| File: | .build-ci/../discoverd/src/main.c |
| Warning: | line 236, column 13 Although the value stored to 'read_devname' is used in the enclosing expression, the value is never actually read from 'read_devname' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | /* |
| 3 | * This file is part of nvme-cli. |
| 4 | * Copyright (c) 2026 Dell Technologies Inc. or its subsidiaries. |
| 5 | * |
| 6 | * Authors: Martin Belanger <martin.belanger@dell.com> |
| 7 | */ |
| 8 | |
| 9 | #include <dirent.h> |
| 10 | #include <errno(*__errno_location ()).h> |
| 11 | #include <getopt.h> |
| 12 | #include <inttypes.h> |
| 13 | #include <stdio.h> |
| 14 | #include <stdlib.h> |
| 15 | #include <string.h> |
| 16 | #include <ifaddrs.h> |
| 17 | |
| 18 | #include <systemd/sd-bus.h> |
| 19 | #include <systemd/sd-daemon.h> |
| 20 | #include <systemd/sd-device.h> |
| 21 | #include <systemd/sd-event.h> |
| 22 | |
| 23 | #include <ccan/list/list.h> |
| 24 | #include <ccan/str/str.h> |
| 25 | |
| 26 | #include <daemon-util/log.h> |
| 27 | #include <daemon-util/resolve.h> |
| 28 | #include <daemon-util/work.h> |
| 29 | #include <daemon-util/signals.h> |
| 30 | #include <shared/array-util.h> |
| 31 | #include <shared/cleanup-util.h> |
| 32 | #include <shared/string-util.h> |
| 33 | #include <shared/time-util.h> |
| 34 | #include <nvme/config.h> |
| 35 | #include <nvme/exclusion.h> |
| 36 | #include <nvme/fabrics.h> |
| 37 | #include <nvme/lib.h> |
| 38 | #include <nvme/registry.h> |
| 39 | #include <nvme/util.h> |
| 40 | |
| 41 | #include "inventory.h" |
| 42 | #include "config.h" |
| 43 | #include "ctx.h" |
| 44 | #include "dlp.h" |
| 45 | #include "events.h" |
| 46 | #include "fc.h" |
| 47 | #include "mdns.h" |
| 48 | #include "state.h" |
| 49 | #include "tid.h" |
| 50 | #include "units.h" |
| 51 | |
| 52 | // Exponential backoff for failed (re)connect attempts: 1s, 2s, 4s, ... capped |
| 53 | // at 5 min. A dynamically-discovered DC (not NBFT- or config-sourced — a |
| 54 | // referral or FC-kickstart find) additionally gives up after |
| 55 | // dc-giveup-timeout (default 72hours) of unbroken failure and is dropped |
| 56 | // from tracking, since nothing but its own retries vouches for it any more. |
| 57 | // Static and NBFT-sourced DCs represent deliberate admin/firmware intent |
| 58 | // and always retry forever. |
| 59 | #define RETRY_INITIAL_DELAY_SEC1 1 |
| 60 | #define RETRY_MAX_DELAY_SEC300 300 |
| 61 | |
| 62 | // Name lookups of the fabrics configuration that run at the same time. |
| 63 | #define RESOLVE_WORKERS8 8 |
| 64 | #define DLP_WORKERS4 4 |
| 65 | |
| 66 | struct active_ctrl { |
| 67 | struct list_node entry; |
| 68 | char *unit_name; // "nvme-discoverd-<12hex>.service" |
| 69 | char *devname; // "nvmeX"; NULL until confirmed in sysfs |
| 70 | struct libnvmf_tid *tid; |
| 71 | bool_Bool is_dc; |
| 72 | unsigned int attempts; // consecutive failed (re)connect attempts |
| 73 | uint64_t giveup_at_usec; // 0 = no deadline armed (ctrl_arm_giveup) |
| 74 | sd_event_source *retry_timer; // NULL when no retry pending |
| 75 | |
| 76 | /* |
| 77 | * EPCSD this DC's own referral entry carried in its parent's |
| 78 | * Discovery Log Page, if this DC was reached via referral. Fallback |
| 79 | * for when this DC's own self entry (SUBTYPE 03h) turns out to be |
| 80 | * absent from its own Discovery Log Page. Unset for primary DCs |
| 81 | * (statically configured, mDNS-discovered, or NBFT) — they have no |
| 82 | * parent. |
| 83 | */ |
| 84 | bool_Bool parent_epcsd_known; |
| 85 | bool_Bool parent_epcsd; // meaningful only if parent_epcsd_known |
| 86 | |
| 87 | // This DC's resolved "persistent" setting is "force": keep it |
| 88 | // connected regardless of what its own EPCSD bit reports. |
| 89 | bool_Bool force_persistent; |
| 90 | bool_Bool force_persistent_logged; // log_info_once() marker |
| 91 | |
| 92 | sd_event_source *epcsd_poll_timer; // NULL when not EPCSD-parked |
| 93 | |
| 94 | // Desired at the last release_undesired() pass, see there. |
| 95 | bool_Bool desired; |
| 96 | |
| 97 | // A DC whose DLP was fetched since nvme-discoverd started. |
| 98 | bool_Bool dlp_fetched; |
| 99 | |
| 100 | struct dmn_work *dlp_read; // NULL when no DLP read runs |
| 101 | bool_Bool dlp_again; // read the DLP again when this read ends |
| 102 | }; |
| 103 | |
| 104 | /* |
| 105 | * A controller from the file of last known desired controllers that is |
| 106 | * not decided yet. See decide_saved(). |
| 107 | */ |
| 108 | struct saved_ctrl { |
| 109 | struct list_node entry; |
| 110 | char *source; |
| 111 | struct libnvmf_tid *tid; |
| 112 | struct libnvmf_tid *parent; // the DC that listed it, for "dlp" |
| 113 | }; |
| 114 | |
| 115 | static LIST_HEAD(g_saved)struct list_head g_saved = { { &(g_saved).n, &(g_saved ).n } }; |
| 116 | |
| 117 | static LIST_HEAD(g_ctrls)struct list_head g_ctrls = { { &(g_ctrls).n, &(g_ctrls ).n } }; |
| 118 | static struct discoverd_ctx ctx; |
| 119 | |
| 120 | static struct active_ctrl *ctrl_find_by_unit(const char *unit_name) |
| 121 | { |
| 122 | struct active_ctrl *e; |
| 123 | |
| 124 | list_for_each(&g_ctrls, e, entry)for ((e) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "124", (&g_ctrls))->n.next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((void *)(e), (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))) != &((&g_ctrls))->n; (e) = list_node_to_off_ (list_node_from_off_((void *)(e), (((__builtin_offsetof(typeof (*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))) { |
| 125 | if (streq(e->unit_name, unit_name)(strcmp((e->unit_name),(unit_name)) == 0)) |
| 126 | return e; |
| 127 | } |
| 128 | return NULL((void*)0); |
| 129 | } |
| 130 | |
| 131 | static struct active_ctrl *ctrl_find_by_devname(const char *devname) |
| 132 | { |
| 133 | struct active_ctrl *e; |
| 134 | |
| 135 | list_for_each(&g_ctrls, e, entry)for ((e) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "135", (&g_ctrls))->n.next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((void *)(e), (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))) != &((&g_ctrls))->n; (e) = list_node_to_off_ (list_node_from_off_((void *)(e), (((__builtin_offsetof(typeof (*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))) { |
| 136 | if (shr_streq0(e->devname, devname)) |
| 137 | return e; |
| 138 | } |
| 139 | return NULL((void*)0); |
| 140 | } |
| 141 | |
| 142 | static void ctrl_free(struct active_ctrl *e) |
| 143 | { |
| 144 | if (!e) |
| 145 | return; |
| 146 | sd_event_source_unref(e->retry_timer); |
| 147 | sd_event_source_unref(e->epcsd_poll_timer); |
| 148 | dmn_work_cancel(e->dlp_read); |
| 149 | tid_free(e->tid); |
| 150 | free(e->unit_name); |
| 151 | free(e->devname); |
| 152 | free(e); |
| 153 | } |
| 154 | |
| 155 | static void schedule_release(bool_Bool check_exclusions); |
| 156 | |
| 157 | static int ctrl_add(const char *unit_name, const struct libnvmf_tid *t, |
| 158 | bool_Bool is_dc, const struct libnvmf_params *params) |
| 159 | { |
| 160 | struct active_ctrl *e; |
| 161 | |
| 162 | if (ctrl_find_by_unit(unit_name)) |
| 163 | return 0; // already tracked |
| 164 | |
| 165 | e = calloc(1, sizeof(*e)); |
| 166 | if (!e) |
| 167 | return -ENOMEM12; |
| 168 | |
| 169 | e->unit_name = strdup(unit_name); |
| 170 | e->tid = libnvmf_tid_dup(t); |
| 171 | e->is_dc = is_dc; |
| 172 | e->desired = inventory_is_desired(ctx.inventory, t); |
| 173 | e->force_persistent = params && |
| 174 | shr_streqcase0(libnvmf_params_get(params, "persistent"), |
| 175 | "force"); |
| 176 | if (!e->unit_name || !e->tid) { |
| 177 | ctrl_free(e); |
| 178 | return -ENOMEM12; |
| 179 | } |
| 180 | |
| 181 | list_add(&g_ctrls, &e->entry)list_add_(&g_ctrls, &e->entry, "../discoverd/src/main.c" ":" "181"); |
| 182 | schedule_release(false0); // the saved desired set changes |
| 183 | return 0; |
| 184 | } |
| 185 | |
| 186 | static void ctrl_remove(struct active_ctrl *entry) |
| 187 | { |
| 188 | list_del_init(&entry->entry)list_del_init_(&entry->entry, "../discoverd/src/main.c" ":" "188"); |
| 189 | ctrl_free(entry); |
| 190 | schedule_release(false0); // the saved desired set changes |
| 191 | } |
| 192 | |
| 193 | /* |
| 194 | * Release @unit_name's controller, and leave its connection up. The state |
| 195 | * goes first, so that the unit's ExecStop= has nothing to disconnect. Then |
| 196 | * the unit is stopped and the registry owner is cleared. @devname may be |
| 197 | * NULL when the controller is not connected. |
| 198 | */ |
| 199 | static void release_unit(const char *unit_name, const char *devname, |
| 200 | const struct libnvmf_tid *tid, const char *reason) |
| 201 | { |
| 202 | state_remove_devid(unit_name); |
| 203 | if (devname) { |
| 204 | __cleanup_free__attribute__((cleanup(shr_freep))) char *owner_unit = state_read_unit(devname); |
| 205 | |
| 206 | if (shr_streq0(owner_unit, unit_name)) |
| 207 | state_remove_ctrl(devname); |
| 208 | } |
| 209 | |
| 210 | if (unit_exists(ctx.umgr, unit_name)) |
| 211 | unit_stop(ctx.umgr, unit_name); |
| 212 | |
| 213 | /* Only clear an owner that is still ours. */ |
| 214 | if (devname && |
| 215 | !libnvmf_registry_attr_equal(ctx.nvme_ctx, devname, "owner", |
| 216 | "discoverd")) { |
| 217 | int r = libnvmf_registry_update(ctx.nvme_ctx, devname, |
| 218 | "owner", NULL((void*)0)); |
| 219 | |
| 220 | if (r < 0) |
| 221 | log_warn("%s | %s - failed to clear registry owner: %s",dmn_log_msg(DMN_LOG_WARN, "%s | %s - failed to clear registry owner: %s" , libnvmf_tid_str(tid), devname, strerror(-r)) |
| 222 | libnvmf_tid_str(tid), devname, strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s | %s - failed to clear registry owner: %s" , libnvmf_tid_str(tid), devname, strerror(-r)); |
| 223 | } |
| 224 | |
| 225 | log_info("%s | %s - %s, released", libnvmf_tid_str(tid),dmn_log_msg(DMN_LOG_INFO, "%s | %s - %s, released", libnvmf_tid_str (tid), devname ? devname : "-", reason) |
| 226 | devname ? devname : "-", reason)dmn_log_msg(DMN_LOG_INFO, "%s | %s - %s, released", libnvmf_tid_str (tid), devname ? devname : "-", reason); |
| 227 | } |
| 228 | |
| 229 | /* Release a tracked controller, and stop tracking it. */ |
| 230 | static void ctrl_release(struct active_ctrl *e, const char *reason) |
| 231 | { |
| 232 | __cleanup_free__attribute__((cleanup(shr_freep))) char *read_devname = NULL((void*)0); |
| 233 | const char *devname = e->devname; |
| 234 | |
| 235 | if (!devname) |
| 236 | devname = read_devname = unit_read_devid(e->unit_name); |
Although the value stored to 'read_devname' is used in the enclosing expression, the value is never actually read from 'read_devname' | |
| 237 | |
| 238 | release_unit(e->unit_name, devname, e->tid, reason); |
| 239 | ctrl_remove(e); |
| 240 | } |
| 241 | |
| 242 | static bool_Bool devname_matches_tid(const char *devname, |
| 243 | const struct libnvmf_tid *tid); |
| 244 | |
| 245 | static void saved_free(struct saved_ctrl *s) |
| 246 | { |
| 247 | list_del_init(&s->entry)list_del_init_(&s->entry, "../discoverd/src/main.c" ":" "247"); |
| 248 | free(s->source); |
| 249 | tid_free(s->tid); |
| 250 | tid_free(s->parent); |
| 251 | free(s); |
| 252 | } |
| 253 | |
| 254 | /* |
