4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License version 2 for more details. A copy is
14 * included in the COPYING file that accompanied this code.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 * Copyright (c) 2011, 2012 Commissariat a l'energie atomique et aux energies
26 * Copyright (c) 2013, 2017, Intel Corporation.
27 * Use is subject to license terms.
30 * lustre/mdt/mdt_coordinator.c
32 * Lustre HSM Coordinator
34 * Author: Jacques-Charles Lafoucriere <jacques-charles.lafoucriere@cea.fr>
35 * Author: Aurelien Degremont <aurelien.degremont@cea.fr>
36 * Author: Thomas Leibovici <thomas.leibovici@cea.fr>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/kthread.h>
42 #include <obd_support.h>
43 #include <lustre_export.h>
45 #include <lprocfs_status.h>
46 #include <lustre_log.h>
47 #include <lustre_kernelcomm.h>
48 #include "mdt_internal.h"
50 static struct lprocfs_vars lprocfs_mdt_hsm_vars[];
53 * get obj and HSM attributes on a fid
54 * \param mti [IN] context
55 * \param fid [IN] object fid
56 * \param hsm [OUT] HSM meta data
57 * \retval obj or error (-ENOENT if not found)
59 struct mdt_object *mdt_hsm_get_md_hsm(struct mdt_thread_info *mti,
60 const struct lu_fid *fid,
64 struct mdt_object *obj;
72 /* find object by FID */
73 obj = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
77 if (!mdt_object_exists(obj)) {
79 mdt_object_put(mti->mti_env, obj);
80 RETURN(ERR_PTR(-ENOENT));
83 rc = mdt_attr_get_complex(mti, obj, ma);
85 mdt_object_put(mti->mti_env, obj);
89 if (ma->ma_valid & MA_HSM)
92 memset(hsm, 0, sizeof(*hsm));
97 void mdt_hsm_dump_hal(int level, const char *prefix,
98 struct hsm_action_list *hal)
101 struct hsm_action_item *hai;
104 CDEBUG(level, "%s: HAL header: version %X count %d"
105 " archive_id %d flags %#llx\n",
106 prefix, hal->hal_version, hal->hal_count,
107 hal->hal_archive_id, hal->hal_flags);
109 hai = hai_first(hal);
110 for (i = 0; i < hal->hal_count; i++) {
111 sz = hai->hai_len - sizeof(*hai);
112 CDEBUG(level, "%s %d: fid="DFID" dfid="DFID
114 " action=%s extent=%#llx-%#llx gid=%#llx"
115 " datalen=%d data=[%s]\n",
117 PFID(&hai->hai_fid), PFID(&hai->hai_dfid),
119 hsm_copytool_action2name(hai->hai_action),
120 hai->hai_extent.offset,
121 hai->hai_extent.length,
123 hai_dump_data_field(hai, buf, sizeof(buf)));
129 * data passed to llog_cat_process() callback
130 * to scan requests and take actions
132 struct hsm_scan_request {
135 struct hsm_action_list *hal;
138 struct hsm_scan_data {
139 struct mdt_thread_info *hsd_mti;
140 char hsd_fsname[MTI_NAME_MAXLEN + 1];
141 /* are we scanning the logs for housekeeping, or just looking
144 bool hsd_housekeeping;
145 int hsd_action_count;
146 int hsd_request_len; /* array alloc len */
147 int hsd_request_count; /* array used count */
148 struct hsm_scan_request *hsd_request;
151 static int mdt_cdt_waiting_cb(const struct lu_env *env,
152 struct mdt_device *mdt,
153 struct llog_handle *llh,
154 struct llog_agent_req_rec *larr,
155 struct hsm_scan_data *hsd)
157 struct coordinator *cdt = &mdt->mdt_coordinator;
158 struct hsm_scan_request *request;
159 struct hsm_action_item *hai;
164 /* Are agents full? */
165 if (hsd->hsd_action_count + atomic_read(&cdt->cdt_request_count) >=
166 cdt->cdt_max_requests) {
167 if (hsd->hsd_housekeeping) {
168 /* Unknown request and no more room for a new
169 * request. Continue to scan to find other
170 * entries for already existing requests. */
173 /* We cannot send and more requests, stop
174 * here. There might be more known requests
175 * that could be merged, but this avoid
176 * analyzing too many llogs for minor
178 RETURN(LLOG_PROC_BREAK);
182 hai_size = cfs_size_round(larr->arr_hai.hai_len);
