Christoph Hellwig | bc50ad7 | 2019-02-18 09:36:29 +0100 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 2 | /* |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 3 | * Copyright (c) 2017-2018 Christoph Hellwig. |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 4 | */ |
| 5 | |
| 6 | #include <linux/moduleparam.h> |
Hannes Reinecke | 2796b56 | 2018-06-07 10:38:47 +0200 | [diff] [blame] | 7 | #include <trace/events/block.h> |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 8 | #include "nvme.h" |
| 9 | |
| 10 | static bool multipath = true; |
Keith Busch | 5cadde8 | 2018-04-26 14:24:29 -0600 | [diff] [blame] | 11 | module_param(multipath, bool, 0444); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 12 | MODULE_PARM_DESC(multipath, |
| 13 | "turn on native support for multiple controllers per subsystem"); |
| 14 | |
Sagi Grimberg | b9156da | 2019-07-31 11:00:26 -0700 | [diff] [blame] | 15 | void nvme_mpath_unfreeze(struct nvme_subsystem *subsys) |
| 16 | { |
| 17 | struct nvme_ns_head *h; |
| 18 | |
| 19 | lockdep_assert_held(&subsys->lock); |
| 20 | list_for_each_entry(h, &subsys->nsheads, entry) |
| 21 | if (h->disk) |
| 22 | blk_mq_unfreeze_queue(h->disk->queue); |
| 23 | } |
| 24 | |
| 25 | void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys) |
| 26 | { |
| 27 | struct nvme_ns_head *h; |
| 28 | |
| 29 | lockdep_assert_held(&subsys->lock); |
| 30 | list_for_each_entry(h, &subsys->nsheads, entry) |
| 31 | if (h->disk) |
| 32 | blk_mq_freeze_queue_wait(h->disk->queue); |
| 33 | } |
| 34 | |
| 35 | void nvme_mpath_start_freeze(struct nvme_subsystem *subsys) |
| 36 | { |
| 37 | struct nvme_ns_head *h; |
| 38 | |
| 39 | lockdep_assert_held(&subsys->lock); |
| 40 | list_for_each_entry(h, &subsys->nsheads, entry) |
| 41 | if (h->disk) |
| 42 | blk_freeze_queue_start(h->disk->queue); |
| 43 | } |
| 44 | |
Keith Busch | a785dbc | 2018-04-26 14:22:41 -0600 | [diff] [blame] | 45 | /* |
| 46 | * If multipathing is enabled we need to always use the subsystem instance |
| 47 | * number for numbering our devices to avoid conflicts between subsystems that |
| 48 | * have multiple controllers and thus use the multipath-aware subsystem node |
| 49 | * and those that have a single controller and use the controller node |
| 50 | * directly. |
| 51 | */ |
| 52 | void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns, |
| 53 | struct nvme_ctrl *ctrl, int *flags) |
| 54 | { |
| 55 | if (!multipath) { |
| 56 | sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance); |
| 57 | } else if (ns->head->disk) { |
| 58 | sprintf(disk_name, "nvme%dc%dn%d", ctrl->subsys->instance, |
Hannes Reinecke | 8a03b27 | 2019-05-03 15:37:35 +0200 | [diff] [blame] | 59 | ctrl->instance, ns->head->instance); |
Keith Busch | a785dbc | 2018-04-26 14:22:41 -0600 | [diff] [blame] | 60 | *flags = GENHD_FL_HIDDEN; |
| 61 | } else { |
| 62 | sprintf(disk_name, "nvme%dn%d", ctrl->subsys->instance, |
| 63 | ns->head->instance); |
| 64 | } |
| 65 | } |
| 66 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 67 | void nvme_failover_req(struct request *req) |
| 68 | { |
| 69 | struct nvme_ns *ns = req->q->queuedata; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 70 | u16 status = nvme_req(req)->status; |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 71 | unsigned long flags; |
| 72 | |
| 73 | spin_lock_irqsave(&ns->head->requeue_lock, flags); |
| 74 | blk_steal_bios(&ns->head->requeue_list, req); |
| 75 | spin_unlock_irqrestore(&ns->head->requeue_lock, flags); |
| 76 | blk_mq_end_request(req, 0); |
| 77 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 78 | switch (status & 0x7ff) { |
| 79 | case NVME_SC_ANA_TRANSITION: |
| 80 | case NVME_SC_ANA_INACCESSIBLE: |
| 81 | case NVME_SC_ANA_PERSISTENT_LOSS: |
| 82 | /* |
| 83 | * If we got back an ANA error we know the controller is alive, |
| 84 | * but not ready to serve this namespaces. The spec suggests |
| 85 | * we should update our general state here, but due to the fact |
| 86 | * that the admin and I/O queues are not serialized that is |
| 87 | * fundamentally racy. So instead just clear the current path, |
