1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2016 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2021 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #ifndef __NVMF_INTERNAL_H__ 8 #define __NVMF_INTERNAL_H__ 9 10 #include "spdk/stdinc.h" 11 12 #include "spdk/likely.h" 13 #include "spdk/nvmf.h" 14 #include "spdk/nvmf_cmd.h" 15 #include "spdk/nvmf_transport.h" 16 #include "spdk/nvmf_spec.h" 17 #include "spdk/assert.h" 18 #include "spdk/bdev.h" 19 #include "spdk/queue.h" 20 #include "spdk/util.h" 21 #include "spdk/thread.h" 22 #include "spdk/tree.h" 23 24 /* The spec reserves cntlid values in the range FFF0h to FFFFh. */ 25 #define NVMF_MIN_CNTLID 1 26 #define NVMF_MAX_CNTLID 0xFFEF 27 28 enum spdk_nvmf_tgt_state { 29 NVMF_TGT_IDLE = 0, 30 NVMF_TGT_RUNNING, 31 NVMF_TGT_PAUSING, 32 NVMF_TGT_PAUSED, 33 NVMF_TGT_RESUMING, 34 }; 35 36 enum spdk_nvmf_subsystem_state { 37 SPDK_NVMF_SUBSYSTEM_INACTIVE = 0, 38 SPDK_NVMF_SUBSYSTEM_ACTIVATING, 39 SPDK_NVMF_SUBSYSTEM_ACTIVE, 40 SPDK_NVMF_SUBSYSTEM_PAUSING, 41 SPDK_NVMF_SUBSYSTEM_PAUSED, 42 SPDK_NVMF_SUBSYSTEM_RESUMING, 43 SPDK_NVMF_SUBSYSTEM_DEACTIVATING, 44 SPDK_NVMF_SUBSYSTEM_NUM_STATES, 45 }; 46 47 RB_HEAD(subsystem_tree, spdk_nvmf_subsystem); 48 49 struct spdk_nvmf_tgt { 50 char name[NVMF_TGT_NAME_MAX_LENGTH]; 51 52 pthread_mutex_t mutex; 53 54 uint64_t discovery_genctr; 55 56 uint32_t max_subsystems; 57 58 enum spdk_nvmf_tgt_discovery_filter discovery_filter; 59 60 enum spdk_nvmf_tgt_state state; 61 62 struct spdk_bit_array *subsystem_ids; 63 64 struct subsystem_tree subsystems; 65 66 TAILQ_HEAD(, spdk_nvmf_transport) transports; 67 TAILQ_HEAD(, spdk_nvmf_poll_group) poll_groups; 68 69 /* Used for round-robin assignment of connections to poll groups */ 70 struct spdk_nvmf_poll_group *next_poll_group; 71 72 spdk_nvmf_tgt_destroy_done_fn *destroy_cb_fn; 73 void *destroy_cb_arg; 74 75 uint16_t crdt[3]; 76 uint16_t num_poll_groups; 77 78 TAILQ_ENTRY(spdk_nvmf_tgt) link; 79 }; 80 81 struct spdk_nvmf_host { 82 char nqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 83 TAILQ_ENTRY(spdk_nvmf_host) link; 84 }; 85 86 struct spdk_nvmf_subsystem_listener { 87 struct spdk_nvmf_subsystem *subsystem; 88 spdk_nvmf_tgt_subsystem_listen_done_fn cb_fn; 89 void *cb_arg; 90 struct spdk_nvme_transport_id *trid; 91 struct spdk_nvmf_transport *transport; 92 enum spdk_nvme_ana_state *ana_state; 93 uint64_t ana_state_change_count; 94 uint16_t id; 95 struct spdk_nvmf_listener_opts opts; 96 TAILQ_ENTRY(spdk_nvmf_subsystem_listener) link; 97 }; 98 99 /* Maximum number of registrants supported per namespace */ 100 #define SPDK_NVMF_MAX_NUM_REGISTRANTS 16 101 102 struct spdk_nvmf_registrant_info { 103 uint64_t rkey; 104 char host_uuid[SPDK_UUID_STRING_LEN]; 105 }; 106 107 struct spdk_nvmf_reservation_info { 108 bool ptpl_activated; 109 enum spdk_nvme_reservation_type rtype; 110 uint64_t crkey; 111 char bdev_uuid[SPDK_UUID_STRING_LEN]; 112 char holder_uuid[SPDK_UUID_STRING_LEN]; 113 uint32_t num_regs; 114 struct