| 255 | * Decide a saved controller that nvme-discoverd did not adopt at startup. |
| 256 | * Returns true when it is decided, and false while it waits for the DLP of |
| 257 | * the DC that listed it. |
| 258 | */ |
| 259 | static bool_Bool decide_saved(struct saved_ctrl *s) |
| 260 | { |
| 261 | __cleanup_free__attribute__((cleanup(shr_freep))) char *unit_name = tid_unit_name(s->tid); |
| 262 | __cleanup_free__attribute__((cleanup(shr_freep))) char *devname = NULL((void*)0); |
| 263 | |
| 264 | if (!unit_name || ctrl_find_by_unit(unit_name) || |
| 265 | inventory_is_desired(ctx.inventory, s->tid)) |
| 266 | return true1; |
| 267 | |
| 268 | if (streq(s->source, "dlp")(strcmp((s->source),("dlp")) == 0) && s->parent && |
| 269 | inventory_is_desired(ctx.inventory, s->parent)) { |
| 270 | __cleanup_free__attribute__((cleanup(shr_freep))) char *dc_unit = tid_unit_name(s->parent); |
| 271 | struct active_ctrl *dc = dc_unit ? |
| 272 | ctrl_find_by_unit(dc_unit) : NULL((void*)0); |
| 273 | |
| 274 | if (!dc || !dc->dlp_fetched) |
| 275 | return false0; |
| 276 | } |
| 277 | |
| 278 | /* The kernel reuses device names: check it is still this one. */ |
| 279 | devname = unit_read_devid(unit_name); |
| 280 | if (devname && !devname_matches_tid(devname, s->tid)) { |
| 281 | free(devname); |
| 282 | devname = NULL((void*)0); |
| 283 | } |
| 284 | |
| 285 | /* No unit and no connection: nothing is left to release. */ |
| 286 | if (!devname && !unit_exists(ctx.umgr, unit_name)) |
| 287 | return true1; |
| 288 | |
| 289 | release_unit(unit_name, devname, s->tid, |
| 290 | "no longer desired since the last run"); |
| 291 | return true1; |
| 292 | } |
| 293 | |
| 294 | /* |
| 295 | * Rewrite the file of last known desired controllers: the tracked |
| 296 | * controllers that are desired, and the saved ones not decided yet. |
| 297 | */ |
| 298 | static void save_desired(void) |
| 299 | { |
| 300 | __cleanup_free__attribute__((cleanup(shr_freep))) char *content = NULL((void*)0); |
| 301 | struct active_ctrl *e; |
| 302 | struct saved_ctrl *s; |
| 303 | size_t size = 0; |
| 304 | FILE *f; |
| 305 | int r; |
| 306 | |
| 307 | f = open_memstream(&content, &size); |
| 308 | if (!f) |
| 309 | return; |
| 310 | |
| 311 | list_for_each(&g_ctrls, e, entry)for ((e) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "311", (&g_ctrls))->n.next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((void *)(e), (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))) != &((&g_ctrls))->n; (e) = list_node_to_off_ (list_node_from_off_((void *)(e), (((__builtin_offsetof(typeof (*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))) { |
| 312 | const struct libnvmf_tid *parent; |
| 313 | const char *source; |
| 314 | |
| 315 | if (!e->desired) |
| 316 | continue; |
| 317 | source = inventory_source(ctx.inventory, e->tid, &parent); |
| 318 | if (!source) |
| 319 | continue; |
| 320 | fprintf(f, "%s\t%s\t%s\n", source, |
| 321 | libnvmf_tid_get_canonical(e->tid), |
| 322 | parent ? libnvmf_tid_get_canonical(parent) : "-"); |
| 323 | } |
| 324 | list_for_each(&g_saved, s, entry)for ((s) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "324", (&g_saved))->n.next, (((__builtin_offsetof (typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((void *)(s), (((__builtin_offsetof (typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0))))) != &((&g_saved))->n; (s) = list_node_to_off_ (list_node_from_off_((void *)(s), (((__builtin_offsetof(typeof (*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0)))))) |
| 325 | fprintf(f, "%s\t%s\t%s\n", s->source, |
| 326 | libnvmf_tid_get_canonical(s->tid), |
| 327 | s->parent ? libnvmf_tid_get_canonical(s->parent) : "-"); |
| 328 | |
| 329 | if (fclose(f) != 0) |
| 330 | return; |
| 331 | |
| 332 | r = state_write_desired(content); |
| 333 | if (r < 0) |
| 334 | log_warn("cannot save the desired controllers: %s",dmn_log_msg(DMN_LOG_WARN, "cannot save the desired controllers: %s" , strerror(-r)) |
| 335 | strerror(-r))dmn_log_msg(DMN_LOG_WARN, "cannot save the desired controllers: %s" , strerror(-r)); |
| 336 | } |
| 337 | |
| 338 | /* |
| 339 | * Read the file of last known desired controllers. Discovered DCs become |
| 340 | * desired again, as they were before the restart. The others wait in |
| 341 | * g_saved for decide_saved(). |
| 342 | */ |
| 343 | static void load_saved(void) |
| 344 | { |
| 345 | __cleanup_free__attribute__((cleanup(shr_freep))) char *content = state_read_desired(); |
| 346 | char *line, *next; |
| 347 | |
| 348 | for (line = content; line && *line; line = next) { |
| 349 | char *source, *canon, *parent; |
| 350 | struct saved_ctrl *s; |
| 351 | |
| 352 | next = strchr(line, '\n')_Generic (0 ? (line) : (void *) 1, const void *: (const char * ) (strchr (line, '\n')), default: strchr (line, '\n')); |
| 353 | if (next) |
| 354 | *next++ = '\0'; |
| 355 | else |
| 356 | next = line + strlen(line); |
| 357 | |
| 358 | source = strtok(line, "\t"); |
| 359 | canon = strtok(NULL((void*)0), "\t"); |
| 360 | parent = strtok(NULL((void*)0), "\t"); |
| 361 | if (!source || !canon || !parent) |
| 362 | continue; |
| 363 | |
| 364 | s = calloc(1, sizeof(*s)); |
| 365 | if (!s) |
| 366 | return; |
| 367 | list_node_init(&s->entry); |
| 368 | s->source = strdup(source); |
| 369 | if (!s->source || |
| 370 | libnvmf_tid_parse(ctx.nvme_ctx, canon, &s->tid) < 0 || |
| 371 | (!streq(parent, "-")(strcmp((parent),("-")) == 0) && |
| 372 | libnvmf_tid_parse(ctx.nvme_ctx, parent, &s->parent) < 0)) { |
| 373 | saved_free(s); |
| 374 | continue; |
| 375 | } |
| 376 | |
| 377 | if (streq(s->source, "discovered")(strcmp((s->source),("discovered")) == 0)) { |
| 378 | inventory_add_discovered_dc(ctx.inventory, s->tid); |
| 379 | saved_free(s); |
| 380 | continue; |
| 381 | } |
| 382 | list_add_tail(&g_saved, &s->entry)list_add_tail_(&g_saved, &s->entry, "../discoverd/src/main.c" ":" "382"); |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | static bool_Bool release_pending; |
| 387 | static bool_Bool release_check_exclusions; |
| 388 | |
| 389 | /* |
| 390 | * Release the controllers that are no longer desired: desired at the last |
| 391 | * pass, and not desired now. A controller that was never desired, such as |
| 392 | * an mDNS DC whose DLP is not fetched yet, is never released here. With |
| 393 | * @release_check_exclusions, also release the controllers that the |
| 394 | * exclusion list now matches. NBFT controllers are never released. |
| 395 | * While a host name of the fabrics configuration is not resolved yet, its |
| 396 | * controller looks not desired, so the pass waits for on_config_lookup(). |
| 397 | */ |
| 398 | static int release_undesired(sd_event_source *src, |
| 399 | void *user_data __attribute__((unused))) |
| 400 | { |
| 401 | struct active_ctrl *e, *next; |
| 402 | struct saved_ctrl *s, *snext; |
| 403 | bool_Bool check_exclusions; |
| 404 | |
| 405 | release_pending = false0; |
| 406 | sd_event_source_disable_unref(src); |
| 407 | |
| 408 | if (inventory_config_pending(ctx.inventory)) |
| 409 | return 0; |
| 410 | |
| 411 | check_exclusions = release_check_exclusions; |
| 412 | release_check_exclusions = false0; |
| 413 | |
| 414 | list_for_each_safe(&g_ctrls, e, next, entry)for ((e) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "414", (&g_ctrls))->n.next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))), (next) = list_node_to_off_(list_node_from_off_((e) , (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry ) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((e), (((__builtin_offsetof(typeof (*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))) != &((&g_ctrls))->n; (e) = (next), (next ) = list_node_to_off_(list_node_from_off_((e), (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))) { |
| 415 | bool_Bool desired = inventory_is_desired(ctx.inventory, e->tid); |
| 416 | |
| 417 | if (inventory_is_nbft(ctx.inventory, e->tid)) |
| 418 | continue; |
| 419 | |
| 420 | if (check_exclusions && |
| 421 | libnvmf_exclusion_match(ctx.nvme_ctx, e->tid)) { |
| 422 | ctrl_release(e, "excluded"); |
| 423 | continue; |
| 424 | } |
| 425 | |
| 426 | if (e->desired && !desired) { |
| 427 | ctrl_release(e, "no longer desired"); |
| 428 | continue; |
| 429 | } |
| 430 | |
| 431 | e->desired = desired; |
| 432 | } |
| 433 | |
| 434 | list_for_each_safe(&g_saved, s, snext, entry)for ((s) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "434", (&g_saved))->n.next, (((__builtin_offsetof (typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0))))), (snext) = list_node_to_off_(list_node_from_off_((s ), (((__builtin_offsetof(typeof(*s), entry) + ((typeof(s-> entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof (typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((s), (((__builtin_offsetof(typeof (*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0))))) != &((&g_saved))->n; (s) = (snext), (snext ) = list_node_to_off_(list_node_from_off_((s), (((__builtin_offsetof (typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*s), entry) + ((typeof(s->entry) *)0 != (struct list_node *)0)))))) { |
| 435 | if (decide_saved(s)) |
| 436 | saved_free(s); |
| 437 | } |
| 438 | |
| 439 | save_desired(); |
| 440 | |
| 441 | return 0; |
| 442 | } |
| 443 | |
| 444 | /* |
| 445 | * Run release_undesired() from the event loop, after the change that |
| 446 | * called this is complete. Several changes before the pass share it. |
| 447 | */ |
| 448 | static void schedule_release(bool_Bool check_exclusions) |
| 449 | { |
| 450 | sd_event_source *src; |
| 451 | int r; |
| 452 | |
| 453 | release_check_exclusions |= check_exclusions; |
| 454 | if (release_pending) |
| 455 | return; |
| 456 | |
| 457 | r = sd_event_add_defer(ctx.event, &src, release_undesired, NULL((void*)0)); |
| 458 | if (r < 0) { |
| 459 | log_warn("cannot schedule the release check: %s",dmn_log_msg(DMN_LOG_WARN, "cannot schedule the release check: %s" , strerror(-r)) |
| 460 | strerror(-r))dmn_log_msg(DMN_LOG_WARN, "cannot schedule the release check: %s" , strerror(-r)); |
| 461 | return; |
| 462 | } |
| 463 | release_pending = true1; |
| 464 | } |
| 465 | |
| 466 | /* |
| 467 | * Take the current desired state as the new reference, without releasing |
| 468 | * anything. For inventory changes that must not release, such as giving up |
| 469 | * on a DC. |
| 470 | */ |
| 471 | static void refresh_desired(void) |
| 472 | { |
| 473 | struct active_ctrl *e; |
| 474 | |
| 475 | list_for_each(&g_ctrls, e, entry)for ((e) = list_node_to_off_(((void)"../discoverd/src/main.c" ":" "475", (&g_ctrls))->n.next, (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))); list_node_from_off_((void *)(e), (((__builtin_offsetof (typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0))))) != &((&g_ctrls))->n; (e) = list_node_to_off_ (list_node_from_off_((void *)(e), (((__builtin_offsetof(typeof (*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))->next, (((__builtin_offsetof(typeof(*e), entry) + ((typeof(e->entry) *)0 != (struct list_node *)0)))))) |
| 476 | e->desired = inventory_is_desired(ctx.inventory, e->tid); |
| 477 | } |
| 478 | |
| 479 | /* One connected controller, as sysfs reports it. */ |
