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