spdk_nvmf_registrant_info registrants[SPDK_NVMF_MAX_NUM_REGISTRANTS]; 115 }; 116 117 struct spdk_nvmf_subsystem_pg_ns_info { 118 struct spdk_io_channel *channel; 119 struct spdk_uuid uuid; 120 /* current reservation key, no reservation if the value is 0 */ 121 uint64_t crkey; 122 /* reservation type */ 123 enum spdk_nvme_reservation_type rtype; 124 /* Host ID which holds the reservation */ 125 struct spdk_uuid holder_id; 126 /* Host ID for the registrants with the namespace */ 127 struct spdk_uuid reg_hostid[SPDK_NVMF_MAX_NUM_REGISTRANTS]; 128 uint64_t num_blocks; 129 130 /* I/O outstanding to this namespace */ 131 uint64_t io_outstanding; 132 enum spdk_nvmf_subsystem_state state; 133 }; 134 135 typedef void(*spdk_nvmf_poll_group_mod_done)(void *cb_arg, int status); 136 137 struct spdk_nvmf_subsystem_poll_group { 138 /* Array of namespace information for each namespace indexed by nsid - 1 */ 139 struct spdk_nvmf_subsystem_pg_ns_info *ns_info; 140 uint32_t num_ns; 141 enum spdk_nvmf_subsystem_state state; 142 143 /* Number of ADMIN and FABRICS requests outstanding */ 144 uint64_t mgmt_io_outstanding; 145 spdk_nvmf_poll_group_mod_done cb_fn; 146 void *cb_arg; 147 148 TAILQ_HEAD(, spdk_nvmf_request) queued; 149 }; 150 151 struct spdk_nvmf_registrant { 152 TAILQ_ENTRY(spdk_nvmf_registrant) link; 153 struct spdk_uuid hostid; 154 /* Registration key */ 155 uint64_t rkey; 156 }; 157 158 struct spdk_nvmf_ns { 159 uint32_t nsid; 160 uint32_t anagrpid; 161 struct spdk_nvmf_subsystem *subsystem; 162 struct spdk_bdev *bdev; 163 struct spdk_bdev_desc *desc; 164 struct spdk_nvmf_ns_opts opts; 165 /* reservation notification mask */ 166 uint32_t mask; 167 /* generation code */ 168 uint32_t gen; 169 /* registrants head */ 170 TAILQ_HEAD(, spdk_nvmf_registrant) registrants; 171 /* current reservation key */ 172 uint64_t crkey; 173 /* reservation type */ 174 enum spdk_nvme_reservation_type rtype; 175 /* current reservation holder, only valid if reservation type can only have one holder */ 176 struct spdk_nvmf_registrant *holder; 177 /* Persist Through Power Loss file which contains the persistent reservation */ 178 char *ptpl_file; 179 /* Persist Through Power Loss feature is enabled */ 180 bool ptpl_activated; 181 /* ZCOPY supported on bdev device */ 182 bool zcopy; 183 /* Command Set Identifier */ 184 enum spdk_nvme_csi csi; 185 }; 186 187 /* 188 * NVMf reservation notification log page. 189 */ 190 struct spdk_nvmf_reservation_log { 191 struct spdk_nvme_reservation_notification_log log; 192 TAILQ_ENTRY(spdk_nvmf_reservation_log) link; 193 struct spdk_nvmf_ctrlr *ctrlr; 194 }; 195 196 /* 197 * NVMf async event completion. 198 */ 199 struct spdk_nvmf_async_event_completion { 200 union spdk_nvme_async_event_completion event; 201 STAILQ_ENTRY(spdk_nvmf_async_event_completion) link; 202 }; 203 204 /* 205 * This structure represents an NVMe-oF controller, 206 * which is like a "session" in networking terms. 