| 480 | struct scanned_ctrl { |
| 481 | char *devname; |
| 482 | struct libnvmf_tid *tid; |
| 483 | bool_Bool is_dc; |
| 484 | bool_Bool live; |
| 485 | }; |
| 486 | |
| 487 | SHR_PTRARRAY_DEFINE(scanned_ctrl_list, struct scanned_ctrl)struct scanned_ctrl_list { struct scanned_ctrl **items; size_t len; size_t cap; }; static inline int scanned_ctrl_list_append (struct scanned_ctrl_list *a, struct scanned_ctrl *item) { do { (void) sizeof(char [1 - 2*!(sizeof(struct scanned_ctrl_list ) == sizeof(struct shr_ptrarray))]); } while(0); return shr_ptrarray_append ( (struct shr_ptrarray *)a, item); } static inline void scanned_ctrl_list_free (struct scanned_ctrl_list *a) { shr_ptrarray_free((struct shr_ptrarray *)a); }; |
| 488 | |
| 489 | static void scanned_ctrl_free(struct scanned_ctrl *sc) |
| 490 | { |
| 491 | if (!sc) |
| 492 | return; |
| 493 | tid_free(sc->tid); |
| 494 | free(sc->devname); |
| 495 | free(sc); |
| 496 | } |
| 497 | |
| 498 | /* |
| 499 | * Snapshot of every connected controller plus the host's interface |
| 500 | * addresses. Matching one candidate needs both, and both are expensive to |
| 501 | * build: a sysfs walk reads six attributes per controller, and getifaddrs() |
| 502 | * dumps the whole address table. A caller about to test many candidates -- |
| 503 | * every entry of a Discovery Log Page, every desired connection at startup |
| 504 | * -- builds one snapshot and passes it down, instead of paying for both per |
| 505 | * candidate. |
| 506 | * |
| 507 | * The snapshot is deliberately not cached across calls. Addresses and |
| 508 | * controllers change under a daemon that runs for weeks, and a stale |
| 509 | * snapshot produces wrong match results rather than merely slow ones. |
| 510 | */ |
| 511 | struct conn_scan { |
| 512 | struct ifaddrs *iface_list; |
| 513 | struct scanned_ctrl_list ctrls; |
| 514 | }; |
| 515 | |
| 516 | static void conn_scan_load(struct conn_scan *scan); |
| 517 | static void conn_scan_free(struct conn_scan *scan); |
| 518 | static const char *find_devname_for_tid(const struct conn_scan *scan, |
| 519 | const struct libnvmf_tid *tid); |
| 520 | |
| 521 | /* Frees a conn_scan on scope exit; a zeroed scan is a safe no-op. */ |
| 522 | #define __cleanup_conn_scan__attribute__((cleanup(conn_scan_free))) __attribute__((cleanup(conn_scan_free))) |
| 523 | |
| 524 | /* |
| 525 | * Exclusion + registry-owner check — called before every connect decision. |
| 526 | * |
| 527 | * Exclusion applies to NBFT-sourced controllers too: the exclusion list is |
| 528 | * the host administrator's explicit, root-only instruction, and is the |
| 529 | * supported way to take a boot path out of service for testing or |
| 530 | * maintenance. owner=nbft still protects a boot-path controller from every |
| 531 | * *other* orchestrator via the registry — it just does not override the |
| 532 | * local admin's own exclusion entry. |
| 533 | * |
| 534 | * The registry check skips a controller another orchestrator (e.g. |
| 535 | * nvme-stas) already owns. "discoverd" and "nbft" are discoverd's own |
| 536 | * registry owner strings (see unit_start_dc()/unit_start_ioc() in units.c), |
| 537 | * so a controller discoverd itself owns is never skipped here. |
| 538 | * |
| 539 | * @scan is a snapshot the caller built before a loop, or NULL to build a |
| 540 | * throwaway one for this call. It is only consulted when @known_devname |
| 541 | * is NULL. |
| 542 | * |
| 543 | * @known_devname is the existing (or just-removed) device name for @tid, |
| 544 | * when the caller already has it in hand (on_nvme_remove()); NULL |
| 545 | * otherwise, in which case this function resolves it itself from @scan |
| 546 | * (find_devname_for_tid()). Either way the registry is checked directly by |
| 547 | * device name, never through libnvme's in-process topology tree: that tree |
| 548 | * is only populated by a caller that has just run libnvme_scan_topology() |
| 549 | * (true for the CLI's one-shot fabrics commands, never true for this |
| 550 | * long-running daemon), so a tree-based match would silently report "no |
| 551 | * owner" for every already-connected controller discoverd didn't itself |
| 552 | * just scan. |
| 553 | */ |
| 554 | static bool_Bool should_connect(const struct conn_scan *scan, |
| 555 | const struct libnvmf_tid *tid, |
| 556 | const char *known_devname) |
| 557 | { |
| 558 | __cleanup_conn_scan__attribute__((cleanup(conn_scan_free))) struct conn_scan local = { 0 }; |
| 559 | __cleanup_free__attribute__((cleanup(shr_freep))) char *owner = NULL((void*)0); |
| 560 | const char *devname = known_devname; |
| 561 | int r; |
| 562 | |
| 563 | if (libnvmf_exclusion_match(ctx.nvme_ctx, tid)) { |
| 564 | log_info("%s - excluded, skipping", libnvmf_tid_str(tid))dmn_log_msg(DMN_LOG_INFO, "%s - excluded, skipping", libnvmf_tid_str (tid)); |
| 565 | return false0; |
| 566 | } |
| 567 | |
| 568 | if (!devname) { |
| 569 | if (!scan) { |
| 570 | conn_scan_load(&local); |
| 571 | scan = &local; |
| 572 | } |
| 573 | devname = find_devname_for_tid(scan, tid); |
| 574 | } |
| 575 | |
| 576 | if (devname) { |
| 577 | r = libnvmf_registry_retrieve(ctx.nvme_ctx, devname, "owner", |
| 578 | &owner); |
| 579 | if (r == -ENOENT2) |
| 580 | r = 0; |
| 581 | } else { |
| 582 | r = 0; // nothing currently connected matching tid |
| 583 | } |
| 584 | |
| 585 | if (r < 0) { |
| 586 | log_warn("%s - failed to check registry owner: %s",dmn_log_msg(DMN_LOG_WARN, "%s - failed to check registry owner: %s" , libnvmf_tid_str(tid), strerror(-r)) |
| 587 | libnvmf_tid_str(tid), strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s - failed to check registry owner: %s" , libnvmf_tid_str(tid), strerror(-r)); |
| 588 | return false0; |
| 589 | } |
| 590 | if (owner && !streq(owner, "discoverd")(strcmp((owner),("discoverd")) == 0) && !streq(owner, "nbft")(strcmp((owner),("nbft")) == 0)) { |
| 591 | log_info("%s - owned by '%s', skipping",dmn_log_msg(DMN_LOG_INFO, "%s - owned by '%s', skipping", libnvmf_tid_str (tid), owner) |
| 592 | libnvmf_tid_str(tid), owner)dmn_log_msg(DMN_LOG_INFO, "%s - owned by '%s', skipping", libnvmf_tid_str (tid), owner); |
| 593 | return false0; |
| 594 | } |
| 595 | |
| 596 | return true1; |
| 597 | } |
| 598 | |
| 599 | /* |
| 600 | * Resolved connect parameters for @t: a statically configured connection's |
| 601 | * own params, or — for anything discoverd found on its own (NBFT, DLP, FC |
| 602 | * kickstart) — the discovered-controller defaults for the scope @via_dc |
| 603 | * was learned through (NULL if @t has no configured parent DC either), |
| 604 | * with the KX-HMAC-CHAP secrets of the section that has @t's NQN. |
| 605 | * See libnvmf_config_get_discovered_params() in <nvme/config.h>. |
| 606 | */ |
| 607 | static const struct libnvmf_params *params_for( |
| 608 | const struct libnvmf_tid *t, bool_Bool is_dc, |
| 609 | const struct libnvmf_config_conn *via_dc) |
| 610 | { |
| 611 | const struct libnvmf_config_conn *conn = |
| 612 | inventory_config_conn_for(ctx.inventory, t); |
| 613 | const struct libnvmf_params *params; |
| 614 | |
| 615 | if (conn) |
| 616 | return libnvmf_config_conn_get_params(conn); |
| 617 | if (!ctx.fabrics_cfg) |
| 618 | return NULL((void*)0); |
| 619 | if (libnvmf_config_get_discovered_params(ctx.fabrics_cfg, |
| 620 | libnvmf_config_conn_get_hostnqn(via_dc), is_dc, |
| 621 | libnvmf_tid_get_subsysnqn(t), ¶ms)) |
| 622 | return NULL((void*)0); |
| 623 | |
| 624 | return params; |
| 625 | } |
| 626 | |
| 627 | static void dlp_fetch(struct active_ctrl *e); |
| 628 | static bool_Bool devname_matches_tid(const char *devname, |
| 629 | const struct libnvmf_tid *tid); |
| 630 | |
| 631 | /* Fetch the DLP of an adopted DC. Runs once, from the event loop. */ |
| 632 | static int adopted_dc_fetch(sd_event_source *src, void *user_data) |
| 633 | { |
| 634 | char *unit_name = user_data; |
| 635 | struct active_ctrl *e = ctrl_find_by_unit(unit_name); |
| 636 | |
| 637 | if (e) |
| 638 | dlp_fetch(e); |
| 639 | |
| 640 | free(unit_name); |
| 641 | sd_event_source_disable_unref(src); |
| 642 | |
| 643 | return 0; |
| 644 | } |
| 645 | |
| 646 | /* |
| 647 | * Track a unit this daemon started before it restarted. The connection is |
| 648 | * already up and @devname names it. Takes ownership of @devname. |
| 649 | */ |
| 650 | static void adopt_ctrl(const char *unit_name, const struct libnvmf_tid *tid, |
| 651 | bool_Bool is_dc, const struct libnvmf_params *params, |
| 652 | char *devname) |
| 653 | { |
| 654 | struct active_ctrl *e; |
| 655 | |
| 656 | ctrl_add(unit_name, tid, is_dc, params); |
| 657 | e = ctrl_find_by_unit(unit_name); |
| 658 | if (!e) { |
| 659 | free(devname); |
| 660 | return; |
| 661 | } |
| 662 | e->devname = devname; |
| 663 | |
| 664 | log_info("%s | %s - adopted, already connected",dmn_log_msg(DMN_LOG_INFO, "%s | %s - adopted, already connected" , libnvmf_tid_str(tid), devname) |
| 665 | libnvmf_tid_str(tid), devname)dmn_log_msg(DMN_LOG_INFO, "%s | %s - adopted, already connected" , libnvmf_tid_str(tid), devname); |
| 666 | |
| 667 | /* |
| 668 | * An adopted DC produces no device-add event, so fetch its DLP here. |
| 669 | * Otherwise its DLP-sourced IOCs never enter the desired set and are |
| 670 | * not reconnected if they drop. Defer the fetch: a referral DC is |
| 671 | * adopted while its parent's DLP is being processed, and the parent's |
| 672 | * entries reach the inventory only when that is done. |
| 673 | */ |
| 674 | if (is_dc) { |
| 675 | sd_event_source *src; |
| 676 | char *name = strdup(unit_name); |
| 677 | int r = -ENOMEM12; |
| 678 | |
| 679 | if (name) |
| 680 | r = sd_event_add_defer(ctx.event, &src, |
| 681 | adopted_dc_fetch, name); |
| 682 | if (r < 0) { |
| 683 | log_warn("%s - cannot defer DLP fetch: %s",dmn_log_msg(DMN_LOG_WARN, "%s - cannot defer DLP fetch: %s", libnvmf_tid_str (tid), strerror(-r)) |
| 684 | libnvmf_tid_str(tid), strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s - cannot defer DLP fetch: %s", libnvmf_tid_str (tid), strerror(-r)); |
| 685 | free(name); |
| 686 | dlp_fetch(e); |
| 687 | } |
| 688 | } |
| 689 | } |
| 690 | |
| 691 | /* |
| 692 | * Start a transient unit for @tid (as a DC or an IOC) and track it, or adopt |
| 693 | * the unit if a previous run of this daemon left it connected. Skipped when |
| 694 | * a unit for this TID is already tracked, so that the same IOC behind two |
| 695 | * DCs does not issue a duplicate StartTransient. The caller is responsible |
| 696 | * for the should_connect() decision. @via_dc is the parent DC's config |
| 697 | * connection, if @tid was learned via that DC's Discovery Log Page; NULL |
| 698 | * otherwise. |
| 699 | */ |
| 700 | static void start_ctrl(const struct libnvmf_tid *tid, bool_Bool is_dc, |
| 701 | const struct libnvmf_config_conn *via_dc) |
| 702 | { |
| 703 | __cleanup_free__attribute__((cleanup(shr_freep))) char *unit_name = tid_unit_name(tid); |
| 704 | const struct libnvmf_params *params; |
| 705 | bool_Bool is_nbft; |
| 706 | int r; |
| 707 | |
| 708 | if (!unit_name) |
| 709 | return; |
| 710 | if (ctrl_find_by_unit(unit_name)) |
| 711 | return; // already tracked — no duplicate connect |
| 712 | |