183 archive_id = larr->arr_archive_id;
185 /* Can we add this action to one of the existing HALs in hsd. */
187 for (i = 0; i < hsd->hsd_request_count; i++) {
188 if (hsd->hsd_request[i].hal->hal_archive_id == archive_id &&
189 hsd->hsd_request[i].hal_used_sz + hai_size <=
191 request = &hsd->hsd_request[i];
197 struct hsm_action_list *hal;
199 if (hsd->hsd_request_count == hsd->hsd_request_len) {
200 /* Logic as above. */
201 if (hsd->hsd_housekeeping)
204 RETURN(LLOG_PROC_BREAK);
207 request = &hsd->hsd_request[hsd->hsd_request_count];
209 /* allocates hai vector size just needs to be large
211 request->hal_sz = sizeof(*request->hal) +
212 cfs_size_round(MTI_NAME_MAXLEN + 1) + 2 * hai_size;
213 OBD_ALLOC_LARGE(hal, request->hal_sz);
217 hal->hal_version = HAL_VERSION;
218 strlcpy(hal->hal_fsname, hsd->hsd_fsname, MTI_NAME_MAXLEN + 1);
219 hal->hal_archive_id = larr->arr_archive_id;
220 hal->hal_flags = larr->arr_flags;
222 request->hal_used_sz = hal_size(hal);
224 hsd->hsd_request_count++;
225 } else if (request->hal_sz < request->hal_used_sz + hai_size) {
226 /* Not enough room, need an extension */
230 sz = min_t(int, 2 * request->hal_sz, LDLM_MAXREQSIZE);
231 LASSERT(request->hal_used_sz + hai_size < sz);
233 OBD_ALLOC_LARGE(hal_buffer, sz);
237 memcpy(hal_buffer, request->hal, request->hal_used_sz);
238 OBD_FREE_LARGE(request->hal, request->hal_sz);
239 request->hal = hal_buffer;
240 request->hal_sz = sz;
243 hai = hai_first(request->hal);
244 for (i = 0; i < request->hal->hal_count; i++)
247 memcpy(hai, &larr->arr_hai, larr->arr_hai.hai_len);
249 request->hal_used_sz += cfs_size_round(hai->hai_len);
250 request->hal->hal_count++;
252 hsd->hsd_action_count++;
254 if (hai->hai_action != HSMA_CANCEL)
255 cdt_agent_record_hash_add(cdt, hai->hai_cookie,
256 llh->lgh_hdr->llh_cat_idx,
257 larr->arr_hdr.lrh_index);
262 static int mdt_cdt_started_cb(const struct lu_env *env,
263 struct mdt_device *mdt,
264 struct llog_handle *llh,
265 struct llog_agent_req_rec *larr,
266 struct hsm_scan_data *hsd)
268 struct coordinator *cdt = &mdt->mdt_coordinator;
269 struct hsm_action_item *hai = &larr->arr_hai;
270 struct cdt_agent_req *car;
271 time64_t now = ktime_get_real_seconds();
276 if (!hsd->hsd_housekeeping)
279 /* we search for a running request
280 * error may happen if coordinator crashes or stopped
281 * with running request
283 car = mdt_cdt_find_request(cdt, hai->hai_cookie);
285 last = larr->arr_req_change;
287 last = car->car_req_update;
290 /* test if request too long, if yes cancel it
291 * the same way the copy tool acknowledge a cancel request */
292 if (now <= last + cdt->cdt_active_req_timeout)
293 GOTO(out_car, rc = 0);
295 dump_llog_agent_req_rec("request timed out, start cleaning", larr);
298 car->car_req_update = now;
299 mdt_hsm_agent_update_statistics(cdt, 0, 1, 0, &car->car_uuid);
300 /* Remove car from memory list (LU-9075) */
301 mdt_cdt_remove_request(cdt, hai->hai_cookie);
304 /* Emit a changelog record for the failed action.*/
306 hsm_set_cl_error(&cl_flags, ECANCELED);
308 switch (hai->hai_action) {
310 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
313 hsm_set_cl_event(&cl_flags, HE_RESTORE);
316 hsm_set_cl_event(&cl_flags, HE_REMOVE);
319 hsm_set_cl_event(&cl_flags, HE_CANCEL);
322 /* Unknown record type, skip changelog. */
328 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
331 if (hai->hai_action == HSMA_RESTORE)
332 cdt_restore_handle_del(hsd->hsd_mti, cdt, &hai->hai_fid);
334 larr->arr_status = ARS_CANCELED;
335 larr->arr_req_change = now;
336 rc = llog_write(hsd->hsd_mti->mti_env, llh, &larr->arr_hdr,
337 larr->arr_hdr.lrh_index);
339 CERROR("%s: cannot update agent log: rc = %d\n",