| 88 | * mark the the path as pending and kick of a re-read of the ANA |
| 89 | * log page ASAP. |
| 90 | */ |
| 91 | nvme_mpath_clear_current_path(ns); |
| 92 | if (ns->ctrl->ana_log_buf) { |
| 93 | set_bit(NVME_NS_ANA_PENDING, &ns->flags); |
| 94 | queue_work(nvme_wq, &ns->ctrl->ana_work); |
| 95 | } |
| 96 | break; |
James Smart | 783f4a4 | 2018-09-27 16:58:54 -0700 | [diff] [blame] | 97 | case NVME_SC_HOST_PATH_ERROR: |
Max Gurtovoy | 2dc3947 | 2019-10-13 19:57:35 +0300 | [diff] [blame^] | 98 | case NVME_SC_HOST_ABORTED_CMD: |
James Smart | 783f4a4 | 2018-09-27 16:58:54 -0700 | [diff] [blame] | 99 | /* |
| 100 | * Temporary transport disruption in talking to the controller. |
| 101 | * Try to send on a new path. |
| 102 | */ |
| 103 | nvme_mpath_clear_current_path(ns); |
| 104 | break; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 105 | default: |
| 106 | /* |
| 107 | * Reset the controller for any non-ANA error as we don't know |
| 108 | * what caused the error. |
| 109 | */ |
| 110 | nvme_reset_ctrl(ns->ctrl); |
| 111 | break; |
| 112 | } |
| 113 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 114 | kblockd_schedule_work(&ns->head->requeue_work); |
| 115 | } |
| 116 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 117 | void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl) |
| 118 | { |
| 119 | struct nvme_ns *ns; |
| 120 | |
Jianchao Wang | 765cc031 | 2018-02-12 20:54:46 +0800 | [diff] [blame] | 121 | down_read(&ctrl->namespaces_rwsem); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 122 | list_for_each_entry(ns, &ctrl->namespaces, list) { |
| 123 | if (ns->head->disk) |
| 124 | kblockd_schedule_work(&ns->head->requeue_work); |
| 125 | } |
Jianchao Wang | 765cc031 | 2018-02-12 20:54:46 +0800 | [diff] [blame] | 126 | up_read(&ctrl->namespaces_rwsem); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 127 | } |
| 128 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 129 | static const char *nvme_ana_state_names[] = { |
| 130 | [0] = "invalid state", |
| 131 | [NVME_ANA_OPTIMIZED] = "optimized", |
| 132 | [NVME_ANA_NONOPTIMIZED] = "non-optimized", |
| 133 | [NVME_ANA_INACCESSIBLE] = "inaccessible", |
| 134 | [NVME_ANA_PERSISTENT_LOSS] = "persistent-loss", |
| 135 | [NVME_ANA_CHANGE] = "change", |
| 136 | }; |
| 137 | |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 138 | bool nvme_mpath_clear_current_path(struct nvme_ns *ns) |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 139 | { |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 140 | struct nvme_ns_head *head = ns->head; |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 141 | bool changed = false; |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 142 | int node; |
| 143 | |
| 144 | if (!head) |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 145 | goto out; |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 146 | |
| 147 | for_each_node(node) { |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 148 | if (ns == rcu_access_pointer(head->current_path[node])) { |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 149 | rcu_assign_pointer(head->current_path[node], NULL); |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 150 | changed = true; |
| 151 | } |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 152 | } |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 153 | out: |
| 154 | return changed; |
| 155 | } |
| 156 | |
| 157 | void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl) |
| 158 | { |
| 159 | struct nvme_ns *ns; |
| 160 | |
| 161 | mutex_lock(&ctrl->scan_lock); |
| 162 | list_for_each_entry(ns, &ctrl->namespaces, list) |
| 163 | if (nvme_mpath_clear_current_path(ns)) |
| 164 | kblockd_schedule_work(&ns->head->requeue_work); |
| 165 | mutex_unlock(&ctrl->scan_lock); |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 166 | } |
| 167 | |
Hannes Reinecke | ca7ae5c | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 168 | static bool nvme_path_is_disabled(struct nvme_ns *ns) |
| 169 | { |
| 170 | return ns->ctrl->state != NVME_CTRL_LIVE || |