| 713 | params = params_for(tid, is_dc, via_dc); |
| 714 | |
| 715 | /* |
| 716 | * A transient unit outlives the daemon. Starting it again fails with |
| 717 | * -EEXIST, and the recovery for that stops the unit, which |
| 718 | * disconnects. So adopt it if its recorded device is still this |
| 719 | * connection. Otherwise the unit is stale, and the -EEXIST recovery |
| 720 | * below replaces it. |
| 721 | * |
| 722 | * The kernel reuses nvmeN names, so the recorded device may now be |
| 723 | * another orchestrator's connection. Delete the stale state first: |
| 724 | * ExecStop= disconnects whatever device it records. |
| 725 | */ |
| 726 | if (unit_exists(ctx.umgr, unit_name)) { |
| 727 | char *devname = unit_read_devid(unit_name); |
| 728 | |
| 729 | if (devname && devname_matches_tid(devname, tid)) { |
| 730 | adopt_ctrl(unit_name, tid, is_dc, params, devname); |
| 731 | return; |
| 732 | } |
| 733 | state_remove_devid(unit_name); |
| 734 | if (devname) { |
| 735 | __cleanup_free__attribute__((cleanup(shr_freep))) char *owner_unit = |
| 736 | state_read_unit(devname); |
| 737 | |
| 738 | if (shr_streq0(owner_unit, unit_name)) |
| 739 | state_remove_ctrl(devname); |
| 740 | } |
| 741 | free(devname); |
| 742 | } |
| 743 | |
| 744 | is_nbft = inventory_is_nbft(ctx.inventory, tid); |
| 745 | |
| 746 | r = is_dc ? unit_start_dc(ctx.umgr, tid, params, is_nbft) |
| 747 | : unit_start_ioc(ctx.umgr, tid, params, is_nbft); |
| 748 | if (r >= 0) { |
| 749 | ctrl_add(unit_name, tid, is_dc, params); |
| 750 | log_dbg("%s: requested %s unit", libnvmf_tid_str(tid),dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s: requested %s unit", __func__, libnvmf_tid_str(tid), is_dc ? "DC" : "IOC") |
| 751 | is_dc ? "DC" : "IOC")dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s: requested %s unit", __func__, libnvmf_tid_str(tid), is_dc ? "DC" : "IOC"); |
| 752 | } else { |
| 753 | log_warn("%s - failed to start %s unit: %s",dmn_log_msg(DMN_LOG_WARN, "%s - failed to start %s unit: %s", libnvmf_tid_str(tid), is_dc ? "DC" : "IOC", strerror(-r)) |
| 754 | libnvmf_tid_str(tid), is_dc ? "DC" : "IOC",dmn_log_msg(DMN_LOG_WARN, "%s - failed to start %s unit: %s", libnvmf_tid_str(tid), is_dc ? "DC" : "IOC", strerror(-r)) |
| 755 | strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s - failed to start %s unit: %s", libnvmf_tid_str(tid), is_dc ? "DC" : "IOC", strerror(-r)); |
| 756 | } |
| 757 | } |
| 758 | |
| 759 | struct dlp_fetch_ctx { |
| 760 | const struct libnvmf_config_conn *via_dc; // dc_tid's own conn, if any |
| 761 | const struct conn_scan *scan; // shared by every entry's connect check |
| 762 | struct tid_list iocs; // for the inventory |
| 763 | struct tid_list referrals; // for the inventory |
| 764 | int hops; // the DC's referral hops from its source, -1 if none |
| 765 | bool_Bool self_seen; |
| 766 | bool_Bool epcsd; // meaningful only if self_seen |
| 767 | }; |
| 768 | |
| 769 | static void dlp_ioc_callback(const struct libnvmf_tid *t, void *user_data) |
| 770 | { |
| 771 | struct dlp_fetch_ctx *fctx = user_data; |
| 772 | struct libnvmf_tid *dup; |
| 773 | |
| 774 | dup = libnvmf_tid_dup(t); |
| 775 | if (!dup || tid_list_append(&fctx->iocs, dup) < 0) { |
| 776 | tid_free(dup); |
| 777 | return; |
| 778 | } |
| 779 | |
| 780 | if (should_connect(fctx->scan, t, NULL((void*)0))) |
| 781 | start_ctrl(t, false0, fctx->via_dc); |
| 782 | } |
| 783 | |
| 784 | /* |
| 785 | * Record the EPCSD bit @t's own referral entry carried in its parent's |
| 786 | * Discovery Log Page, as a fallback for when @t is later connected to and |
| 787 | * its own self entry turns out to be absent. A no-op if @t is not tracked |
| 788 | * yet (should_connect() rejected it, or start_ctrl() failed). |
| 789 | */ |
| 790 | static void record_parent_epcsd(const struct libnvmf_tid *t, bool_Bool epcsd) |
| 791 | { |
| 792 | char *unit_name = tid_unit_name(t); |
| 793 | struct active_ctrl *e; |
| 794 | |
| 795 | if (!unit_name) |
| 796 | return; |
| 797 | |
| 798 | e = ctrl_find_by_unit(unit_name); |
| 799 | if (e) { |
| 800 | e->parent_epcsd_known = true1; |
| 801 | e->parent_epcsd = epcsd; |
| 802 | } |
| 803 | free(unit_name); |
| 804 | } |
| 805 | |
| 806 | static void dlp_dc_callback(const struct libnvmf_tid *t, bool_Bool epcsd, |
| 807 | void *user_data) |
| 808 | { |
| 809 | struct dlp_fetch_ctx *fctx = user_data; |
| 810 | struct libnvmf_tid *dup; |
| 811 | |
| 812 | dup = libnvmf_tid_dup(t); |
| 813 | if (!dup || tid_list_append(&fctx->referrals, dup) < 0) |
| 814 | tid_free(dup); |
| 815 | |
| 816 | /* |
| 817 | * Follow a referral only within the hop limit, as the libnvme |
| 818 | * discovery walk does. The referrals of a DC that is no longer |
| 819 | * desired are not followed either. |
| 820 | */ |
| 821 | if (fctx->hops < 0) |
| 822 | return; |
| 823 | if (fctx->hops >= INVENTORY_MAX_REFERRAL_HOPS8) { |
| 824 | log_info("%s - referral %d hops from its source, not followed",dmn_log_msg(DMN_LOG_INFO, "%s - referral %d hops from its source, not followed" , libnvmf_tid_str(t), fctx->hops + 1) |
| 825 | libnvmf_tid_str(t), fctx->hops + 1)dmn_log_msg(DMN_LOG_INFO, "%s - referral %d hops from its source, not followed" , libnvmf_tid_str(t), fctx->hops + 1); |
| 826 | return; |
| 827 | } |
| 828 | |
| 829 | if (should_connect(fctx->scan, t, NULL((void*)0))) { |
| 830 | start_ctrl(t, true1, fctx->via_dc); |
| 831 | record_parent_epcsd(t, epcsd); |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | static void dlp_self_callback(bool_Bool epcsd, void *user_data) |
| 836 | { |
| 837 | struct dlp_fetch_ctx *fctx = user_data; |
| 838 | |
| 839 | fctx->self_seen = true1; |
| 840 | fctx->epcsd = epcsd; |
| 841 | } |
| 842 | |
| 843 | /* |
| 844 | * Self entry seen: its own EPCSD. Else the parent's referral-entry view |
| 845 | * of this DC, if any. Else assume EPCSD=0, matching core libnvme's |
| 846 | * dc_decide(). |
| 847 | */ |
| 848 | static bool_Bool dc_effective_epcsd(const struct dlp_fetch_ctx *fctx, |
| 849 | const struct active_ctrl *e) |
| 850 | { |
| 851 | if (fctx->self_seen) |
| 852 | return fctx->epcsd; |
| 853 | if (e && e->parent_epcsd_known) |
| 854 | return e->parent_epcsd; |
| 855 | return false0; |
| 856 | } |
| 857 | |
| 858 | static void epcsd_park(struct active_ctrl *e); |
| 859 | |
| 860 | // Process the DLP of the DC @e. |
| 861 | static void process_dlp(struct active_ctrl *e, |
| 862 | const struct nvmf_discovery_log *log) |
| 863 | { |
| 864 | __cleanup_conn_scan__attribute__((cleanup(conn_scan_free))) struct conn_scan scan = { 0 }; |
| 865 | const struct libnvmf_tid *dc_tid = e->tid; |
| 866 | struct dlp_fetch_ctx fctx = { |
| 867 | .via_dc = inventory_config_conn_for(ctx.inventory, dc_tid), |
| 868 | .hops = inventory_referral_hops(ctx.inventory, dc_tid), |
| 869 | }; |
| 870 | bool_Bool epcsd; |
| 871 | |
| 872 | // One snapshot for every entry this log page turns out to hold. |
| 873 | conn_scan_load(&scan); |
| 874 | fctx.scan = &scan; |
| 875 | |
| 876 | dlp_process_log(log, dc_tid, dlp_ioc_callback, dlp_dc_callback, |
| 877 | dlp_self_callback, &fctx); |
| 878 | |
| 879 | epcsd = dc_effective_epcsd(&fctx, e); |
| 880 | log_dbg("%s: self entry %s, effective EPCSD=%d",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s: self entry %s, effective EPCSD=%d" , __func__, libnvmf_tid_str(dc_tid), fctx.self_seen ? "seen" : "absent", epcsd) |
| 881 | libnvmf_tid_str(dc_tid), fctx.self_seen ? "seen" : "absent",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s: self entry %s, effective EPCSD=%d" , __func__, libnvmf_tid_str(dc_tid), fctx.self_seen ? "seen" : "absent", epcsd) |
| 882 | epcsd)dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s: self entry %s, effective EPCSD=%d" , __func__, libnvmf_tid_str(dc_tid), fctx.self_seen ? "seen" : "absent", epcsd); |
| 883 | |
| 884 | if (e->is_dc && !epcsd) { |
| 885 | if (e->force_persistent) |
| 886 | log_info_once(&e->force_persistent_logged,dmn_log_msg_once(&e->force_persistent_logged, DMN_LOG_INFO , "%s - EPCSD=0, but persistent=force: staying connected", libnvmf_tid_str (dc_tid)) |
| 887 | "%s - EPCSD=0, but persistent=force: staying connected",dmn_log_msg_once(&e->force_persistent_logged, DMN_LOG_INFO , "%s - EPCSD=0, but persistent=force: staying connected", libnvmf_tid_str (dc_tid)) |
| 888 | libnvmf_tid_str(dc_tid))dmn_log_msg_once(&e->force_persistent_logged, DMN_LOG_INFO , "%s - EPCSD=0, but persistent=force: staying connected", libnvmf_tid_str (dc_tid)); |
| 889 | else |
| 890 | epcsd_park(e); |
| 891 | } |
| 892 | |
| 893 | // A log page without IOC entries replaces the last list. |
| 894 | if (tid_list_append(&fctx.iocs, NULL((void*)0)) == 0 && |
| 895 | tid_list_append(&fctx.referrals, NULL((void*)0)) == 0) { |
| 896 | inventory_update_dlp(ctx.inventory, dc_tid, fctx.iocs.items, |
| 897 | fctx.referrals.items); |
| 898 | e->dlp_fetched = true1; |
| 899 | schedule_release(false0); |
| 900 | } else { |
| 901 | tid_list_free_items(&fctx.iocs); |
| 902 | tid_list_free_items(&fctx.referrals); |
| 903 | } |
| 904 | } |
| 905 | |
| 906 | static int log_level(void) |
| 907 | { |
| 908 | return ctx.force_debug ? DMN_LOG_DEBUG : ctx.cfg->debug_level; |
| 909 | } |
| 910 | |
| 911 | static void on_dlp_read(int ret, const struct nvmf_discovery_log *log, |
| 912 | void *user_data) |
| 913 | { |
| 914 | struct active_ctrl *e = user_data; |
| 915 | |
| 916 | e->dlp_read = NULL((void*)0); |
| 917 | |
| 918 | /* |
| 919 | * A failed read tells nothing about EPCSD or the DC's entries, so |
| 920 | * keep the DC and its last list. |
| 921 | */ |
| 922 | if (ret < 0) |
| 923 | log_warn("%s | %s - cannot read the discovery log: %s",dmn_log_msg(DMN_LOG_WARN, "%s | %s - cannot read the discovery log: %s" , libnvmf_tid_str(e->tid), e->devname ? e->devname : "-", strerror(-ret)) |
| 924 | libnvmf_tid_str(e->tid), e->devname ? e->devname : "-",dmn_log_msg(DMN_LOG_WARN, "%s | %s - cannot read the discovery log: %s" , libnvmf_tid_str(e->tid), e->devname ? e->devname : "-", strerror(-ret)) |
| 925 | strerror(-ret))dmn_log_msg(DMN_LOG_WARN, "%s | %s - cannot read the discovery log: %s" , libnvmf_tid_str(e->tid), e->devname ? e->devname : "-", strerror(-ret)); |
| 926 | else |
| 927 | process_dlp(e, log); |
| 928 | |
| 929 | if (e->dlp_again) { |
| 930 | e->dlp_again = false0; |
| 931 | dlp_fetch(e); |
| 932 | } |
| 933 | } |
| 934 | |
| 935 | /* |
| 936 | * Read the DLP of the DC @e, then process it. The read waits for the DC, |
| 937 | * so it runs in a worker and the event loop goes on. One read at a time |
| 938 | * per DC: a request during a read reads the DLP again when it ends. |
| 939 | */ |
| 940 | static void dlp_fetch(struct active_ctrl *e) |
| 941 | { |
| 942 | int r; |
| 943 | |
| 944 | if (!e->devname) |
| 945 | return; |
| 946 | |
| 947 | if (e->dlp_read) { |
| 948 | e->dlp_again = true1; |
| 949 | return; |
| 950 | } |
| 951 | |
| 952 | r = dlp_read(ctx.dlp_pool, e->devname, log_level(), on_dlp_read, e, |
| 953 | &e->dlp_read); |
| 954 | if (r < 0) |
| 955 | log_warn("%s | %s - cannot read the discovery log: %s",dmn_log_msg(DMN_LOG_WARN, "%s | %s - cannot read the discovery log: %s" , libnvmf_tid_str(e->tid), e->devname, strerror(-r)) |