340 mdt_obd_name(mdt), rc);
341 rc = LLOG_DEL_RECORD;
344 /* ct has completed a request, so a slot is available,
345 * signal the coordinator to find new work */
346 mdt_hsm_cdt_event(cdt);
349 mdt_cdt_put_request(car);
355 * llog_cat_process() callback, used to:
356 * - find waiting request and start action
357 * - purge canceled and done requests
358 * \param env [IN] environment
359 * \param llh [IN] llog handle
360 * \param hdr [IN] llog record
361 * \param data [IN/OUT] cb data = struct hsm_scan_data
363 * \retval -ve failure
365 static int mdt_coordinator_cb(const struct lu_env *env,
366 struct llog_handle *llh,
367 struct llog_rec_hdr *hdr,
370 struct llog_agent_req_rec *larr = (struct llog_agent_req_rec *)hdr;
371 struct hsm_scan_data *hsd = data;
372 struct mdt_device *mdt = hsd->hsd_mti->mti_mdt;
373 struct coordinator *cdt = &mdt->mdt_coordinator;
376 larr = (struct llog_agent_req_rec *)hdr;
377 dump_llog_agent_req_rec("mdt_coordinator_cb(): ", larr);
378 switch (larr->arr_status) {
380 RETURN(mdt_cdt_waiting_cb(env, mdt, llh, larr, hsd));
382 RETURN(mdt_cdt_started_cb(env, mdt, llh, larr, hsd));
384 if (!hsd->hsd_housekeeping)
387 if ((larr->arr_req_change + cdt->cdt_grace_delay) <
388 ktime_get_real_seconds()) {
389 cdt_agent_record_hash_del(cdt,
390 larr->arr_hai.hai_cookie);
391 RETURN(LLOG_DEL_RECORD);
399 * create /proc entries for coordinator
402 * \retval -ve failure
404 int hsm_cdt_procfs_init(struct mdt_device *mdt)
406 struct coordinator *cdt = &mdt->mdt_coordinator;
410 /* init /proc entries, failure is not critical */
411 cdt->cdt_proc_dir = lprocfs_register("hsm",
412 mdt2obd_dev(mdt)->obd_proc_entry,
413 lprocfs_mdt_hsm_vars, mdt);
414 if (IS_ERR(cdt->cdt_proc_dir)) {
415 rc = PTR_ERR(cdt->cdt_proc_dir);
416 CERROR("%s: Cannot create 'hsm' directory in mdt proc dir,"
417 " rc=%d\n", mdt_obd_name(mdt), rc);
418 cdt->cdt_proc_dir = NULL;
426 * remove /proc entries for coordinator
429 void hsm_cdt_procfs_fini(struct mdt_device *mdt)
431 struct coordinator *cdt = &mdt->mdt_coordinator;
433 if (cdt->cdt_proc_dir != NULL)
434 lprocfs_remove(&cdt->cdt_proc_dir);
438 * get vector of hsm cdt /proc vars
442 struct lprocfs_vars *hsm_cdt_get_proc_vars(void)
444 return lprocfs_mdt_hsm_vars;
447 /* Release the ressource used by the coordinator. Called when the
448 * coordinator is stopping. */
449 static void mdt_hsm_cdt_cleanup(struct mdt_device *mdt)
451 struct coordinator *cdt = &mdt->mdt_coordinator;
452 struct cdt_agent_req *car, *tmp1;
453 struct hsm_agent *ha, *tmp2;
454 struct cdt_restore_handle *crh, *tmp3;
455 struct mdt_thread_info *cdt_mti;
458 down_write(&cdt->cdt_request_lock);
459 list_for_each_entry_safe(car, tmp1, &cdt->cdt_request_list,
461 cfs_hash_del(cdt->cdt_request_cookie_hash,
462 &car->car_hai->hai_cookie,
463 &car->car_cookie_hash);
464 list_del(&car->car_request_list);
465 mdt_cdt_put_request(car);
467 up_write(&cdt->cdt_request_lock);
469 down_write(&cdt->cdt_agent_lock);
470 list_for_each_entry_safe(ha, tmp2, &cdt->cdt_agents, ha_list) {
471 list_del(&ha->ha_list);
474 up_write(&cdt->cdt_agent_lock);
476 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
477 mutex_lock(&cdt->cdt_restore_lock);
478 list_for_each_entry_safe(crh, tmp3, &cdt->cdt_restore_handle_list,
480 list_del(&crh->crh_list);
481 /* give back layout lock */
482 mdt_object_unlock(cdt_mti, NULL, &crh->crh_lh, 1);
483 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
485 mutex_unlock(&cdt->cdt_restore_lock);
489 * Coordinator state transition table, indexed on enum cdt_states, taking
490 * from and to states. For instance since CDT_INIT to CDT_RUNNING is a
491 * valid transition, cdt_transition[CDT_INIT][CDT_RUNNING] is true.