207 */ 208 struct spdk_nvmf_ctrlr { 209 uint16_t cntlid; 210 char hostnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 211 struct spdk_nvmf_subsystem *subsys; 212 213 struct spdk_nvmf_ctrlr_data cdata; 214 215 struct spdk_nvmf_registers vcprop; 216 217 struct spdk_nvmf_ctrlr_feat feat; 218 219 struct spdk_nvmf_qpair *admin_qpair; 220 struct spdk_thread *thread; 221 struct spdk_bit_array *qpair_mask; 222 223 const struct spdk_nvmf_subsystem_listener *listener; 224 225 struct spdk_nvmf_request *aer_req[SPDK_NVMF_MAX_ASYNC_EVENTS]; 226 STAILQ_HEAD(, spdk_nvmf_async_event_completion) async_events; 227 uint64_t notice_aen_mask; 228 uint8_t nr_aer_reqs; 229 struct spdk_uuid hostid; 230 231 uint32_t association_timeout; /* in milliseconds */ 232 uint16_t changed_ns_list_count; 233 struct spdk_nvme_ns_list changed_ns_list; 234 uint64_t log_page_count; 235 uint8_t num_avail_log_pages; 236 TAILQ_HEAD(log_page_head, spdk_nvmf_reservation_log) log_head; 237 238 /* Time to trigger keep-alive--poller_time = now_tick + period */ 239 uint64_t last_keep_alive_tick; 240 struct spdk_poller *keep_alive_poller; 241 242 struct spdk_poller *association_timer; 243 244 struct spdk_poller *cc_timer; 245 uint64_t cc_timeout_tsc; 246 struct spdk_poller *cc_timeout_timer; 247 248 bool dif_insert_or_strip; 249 bool in_destruct; 250 bool disconnect_in_progress; 251 /* valid only when disconnect_in_progress is true */ 252 bool disconnect_is_shn; 253 bool acre_enabled; 254 bool dynamic_ctrlr; 255 256 TAILQ_ENTRY(spdk_nvmf_ctrlr) link; 257 }; 258 259 #define NVMF_MAX_LISTENERS_PER_SUBSYSTEM 16 260 261 struct spdk_nvmf_subsystem { 262 struct spdk_thread *thread; 263 264 uint32_t id; 265 266 enum spdk_nvmf_subsystem_state state; 267 enum spdk_nvmf_subtype subtype; 268 269 uint16_t next_cntlid; 270 struct { 271 uint8_t allow_any_host : 1; 272 uint8_t allow_any_listener : 1; 273 uint8_t ana_reporting : 1; 274 uint8_t reserved : 5; 275 } flags; 276 277 /* boolean for state change synchronization */ 278 bool changing_state; 279 280 bool destroying; 281 bool async_destroy; 282 283 /* Zoned storage related fields */ 284 bool zone_append_supported; 285 uint64_t max_zone_append_size_kib; 286 287 struct spdk_nvmf_tgt *tgt; 288 RB_ENTRY(spdk_nvmf_subsystem) link; 289 290 /* Array of pointers to namespaces of size max_nsid indexed by nsid - 1 */ 291 struct spdk_nvmf_ns **ns; 292 uint32_t max_nsid; 293 294 uint16_t min_cntlid; 295 uint16_t max_cntlid; 296 297 TAILQ_HEAD(, spdk_nvmf_ctrlr) ctrlrs; 298 299 /* A mutex used to protect the hosts list and allow_any_host flag. Unlike the namespace 300 * array, this list is not used on the I/O path (it's needed for handling things like 301 * the CONNECT command), so use a mutex to protect it instead of requiring the subsystem 302 * state to be paused. This removes the requirement to pause the subsystem when hosts 303 * are added or removed dynamically. */ 304 pthread_mutex_t mutex; 305 TAILQ_HEAD(, spdk_nvmf_host) hosts; 306 TAILQ_HEAD(, spdk_nvmf_subsystem_listener) listeners; 307 struct spdk_bit_array *used_listener_ids; 308 309 TAILQ_ENTRY(spdk_nvmf_subsystem) entries; 310 311 nvmf_subsystem_destroy_cb async_destroy_cb; 312 void *async_destroy_cb_arg; 313 314 char sn[SPDK_NVME_CTRLR_SN_LEN + 1]; 315 char mn[SPDK_NVME_CTRLR_MN_LEN + 1]; 316 char subnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 317 318 /* Array of namespace count per ANA group of size max_nsid indexed anagrpid - 1 319 * It will be enough for ANA group to use the same size as namespaces. 