| 956 | libnvmf_tid_str(e->tid), e->devname, strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s | %s - cannot read the discovery log: %s" , libnvmf_tid_str(e->tid), e->devname, strerror(-r)); |
| 957 | } |
| 958 | |
| 959 | /* |
| 960 | * Arm a dynamically-discovered DC's give-up deadline the first time it is |
| 961 | * seen failing to (re)connect. Static/NBFT-sourced DCs and IOCs never get |
| 962 | * one and retry forever, and so does a dynamic DC when dc-giveup-timeout |
| 963 | * is configured to infinity. |
| 964 | * |
| 965 | * Note the two distinct zeros: giveup_at_usec == 0 means no deadline is |
| 966 | * armed, while a dc-giveup-timeout of 0 means give up on the first |
| 967 | * failure. The latter still arms, because now + 0 is not 0. |
| 968 | */ |
| 969 | static void ctrl_arm_giveup(struct active_ctrl *e) |
| 970 | { |
| 971 | uint64_t now; |
| 972 | |
| 973 | if (e->giveup_at_usec || !e->is_dc) |
| 974 | return; |
| 975 | if (inventory_is_nbft(ctx.inventory, e->tid) || |
| 976 | inventory_config_conn_for(ctx.inventory, e->tid)) |
| 977 | return; |
| 978 | if (ctx.cfg->dc_giveup_timeout_usec == SHR_USEC_INFINITY(18446744073709551615UL)) |
| 979 | return; |
| 980 | |
| 981 | if (sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now) < 0) |
| 982 | return; |
| 983 | |
| 984 | // A span so large the deadline would wrap means forever, not now. |
| 985 | if (ctx.cfg->dc_giveup_timeout_usec >= SHR_USEC_INFINITY(18446744073709551615UL) - now) |
| 986 | return; |
| 987 | |
| 988 | e->giveup_at_usec = now + ctx.cfg->dc_giveup_timeout_usec; |
| 989 | } |
| 990 | |
| 991 | static uint64_t backoff_delay_usec(unsigned int attempts) |
| 992 | { |
| 993 | uint64_t sec = RETRY_INITIAL_DELAY_SEC1; |
| 994 | unsigned int i; |
| 995 | |
| 996 | for (i = 0; i < attempts && sec < RETRY_MAX_DELAY_SEC300; i++) |
| 997 | sec *= 2; |
| 998 | if (sec > RETRY_MAX_DELAY_SEC300) |
| 999 | sec = RETRY_MAX_DELAY_SEC300; |
| 1000 | return sec * UINT64_C(1000000)1000000UL; |
| 1001 | } |
| 1002 | |
| 1003 | static int retry_timeout(sd_event_source *src, uint64_t usec, void *user_data); |
| 1004 | |
| 1005 | static int schedule_retry(struct active_ctrl *e) |
| 1006 | { |
| 1007 | uint64_t now, delay; |
| 1008 | int r; |
| 1009 | |
| 1010 | if (e->retry_timer) |
| 1011 | return 0; // already scheduled |
| 1012 | |
| 1013 | ctrl_arm_giveup(e); |
| 1014 | |
| 1015 | r = sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now); |
| 1016 | if (r < 0) |
| 1017 | return r; |
| 1018 | |
| 1019 | delay = backoff_delay_usec(e->attempts); |
| 1020 | e->attempts++; |
| 1021 | |
| 1022 | return sd_event_add_time(ctx.event, &e->retry_timer, CLOCK_BOOTTIME7, |
| 1023 | now + delay, 0, retry_timeout, e); |
| 1024 | } |
| 1025 | |
| 1026 | /* |
| 1027 | * Restart @e's unit (parameters are baked in). Falls back to a fresh |
| 1028 | * StartTransient — using the top-level discovered-controller scope, not |
| 1029 | * @e's original parent DC's scope, a deliberate simplification since |
| 1030 | * active_ctrl does not track that borrowed config_conn across a possible |
| 1031 | * SIGHUP config reload — if the unit was garbage-collected. |
| 1032 | */ |
| 1033 | static int restart_or_start(struct active_ctrl *e) |
| 1034 | { |
| 1035 | int r = unit_restart(ctx.umgr, e->unit_name); |
| 1036 | |
| 1037 | if (r < 0) { |
| 1038 | const struct libnvmf_params *params = |
| 1039 | params_for(e->tid, e->is_dc, NULL((void*)0)); |
| 1040 | bool_Bool is_nbft = inventory_is_nbft(ctx.inventory, e->tid); |
| 1041 | |
| 1042 | if (e->is_dc) |
| 1043 | r = unit_start_dc(ctx.umgr, e->tid, params, is_nbft); |
| 1044 | else |
| 1045 | r = unit_start_ioc(ctx.umgr, e->tid, params, is_nbft); |
| 1046 | } |
| 1047 | return r; |
| 1048 | } |
| 1049 | |
| 1050 | static int retry_timeout(sd_event_source *src, |
| 1051 | uint64_t usec __attribute__((unused)), |
| 1052 | void *user_data) |
| 1053 | { |
| 1054 | struct active_ctrl *e = user_data; |
| 1055 | uint64_t now; |
| 1056 | int r; |
| 1057 | |
| 1058 | sd_event_source_unref(src); |
| 1059 | e->retry_timer = NULL((void*)0); |
| 1060 | |
| 1061 | if (!inventory_is_desired(ctx.inventory, e->tid)) { |
| 1062 | log_info("%s - no longer desired, not retrying",dmn_log_msg(DMN_LOG_INFO, "%s - no longer desired, not retrying" , libnvmf_tid_str(e->tid)) |
| 1063 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_INFO, "%s - no longer desired, not retrying" , libnvmf_tid_str(e->tid)); |
| 1064 | ctrl_remove(e); |
| 1065 | return 0; |
| 1066 | } |
| 1067 | |
| 1068 | if (e->giveup_at_usec && |
| 1069 | sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now) >= 0 && |
| 1070 | now >= e->giveup_at_usec) { |
| 1071 | log_warn("%s - giving up after repeated failures",dmn_log_msg(DMN_LOG_WARN, "%s - giving up after repeated failures" , libnvmf_tid_str(e->tid)) |
| 1072 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_WARN, "%s - giving up after repeated failures" , libnvmf_tid_str(e->tid)); |
| 1073 | if (e->is_dc) { |
| 1074 | inventory_forget_dc(ctx.inventory, e->tid); |
| 1075 | refresh_desired(); // giving up releases nothing |
| 1076 | } |
| 1077 | ctrl_remove(e); |
| 1078 | return 0; |
| 1079 | } |
| 1080 | |
| 1081 | r = restart_or_start(e); |
| 1082 | if (r < 0) { |
| 1083 | log_err("%s - retry failed: %s",dmn_log_msg(DMN_LOG_ERR, "%s - retry failed: %s", libnvmf_tid_str (e->tid), strerror(-r)) |
| 1084 | libnvmf_tid_str(e->tid), strerror(-r))dmn_log_msg(DMN_LOG_ERR, "%s - retry failed: %s", libnvmf_tid_str (e->tid), strerror(-r)); |
| 1085 | schedule_retry(e); |
| 1086 | } |
| 1087 | return 0; |
| 1088 | } |
| 1089 | |
| 1090 | static int epcsd_poll_timeout(sd_event_source *src, |
| 1091 | uint64_t usec __attribute__((unused)), |
| 1092 | void *user_data) |
| 1093 | { |
| 1094 | struct active_ctrl *e = user_data; |
| 1095 | int r; |
| 1096 | |
| 1097 | sd_event_source_unref(src); |
| 1098 | e->epcsd_poll_timer = NULL((void*)0); |
| 1099 | |
| 1100 | if (!inventory_is_desired(ctx.inventory, e->tid)) { |
| 1101 | log_info("%s - no longer desired, dropping",dmn_log_msg(DMN_LOG_INFO, "%s - no longer desired, dropping", libnvmf_tid_str(e->tid)) |
| 1102 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_INFO, "%s - no longer desired, dropping", libnvmf_tid_str(e->tid)); |
| 1103 | ctrl_remove(e); |
| 1104 | return 0; |
| 1105 | } |
| 1106 | |
| 1107 | log_dbg("%s - EPCSD poll: reconnecting to re-check",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - EPCSD poll: reconnecting to re-check" , __func__, libnvmf_tid_str(e->tid)) |
| 1108 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - EPCSD poll: reconnecting to re-check" , __func__, libnvmf_tid_str(e->tid)); |
| 1109 | r = restart_or_start(e); |
| 1110 | if (r < 0) { |
| 1111 | log_err("%s - EPCSD poll reconnect failed: %s",dmn_log_msg(DMN_LOG_ERR, "%s - EPCSD poll reconnect failed: %s" , libnvmf_tid_str(e->tid), strerror(-r)) |
| 1112 | libnvmf_tid_str(e->tid), strerror(-r))dmn_log_msg(DMN_LOG_ERR, "%s - EPCSD poll reconnect failed: %s" , libnvmf_tid_str(e->tid), strerror(-r)); |
| 1113 | schedule_retry(e); |
| 1114 | } |
| 1115 | return 0; |
| 1116 | } |
| 1117 | |
| 1118 | /* |
| 1119 | * Disconnect a DC whose effective EPCSD is 0 and arm a long poll timer to |
| 1120 | * reconnect and re-check it later. Never retried with the failure |
| 1121 | * backoff/give-up path — EPCSD=0 is an expected, stable outcome, not a |
| 1122 | * connect failure. |
| 1123 | */ |
| 1124 | static void epcsd_park(struct active_ctrl *e) |
| 1125 | { |
| 1126 | uint64_t now, interval; |
| 1127 | |
| 1128 | if (e->epcsd_poll_timer) |
| 1129 | return; // already parked |
| 1130 | |
| 1131 | unit_stop(ctx.umgr, e->unit_name); |
| 1132 | |
| 1133 | if (sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now) < 0) |
| 1134 | return; |
| 1135 | |
| 1136 | interval = (uint64_t)ctx.cfg->epcsd_poll_interval_minutes * |
| 1137 | 60 * UINT64_C(1000000)1000000UL; |
| 1138 | |
| 1139 | if (sd_event_add_time(ctx.event, &e->epcsd_poll_timer, CLOCK_BOOTTIME7, |
| 1140 | now + interval, 0, epcsd_poll_timeout, e) < 0) { |
| 1141 | log_warn("%s - failed to arm EPCSD poll timer",dmn_log_msg(DMN_LOG_WARN, "%s - failed to arm EPCSD poll timer" , libnvmf_tid_str(e->tid)) |
| 1142 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_WARN, "%s - failed to arm EPCSD poll timer" , libnvmf_tid_str(e->tid)); |
| 1143 | return; |
| 1144 | } |
| 1145 | |
| 1146 | log_info("%s - EPCSD=0, disconnecting; re-checking in %u min",dmn_log_msg(DMN_LOG_INFO, "%s - EPCSD=0, disconnecting; re-checking in %u min" , libnvmf_tid_str(e->tid), ctx.cfg->epcsd_poll_interval_minutes ) |
| 1147 | libnvmf_tid_str(e->tid), ctx.cfg->epcsd_poll_interval_minutes)dmn_log_msg(DMN_LOG_INFO, "%s - EPCSD=0, disconnecting; re-checking in %u min" , libnvmf_tid_str(e->tid), ctx.cfg->epcsd_poll_interval_minutes ); |
| 1148 | } |
| 1149 | |
| 1150 | // Periodic FC kickstart (opt-in via fc-kickstart-interval-minutes). |
| 1151 | static int fc_kickstart_timeout(sd_event_source *src, |
| 1152 | uint64_t usec __attribute__((unused)), |
| 1153 | void *user_data __attribute__((unused))) |
| 1154 | { |
| 1155 | uint64_t now; |
| 1156 | |
| 1157 | fc_kickstart(); |
| 1158 | |
| 1159 | if (sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now) >= 0) { |
| 1160 | uint64_t interval = |
| 1161 | (uint64_t)ctx.cfg->fc_kickstart_interval_minutes * |
| 1162 | 60 * UINT64_C(1000000)1000000UL; |
| 1163 | |
| 1164 | sd_event_source_set_time(src, now + interval); |
| 1165 | sd_event_source_set_enabled(src, SD_EVENT_ONESHOT); |
| 1166 | } |
| 1167 | return 0; |
| 1168 | } |
| 1169 | |
| 1170 | /* |
| 1171 | * Link a newly connected DC to its device and fetch its DLP. Both the |
| 1172 | * device's add event and the unit's job completion call this, in either |
| 1173 | * order. Only the first call does the work. |
| 1174 | */ |
| 1175 | static void dc_connected(struct active_ctrl *e, const char *devname) |
| 1176 | { |
| 1177 | if (e->devname) |
| 1178 | return; |
| 1179 | |
| 1180 | e->devname = strdup(devname); |
| 1181 | |
| 1182 | /* |
| 1183 | * DLP entries inherit host-side fields from the DC's TID. Use the |
| 1184 | * candidate TID, not one read from sysfs: sysfs reports the source |
| 1185 | * address the kernel selected, which the configuration did not ask |
| 1186 | * for. |
| 1187 | */ |
| 1188 | dlp_fetch(e); |
| 1189 | } |
| 1190 | |
| 1191 | static void on_job_done(const char *unit_name, bool_Bool success, |
| 1192 | void *user_data __attribute__((unused))) |
| 1193 | { |
| 1194 | struct active_ctrl *e; |
| 1195 | |
| 1196 | e = ctrl_find_by_unit(unit_name); |
| 1197 | if (!e) { |
| 1198 | if (!success) |
| 1199 | log_warn("unit %s failed (untracked)", unit_name)dmn_log_msg(DMN_LOG_WARN, "unit %s failed (untracked)", unit_name ); |
| 1200 | return; |
| 1201 | } |
| 1202 | |
| 1203 | if (success) { |
| 1204 | e->attempts = 0; |
| 1205 | e->giveup_at_usec = 0; |
| 1206 | |
| 1207 | /* |
| 1208 | * The device's add event is soaked for about a second and |
| 1209 | * can be processed before ExecStartPost= wrote the state |
| 1210 | * file. The job completes only after ExecStartPost=. |
| 1211 | */ |