493 static bool cdt_transition[CDT_STATES_COUNT][CDT_STATES_COUNT] = {
494 /* from -> to: stopped init running disable stopping */
495 /* stopped */ { true, true, false, false, false },
496 /* init */ { true, false, true, false, false },
497 /* running */ { false, false, true, true, true },
498 /* disable */ { false, false, true, true, true },
499 /* stopping */ { true, false, false, false, false }
503 * Change coordinator thread state
504 * Some combinations are not valid, so catch them here.
506 * Returns 0 on success, with old_state set if not NULL, or -EINVAL if
507 * the transition was not possible.
509 static int set_cdt_state_locked(struct coordinator *cdt,
510 enum cdt_states new_state)
513 enum cdt_states state;
515 state = cdt->cdt_state;
517 if (cdt_transition[state][new_state]) {
518 cdt->cdt_state = new_state;
522 "unexpected coordinator transition, from=%s, to=%s\n",
523 cdt_mdt_state2str(state), cdt_mdt_state2str(new_state));
530 static int set_cdt_state(struct coordinator *cdt, enum cdt_states new_state)
534 mutex_lock(&cdt->cdt_state_lock);
535 rc = set_cdt_state_locked(cdt, new_state);
536 mutex_unlock(&cdt->cdt_state_lock);
545 * \param data [IN] obd device
547 * \retval -ve failure
549 static int mdt_coordinator(void *data)
551 struct mdt_thread_info *mti = data;
552 struct mdt_device *mdt = mti->mti_mdt;
553 struct coordinator *cdt = &mdt->mdt_coordinator;
554 struct hsm_scan_data hsd = { NULL };
555 time64_t last_housekeeping = 0;
556 size_t request_sz = 0;
560 CDEBUG(D_HSM, "%s: coordinator thread starting, pid=%d\n",
561 mdt_obd_name(mdt), current_pid());
564 obd_uuid2fsname(hsd.hsd_fsname, mdt_obd_name(mdt),
565 sizeof(hsd.hsd_fsname));
567 set_cdt_state(cdt, CDT_RUNNING);
569 /* Inform mdt_hsm_cdt_start(). */
570 wake_up_all(&cdt->cdt_waitq);
577 struct hsm_record_update *updates;
579 /* Limit execution of the expensive requests traversal
580 * to at most one second. This prevents repeatedly
581 * locking/unlocking the catalog for each request
582 * and preventing other HSM operations from happening
584 wait_event_interruptible_timeout(cdt->cdt_waitq,
585 kthread_should_stop() ||
586 cdt->cdt_wakeup_coordinator,
587 cfs_time_seconds(1));
589 cdt->cdt_wakeup_coordinator = false;
590 CDEBUG(D_HSM, "coordinator resumes\n");
592 if (kthread_should_stop()) {
593 CDEBUG(D_HSM, "Coordinator stops\n");
598 /* if coordinator is suspended continue to wait */
599 if (cdt->cdt_state == CDT_DISABLE) {
600 CDEBUG(D_HSM, "disable state, coordinator sleeps\n");
604 /* If no event, and no housekeeping to do, continue to
606 if (last_housekeeping + cdt->cdt_loop_period <=
607 ktime_get_real_seconds()) {
608 last_housekeeping = ktime_get_real_seconds();
609 hsd.hsd_housekeeping = true;
610 } else if (cdt->cdt_event) {
611 hsd.hsd_housekeeping = false;
616 cdt->cdt_event = false;
618 CDEBUG(D_HSM, "coordinator starts reading llog\n");
620 if (hsd.hsd_request_len != cdt->cdt_max_requests) {
621 /* cdt_max_requests has changed,
622 * we need to allocate a new buffer
624 struct hsm_scan_request *tmp = NULL;
625 int max_requests = cdt->cdt_max_requests;
626 OBD_ALLOC_LARGE(tmp, max_requests *
627 sizeof(struct hsm_scan_request));
629 CERROR("Failed to resize request buffer, "
630 "keeping it at %d\n",
631 hsd.hsd_request_len);
633 if (hsd.hsd_request != NULL)
634 OBD_FREE_LARGE(hsd.hsd_request,
637 hsd.hsd_request_len = max_requests;
638 request_sz = hsd.hsd_request_len *
639 sizeof(struct hsm_scan_request);
640 hsd.hsd_request = tmp;
644 hsd.hsd_action_count = 0;
645 hsd.hsd_request_count = 0;
647 rc = cdt_llog_process(mti->mti_env, mdt, mdt_coordinator_cb,