320 */ 321 uint32_t *ana_group; 322 }; 323 324 static int 325 subsystem_cmp(struct spdk_nvmf_subsystem *subsystem1, struct spdk_nvmf_subsystem *subsystem2) 326 { 327 return strncmp(subsystem1->subnqn, subsystem2->subnqn, sizeof(subsystem1->subnqn)); 328 } 329 330 RB_GENERATE_STATIC(subsystem_tree, spdk_nvmf_subsystem, link, subsystem_cmp); 331 332 int nvmf_poll_group_update_subsystem(struct spdk_nvmf_poll_group *group, 333 struct spdk_nvmf_subsystem *subsystem); 334 int nvmf_poll_group_add_subsystem(struct spdk_nvmf_poll_group *group, 335 struct spdk_nvmf_subsystem *subsystem, 336 spdk_nvmf_poll_group_mod_done cb_fn, void *cb_arg); 337 void nvmf_poll_group_remove_subsystem(struct spdk_nvmf_poll_group *group, 338 struct spdk_nvmf_subsystem *subsystem, spdk_nvmf_poll_group_mod_done cb_fn, void *cb_arg); 339 void nvmf_poll_group_pause_subsystem(struct spdk_nvmf_poll_group *group, 340 struct spdk_nvmf_subsystem *subsystem, 341 uint32_t nsid, 342 spdk_nvmf_poll_group_mod_done cb_fn, void *cb_arg); 343 void nvmf_poll_group_resume_subsystem(struct spdk_nvmf_poll_group *group, 344 struct spdk_nvmf_subsystem *subsystem, spdk_nvmf_poll_group_mod_done cb_fn, void *cb_arg); 345 346 void nvmf_update_discovery_log(struct spdk_nvmf_tgt *tgt, const char *hostnqn); 347 void nvmf_get_discovery_log_page(struct spdk_nvmf_tgt *tgt, const char *hostnqn, struct iovec *iov, 348 uint32_t iovcnt, uint64_t offset, uint32_t length, 349 struct spdk_nvme_transport_id *cmd_source_trid); 350 351 void nvmf_ctrlr_destruct(struct spdk_nvmf_ctrlr *ctrlr); 352 int nvmf_ctrlr_process_admin_cmd(struct spdk_nvmf_request *req); 353 int nvmf_ctrlr_process_io_cmd(struct spdk_nvmf_request *req); 354 bool nvmf_ctrlr_dsm_supported(struct spdk_nvmf_ctrlr *ctrlr); 355 bool nvmf_ctrlr_write_zeroes_supported(struct spdk_nvmf_ctrlr *ctrlr); 356 bool nvmf_ctrlr_copy_supported(struct spdk_nvmf_ctrlr *ctrlr); 357 void nvmf_ctrlr_ns_changed(struct spdk_nvmf_ctrlr *ctrlr, uint32_t nsid); 358 bool nvmf_ctrlr_use_zcopy(struct spdk_nvmf_request *req); 359 360 void nvmf_bdev_ctrlr_identify_ns(struct spdk_nvmf_ns *ns, struct spdk_nvme_ns_data *nsdata, 361 bool dif_insert_or_strip); 362 int nvmf_bdev_ctrlr_read_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 363 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 364 int nvmf_bdev_ctrlr_write_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 365 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 366 int nvmf_bdev_ctrlr_compare_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 367 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 