ANA_PENDING, &ns->flags); 650 /* 651 * nvme_mpath_set_live() will trigger I/O to the multipath path device 652 * and in turn to this path device. However we cannot accept this I/O 653 * if the controller is not live. This may deadlock if called from 654 * nvme_mpath_init_identify() and the ctrl will never complete 655 * initialization, preventing I/O from completing. For this case we 656 * will reprocess the ANA log page in nvme_mpath_update() once the 657 * controller is ready. 658 */ 659 if (nvme_state_is_live(ns->ana_state) && 660 ns->ctrl->state == NVME_CTRL_LIVE) 661 nvme_mpath_set_live(ns); 662 } 663 664 static int nvme_update_ana_state(struct nvme_ctrl *ctrl, 665 struct nvme_ana_group_desc *desc, void *data) 666 { 667 u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0; 668 unsigned *nr_change_groups = data; 669 struct nvme_ns *ns; 670 671 dev_dbg(ctrl->device, "ANA group %d: %s.\n", 672 le32_to_cpu(desc->grpid), 673 nvme_ana_state_names[desc->state]); 674 675 if (desc->state == NVME_ANA_CHANGE) 676 (*nr_change_groups)++; 677 678 if (!nr_nsids) 679 return 0; 680 681 down_read(&ctrl->namespaces_rwsem); 682 list_for_each_entry(ns, &ctrl->namespaces, list) { 683 unsigned nsid; 684 again: 685 nsid = le32_to_cpu(desc->nsids[n]); 686 if (ns->head->ns_id < nsid) 687 continue; 688 if (ns->head->ns_id == nsid) 689 nvme_update_ns_ana_state(desc, ns); 690 if (++n == nr_nsids) 691 break; 692 if (ns->head->ns_id > nsid) 693 goto again; 694 } 695 up_read(&ctrl->namespaces_rwsem); 696 return 0; 697 } 698 699 static int nvme_read_ana_log(struct nvme_ctrl *ctrl) 700 { 701 u32 nr_change_groups = 0; 702 int error; 703 704 mutex_lock(&ctrl->ana_lock); 705 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, 0, NVME_CSI_NVM, 706 ctrl->ana_log_buf, ctrl->ana_log_size, 0); 707 if (error) { 708 dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error); 709 goto out_unlock; 710 } 711 712 error = nvme_parse_ana_log(ctrl, &nr_change_groups, 713 nvme_update_ana_state); 714 if (error) 715 goto out_unlock; 716 717 /* 718 * In theory we should have an ANATT timer per group as they might enter 719 * the change state at different times. But that is a lot of overhead 720 * just to protect against a target that keeps entering new changes 721 * states while never finishing previous ones. But we'll still 722 * eventually time out once all groups are in change state, so this 723 * isn't a big deal. 724 * 725 * We also double the ANATT value to provide some slack for transports 726 * or AEN processing overhead. 727 */ 728 if (nr_change_groups) 729 mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies); 730 else 731 del_timer_sync(&ctrl->anatt_timer); 732 out_unlock: 733 mutex_unlock(&ctrl->ana_lock); 734 return error; 735 } 736 737 static void nvme_ana_work(struct work_struct *work) 738 { 739 struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work); 740 741 if (ctrl->state != NVME_CTRL_LIVE) 742 return; 743 744 nvme_read_ana_log(ctrl); 745 } 746 747 void nvme_mpath_update(struct nvme_ctrl *ctrl) 748 { 749 u32 nr_change_groups = 0; 750 751 if (!ctrl->ana_log_buf) 752 return; 753 754 mutex_lock(&ctrl->ana_lock); 755 nvme_parse_ana_log(ctrl, &nr_change_groups, nvme_update_ana_state); 756 mutex_unlock(&ctrl->ana_lock); 757 } 758 759 static void nvme_anatt_timeout(struct timer_list *t) 760 { 761 struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer); 762 763 dev_info(ctrl->device, "ANATT timeout, resetting controller.\n"); 764 nvme_reset_ctrl(ctrl); 765 } 766 767 void nvme_mpath_stop(struct nvme_ctrl *ctrl) 768 { 769 if (!nvme_ctrl_use_ana(ctrl)) 770 return; 771 del_timer_sync(&ctrl->anatt_timer); 772 cancel_work_sync(&ctrl->ana_work); 773 } 774 775 #define SUBSYS_ATTR_RW(_name, _mode, _show, _store) \ 776 struct device_attribute subsys_attr_##_name = \ 777 __ATTR(_name, _mode, _show, _store) 778 779 static ssize_t nvme_subsys_iopolicy_show(struct device *dev, 780 struct device_attribute *attr, char *buf) 781 { 782 struct nvme_subsystem *subsys = 783 container_of(dev, struct nvme_subsystem, dev); 784 785 return sysfs_emit(buf, "%s\n", 786 nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]); 787 } 788 789 static ssize_t nvme_subsys_iopolicy_store(struct device *dev, 790 struct device_attribute *attr, const char *buf, size_t count) 791 { 792 struct nvme_subsystem *subsys = 793 container_of(dev, struct nvme_subsystem, dev); 794 int i; 795 796 for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) { 797 if (sysfs_streq(buf, nvme_iopolicy_names[i])) { 798 WRITE_ONCE(subsys->iopolicy, i); 