| 1212 | if (e->is_dc) { |
| 1213 | __cleanup_free__attribute__((cleanup(shr_freep))) char *devname = |
| 1214 | unit_read_devid(unit_name); |
| 1215 | |
| 1216 | if (devname) |
| 1217 | dc_connected(e, devname); |
| 1218 | } |
| 1219 | return; |
| 1220 | } |
| 1221 | |
| 1222 | log_warn("%s - connection unit failed", libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_WARN, "%s - connection unit failed", libnvmf_tid_str (e->tid)); |
| 1223 | if (schedule_retry(e) < 0) |
| 1224 | log_err("%s - failed to schedule retry",dmn_log_msg(DMN_LOG_ERR, "%s - failed to schedule retry", libnvmf_tid_str (e->tid)) |
| 1225 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_ERR, "%s - failed to schedule retry", libnvmf_tid_str (e->tid)); |
| 1226 | } |
| 1227 | |
| 1228 | static void on_dc_add(const char *devname, |
| 1229 | void *user_data __attribute__((unused))) |
| 1230 | { |
| 1231 | struct active_ctrl *e; |
| 1232 | char *unit_name; |
| 1233 | |
| 1234 | // Only a DC that nvme-discoverd connected has a state file. |
| 1235 | unit_name = state_read_unit(devname); |
| 1236 | if (!unit_name) |
| 1237 | return; |
| 1238 | |
| 1239 | e = ctrl_find_by_unit(unit_name); |
| 1240 | free(unit_name); |
| 1241 | if (e) |
| 1242 | dc_connected(e, devname); |
| 1243 | } |
| 1244 | |
| 1245 | /* |
| 1246 | * A DC that nobody owns, e.g. from a manual "nvme discover --persistent". |
| 1247 | * Without the legacy udev rules, nothing else handles its events. Run |
| 1248 | * connect-all for it, as those rules do. The connections it makes stay |
| 1249 | * unowned, like the DC. |
| 1250 | */ |
| 1251 | static void dc_changed_unowned(const char *devname, sd_device *dev) |
| 1252 | { |
| 1253 | __cleanup_free__attribute__((cleanup(shr_freep))) char *owner = NULL((void*)0); |
| 1254 | struct dc_addr addr = { }; |
| 1255 | int r; |
| 1256 | |
| 1257 | r = libnvmf_registry_retrieve(ctx.nvme_ctx, devname, "owner", &owner); |
| 1258 | if (r < 0 && r != -ENOENT2) { |
| 1259 | log_warn("%s - failed to check registry owner: %s",dmn_log_msg(DMN_LOG_WARN, "%s - failed to check registry owner: %s" , devname, strerror(-r)) |
| 1260 | devname, strerror(-r))dmn_log_msg(DMN_LOG_WARN, "%s - failed to check registry owner: %s" , devname, strerror(-r)); |
| 1261 | return; |
| 1262 | } |
| 1263 | if (owner) { |
| 1264 | log_dbg("%s - discovery log changed, owned by '%s', skipping",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - discovery log changed, owned by '%s', skipping" , __func__, devname, owner) |
| 1265 | devname, owner)dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - discovery log changed, owned by '%s', skipping" , __func__, devname, owner); |
| 1266 | return; |
| 1267 | } |
| 1268 | |
| 1269 | sd_device_get_property_value(dev, "NVME_TRTYPE", &addr.transport); |
| 1270 | sd_device_get_property_value(dev, "NVME_TRADDR", &addr.traddr); |
| 1271 | sd_device_get_property_value(dev, "NVME_TRSVCID", &addr.trsvcid); |
| 1272 | sd_device_get_property_value(dev, "NVME_HOST_TRADDR", |
| 1273 | &addr.host_traddr); |
| 1274 | sd_device_get_property_value(dev, "NVME_HOST_IFACE", |
| 1275 | &addr.host_iface); |
| 1276 | |
| 1277 | log_info("%s - discovery log changed on an unowned DC, running connect-all",dmn_log_msg(DMN_LOG_INFO, "%s - discovery log changed on an unowned DC, running connect-all" , devname) |
| 1278 | devname)dmn_log_msg(DMN_LOG_INFO, "%s - discovery log changed on an unowned DC, running connect-all" , devname); |
| 1279 | r = unit_start_connect_all(ctx.umgr, devname, &addr); |
| 1280 | if (r < 0) |
| 1281 | log_err("%s - failed to start connect-all: %s",dmn_log_msg(DMN_LOG_ERR, "%s - failed to start connect-all: %s" , devname, strerror(-r)) |
| 1282 | devname, strerror(-r))dmn_log_msg(DMN_LOG_ERR, "%s - failed to start connect-all: %s" , devname, strerror(-r)); |
| 1283 | } |
| 1284 | |
| 1285 | static void on_dc_changed(const char *devname, sd_device *dev, |
| 1286 | void *user_data __attribute__((unused))) |
| 1287 | { |
| 1288 | struct active_ctrl *e; |
| 1289 | |
| 1290 | e = ctrl_find_by_devname(devname); |
| 1291 | if (!e) { |
| 1292 | dc_changed_unowned(devname, dev); |
| 1293 | return; |
| 1294 | } |
| 1295 | if (!e->tid) { |
| 1296 | log_warn("%s - dc_changed for untracked device", devname)dmn_log_msg(DMN_LOG_WARN, "%s - dc_changed for untracked device" , devname); |
| 1297 | return; |
| 1298 | } |
| 1299 | |
| 1300 | log_dbg("%s | %s: discovery log changed, re-fetching",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s | %s: discovery log changed, re-fetching" , __func__, libnvmf_tid_str(e->tid), devname) |
| 1301 | libnvmf_tid_str(e->tid), devname)dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s | %s: discovery log changed, re-fetching" , __func__, libnvmf_tid_str(e->tid), devname); |
| 1302 | dlp_fetch(e); |
| 1303 | } |
| 1304 | |
| 1305 | static void on_ioc_add(const char *devname, |
| 1306 | void *user_data __attribute__((unused))) |
| 1307 | { |
| 1308 | struct active_ctrl *e; |
| 1309 | char *unit_name; |
| 1310 | |
| 1311 | unit_name = state_read_unit(devname); |
| 1312 | if (!unit_name) |
| 1313 | return; |
| 1314 | |
| 1315 | e = ctrl_find_by_unit(unit_name); |
| 1316 | if (e && !e->devname) |
| 1317 | e->devname = strdup(devname); |
| 1318 | free(unit_name); |
| 1319 | } |
| 1320 | |
| 1321 | static bool_Bool sysfs_ctrl_deleting(const char *devname) |
| 1322 | { |
| 1323 | sd_device *dev = NULL((void*)0); |
| 1324 | const char *state = NULL((void*)0); |
| 1325 | char syspath[256]; |
| 1326 | bool_Bool deleting; |
| 1327 | |
| 1328 | snprintf(syspath, sizeof(syspath), SYSFS_NVME_DIR"/sys/class/nvme" "/%s", devname); |
| 1329 | if (sd_device_new_from_syspath(&dev, syspath) < 0) |
| 1330 | return true1; |
| 1331 | |
| 1332 | sd_device_get_sysattr_value(dev, "state", &state); |
| 1333 | deleting = !state || !strncmp(state, "deleting", 8); |
| 1334 | sd_device_unref(dev); |
| 1335 | |
| 1336 | return deleting; |
| 1337 | } |
| 1338 | |
| 1339 | static void on_nvme_remove(const char *devname, |
| 1340 | void *user_data __attribute__((unused))) |
| 1341 | { |
| 1342 | struct active_ctrl *e; |
| 1343 | bool_Bool is_fc; |
| 1344 | |
| 1345 | /* |
| 1346 | * The kernel reuses a device name as soon as it is free. A "remove" |
| 1347 | * can arrive after a new device took the name, and the new device |
| 1348 | * must be kept. The kernel sends "remove" before it deletes the |
| 1349 | * sysfs directory, so a device that is being deleted is not new. |
| 1350 | */ |
| 1351 | if (state_ctrl_present(devname) && !sysfs_ctrl_deleting(devname)) { |
| 1352 | log_dbg("%s - remove for an earlier device of this name, ignored",dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - remove for an earlier device of this name, ignored" , __func__, devname) |
| 1353 | devname)dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "%s - remove for an earlier device of this name, ignored" , __func__, devname); |
| 1354 | return; |
| 1355 | } |
| 1356 | |
| 1357 | e = ctrl_find_by_devname(devname); |
| 1358 | state_remove_ctrl(devname); |
| 1359 | if (!e) |
| 1360 | return; // might be a controller we didn't start — ignore |
| 1361 | |
| 1362 | state_remove_devid(e->unit_name); // ExecStopPost usually beat us to it |
| 1363 | |
| 1364 | free(e->devname); |
| 1365 | e->devname = NULL((void*)0); |
| 1366 | |
| 1367 | /* |
| 1368 | * Both checks must pass before reconnecting: a matching exclusion |
| 1369 | * entry wins even over a still-desired controller — it is the |
| 1370 | * administrator's explicit override. |
| 1371 | */ |
| 1372 | if (!should_connect(NULL((void*)0), e->tid, devname)) { |
| 1373 | unit_stop(ctx.umgr, e->unit_name); |
| 1374 | ctrl_remove(e); |
| 1375 | return; |
| 1376 | } |
| 1377 | |
| 1378 | /* |
| 1379 | * Checked before "not desired": this removal is discoverd's own |
| 1380 | * doing, and the poll timer owns what happens next. Checking |
| 1381 | * "desired" first can call ctrl_remove(), which frees the timer |
| 1382 | * epcsd_park() armed moments ago. The timer re-checks "desired" |
| 1383 | * itself when it fires, so nothing is skipped by deferring. |
| 1384 | */ |
| 1385 | if (e->epcsd_poll_timer) |
| 1386 | return; |
| 1387 | |
| 1388 | if (!inventory_is_desired(ctx.inventory, e->tid)) { |
| 1389 | log_info("%s | %s - removed, not desired, dropping",dmn_log_msg(DMN_LOG_INFO, "%s | %s - removed, not desired, dropping" , libnvmf_tid_str(e->tid), devname) |
| 1390 | libnvmf_tid_str(e->tid), devname)dmn_log_msg(DMN_LOG_INFO, "%s | %s - removed, not desired, dropping" , libnvmf_tid_str(e->tid), devname); |
| 1391 | ctrl_remove(e); |
| 1392 | return; |
| 1393 | } |
| 1394 | |
| 1395 | log_info("%s | %s - removed but still desired, reconnecting",dmn_log_msg(DMN_LOG_INFO, "%s | %s - removed but still desired, reconnecting" , libnvmf_tid_str(e->tid), devname) |
| 1396 | libnvmf_tid_str(e->tid), devname)dmn_log_msg(DMN_LOG_INFO, "%s | %s - removed but still desired, reconnecting" , libnvmf_tid_str(e->tid), devname); |
| 1397 | |
| 1398 | is_fc = shr_streq0(libnvmf_tid_get_transport(e->tid), "fc"); |
| 1399 | |
| 1400 | if (is_fc) { |
| 1401 | // FC: stop old unit, re-issue kickstart. |
| 1402 | unit_stop(ctx.umgr, e->unit_name); |
| 1403 | ctrl_remove(e); |
| 1404 | fc_kickstart(); |
| 1405 | } else if (restart_or_start(e) < 0) { |
| 1406 | if (schedule_retry(e) < 0) |
| 1407 | log_err("%s - failed to schedule retry",dmn_log_msg(DMN_LOG_ERR, "%s - failed to schedule retry", libnvmf_tid_str (e->tid)) |
| 1408 | libnvmf_tid_str(e->tid))dmn_log_msg(DMN_LOG_ERR, "%s - failed to schedule retry", libnvmf_tid_str (e->tid)); |
| 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | static void on_fc_discovery(const struct libnvmf_tid *t, |
| 1413 | void *user_data __attribute__((unused))) |
| 1414 | { |
| 1415 | __cleanup_tid__attribute__((cleanup(cleanup_tid))) struct libnvmf_tid *tid = libnvmf_tid_dup(t); |
| 1416 | |
| 1417 | // The uevent names no host: connect as the default host. |
| 1418 | if (!tid || |
| 1419 | tid_set_default_host_if_unset(tid, ctx.hostnqn, ctx.hostid) < 0) |
| 1420 | return; |
| 1421 | |
| 1422 | /* |
| 1423 | * fc_monitor_handler() is the only producer of fc_discovery |
| 1424 | * callbacks, and the kernel only fires that uevent for an FC |
| 1425 | * remote port advertising FC_PORT_ROLE_NVME_DISCOVERY - so tid is |
| 1426 | * always a DC here, never an IOC. Connect as a DC; we fetch its |
| 1427 | * DLP (and discover any IOCs behind it) once the device appears. |
| 1428 | */ |
| 1429 | inventory_add_discovered_dc(ctx.inventory, tid); |
| 1430 | if (should_connect(NULL((void*)0), tid, NULL((void*)0))) |
| 1431 | start_ctrl(tid, true1, NULL((void*)0)); |
| 1432 | } |
| 1433 | |
| 1434 | /* |
| 1435 | * Whether the kernel accepts a discovery connect to a DC's own NQN (TP8013). |
| 1436 | * On any error, answer no: the well-known discovery NQN always works. |
| 1437 | */ |
| 1438 | static bool_Bool kernel_supports_discovery_nqn(void) |
| 1439 | { |
| 1440 | bool_Bool supported; |
| 1441 | int r; |
| 1442 | |
| 1443 | r = libnvmf_kernel_option_supported(ctx.nvme_ctx, "discovery", |
| 1444 | &supported); |
| 1445 | if (r < 0) { |