368 int nvmf_bdev_ctrlr_compare_and_write_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 369 struct spdk_io_channel *ch, struct spdk_nvmf_request *cmp_req, struct spdk_nvmf_request *write_req); 370 int nvmf_bdev_ctrlr_write_zeroes_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 371 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 372 int nvmf_bdev_ctrlr_flush_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 373 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 374 int nvmf_bdev_ctrlr_dsm_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 375 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 376 int nvmf_bdev_ctrlr_copy_cmd(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 377 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 378 int nvmf_bdev_ctrlr_nvme_passthru_io(struct spdk_bdev *bdev, struct spdk_bdev_desc *desc, 379 struct spdk_io_channel *ch, struct spdk_nvmf_request *req); 380 bool nvmf_bdev_ctrlr_get_dif_ctx(struct spdk_bdev *bdev, struct spdk_nvme_cmd *cmd, 381 struct spdk_dif_ctx *dif_ctx); 382 bool nvmf_bdev_zcopy_enabled(struct spdk_bdev *bdev); 383 384 int nvmf_subsystem_add_ctrlr(struct spdk_nvmf_subsystem *subsystem, 385 struct spdk_nvmf_ctrlr *ctrlr); 386 void nvmf_subsystem_remove_ctrlr(struct spdk_nvmf_subsystem *subsystem, 387 struct spdk_nvmf_ctrlr *ctrlr); 388 void nvmf_subsystem_remove_all_listeners(struct spdk_nvmf_subsystem *subsystem, 389 bool stop); 390 struct spdk_nvmf_ctrlr *nvmf_subsystem_get_ctrlr(struct spdk_nvmf_subsystem *subsystem, 391 uint16_t cntlid); 392 struct spdk_nvmf_subsystem_listener *nvmf_subsystem_find_listener( 393 struct spdk_nvmf_subsystem *subsystem, 394 const struct spdk_nvme_transport_id *trid); 395 struct spdk_nvmf_listener *nvmf_transport_find_listener( 396 struct spdk_nvmf_transport *transport, 397 const struct spdk_nvme_transport_id *trid); 398 void nvmf_transport_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w, 399 bool named); 400 void nvmf_transport_listen_dump_opts(struct spdk_nvmf_transport *transport, 401 const struct spdk_nvme_transport_id *trid, struct spdk_json_write_ctx *w); 402 void nvmf_subsystem_set_ana_state(struct spdk_nvmf_subsystem *subsystem, 403 const struct spdk_nvme_transport_id *trid, 404 enum spdk_nvme_ana_state ana_state, uint32_t anagrpid, 405 spdk_nvmf_tgt_subsystem_listen_done_fn cb_fn, void *cb_arg); 406 bool nvmf_subsystem_get_ana_reporting(struct spdk_nvmf_subsystem *subsystem); 407 408 /** 409 * Sets the controller ID range for a subsystem. 410 * Valid range is [1, 0xFFEF]. 411 * 412 * May only be performed on subsystems in the INACTIVE state. 413 * 414 * \param subsystem Subsystem to modify. 415 * \param min_cntlid Minimum controller ID. 416 * \param max_cntlid Maximum controller ID. 417 * 418 * \return 0 on success, or negated errno value on failure. 