652 CDEBUG(D_HSM, "found %d requests to send\n",
653 hsd.hsd_request_count);
655 if (list_empty(&cdt->cdt_agents)) {
656 CDEBUG(D_HSM, "no agent available, "
657 "coordinator sleeps\n");
661 /* Compute how many HAI we have in all the requests */
663 for (i = 0; i < hsd.hsd_request_count; i++) {
664 const struct hsm_scan_request *request =
667 updates_cnt += request->hal->hal_count;
670 /* Allocate a temporary array to store the cookies to
671 * update, and their status. */
672 updates_sz = updates_cnt * sizeof(*updates);
673 OBD_ALLOC_LARGE(updates, updates_sz);
674 if (updates == NULL) {
675 CERROR("%s: Cannot allocate memory (%d o) "
677 mdt_obd_name(mdt), updates_sz, updates_cnt);
681 /* here hsd contains a list of requests to be started */
682 for (i = 0; i < hsd.hsd_request_count; i++) {
683 struct hsm_scan_request *request = &hsd.hsd_request[i];
684 struct hsm_action_list *hal = request->hal;
685 struct hsm_action_item *hai;
688 /* still room for work ? */
689 if (atomic_read(&cdt->cdt_request_count) >=
690 cdt->cdt_max_requests)
693 rc = mdt_hsm_agent_send(mti, hal, 0);
694 /* if failure, we suppose it is temporary
695 * if the copy tool failed to do the request
696 * it has to use hsm_progress
699 /* set up cookie vector to set records status
700 * after copy tools start or failed
702 hai = hai_first(hal);
703 for (j = 0; j < hal->hal_count; j++) {
704 updates[update_idx].cookie = hai->hai_cookie;
705 updates[update_idx].status =
706 (rc ? ARS_WAITING : ARS_STARTED);
713 rc = mdt_agent_record_update(mti->mti_env, mdt,
714 updates, update_idx);
716 CERROR("%s: mdt_agent_record_update() failed, "
717 "rc=%d, cannot update records "
719 mdt_obd_name(mdt), rc, update_idx);
722 OBD_FREE_LARGE(updates, updates_sz);
725 /* free hal allocated by callback */
726 for (i = 0; i < hsd.hsd_request_count; i++) {
727 struct hsm_scan_request *request = &hsd.hsd_request[i];
729 OBD_FREE_LARGE(request->hal, request->hal_sz);
733 if (hsd.hsd_request != NULL)
734 OBD_FREE_LARGE(hsd.hsd_request, request_sz);
736 mdt_hsm_cdt_cleanup(mdt);
739 CERROR("%s: coordinator thread exiting, process=%d, rc=%d\n",
740 mdt_obd_name(mdt), current_pid(), rc);
742 CDEBUG(D_HSM, "%s: coordinator thread exiting, process=%d,"
744 mdt_obd_name(mdt), current_pid());
749 int cdt_restore_handle_add(struct mdt_thread_info *mti, struct coordinator *cdt,
750 const struct lu_fid *fid,
751 const struct hsm_extent *he)
753 struct cdt_restore_handle *crh;
754 struct mdt_object *obj;
758 OBD_SLAB_ALLOC_PTR(crh, mdt_hsm_cdt_kmem);
763 /* in V1 all file is restored
764 * crh->extent.start = he->offset;
765 * crh->extent.end = he->offset + he->length;
767 crh->crh_extent.start = 0;
768 crh->crh_extent.end = he->length;
769 /* get the layout lock */
770 mdt_lock_reg_init(&crh->crh_lh, LCK_EX);
771 obj = mdt_object_find_lock(mti, &crh->crh_fid, &crh->crh_lh,
772 MDS_INODELOCK_LAYOUT);
774 GOTO(out_crh, rc = PTR_ERR(obj));
776 /* We do not keep a reference on the object during the restore
777 * which can be very long. */
778 mdt_object_put(mti->mti_env, obj);
780 mutex_lock(&cdt->cdt_restore_lock);
781 if (unlikely(cdt->cdt_state == CDT_STOPPED ||
782 cdt->cdt_state == CDT_STOPPING)) {
783 mutex_unlock(&cdt->cdt_restore_lock);
784 GOTO(out_lh, rc = -EAGAIN);
787 list_add_tail(&crh->crh_list, &cdt->cdt_restore_handle_list);
788 mutex_unlock(&cdt->cdt_restore_lock);
792 mdt_object_unlock(mti, NULL, &crh->crh_lh, 1);
794 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
800 * lookup a restore handle by FID
801 * caller needs to hold cdt_restore_lock
802 * \param cdt [IN] coordinator
803 * \param fid [IN] FID
804 * \retval cdt_restore_handle found