| 1446 | log_dbg("kernel fabrics options: %s", libnvme_strerror(-r))dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "kernel fabrics options: %s" , __func__, libnvme_strerror(-r)); |
| 1447 | return false0; |
| 1448 | } |
| 1449 | |
| 1450 | return supported; |
| 1451 | } |
| 1452 | |
| 1453 | /* |
| 1454 | * mDNS found a DC. Connect with its advertised NQN if the kernel allows |
| 1455 | * it, else with the well-known discovery NQN. The DLP gives the real NQN |
| 1456 | * either way. host_iface is only valid for tcp. A link-local traddr gets |
| 1457 | * the interface as its scope, because rdma has no host_iface. |
| 1458 | */ |
| 1459 | static void on_mdns_add(const char *traddr, const char *trsvcid, |
| 1460 | const char *transport, const char *nqn, |
| 1461 | const char *ifname, |
| 1462 | int ifindex __attribute__((unused)), |
| 1463 | void *user_data __attribute__((unused))) |
| 1464 | { |
| 1465 | const char *subsysnqn = NVME_DISC_SUBSYS_NAME"nqn.2014-08.org.nvmexpress.discovery"; |
| 1466 | __cleanup_tid__attribute__((cleanup(cleanup_tid))) struct libnvmf_tid *tid = NULL((void*)0); |
| 1467 | bool_Bool is_tcp = shr_streq0(transport, "tcp"); |
| 1468 | __cleanup_free__attribute__((cleanup(shr_freep))) char *scoped = NULL((void*)0); |
| 1469 | |
| 1470 | if (nqn && kernel_supports_discovery_nqn()) |
| 1471 | subsysnqn = nqn; |
| 1472 | |
| 1473 | scoped = tid_scope_link_local(traddr, ifname); |
| 1474 | if (!scoped) |
| 1475 | return; |
| 1476 | |
| 1477 | tid = tid_new(transport, scoped, trsvcid, subsysnqn, NULL((void*)0), |
| 1478 | is_tcp ? ifname : NULL((void*)0), NULL((void*)0), NULL((void*)0), true1); |
| 1479 | if (!tid || |
| 1480 | tid_set_default_host_if_unset(tid, ctx.hostnqn, ctx.hostid) < 0) |
| 1481 | return; |
| 1482 | |
| 1483 | inventory_add_discovered_dc(ctx.inventory, tid); |
| 1484 | if (should_connect(NULL((void*)0), tid, NULL((void*)0))) |
| 1485 | start_ctrl(tid, true1, NULL((void*)0)); |
| 1486 | } |
| 1487 | |
| 1488 | /* |
| 1489 | * No service_remove callback: an mDNS announcement can disappear briefly |
| 1490 | * (cache expiry, link flap). A connected DC is managed by the same paths |
| 1491 | * as any other DC, not by its announcement. |
| 1492 | */ |
| 1493 | static const struct mdns_callbacks mdns_callbacks = { |
| 1494 | .service_add = on_mdns_add, |
| 1495 | }; |
| 1496 | |
| 1497 | /* Start or stop mDNS to match the zeroconf setting. */ |
| 1498 | static void apply_zeroconf(void) |
| 1499 | { |
| 1500 | int r; |
| 1501 | |
| 1502 | if (ctx.cfg->zeroconf && !ctx.mdns) { |
| 1503 | r = mdns_start(ctx.event, &mdns_callbacks, NULL((void*)0), &ctx.mdns); |
| 1504 | if (r == -ENOSYS38) |
| 1505 | log_warn("mdns: zeroconf is enabled, but nvme-discoverd was built without mDNS support")dmn_log_msg(DMN_LOG_WARN, "mdns: zeroconf is enabled, but nvme-discoverd was built without mDNS support" ); |
| 1506 | else if (r == -EOPNOTSUPP95) |
| 1507 | log_warn("mdns: zeroconf is enabled, but systemd-resolved has no BrowseServices (needs >= v258)")dmn_log_msg(DMN_LOG_WARN, "mdns: zeroconf is enabled, but systemd-resolved has no BrowseServices (needs >= v258)" ); |
| 1508 | else if (r < 0) |
| 1509 | log_warn("mdns: zeroconf is enabled, but mDNS cannot start: %s",dmn_log_msg(DMN_LOG_WARN, "mdns: zeroconf is enabled, but mDNS cannot start: %s" , strerror(-r)) |
| 1510 | strerror(-r))dmn_log_msg(DMN_LOG_WARN, "mdns: zeroconf is enabled, but mDNS cannot start: %s" , strerror(-r)); |
| 1511 | } else if (!ctx.cfg->zeroconf && ctx.mdns) { |
| 1512 | mdns_stop(ctx.mdns); |
| 1513 | ctx.mdns = NULL((void*)0); |
| 1514 | } |
| 1515 | } |
| 1516 | |
| 1517 | static struct libnvmf_tid *sysfs_read_tid(const char *devname, bool_Bool *is_dc, |
| 1518 | bool_Bool *live) |
| 1519 | { |
| 1520 | sd_device *dev = NULL((void*)0); |
| 1521 | const char *state = NULL((void*)0); |
| 1522 | char syspath[256]; |
| 1523 | struct libnvmf_tid *t; |
| 1524 | |
| 1525 | if (is_dc) |
| 1526 | *is_dc = false0; |
| 1527 | *live = false0; |
| 1528 | |
| 1529 | snprintf(syspath, sizeof(syspath), "/sys/class/nvme/%s", devname); |
| 1530 | if (sd_device_new_from_syspath(&dev, syspath) < 0) |
| 1531 | return NULL((void*)0); |
| 1532 | |
| 1533 | t = tid_from_sysfs(dev, is_dc); |
| 1534 | sd_device_get_sysattr_value(dev, "state", &state); |
| 1535 | *live = shr_streq0(state, "live"); |
| 1536 | sd_device_unref(dev); |
| 1537 | return t; |
| 1538 | } |
| 1539 | |
| 1540 | /* |
| 1541 | * Is @devname a connection that satisfies @tid? A device name alone does not |
| 1542 | * identify a controller, because the kernel reuses nvmeN names. |
| 1543 | */ |
| 1544 | static bool_Bool devname_matches_tid(const char *devname, |
| 1545 | const struct libnvmf_tid *tid) |
| 1546 | { |
| 1547 | __cleanup_tid__attribute__((cleanup(cleanup_tid))) struct libnvmf_tid *existing = NULL((void*)0); |
| 1548 | struct ifaddrs *iface_list = NULL((void*)0); |
| 1549 | bool_Bool is_dc, live, match; |
| 1550 | |
| 1551 | existing = sysfs_read_tid(devname, &is_dc, &live); |
| 1552 | if (!existing) |
| 1553 | return false0; |
| 1554 | |
| 1555 | if (getifaddrs(&iface_list) < 0) { |
| 1556 | log_warn("getifaddrs: %s", strerror(errno))dmn_log_msg(DMN_LOG_WARN, "getifaddrs: %s", strerror((*__errno_location ()))); |
| 1557 | iface_list = NULL((void*)0); |
| 1558 | } |
| 1559 | |
| 1560 | match = tid_matches_existing(tid, existing, is_dc, live, iface_list); |
| 1561 | freeifaddrs(iface_list); |
| 1562 | |
| 1563 | return match; |
| 1564 | } |
| 1565 | |
| 1566 | static void conn_scan_free(struct conn_scan *scan) |
| 1567 | { |
| 1568 | size_t i; |
| 1569 | |
| 1570 | for (i = 0; i < scan->ctrls.len; i++) |
| 1571 | scanned_ctrl_free(scan->ctrls.items[i]); |
| 1572 | scanned_ctrl_list_free(&scan->ctrls); |
| 1573 | freeifaddrs(scan->iface_list); |
| 1574 | memset(scan, 0, sizeof(*scan)); |
| 1575 | } |
| 1576 | |
| 1577 | /* |
| 1578 | * Read every connected controller out of sysfs and snapshot the host's |
| 1579 | * interface addresses. A controller whose TID cannot be built is left out: |
| 1580 | * it cannot be matched against anyway. On failure the scan is left short |
| 1581 | * rather than failed, which costs a missed match, not a crash. |
| 1582 | */ |
| 1583 | static void conn_scan_load(struct conn_scan *scan) |
| 1584 | { |
| 1585 | DIR *d; |
| 1586 | struct dirent *ent; |
| 1587 | |
| 1588 | memset(scan, 0, sizeof(*scan)); |
| 1589 | |
| 1590 | if (getifaddrs(&scan->iface_list) < 0) { |
| 1591 | log_warn("getifaddrs: %s", strerror(errno))dmn_log_msg(DMN_LOG_WARN, "getifaddrs: %s", strerror((*__errno_location ()))); |
| 1592 | scan->iface_list = NULL((void*)0); |
| 1593 | } |
| 1594 | |
| 1595 | d = opendir("/sys/class/nvme"); |
| 1596 | if (!d) { |
| 1597 | if (errno(*__errno_location ()) != ENOENT2) |
| 1598 | log_warn("opendir /sys/class/nvme: %s",dmn_log_msg(DMN_LOG_WARN, "opendir /sys/class/nvme: %s", strerror ((*__errno_location ()))) |
| 1599 | strerror(errno))dmn_log_msg(DMN_LOG_WARN, "opendir /sys/class/nvme: %s", strerror ((*__errno_location ()))); |
| 1600 | return; |
| 1601 | } |
| 1602 | |
| 1603 | while ((ent = readdir(d))) { |
| 1604 | struct scanned_ctrl *sc; |
| 1605 | |
| 1606 | if (ent->d_name[0] == '.') |
| 1607 | continue; |
| 1608 | |
| 1609 | sc = calloc(1, sizeof(*sc)); |
| 1610 | if (!sc) { |
| 1611 | log_warn("connection scan: out of memory")dmn_log_msg(DMN_LOG_WARN, "connection scan: out of memory"); |
| 1612 | break; |
| 1613 | } |
| 1614 | |
| 1615 | sc->tid = sysfs_read_tid(ent->d_name, &sc->is_dc, &sc->live); |
| 1616 | sc->devname = shr_xstrdup(ent->d_name); |
| 1617 | if (!sc->tid || !sc->devname) { |
| 1618 | scanned_ctrl_free(sc); |
| 1619 | continue; |
| 1620 | } |
| 1621 | |
| 1622 | if (scanned_ctrl_list_append(&scan->ctrls, sc) < 0) { |
| 1623 | log_warn("connection scan: out of memory")dmn_log_msg(DMN_LOG_WARN, "connection scan: out of memory"); |
| 1624 | scanned_ctrl_free(sc); |
| 1625 | break; |
| 1626 | } |
| 1627 | } |
| 1628 | closedir(d); |
| 1629 | } |
| 1630 | |
| 1631 | /* |
| 1632 | * Find the device name (e.g. "nvme3") of a connected controller that |
| 1633 | * satisfies @tid, using the snapshot rather than libnvme's topology tree |
| 1634 | * (see should_connect()'s comment for why that tree can't be relied on |
| 1635 | * here). Returns NULL if nothing in the snapshot matches. The returned |
| 1636 | * string belongs to the scan. |
| 1637 | */ |
| 1638 | static const char *find_devname_for_tid(const struct conn_scan *scan, |
| 1639 | const struct libnvmf_tid *tid) |
| 1640 | { |
| 1641 | size_t i; |
| 1642 | |
| 1643 | for (i = 0; i < scan->ctrls.len; i++) { |
| 1644 | const struct scanned_ctrl *sc = scan->ctrls.items[i]; |
| 1645 | |
| 1646 | if (tid_matches_existing(tid, sc->tid, sc->is_dc, sc->live, |
| 1647 | scan->iface_list)) |
| 1648 | return sc->devname; |
| 1649 | } |
| 1650 | return NULL((void*)0); |
| 1651 | } |
| 1652 | |
| 1653 | static void connect_desired(void) |
| 1654 | { |
| 1655 | __cleanup_conn_scan__attribute__((cleanup(conn_scan_free))) struct conn_scan scan = { 0 }; |
| 1656 | struct libnvmf_tid **dcs, **iocs; |
| 1657 | int i; |
| 1658 | |
| 1659 | conn_scan_load(&scan); |
| 1660 | |
| 1661 | dcs = inventory_desired_dcs(ctx.inventory); |
| 1662 | if (dcs) { |
| 1663 | for (i = 0; dcs[i]; i++) { |
| 1664 | if (should_connect(&scan, dcs[i], NULL((void*)0))) |
| 1665 | start_ctrl(dcs[i], true1, NULL((void*)0)); |
| 1666 | tid_free(dcs[i]); |
| 1667 | } |
| 1668 | free(dcs); |
| 1669 | } |
| 1670 | |
| 1671 | iocs = inventory_desired_iocs(ctx.inventory); |
| 1672 | if (iocs) { |
| 1673 | for (i = 0; iocs[i]; i++) { |
| 1674 | if (should_connect(&scan, iocs[i], NULL((void*)0))) |
| 1675 | start_ctrl(iocs[i], false0, NULL((void*)0)); |
| 1676 | tid_free(iocs[i]); |
| 1677 | } |
| 1678 | free(iocs); |
| 1679 | } |
| 1680 | } |
| 1681 | |
| 1682 | /* |
| 1683 | * Apply the effective log level to both discoverd and its in-process libnvme |
| 1684 | * context. A command-line --debug (ctx.force_debug) forces DEBUG and |
| 1685 | * overrides the config; otherwise the configured debug-level (default INFO) |
| 1686 | * is used. Called at startup after config_load() and again after a SIGHUP |
| 1687 | * reload, since the libnvme context outlives any single config. |
| 1688 | */ |
| 1689 | /* |
| 1690 | * A name lookup of the fabrics configuration ended. Connect what it added, |
| 1691 | * and run the release pass that waited for it. |
| 1692 | */ |
| 1693 | static void on_config_lookup(bool_Bool added, |
| 1694 | void *user_data __attribute__((unused))) |
| 1695 | { |
| 1696 | if (added) |
| 1697 | connect_desired(); |
| 1698 | schedule_release(false0); |
| 1699 | } |
| 1700 | |
| 1701 | static void apply_log_level(void) |
| 1702 | { |
| 1703 | int level = log_level(); |
| 1704 | |
| 1705 | dmn_log_set_level(level); |
| 1706 | libnvme_set_logging_level(ctx.nvme_ctx, level, false0, false0); |
| 1707 | } |
| 1708 | |
| 1709 | static void reload_config(void *user_data __attribute__((unused))) |
| 1710 | { |
| 1711 | struct discoverd_config *new_cfg; |
| 1712 | struct libnvmf_config *new_fabrics_cfg; |
| 1713 | |
| 1714 | new_cfg = config_load(ctx.conf_path); |
| 1715 | if (!new_cfg) { |