419 */ 420 int nvmf_subsystem_set_cntlid_range(struct spdk_nvmf_subsystem *subsystem, 421 uint16_t min_cntlid, uint16_t max_cntlid); 422 423 int nvmf_ctrlr_async_event_ns_notice(struct spdk_nvmf_ctrlr *ctrlr); 424 int nvmf_ctrlr_async_event_ana_change_notice(struct spdk_nvmf_ctrlr *ctrlr); 425 void nvmf_ctrlr_async_event_discovery_log_change_notice(void *ctx); 426 void nvmf_ctrlr_async_event_reservation_notification(struct spdk_nvmf_ctrlr *ctrlr); 427 428 void nvmf_ns_reservation_request(void *ctx); 429 void nvmf_ctrlr_reservation_notice_log(struct spdk_nvmf_ctrlr *ctrlr, 430 struct spdk_nvmf_ns *ns, 431 enum spdk_nvme_reservation_notification_log_page_type type); 432 433 434 /* 435 * Abort zero-copy requests that already got the buffer (received zcopy_start cb), but haven't 436 * started zcopy_end. These requests are kept on the outstanding queue, but are not waiting for a 437 * completion from the bdev layer, so, when a qpair is being disconnected, we need to kick them to 438 * force their completion. 439 */ 440 void nvmf_qpair_abort_pending_zcopy_reqs(struct spdk_nvmf_qpair *qpair); 441 442 /* 443 * Free aer simply frees the rdma resources for the aer without informing the host. 444 * This function should be called when deleting a qpair when one wants to make sure 445 * the qpair is completely empty before freeing the request. The reason we free the 446 * AER without sending a completion is to prevent the host from sending another AER. 447 */ 448 void nvmf_qpair_free_aer(struct spdk_nvmf_qpair *qpair); 449 450 int nvmf_ctrlr_abort_request(struct spdk_nvmf_request *req); 451 452 void nvmf_ctrlr_set_fatal_status(struct spdk_nvmf_ctrlr *ctrlr); 453 454 static inline struct spdk_nvmf_ns * 455 _nvmf_subsystem_get_ns(struct spdk_nvmf_subsystem *subsystem, uint32_t nsid) 456 { 457 /* NOTE: This implicitly also checks for 0, since 0 - 1 wraps around to UINT32_MAX. */ 458 if (spdk_unlikely(nsid - 1 >= subsystem->max_nsid)) { 459 return NULL; 460 } 461 462 return subsystem->ns[nsid - 1]; 463 } 464 465 static inline bool 466 nvmf_qpair_is_admin_queue(struct spdk_nvmf_qpair *qpair) 467 { 468 return qpair->qid == 0; 469 } 470 471 /** 472 * Initiates a zcopy start operation 473 * 474 * \param bdev The \ref spdk_bdev 475 * \param desc The \ref spdk_bdev_desc 476 * \param ch The \ref spdk_io_channel 477 * \param req The \ref spdk_nvmf_request passed to the bdev for processing 478 * 479 * \return SPDK_NVMF_REQUEST_EXEC_STATUS_COMPLETE if the command was completed immediately or 480 * SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS if the command was submitted and will be 481 * completed asynchronously. Asynchronous completions are notified through 482 * spdk_nvmf_request_complete(). 483 */ 484 int nvmf_bdev_ctrlr_zcopy_start(struct spdk_bdev *bdev, 485 struct spdk_bdev_desc *desc, 486 struct spdk_io_channel *ch, 487 struct spdk_nvmf_request *req); 488 489 /** 490 * Ends a zcopy operation 491 * 492 * \param req The NVMe-oF request 493 * \param commit Flag indicating whether the buffers should be committed 494 */ 495 void nvmf_bdev_ctrlr_zcopy_end(struct spdk_nvmf_request *req, bool commit); 496 497 #endif /* __NVMF_INTERNAL_H__ */ 498