805 * \retval NULL not found
807 struct cdt_restore_handle *cdt_restore_handle_find(struct coordinator *cdt,
808 const struct lu_fid *fid)
810 struct cdt_restore_handle *crh;
813 list_for_each_entry(crh, &cdt->cdt_restore_handle_list, crh_list) {
814 if (lu_fid_eq(&crh->crh_fid, fid))
821 void cdt_restore_handle_del(struct mdt_thread_info *mti,
822 struct coordinator *cdt, const struct lu_fid *fid)
824 struct cdt_restore_handle *crh;
826 /* give back layout lock */
827 mutex_lock(&cdt->cdt_restore_lock);
828 crh = cdt_restore_handle_find(cdt, fid);
830 list_del(&crh->crh_list);
831 mutex_unlock(&cdt->cdt_restore_lock);
836 /* XXX We pass a NULL object since the restore handle does not
837 * keep a reference on the object being restored. */
838 mdt_object_unlock(mti, NULL, &crh->crh_lh, 1);
839 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
843 * data passed to llog_cat_process() callback
844 * to scan requests and take actions
846 struct hsm_restore_data {
847 struct mdt_thread_info *hrd_mti;
851 * llog_cat_process() callback, used to:
852 * - find restore request and allocate the restore handle
853 * \param env [IN] environment
854 * \param llh [IN] llog handle
855 * \param hdr [IN] llog record
856 * \param data [IN/OUT] cb data = struct hsm_restore_data
858 * \retval -ve failure
860 static int hsm_restore_cb(const struct lu_env *env,
861 struct llog_handle *llh,
862 struct llog_rec_hdr *hdr, void *data)
864 struct llog_agent_req_rec *larr;
865 struct hsm_restore_data *hrd;
866 struct hsm_action_item *hai;
867 struct mdt_thread_info *mti;
868 struct coordinator *cdt;
874 cdt = &mti->mti_mdt->mdt_coordinator;
876 larr = (struct llog_agent_req_rec *)hdr;
877 hai = &larr->arr_hai;
878 if (hai->hai_cookie > cdt->cdt_last_cookie)
879 /* update the cookie to avoid collision */
880 cdt->cdt_last_cookie = hai->hai_cookie + 1;
882 if (hai->hai_action != HSMA_RESTORE ||
883 agent_req_in_final_state(larr->arr_status))
886 /* restore request not in a final state */
888 /* force replay of restore requests left in started state from previous
889 * CDT context, to be canceled later if finally found to be incompatible
890 * when being re-started */
891 if (larr->arr_status == ARS_STARTED) {
892 larr->arr_status = ARS_WAITING;
893 larr->arr_req_change = ktime_get_real_seconds();
894 rc = llog_write(env, llh, hdr, hdr->lrh_index);
899 rc = cdt_restore_handle_add(mti, cdt, &hai->hai_fid, &hai->hai_extent);
905 * restore coordinator state at startup
906 * the goal is to take a layout lock for each registered restore request
907 * \param mti [IN] context
909 static int mdt_hsm_pending_restore(struct mdt_thread_info *mti)
911 struct hsm_restore_data hrd;
917 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, hsm_restore_cb, &hrd,
923 static int hsm_init_ucred(struct lu_ucred *uc)
927 uc->uc_valid = UCRED_OLD;
936 uc->uc_suppgids[0] = -1;
937 uc->uc_suppgids[1] = -1;
938 uc->uc_cap = CFS_CAP_FS_MASK;
941 uc->uc_identity = NULL;
942 /* always record internal HSM activity if also enabled globally */
943 uc->uc_enable_audit = 1;
949 * initialize coordinator struct
950 * \param mdt [IN] device
952 * \retval -ve failure
954 int mdt_hsm_cdt_init(struct mdt_device *mdt)
956 struct coordinator *cdt = &mdt->mdt_coordinator;
957 struct mdt_thread_info *cdt_mti = NULL;
961 init_waitqueue_head(&cdt->cdt_waitq);
962 init_rwsem(&cdt->cdt_llog_lock);
963 init_rwsem(&cdt->cdt_agent_lock);
964 init_rwsem(&cdt->cdt_request_lock);
965 mutex_init(&cdt->cdt_restore_lock);
966 mutex_init(&cdt->cdt_state_lock);
967 set_cdt_state(cdt, CDT_STOPPED);
969 INIT_LIST_HEAD(&cdt->cdt_request_list);
970 INIT_LIST_HEAD(&cdt->cdt_agents);