799 return count; 800 } 801 } 802 803 return -EINVAL; 804 } 805 SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR, 806 nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store); 807 808 static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr, 809 char *buf) 810 { 811 return sysfs_emit(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid); 812 } 813 DEVICE_ATTR_RO(ana_grpid); 814 815 static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr, 816 char *buf) 817 { 818 struct nvme_ns *ns = nvme_get_ns_from_dev(dev); 819 820 return sysfs_emit(buf, "%s\n", nvme_ana_state_names[ns->ana_state]); 821 } 822 DEVICE_ATTR_RO(ana_state); 823 824 static int nvme_lookup_ana_group_desc(struct nvme_ctrl *ctrl, 825 struct nvme_ana_group_desc *desc, void *data) 826 { 827 struct nvme_ana_group_desc *dst = data; 828 829 if (desc->grpid != dst->grpid) 830 return 0; 831 832 *dst = *desc; 833 return -ENXIO; /* just break out of the loop */ 834 } 835 836 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid) 837 { 838 if (nvme_ctrl_use_ana(ns->ctrl)) { 839 struct nvme_ana_group_desc desc = { 840 .grpid = anagrpid, 841 .state = 0, 842 }; 843 844 mutex_lock(&ns->ctrl->ana_lock); 845 ns->ana_grpid = le32_to_cpu(anagrpid); 846 nvme_parse_ana_log(ns->ctrl, &desc, nvme_lookup_ana_group_desc); 847 mutex_unlock(&ns->ctrl->ana_lock); 848 if (desc.state) { 849 /* found the group desc: update */ 850 nvme_update_ns_ana_state(&desc, ns); 851 } else { 852 /* group desc not found: trigger a re-read */ 853 set_bit(NVME_NS_ANA_PENDING, &ns->flags); 854 queue_work(nvme_wq, &ns->ctrl->ana_work); 855 } 856 } else { 857 ns->ana_state = NVME_ANA_OPTIMIZED; 858 nvme_mpath_set_live(ns); 859 } 860 861 if (blk_queue_stable_writes(ns->queue) && ns->head->disk) 862 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, 863 ns->head->disk->queue); 864 #ifdef CONFIG_BLK_DEV_ZONED 865 if (blk_queue_is_zoned(ns->queue) && ns->head->disk) 866 ns->head->disk->nr_zones = ns->disk->nr_zones; 867 #endif 868 } 869 870 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head) 871 { 872 if (!head->disk) 873 return; 874 kblockd_schedule_work(&head->requeue_work); 875 if (test_bit(NVME_NSHEAD_DISK_LIVE, &head->flags)) { 876 nvme_cdev_del(&head->cdev, &head->cdev_device); 877 del_gendisk(head->disk); 878 } 879 } 880 881 void nvme_mpath_remove_disk(struct nvme_ns_head *head) 882 { 883 if (!head->disk) 884 return; 885 blk_mark_disk_dead(head->disk); 886 /* make sure all pending bios are cleaned up */ 887 kblockd_schedule_work(&head->requeue_work); 888 flush_work(&head->requeue_work); 889 put_disk(head->disk); 890 } 891 892 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl) 893 { 894 mutex_init(&ctrl->ana_lock); 895 timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0); 896 INIT_WORK(&ctrl->ana_work, nvme_ana_work); 897 } 898 899 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id) 900 { 901 size_t max_transfer_size = ctrl->max_hw_sectors << SECTOR_SHIFT; 902 size_t ana_log_size; 903 int error = 0; 904 905 /* check if multipath is enabled and we have the capability */ 906 if (!multipath || !ctrl->subsys || 907 !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA)) 908 return 0; 909 910 if (!ctrl->max_namespaces || 911 ctrl->max_namespaces > le32_to_cpu(id->nn)) { 912 dev_err(ctrl->device, 913 "Invalid MNAN value %u\n", ctrl->max_namespaces); 914 return -EINVAL; 915 } 916 917 ctrl->anacap = id->anacap; 918 ctrl->anatt = id->anatt; 919 ctrl->nanagrpid = le32_to_cpu(id->nanagrpid); 920 ctrl->anagrpmax = le32_to_cpu(id->anagrpmax); 921 922 ana_log_size = sizeof(struct nvme_ana_rsp_hdr) + 923 ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc) + 924 ctrl->max_namespaces * sizeof(__le32); 925 if (ana_log_size > max_transfer_size) { 926 dev_err(ctrl->device, 927 "ANA log page size (%zd) larger than MDTS (%zd).\n", 928 ana_log_size, max_transfer_size); 929 dev_err(ctrl->device, "disabling ANA support.\n"); 930 goto out_uninit; 931 } 932 if (ana_log_size > ctrl->ana_log_size) { 933 nvme_mpath_stop(ctrl); 934 nvme_mpath_uninit(ctrl); 935 ctrl->ana_log_buf = kvmalloc(ana_log_size, GFP_KERNEL); 936 if (!ctrl->ana_log_buf) 937 return -ENOMEM; 938 } 939 ctrl->ana_log_size = ana_log_size; 940 error = nvme_read_ana_log(ctrl); 941 if (error) 942 goto out_uninit; 943 return 0; 944 945 out_uninit: 946 nvme_mpath_uninit(ctrl); 947 return error; 948 } 949 950 void nvme_mpath_uninit(struct nvme_ctrl *ctrl) 951 { 952 kvfree(ctrl->ana_log_buf); 953 ctrl->ana_log_buf = NULL; 954 ctrl->ana_log_size = 0; 955 } 956