Hannes Reinecke | 04e70bd | 2019-07-04 08:10:47 +0200 | [diff] [blame] | 171 | test_bit(NVME_NS_ANA_PENDING, &ns->flags) || |
| 172 | test_bit(NVME_NS_REMOVING, &ns->flags); |
Hannes Reinecke | ca7ae5c | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 173 | } |
| 174 | |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 175 | static struct nvme_ns *__nvme_find_path(struct nvme_ns_head *head, int node) |
| 176 | { |
| 177 | int found_distance = INT_MAX, fallback_distance = INT_MAX, distance; |
| 178 | struct nvme_ns *found = NULL, *fallback = NULL, *ns; |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 179 | |
| 180 | list_for_each_entry_rcu(ns, &head->list, siblings) { |
Hannes Reinecke | ca7ae5c | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 181 | if (nvme_path_is_disabled(ns)) |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 182 | continue; |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 183 | |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 184 | if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_NUMA) |
| 185 | distance = node_distance(node, ns->ctrl->numa_node); |
| 186 | else |
| 187 | distance = LOCAL_DISTANCE; |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 188 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 189 | switch (ns->ana_state) { |
| 190 | case NVME_ANA_OPTIMIZED: |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 191 | if (distance < found_distance) { |
| 192 | found_distance = distance; |
| 193 | found = ns; |
| 194 | } |
| 195 | break; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 196 | case NVME_ANA_NONOPTIMIZED: |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 197 | if (distance < fallback_distance) { |
| 198 | fallback_distance = distance; |
| 199 | fallback = ns; |
| 200 | } |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 201 | break; |
| 202 | default: |
| 203 | break; |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 204 | } |
| 205 | } |
| 206 | |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 207 | if (!found) |
| 208 | found = fallback; |
| 209 | if (found) |
| 210 | rcu_assign_pointer(head->current_path[node], found); |
| 211 | return found; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 212 | } |
| 213 | |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 214 | static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head, |
| 215 | struct nvme_ns *ns) |
| 216 | { |
| 217 | ns = list_next_or_null_rcu(&head->list, &ns->siblings, struct nvme_ns, |
| 218 | siblings); |
| 219 | if (ns) |
| 220 | return ns; |
| 221 | return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings); |
| 222 | } |
| 223 | |
| 224 | static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head, |
| 225 | int node, struct nvme_ns *old) |
| 226 | { |
| 227 | struct nvme_ns *ns, *found, *fallback = NULL; |
| 228 | |
Hannes Reinecke | 2032d07 | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 229 | if (list_is_singular(&head->list)) { |
| 230 | if (nvme_path_is_disabled(old)) |
| 231 | return NULL; |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 232 | return old; |
Hannes Reinecke | 2032d07 | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 233 | } |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 234 | |
| 235 | for (ns = nvme_next_ns(head, old); |
| 236 | ns != old; |
| 237 | ns = nvme_next_ns(head, ns)) { |
Hannes Reinecke | ca7ae5c | 2019-07-04 08:10:46 +0200 | [diff] [blame] | 238 | if (nvme_path_is_disabled(ns)) |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 239 | continue; |
| 240 | |
| 241 | if (ns->ana_state == NVME_ANA_OPTIMIZED) { |
| 242 | found = ns; |
| 243 | goto out; |
| 244 | } |
| 245 | if (ns->ana_state == NVME_ANA_NONOPTIMIZED) |
| 246 | fallback = ns; |
| 247 | } |
| 248 | |
| 249 | if (!fallback) |
| 250 | return NULL; |
| 251 | found = fallback; |
| 252 | out: |
| 253 | rcu_assign_pointer(head->current_path[node], found); |
| 254 | return found; |
| 255 | } |
| 256 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 257 | static inline bool nvme_path_is_optimized(struct nvme_ns *ns) |
| 258 | { |