971 INIT_LIST_HEAD(&cdt->cdt_restore_handle_list);
973 cdt->cdt_request_cookie_hash = cfs_hash_create("REQUEST_COOKIE_HASH",
980 &cdt_request_cookie_hash_ops,
982 if (cdt->cdt_request_cookie_hash == NULL)
985 cdt->cdt_agent_record_hash = cfs_hash_create("AGENT_RECORD_HASH",
992 &cdt_agent_record_hash_ops,
994 if (cdt->cdt_agent_record_hash == NULL)
995 GOTO(out_request_cookie_hash, rc = -ENOMEM);
997 rc = lu_env_init(&cdt->cdt_env, LCT_MD_THREAD);
999 GOTO(out_agent_record_hash, rc);
1001 /* for mdt_ucred(), lu_ucred stored in lu_ucred_key */
1002 rc = lu_context_init(&cdt->cdt_session, LCT_SERVER_SESSION);
1006 lu_context_enter(&cdt->cdt_session);
1007 cdt->cdt_env.le_ses = &cdt->cdt_session;
1009 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
1010 LASSERT(cdt_mti != NULL);
1012 cdt_mti->mti_env = &cdt->cdt_env;
1013 cdt_mti->mti_mdt = mdt;
1015 hsm_init_ucred(mdt_ucred(cdt_mti));
1017 /* default values for /proc tunnables
1018 * can be override by MGS conf */
1019 cdt->cdt_default_archive_id = 1;
1020 cdt->cdt_grace_delay = 60;
1021 cdt->cdt_loop_period = 10;
1022 cdt->cdt_max_requests = 3;
1023 cdt->cdt_policy = CDT_DEFAULT_POLICY;
1024 cdt->cdt_active_req_timeout = 3600;
1026 /* by default do not remove archives on last unlink */
1027 cdt->cdt_remove_archive_on_last_unlink = false;
1032 lu_env_fini(&cdt->cdt_env);
1033 out_agent_record_hash:
1034 cfs_hash_putref(cdt->cdt_agent_record_hash);
1035 cdt->cdt_agent_record_hash = NULL;
1036 out_request_cookie_hash:
1037 cfs_hash_putref(cdt->cdt_request_cookie_hash);
1038 cdt->cdt_request_cookie_hash = NULL;
1044 * free a coordinator thread
1045 * \param mdt [IN] device
1047 int mdt_hsm_cdt_fini(struct mdt_device *mdt)
1049 struct coordinator *cdt = &mdt->mdt_coordinator;
1052 lu_context_exit(cdt->cdt_env.le_ses);
1053 lu_context_fini(cdt->cdt_env.le_ses);
1055 lu_env_fini(&cdt->cdt_env);
1057 cfs_hash_putref(cdt->cdt_agent_record_hash);
1058 cdt->cdt_agent_record_hash = NULL;
1060 cfs_hash_putref(cdt->cdt_request_cookie_hash);
1061 cdt->cdt_request_cookie_hash = NULL;
1067 * start a coordinator thread
1068 * \param mdt [IN] device
1070 * \retval -ve failure
1072 static int mdt_hsm_cdt_start(struct mdt_device *mdt)
1074 struct coordinator *cdt = &mdt->mdt_coordinator;
1075 struct mdt_thread_info *cdt_mti;
1078 struct task_struct *task;
1081 /* functions defined but not yet used
1082 * this avoid compilation warning
1084 ptr = dump_requests;
1086 rc = set_cdt_state(cdt, CDT_INIT);
1088 CERROR("%s: Coordinator already started or stopping\n",
1093 CLASSERT(1 << (CDT_POLICY_SHIFT_COUNT - 1) == CDT_POLICY_LAST);
1094 cdt->cdt_policy = CDT_DEFAULT_POLICY;
1096 /* just need to be larger than previous one */
1097 /* cdt_last_cookie is protected by cdt_llog_lock */
1098 cdt->cdt_last_cookie = ktime_get_real_seconds();
1099 atomic_set(&cdt->cdt_request_count, 0);
1100 atomic_set(&cdt->cdt_archive_count, 0);
1101 atomic_set(&cdt->cdt_restore_count, 0);
1102 atomic_set(&cdt->cdt_remove_count, 0);
1103 cdt->cdt_user_request_mask = (1UL << HSMA_RESTORE);
1104 cdt->cdt_group_request_mask = (1UL << HSMA_RESTORE);
1105 cdt->cdt_other_request_mask = (1UL << HSMA_RESTORE);
1107 /* to avoid deadlock when start is made through /proc
1108 * /proc entries are created by the coordinator thread */
1110 /* set up list of started restore requests */
1111 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
1112 rc = mdt_hsm_pending_restore(cdt_mti);
1114 CERROR("%s: cannot take the layout locks needed"
1115 " for registered restore: %d\n",
1116 mdt_obd_name(mdt), rc);
1118 if (mdt->mdt_bottom->dd_rdonly)
1121 task = kthread_run(mdt_coordinator, cdt_mti, "hsm_cdtr");