| 1716 | log_err("failed to reload config")dmn_log_msg(DMN_LOG_ERR, "failed to reload config"); |
| 1717 | return; |
| 1718 | } |
| 1719 | config_free(ctx.cfg); |
| 1720 | ctx.cfg = new_cfg; |
| 1721 | apply_log_level(); |
| 1722 | |
| 1723 | if (libnvmf_config_read(ctx.nvme_ctx, NULL((void*)0), &new_fabrics_cfg) == 0) { |
| 1724 | libnvmf_config_free(ctx.fabrics_cfg); |
| 1725 | ctx.fabrics_cfg = new_fabrics_cfg; |
| 1726 | } else { |
| 1727 | log_err("failed to reload fabrics config, keeping last-good")dmn_log_msg(DMN_LOG_ERR, "failed to reload fabrics config, keeping last-good" ); |
| 1728 | } |
| 1729 | inventory_load_config(ctx.inventory, &ctx); |
| 1730 | |
| 1731 | /* |
| 1732 | * Connect any newly added desired controllers. Release those the |
| 1733 | * configuration or the exclusion list removed, without disconnecting |
| 1734 | * them. |
| 1735 | */ |
| 1736 | connect_desired(); |
| 1737 | schedule_release(true1); |
| 1738 | apply_zeroconf(); |
| 1739 | } |
| 1740 | |
| 1741 | /* |
| 1742 | * Resolve the identity a connect without --hostnqn uses. Must run before |
| 1743 | * anything scans a controller into the libnvme tree: |
| 1744 | * libnvmf_host_get_ids() prefers a host already in the tree. |
| 1745 | */ |
| 1746 | static int resolve_default_host(void) |
| 1747 | { |
| 1748 | int r; |
| 1749 | |
| 1750 | r = libnvmf_host_get_ids(ctx.nvme_ctx, NULL((void*)0), NULL((void*)0), |
| 1751 | &ctx.hostnqn, &ctx.hostid); |
| 1752 | if (r < 0) |
| 1753 | return r; |
| 1754 | |
| 1755 | log_dbg("default host identity: %s, %s", ctx.hostnqn, ctx.hostid)dmn_log_msg(DMN_LOG_DEBUG, "%s() - " "default host identity: %s, %s" , __func__, ctx.hostnqn, ctx.hostid); |
| 1756 | |
| 1757 | return 0; |
| 1758 | } |
| 1759 | |
| 1760 | int main(int argc, char **argv) |
| 1761 | { |
| 1762 | static const struct events_callbacks callbacks = { |
| 1763 | .dc_add = on_dc_add, |
| 1764 | .dc_changed = on_dc_changed, |
| 1765 | .ioc_add = on_ioc_add, |
| 1766 | .nvme_remove = on_nvme_remove, |
| 1767 | .fc_discovery = on_fc_discovery, |
| 1768 | }; |
| 1769 | static const struct option long_opts[] = { |
| 1770 | { "config", required_argument1, NULL((void*)0), 'c' }, |
| 1771 | { "nvme-path", required_argument1, NULL((void*)0), 'N' }, |
| 1772 | { "debug", no_argument0, NULL((void*)0), 'd' }, |
| 1773 | { "help", no_argument0, NULL((void*)0), 'h' }, |
| 1774 | { NULL((void*)0), 0, NULL((void*)0), 0 }, |
| 1775 | }; |
| 1776 | const char *nvme_path = NULL((void*)0), *config_path = NULL((void*)0); |
| 1777 | char *nvme_path_abs = NULL((void*)0), *config_path_abs = NULL((void*)0); |
| 1778 | bool_Bool debug = false0; |
| 1779 | int r, c; |
| 1780 | |
| 1781 | while ((c = getopt_long(argc, argv, "c:dh", long_opts, NULL((void*)0))) != -1) { |
| 1782 | switch (c) { |
| 1783 | case 'c': |
| 1784 | config_path = optarg; |
| 1785 | break; |
| 1786 | case 'N': |
| 1787 | nvme_path = optarg; |
| 1788 | break; |
| 1789 | case 'd': |
| 1790 | debug = true1; |
| 1791 | break; |
| 1792 | case 'h': |
| 1793 | printf("Usage: %s [OPTIONS]\n" |
| 1794 | "\n" |
| 1795 | "All options are optional; specify one only to override its default.\n" |
| 1796 | " --config FILE, -c FILE discoverd configuration file\n" |
| 1797 | " (default: " DISCOVERD_CONF_PATH"/usr/local/etc" "/nvme/nvme-discoverd.conf" ")\n" |
| 1798 | " --nvme-path PATH nvme binary the connection units exec\n" |
| 1799 | " (default: <sbindir>/nvme)\n" |
| 1800 | " --debug, -d enable debug logging (journal + libnvme)\n" |
| 1801 | " --help, -h show this help and exit\n", |
| 1802 | argv[0]); |
| 1803 | return 0; |
| 1804 | default: |
| 1805 | fprintf(stderrstderr, "Try '%s --help'.\n", argv[0]); |
| 1806 | return 1; |
| 1807 | } |
| 1808 | } |
| 1809 | |
| 1810 | if (debug) |
| 1811 | dmn_log_set_level(DMN_LOG_DEBUG); |
| 1812 | ctx.force_debug = debug; |
| 1813 | |
| 1814 | /* |
| 1815 | * The nvme path is baked into each transient unit's ExecStart=, which |
| 1816 | * systemd requires to be absolute. Canonicalize it (resolving a |
| 1817 | * relative path against the current directory) so the daemon can be |
| 1818 | * launched as, e.g., --nvme-path ./.build/nvme. realpath() also |
| 1819 | * confirms the binary exists, failing fast on a typo. |
| 1820 | */ |
| 1821 | if (nvme_path) { |
| 1822 | nvme_path_abs = realpath(nvme_path, NULL((void*)0)); |
| 1823 | if (!nvme_path_abs) { |
| 1824 | fprintf(stderrstderr, "--nvme-path: cannot resolve '%s': %s\n", |
| 1825 | nvme_path, strerror(errno(*__errno_location ()))); |
| 1826 | return 1; |
| 1827 | } |
| 1828 | } |
| 1829 | |
| 1830 | if (config_path) { |
| 1831 | config_path_abs = realpath(config_path, NULL((void*)0)); |
| 1832 | if (!config_path_abs) { |
| 1833 | fprintf(stderrstderr, "--config: cannot resolve '%s': %s\n", |
| 1834 | config_path, strerror(errno(*__errno_location ()))); |
| 1835 | return 1; |
| 1836 | } |
| 1837 | } |
| 1838 | ctx.conf_path = config_path_abs ? config_path_abs : DISCOVERD_CONF_PATH"/usr/local/etc" "/nvme/nvme-discoverd.conf"; |
| 1839 | |
| 1840 | r = sd_event_default(&ctx.event); |
| 1841 | if (r < 0) { |
| 1842 | log_err("sd_event_default: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "sd_event_default: %s", strerror(-r) ); |
| 1843 | return 1; |
| 1844 | } |
| 1845 | |
| 1846 | r = sd_bus_open_system(&ctx.bus); |
| 1847 | if (r < 0) { |
| 1848 | log_err("sd_bus_open_system: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "sd_bus_open_system: %s", strerror(- r)); |
| 1849 | return 1; |
| 1850 | } |
| 1851 | |
| 1852 | r = sd_bus_attach_event(ctx.bus, ctx.event, SD_EVENT_PRIORITY_NORMAL); |
| 1853 | if (r < 0) { |
| 1854 | log_err("sd_bus_attach_event: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "sd_bus_attach_event: %s", strerror( -r)); |
| 1855 | return 1; |
| 1856 | } |
| 1857 | |
| 1858 | r = state_init(); |
| 1859 | if (r < 0) { |
| 1860 | log_err("state_init: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "state_init: %s", strerror(-r)); |
| 1861 | return 1; |
| 1862 | } |
| 1863 | state_gc(); |
| 1864 | |
| 1865 | ctx.nvme_ctx = libnvme_create_global_ctx(); |
| 1866 | if (!ctx.nvme_ctx) { |
| 1867 | log_err("libnvme_create_global_ctx: failed")dmn_log_msg(DMN_LOG_ERR, "libnvme_create_global_ctx: failed"); |
| 1868 | return 1; |
| 1869 | } |
| 1870 | libnvme_set_logging_level(ctx.nvme_ctx, |
| 1871 | debug ? LIBNVME_LOG_DEBUG : LIBNVME_LOG_ERR, |
| 1872 | false0, false0); |
| 1873 | |
| 1874 | ctx.cfg = config_load(ctx.conf_path); |
| 1875 | if (!ctx.cfg) { |
| 1876 | log_err("config_load: failed")dmn_log_msg(DMN_LOG_ERR, "config_load: failed"); |
| 1877 | return 1; |
| 1878 | } |
| 1879 | apply_log_level(); |
| 1880 | |
| 1881 | r = libnvmf_config_read(ctx.nvme_ctx, NULL((void*)0), &ctx.fabrics_cfg); |
| 1882 | if (r < 0) { |
| 1883 | log_err("libnvmf_config_read: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "libnvmf_config_read: %s", strerror( -r)); |
| 1884 | return 1; |
| 1885 | } |
| 1886 | |
| 1887 | r = resolve_default_host(); |
| 1888 | if (r < 0) { |
| 1889 | log_err("failed to resolve the host identity: %s",dmn_log_msg(DMN_LOG_ERR, "failed to resolve the host identity: %s" , strerror(-r)) |
| 1890 | strerror(-r))dmn_log_msg(DMN_LOG_ERR, "failed to resolve the host identity: %s" , strerror(-r)); |
| 1891 | return 1; |
| 1892 | } |
| 1893 | |
| 1894 | r = dmn_resolver_new(ctx.event, RESOLVE_WORKERS8, &ctx.resolver); |
| 1895 | if (r < 0) { |
| 1896 | log_err("failed to create the resolver: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "failed to create the resolver: %s", strerror(-r)); |
| 1897 | return 1; |
| 1898 | } |
| 1899 | |
| 1900 | r = dmn_work_pool_new(ctx.event, DLP_WORKERS4, &ctx.dlp_pool); |
| 1901 | if (r < 0) { |
| 1902 | log_err("failed to create the DLP workers: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "failed to create the DLP workers: %s" , strerror(-r)); |
| 1903 | return 1; |
| 1904 | } |
| 1905 | |
| 1906 | ctx.inventory = inventory_new(); |
| 1907 | if (!ctx.inventory) |
| 1908 | return 1; |
| 1909 | inventory_set_lookup_callback(ctx.inventory, on_config_lookup, NULL((void*)0)); |
| 1910 | |
| 1911 | if (ctx.cfg->nbft) |
| 1912 | inventory_load_nbft(ctx.inventory, &ctx, NBFT_SYSFS_PATH"/sys/firmware/acpi/tables"); |
| 1913 | inventory_load_config(ctx.inventory, &ctx); |
| 1914 | load_saved(); |
| 1915 | |
| 1916 | ctx.umgr = unit_mgr_new(ctx.bus, ctx.event, on_job_done, NULL((void*)0), |
| 1917 | nvme_path_abs); |
| 1918 | if (!ctx.umgr) |
| 1919 | return 1; |
| 1920 | |
| 1921 | ctx.evts = events_start(ctx.event, &callbacks, NULL((void*)0)); |
| 1922 | if (!ctx.evts) |
| 1923 | return 1; |
| 1924 | |
| 1925 | if (dmn_add_signal_handlers(ctx.event, reload_config, NULL((void*)0)) < 0) |
| 1926 | return 1; |
| 1927 | |
| 1928 | /* |
| 1929 | * Connect the desired set. start_ctrl() adopts anything a previous |
| 1930 | * run of this daemon left connected. The release pass then decides |
| 1931 | * what the last run wanted and this one does not. |
| 1932 | */ |
| 1933 | connect_desired(); |
| 1934 | schedule_release(false0); |
| 1935 | |
| 1936 | /* |
| 1937 | * One-shot startup FC kickstart: mimics nvmefc-boot-connections.service |
| 1938 | * (which discoverd replaces). Always issued — fc_kickstart() is a |
| 1939 | * no-op (returns 0 on ENOENT) when no FC HBA is present, so it is |
| 1940 | * harmless on non-FC hosts. Independent of the periodic-kickstart |
| 1941 | * knob below. |
| 1942 | */ |
| 1943 | fc_kickstart(); |
| 1944 | |
| 1945 | apply_zeroconf(); |
| 1946 | |
| 1947 | // Periodic FC kickstart: opt-in (default 0 = disabled). |
| 1948 | if (ctx.cfg->fc_kickstart_interval_minutes > 0) { |
| 1949 | uint64_t now, interval; |
| 1950 | |
| 1951 | interval = (uint64_t)ctx.cfg->fc_kickstart_interval_minutes * |
| 1952 | 60 * UINT64_C(1000000)1000000UL; |
| 1953 | r = sd_event_now(ctx.event, CLOCK_BOOTTIME7, &now); |
| 1954 | if (r >= 0) |
| 1955 | r = sd_event_add_time(ctx.event, NULL((void*)0), CLOCK_BOOTTIME7, |
| 1956 | now + interval, 0, |
| 1957 | fc_kickstart_timeout, NULL((void*)0)); |
| 1958 | if (r < 0) |
| 1959 | log_err("failed to arm FC kickstart timer: %s",dmn_log_msg(DMN_LOG_ERR, "failed to arm FC kickstart timer: %s" , strerror(-r)) |
| 1960 | strerror(-r))dmn_log_msg(DMN_LOG_ERR, "failed to arm FC kickstart timer: %s" , strerror(-r)); |
| 1961 | } |
| 1962 | |
| 1963 | sd_notify(0, "READY=1"); |
| 1964 | |
| 1965 | r = sd_event_loop(ctx.event); |
| 1966 | if (r < 0) |
| 1967 | log_err("sd_event_loop: %s", strerror(-r))dmn_log_msg(DMN_LOG_ERR, "sd_event_loop: %s", strerror(-r)); |
| 1968 | |
| 1969 | mdns_stop(ctx.mdns); |
| 1970 | events_stop(ctx.evts); |
| 1971 | unit_mgr_free(ctx.umgr); |
| 1972 | free(nvme_path_abs); |
| 1973 | free(config_path_abs); |
| 1974 | inventory_free(ctx.inventory); |
| 1975 | dmn_resolver_free(ctx.resolver); |
| 1976 | dmn_work_pool_free(ctx.dlp_pool); |
| 1977 | free(ctx.hostnqn); |
| 1978 | free(ctx.hostid); |
| 1979 | config_free(ctx.cfg); |
| 1980 | libnvmf_config_free(ctx.fabrics_cfg); |
| 1981 | libnvme_free_global_ctx(ctx.nvme_ctx); |
| 1982 | sd_bus_unref(ctx.bus); |
| 1983 | sd_event_unref(ctx.event); |
| 1984 | |
| 1985 | return r < 0 ? 1 : 0; |
| 1986 | } |