| 259 | return ns->ctrl->state == NVME_CTRL_LIVE && |
| 260 | ns->ana_state == NVME_ANA_OPTIMIZED; |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 261 | } |
| 262 | |
| 263 | inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head) |
| 264 | { |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 265 | int node = numa_node_id(); |
| 266 | struct nvme_ns *ns; |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 267 | |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 268 | ns = srcu_dereference(head->current_path[node], &head->srcu); |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 269 | if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR && ns) |
| 270 | ns = nvme_round_robin_path(head, node, ns); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 271 | if (unlikely(!ns || !nvme_path_is_optimized(ns))) |
Christoph Hellwig | f333444 | 2018-09-11 09:51:29 +0200 | [diff] [blame] | 272 | ns = __nvme_find_path(head, node); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 273 | return ns; |
| 274 | } |
| 275 | |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 276 | static bool nvme_available_path(struct nvme_ns_head *head) |
| 277 | { |
| 278 | struct nvme_ns *ns; |
| 279 | |
| 280 | list_for_each_entry_rcu(ns, &head->list, siblings) { |
| 281 | switch (ns->ctrl->state) { |
| 282 | case NVME_CTRL_LIVE: |
| 283 | case NVME_CTRL_RESETTING: |
| 284 | case NVME_CTRL_CONNECTING: |
| 285 | /* fallthru */ |
| 286 | return true; |
| 287 | default: |
| 288 | break; |
| 289 | } |
| 290 | } |
| 291 | return false; |
| 292 | } |
| 293 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 294 | static blk_qc_t nvme_ns_head_make_request(struct request_queue *q, |
| 295 | struct bio *bio) |
| 296 | { |
| 297 | struct nvme_ns_head *head = q->queuedata; |
| 298 | struct device *dev = disk_to_dev(head->disk); |
| 299 | struct nvme_ns *ns; |
| 300 | blk_qc_t ret = BLK_QC_T_NONE; |
| 301 | int srcu_idx; |
| 302 | |
Hannes Reinecke | 525aa5a | 2019-04-30 18:57:09 +0200 | [diff] [blame] | 303 | /* |
| 304 | * The namespace might be going away and the bio might |
| 305 | * be moved to a different queue via blk_steal_bios(), |
| 306 | * so we need to use the bio_split pool from the original |
| 307 | * queue to allocate the bvecs from. |
| 308 | */ |
| 309 | blk_queue_split(q, &bio); |
| 310 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 311 | srcu_idx = srcu_read_lock(&head->srcu); |
| 312 | ns = nvme_find_path(head); |
| 313 | if (likely(ns)) { |
| 314 | bio->bi_disk = ns->disk; |
| 315 | bio->bi_opf |= REQ_NVME_MPATH; |
Hannes Reinecke | 2796b56 | 2018-06-07 10:38:47 +0200 | [diff] [blame] | 316 | trace_block_bio_remap(bio->bi_disk->queue, bio, |
| 317 | disk_devt(ns->head->disk), |
| 318 | bio->bi_iter.bi_sector); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 319 | ret = direct_make_request(bio); |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 320 | } else if (nvme_available_path(head)) { |
| 321 | dev_warn_ratelimited(dev, "no usable path - requeuing I/O\n"); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 322 | |
| 323 | spin_lock_irq(&head->requeue_lock); |
| 324 | bio_list_add(&head->requeue_list, bio); |
| 325 | spin_unlock_irq(&head->requeue_lock); |
| 326 | } else { |
Sagi Grimberg | 0157ec8 | 2019-07-25 11:56:57 -0700 | [diff] [blame] | 327 | dev_warn_ratelimited(dev, "no available path - failing I/O\n"); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 328 | |
| 329 | bio->bi_status = BLK_STS_IOERR; |
| 330 | bio_endio(bio); |
| 331 | } |
| 332 | |
| 333 | srcu_read_unlock(&head->srcu, srcu_idx); |
| 334 | return ret; |
| 335 | } |
| 336 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 337 | static void nvme_requeue_work(struct work_struct *work) |
| 338 | { |
| 339 | struct nvme_ns_head *head = |
| 340 | container_of(work, struct nvme_ns_head, requeue_work); |
| 341 | struct bio *bio, *next; |
| 342 | |
| 343 | spin_lock_irq(&head->requeue_lock); |
| 344 | next = bio_list_get(&head->requeue_list); |
| 345 | spin_unlock_irq(&head->requeue_lock); |
| 346 | |
| 347 | while ((bio = next) != NULL) { |
| 348 | next = bio->bi_next; |
| 349 | bio->bi_next = NULL; |
| 350 | |
| 351 | /* |
| 352 | * Reset disk to the mpath node and resubmit to select a new |
| 353 | * path. |
| 354 | */ |
| 355 | bio->bi_disk = head->disk; |
| 356 | generic_make_request(bio); |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head) |
| 361 | { |
| 362 | struct request_queue *q; |
| 363 | bool vwc = false; |
| 364 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 365 | mutex_init(&head->lock); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 366 | bio_list_init(&head->requeue_list); |
| 367 | spin_lock_init(&head->requeue_lock); |
| 368 | INIT_WORK(&head->requeue_work, nvme_requeue_work); |
| 369 | |
| 370 | /* |
| 371 | * Add a multipath node if the subsystems supports multiple controllers. |
| 372 | * We also do this for private namespaces as the namespace sharing data could |
| 373 | * change after a rescan. |
| 374 | */ |
| 375 | if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath) |
| 376 | return 0; |
| 377 | |
Hannes Reinecke | 103e515 | 2018-11-16 09:22:29 +0100 | [diff] [blame] | 378 | q = blk_alloc_queue_node(GFP_KERNEL, ctrl->numa_node); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 379 | if (!q) |
| 380 | goto out; |
| 381 | q->queuedata = head; |
| 382 | blk_queue_make_request(q, nvme_ns_head_make_request); |
Bart Van Assche | 8b904b5 | 2018-03-07 17:10:10 -0800 | [diff] [blame] | 383 | blk_queue_flag_set(QUEUE_FLAG_NONROT, q); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 384 | /* set to a default value for 512 until disk is validated */ |
| 385 | blk_queue_logical_block_size(q, 512); |
Sagi Grimberg | 8f676b85 | 2018-11-02 11:22:13 -0700 | [diff] [blame] | 386 | blk_set_stacking_limits(&q->limits); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 387 | |
| 388 | /* we need to propagate up the VMC settings */ |
| 389 | if (ctrl->vwc & NVME_CTRL_VWC_PRESENT) |
| 390 | vwc = true; |
| 391 | blk_queue_write_cache(q, vwc, vwc); |
| 392 | |
| 393 | head->disk = alloc_disk(0); |
| 394 | if (!head->disk) |
| 395 | goto out_cleanup_queue; |
| 396 | head->disk->fops = &nvme_ns_head_ops; |
| 397 | head->disk->private_data = head; |
| 398 | head->disk->queue = q; |
| 399 | head->disk->flags = GENHD_FL_EXT_DEVT; |
| 400 | sprintf(head->disk->disk_name, "nvme%dn%d", |
| 401 | ctrl->subsys->instance, head->instance); |
| 402 | return 0; |
| 403 | |
| 404 | out_cleanup_queue: |
| 405 | blk_cleanup_queue(q); |
| 406 | out: |
| 407 | return -ENOMEM; |
| 408 | } |
| 409 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 410 | static void nvme_mpath_set_live(struct nvme_ns *ns) |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 411 | { |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 412 | struct nvme_ns_head *head = ns->head; |
| 413 | |
| 414 | lockdep_assert_held(&ns->head->lock); |
| 415 | |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 416 | if (!head->disk) |
| 417 | return; |
Baegjae Sung | 9bd82b1 | 2018-02-28 16:06:04 +0900 | [diff] [blame] | 418 | |
Hannes Reinecke | 33b14f67 | 2018-09-28 08:17:20 +0200 | [diff] [blame] | 419 | if (!(head->disk->flags & GENHD_FL_UP)) |
| 420 | device_add_disk(&head->subsys->dev, head->disk, |
| 421 | nvme_ns_id_attr_groups); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 422 | |
Keith Busch | 886fabf | 2018-10-05 09:49:37 -0600 | [diff] [blame] | 423 | if (nvme_path_is_optimized(ns)) { |
| 424 | int node, srcu_idx; |
| 425 | |
| 426 | srcu_idx = srcu_read_lock(&head->srcu); |
| 427 | for_each_node(node) |
| 428 | __nvme_find_path(head, node); |
| 429 | srcu_read_unlock(&head->srcu, srcu_idx); |
| 430 | } |
| 431 | |
Anton Eidelman | 504db08 | 2019-08-12 23:00:36 +0300 | [diff] [blame] | 432 | synchronize_srcu(&ns->head->srcu); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 433 | kblockd_schedule_work(&ns->head->requeue_work); |
| 434 | } |
| 435 | |
| 436 | static int nvme_parse_ana_log(struct nvme_ctrl *ctrl, void *data, |
| 437 | int (*cb)(struct nvme_ctrl *ctrl, struct nvme_ana_group_desc *, |
| 438 | void *)) |
| 439 | { |
| 440 | void *base = ctrl->ana_log_buf; |
| 441 | size_t offset = sizeof(struct nvme_ana_rsp_hdr); |
| 442 | int error, i; |
| 443 | |
| 444 | lockdep_assert_held(&ctrl->ana_lock); |
| 445 | |
| 446 | for (i = 0; i < le16_to_cpu(ctrl->ana_log_buf->ngrps); i++) { |
| 447 | struct nvme_ana_group_desc *desc = base + offset; |
| 448 | u32 nr_nsids = le32_to_cpu(desc->nnsids); |
| 449 | size_t nsid_buf_size = nr_nsids * sizeof(__le32); |
| 450 | |
| 451 | if (WARN_ON_ONCE(desc->grpid == 0)) |
| 452 | return -EINVAL; |
| 453 | if (WARN_ON_ONCE(le32_to_cpu(desc->grpid) > ctrl->anagrpmax)) |
| 454 | return -EINVAL; |
| 455 | if (WARN_ON_ONCE(desc->state == 0)) |
| 456 | return -EINVAL; |
| 457 | if (WARN_ON_ONCE(desc->state > NVME_ANA_CHANGE)) |
| 458 | return -EINVAL; |
| 459 | |
| 460 | offset += sizeof(*desc); |
| 461 | if (WARN_ON_ONCE(offset > ctrl->ana_log_size - nsid_buf_size)) |
| 462 | return -EINVAL; |
| 463 | |
| 464 | error = cb(ctrl, desc, data); |
| 465 | if (error) |
| 466 | return error; |
| 467 | |
| 468 | offset += nsid_buf_size; |
| 469 | if (WARN_ON_ONCE(offset > ctrl->ana_log_size - sizeof(*desc))) |
| 470 | return -EINVAL; |
| 471 | } |
| 472 | |
| 473 | return 0; |
| 474 | } |
| 475 | |
| 476 | static inline bool nvme_state_is_live(enum nvme_ana_state state) |
| 477 | { |
| 478 | return state == NVME_ANA_OPTIMIZED || state == NVME_ANA_NONOPTIMIZED; |
| 479 | } |
| 480 | |
| 481 | static void nvme_update_ns_ana_state(struct nvme_ana_group_desc *desc, |
| 482 | struct nvme_ns *ns) |
| 483 | { |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 484 | mutex_lock(&ns->head->lock); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 485 | ns->ana_grpid = le32_to_cpu(desc->grpid); |
| 486 | ns->ana_state = desc->state; |
| 487 | clear_bit(NVME_NS_ANA_PENDING, &ns->flags); |
| 488 | |
Martin George | cc2278c | 2019-03-27 09:52:56 +0100 | [diff] [blame] | 489 | if (nvme_state_is_live(ns->ana_state)) |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 490 | nvme_mpath_set_live(ns); |
| 491 | mutex_unlock(&ns->head->lock); |
| 492 | } |
| 493 | |
| 494 | static int nvme_update_ana_state(struct nvme_ctrl *ctrl, |
| 495 | struct nvme_ana_group_desc *desc, void *data) |
| 496 | { |
| 497 | u32 nr_nsids = le32_to_cpu(desc->nnsids), n = 0; |
| 498 | unsigned *nr_change_groups = data; |
| 499 | struct nvme_ns *ns; |
| 500 | |
Hannes Reinecke | 592b6e7 | 2019-04-28 20:24:42 -0700 | [diff] [blame] | 501 | dev_dbg(ctrl->device, "ANA group %d: %s.\n", |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 502 | le32_to_cpu(desc->grpid), |
| 503 | nvme_ana_state_names[desc->state]); |
| 504 | |
| 505 | if (desc->state == NVME_ANA_CHANGE) |
| 506 | (*nr_change_groups)++; |
| 507 | |
| 508 | if (!nr_nsids) |
| 509 | return 0; |
| 510 | |
| 511 | down_write(&ctrl->namespaces_rwsem); |
| 512 | list_for_each_entry(ns, &ctrl->namespaces, list) { |
Anton Eidelman | e01f91d | 2019-08-16 13:00:10 -0700 | [diff] [blame] | 513 | unsigned nsid = le32_to_cpu(desc->nsids[n]); |
| 514 | |
| 515 | if (ns->head->ns_id < nsid) |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 516 | continue; |
Anton Eidelman | e01f91d | 2019-08-16 13:00:10 -0700 | [diff] [blame] | 517 | if (ns->head->ns_id == nsid) |
| 518 | nvme_update_ns_ana_state(desc, ns); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 519 | if (++n == nr_nsids) |
| 520 | break; |
| 521 | } |
| 522 | up_write(&ctrl->namespaces_rwsem); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 523 | return 0; |
| 524 | } |
| 525 | |
| 526 | static int nvme_read_ana_log(struct nvme_ctrl *ctrl, bool groups_only) |
| 527 | { |
| 528 | u32 nr_change_groups = 0; |
| 529 | int error; |
| 530 | |
| 531 | mutex_lock(&ctrl->ana_lock); |
| 532 | error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_ANA, |
| 533 | groups_only ? NVME_ANA_LOG_RGO : 0, |
| 534 | ctrl->ana_log_buf, ctrl->ana_log_size, 0); |
| 535 | if (error) { |
| 536 | dev_warn(ctrl->device, "Failed to get ANA log: %d\n", error); |
| 537 | goto out_unlock; |
| 538 | } |
| 539 | |
| 540 | error = nvme_parse_ana_log(ctrl, &nr_change_groups, |
| 541 | nvme_update_ana_state); |
| 542 | if (error) |
| 543 | goto out_unlock; |
| 544 | |
| 545 | /* |
| 546 | * In theory we should have an ANATT timer per group as they might enter |
| 547 | * the change state at different times. But that is a lot of overhead |
| 548 | * just to protect against a target that keeps entering new changes |
| 549 | * states while never finishing previous ones. But we'll still |
| 550 | * eventually time out once all groups are in change state, so this |
| 551 | * isn't a big deal. |
| 552 | * |
| 553 | * We also double the ANATT value to provide some slack for transports |
| 554 | * or AEN processing overhead. |
| 555 | */ |
| 556 | if (nr_change_groups) |
| 557 | mod_timer(&ctrl->anatt_timer, ctrl->anatt * HZ * 2 + jiffies); |
| 558 | else |
| 559 | del_timer_sync(&ctrl->anatt_timer); |
| 560 | out_unlock: |
| 561 | mutex_unlock(&ctrl->ana_lock); |
| 562 | return error; |
| 563 | } |
| 564 | |
| 565 | static void nvme_ana_work(struct work_struct *work) |
| 566 | { |
| 567 | struct nvme_ctrl *ctrl = container_of(work, struct nvme_ctrl, ana_work); |
| 568 | |
| 569 | nvme_read_ana_log(ctrl, false); |
| 570 | } |
| 571 | |
| 572 | static void nvme_anatt_timeout(struct timer_list *t) |
| 573 | { |
| 574 | struct nvme_ctrl *ctrl = from_timer(ctrl, t, anatt_timer); |
| 575 | |
| 576 | dev_info(ctrl->device, "ANATT timeout, resetting controller.\n"); |
| 577 | nvme_reset_ctrl(ctrl); |
| 578 | } |
| 579 | |
| 580 | void nvme_mpath_stop(struct nvme_ctrl *ctrl) |
| 581 | { |
| 582 | if (!nvme_ctrl_use_ana(ctrl)) |
| 583 | return; |
| 584 | del_timer_sync(&ctrl->anatt_timer); |
| 585 | cancel_work_sync(&ctrl->ana_work); |
| 586 | } |
| 587 | |
Hannes Reinecke | 75c10e7 | 2019-02-18 11:43:26 +0100 | [diff] [blame] | 588 | #define SUBSYS_ATTR_RW(_name, _mode, _show, _store) \ |
| 589 | struct device_attribute subsys_attr_##_name = \ |
| 590 | __ATTR(_name, _mode, _show, _store) |
| 591 | |
| 592 | static const char *nvme_iopolicy_names[] = { |
| 593 | [NVME_IOPOLICY_NUMA] = "numa", |
| 594 | [NVME_IOPOLICY_RR] = "round-robin", |
| 595 | }; |
| 596 | |
| 597 | static ssize_t nvme_subsys_iopolicy_show(struct device *dev, |
| 598 | struct device_attribute *attr, char *buf) |
| 599 | { |
| 600 | struct nvme_subsystem *subsys = |
| 601 | container_of(dev, struct nvme_subsystem, dev); |
| 602 | |
| 603 | return sprintf(buf, "%s\n", |
| 604 | nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]); |
| 605 | } |
| 606 | |
| 607 | static ssize_t nvme_subsys_iopolicy_store(struct device *dev, |
| 608 | struct device_attribute *attr, const char *buf, size_t count) |
| 609 | { |
| 610 | struct nvme_subsystem *subsys = |
| 611 | container_of(dev, struct nvme_subsystem, dev); |
| 612 | int i; |
| 613 | |
| 614 | for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) { |
| 615 | if (sysfs_streq(buf, nvme_iopolicy_names[i])) { |
| 616 | WRITE_ONCE(subsys->iopolicy, i); |
| 617 | return count; |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | return -EINVAL; |
| 622 | } |
| 623 | SUBSYS_ATTR_RW(iopolicy, S_IRUGO | S_IWUSR, |
| 624 | nvme_subsys_iopolicy_show, nvme_subsys_iopolicy_store); |
| 625 | |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 626 | static ssize_t ana_grpid_show(struct device *dev, struct device_attribute *attr, |
| 627 | char *buf) |
| 628 | { |
| 629 | return sprintf(buf, "%d\n", nvme_get_ns_from_dev(dev)->ana_grpid); |
| 630 | } |
| 631 | DEVICE_ATTR_RO(ana_grpid); |
| 632 | |
| 633 | static ssize_t ana_state_show(struct device *dev, struct device_attribute *attr, |
| 634 | char *buf) |
| 635 | { |
| 636 | struct nvme_ns *ns = nvme_get_ns_from_dev(dev); |
| 637 | |
| 638 | return sprintf(buf, "%s\n", nvme_ana_state_names[ns->ana_state]); |
| 639 | } |
| 640 | DEVICE_ATTR_RO(ana_state); |
| 641 | |
| 642 | static int nvme_set_ns_ana_state(struct nvme_ctrl *ctrl, |
| 643 | struct nvme_ana_group_desc *desc, void *data) |
| 644 | { |
| 645 | struct nvme_ns *ns = data; |
| 646 | |
| 647 | if (ns->ana_grpid == le32_to_cpu(desc->grpid)) { |
| 648 | nvme_update_ns_ana_state(desc, ns); |
| 649 | return -ENXIO; /* just break out of the loop */ |
| 650 | } |
| 651 | |
| 652 | return 0; |
| 653 | } |
| 654 | |
| 655 | void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id) |
| 656 | { |
| 657 | if (nvme_ctrl_use_ana(ns->ctrl)) { |
| 658 | mutex_lock(&ns->ctrl->ana_lock); |
| 659 | ns->ana_grpid = le32_to_cpu(id->anagrpid); |
| 660 | nvme_parse_ana_log(ns->ctrl, ns, nvme_set_ns_ana_state); |
| 661 | mutex_unlock(&ns->ctrl->ana_lock); |
| 662 | } else { |
| 663 | mutex_lock(&ns->head->lock); |
| 664 | ns->ana_state = NVME_ANA_OPTIMIZED; |
| 665 | nvme_mpath_set_live(ns); |
| 666 | mutex_unlock(&ns->head->lock); |
| 667 | } |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 668 | } |
| 669 | |
| 670 | void nvme_mpath_remove_disk(struct nvme_ns_head *head) |
| 671 | { |
| 672 | if (!head->disk) |
| 673 | return; |
Hannes Reinecke | 33b14f67 | 2018-09-28 08:17:20 +0200 | [diff] [blame] | 674 | if (head->disk->flags & GENHD_FL_UP) |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 675 | del_gendisk(head->disk); |
Christoph Hellwig | 32acab3 | 2017-11-02 12:59:30 +0100 | [diff] [blame] | 676 | blk_set_queue_dying(head->disk->queue); |
| 677 | /* make sure all pending bios are cleaned up */ |
| 678 | kblockd_schedule_work(&head->requeue_work); |
| 679 | flush_work(&head->requeue_work); |
| 680 | blk_cleanup_queue(head->disk->queue); |
| 681 | put_disk(head->disk); |
| 682 | } |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 683 | |
| 684 | int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id) |
| 685 | { |
| 686 | int error; |
| 687 | |
Marta Rybczynska | 66b20ac | 2019-07-23 07:41:20 +0200 | [diff] [blame] | 688 | /* check if multipath is enabled and we have the capability */ |
| 689 | if (!multipath || !ctrl->subsys || !(ctrl->subsys->cmic & (1 << 3))) |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 690 | return 0; |
| 691 | |
| 692 | ctrl->anacap = id->anacap; |
| 693 | ctrl->anatt = id->anatt; |
| 694 | ctrl->nanagrpid = le32_to_cpu(id->nanagrpid); |
| 695 | ctrl->anagrpmax = le32_to_cpu(id->anagrpmax); |
| 696 | |
| 697 | mutex_init(&ctrl->ana_lock); |
| 698 | timer_setup(&ctrl->anatt_timer, nvme_anatt_timeout, 0); |
| 699 | ctrl->ana_log_size = sizeof(struct nvme_ana_rsp_hdr) + |
| 700 | ctrl->nanagrpid * sizeof(struct nvme_ana_group_desc); |
Hannes Reinecke | 78a61cd | 2019-01-09 09:45:15 +0100 | [diff] [blame] | 701 | ctrl->ana_log_size += ctrl->max_namespaces * sizeof(__le32); |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 702 | |
| 703 | if (ctrl->ana_log_size > ctrl->max_hw_sectors << SECTOR_SHIFT) { |
| 704 | dev_err(ctrl->device, |
| 705 | "ANA log page size (%zd) larger than MDTS (%d).\n", |
| 706 | ctrl->ana_log_size, |
| 707 | ctrl->max_hw_sectors << SECTOR_SHIFT); |
| 708 | dev_err(ctrl->device, "disabling ANA support.\n"); |
| 709 | return 0; |
| 710 | } |
| 711 | |
| 712 | INIT_WORK(&ctrl->ana_work, nvme_ana_work); |
| 713 | ctrl->ana_log_buf = kmalloc(ctrl->ana_log_size, GFP_KERNEL); |
Susobhan Dey | bb830ad | 2018-09-25 12:29:15 -0700 | [diff] [blame] | 714 | if (!ctrl->ana_log_buf) { |
| 715 | error = -ENOMEM; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 716 | goto out; |
Susobhan Dey | bb830ad | 2018-09-25 12:29:15 -0700 | [diff] [blame] | 717 | } |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 718 | |
| 719 | error = nvme_read_ana_log(ctrl, true); |
| 720 | if (error) |
| 721 | goto out_free_ana_log_buf; |
| 722 | return 0; |
| 723 | out_free_ana_log_buf: |
| 724 | kfree(ctrl->ana_log_buf); |
Hannes Reinecke | c7055fd | 2019-01-08 12:46:58 +0100 | [diff] [blame] | 725 | ctrl->ana_log_buf = NULL; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 726 | out: |
Susobhan Dey | bb830ad | 2018-09-25 12:29:15 -0700 | [diff] [blame] | 727 | return error; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 728 | } |
| 729 | |
| 730 | void nvme_mpath_uninit(struct nvme_ctrl *ctrl) |
| 731 | { |
| 732 | kfree(ctrl->ana_log_buf); |
Hannes Reinecke | c7055fd | 2019-01-08 12:46:58 +0100 | [diff] [blame] | 733 | ctrl->ana_log_buf = NULL; |
Christoph Hellwig | 0d0b660 | 2018-05-14 08:48:54 +0200 | [diff] [blame] | 734 | } |
| 735 | |