1124 set_cdt_state(cdt, CDT_STOPPED);
1125 CERROR("%s: error starting coordinator thread: %d\n",
1126 mdt_obd_name(mdt), rc);
1128 cdt->cdt_task = task;
1129 wait_event(cdt->cdt_waitq,
1130 cdt->cdt_state != CDT_INIT);
1131 CDEBUG(D_HSM, "%s: coordinator thread started\n",
1140 * stop a coordinator thread
1141 * \param mdt [IN] device
1143 int mdt_hsm_cdt_stop(struct mdt_device *mdt)
1145 struct coordinator *cdt = &mdt->mdt_coordinator;
1149 /* stop coordinator thread */
1150 rc = set_cdt_state(cdt, CDT_STOPPING);
1152 kthread_stop(cdt->cdt_task);
1153 cdt->cdt_task = NULL;
1154 set_cdt_state(cdt, CDT_STOPPED);
1160 static int mdt_hsm_set_exists(struct mdt_thread_info *mti,
1161 const struct lu_fid *fid,
1164 struct mdt_object *obj;
1168 obj = mdt_hsm_get_md_hsm(mti, fid, &mh);
1170 GOTO(out, rc = PTR_ERR(obj));
1172 if (mh.mh_flags & HS_EXISTS &&
1173 mh.mh_arch_id == archive_id)
1174 GOTO(out_obj, rc = 0);
1176 mh.mh_flags |= HS_EXISTS;
1177 mh.mh_arch_id = archive_id;
1178 rc = mdt_hsm_attr_set(mti, obj, &mh);
1181 mdt_object_put(mti->mti_env, obj);
1187 * register all requests from an hal in the memory list
1188 * \param mti [IN] context
1189 * \param hal [IN] request
1190 * \param uuid [OUT] in case of CANCEL, the uuid of the agent
1191 * which is running the CT
1193 * \retval -ve failure
1195 int mdt_hsm_add_hal(struct mdt_thread_info *mti,
1196 struct hsm_action_list *hal, struct obd_uuid *uuid)
1198 struct mdt_device *mdt = mti->mti_mdt;
1199 struct coordinator *cdt = &mdt->mdt_coordinator;
1200 struct hsm_action_item *hai;
1204 /* register request in memory list */
1205 hai = hai_first(hal);
1206 for (i = 0; i < hal->hal_count; i++, hai = hai_next(hai)) {
1207 struct cdt_agent_req *car;
1209 /* in case of a cancel request, we first mark the ondisk
1210 * record of the request we want to stop as canceled
1211 * this does not change the cancel record
1212 * it will be done when updating the request status
1214 if (hai->hai_action == HSMA_CANCEL) {
1215 struct hsm_record_update update = {
1216 .cookie = hai->hai_cookie,
1217 .status = ARS_CANCELED,
1220 rc = mdt_agent_record_update(mti->mti_env, mti->mti_mdt,
1223 CERROR("%s: mdt_agent_record_update() failed, "
1224 "rc=%d, cannot update status to %s "
1225 "for cookie %#llx\n",
1226 mdt_obd_name(mdt), rc,
1227 agent_req_status2name(ARS_CANCELED),
1232 /* find the running request to set it canceled */
1233 car = mdt_cdt_find_request(cdt, hai->hai_cookie);
1235 car->car_canceled = 1;
1236 /* uuid has to be changed to the one running the
1237 * request to cancel */
1238 *uuid = car->car_uuid;
1239 mdt_cdt_put_request(car);
1241 /* no need to memorize cancel request
1242 * this also avoid a deadlock when we receive
1243 * a purge all requests command
1248 if (hai->hai_action == HSMA_ARCHIVE) {
1249 rc = mdt_hsm_set_exists(mti, &hai->hai_fid,
1250 hal->hal_archive_id);
1257 car = mdt_cdt_alloc_request(hal->hal_archive_id, hal->hal_flags,
1260 GOTO(out, rc = PTR_ERR(car));
1262 rc = mdt_cdt_add_request(cdt, car);
1264 mdt_cdt_free_request(car);
1271 * swap layouts between 2 fids
1272 * \param mti [IN] context
1275 * \param mh_common [IN] MD HSM
1277 static int hsm_swap_layouts(struct mdt_thread_info *mti,
1278 struct mdt_object *obj, const struct lu_fid *dfid,
1279 struct md_hsm *mh_common)
1281 struct mdt_object *dobj;
1282 struct mdt_lock_handle *dlh;
1286 if (!mdt_object_exists(obj))
1287 GOTO(out, rc = -ENOENT);
1289 /* we already have layout lock on obj so take only
1291 dlh = &mti->mti_lh[MDT_LH_OLD];
1292 mdt_lock_reg_init(dlh, LCK_EX);
1293 dobj = mdt_object_find_lock(mti, dfid, dlh, MDS_INODELOCK_LAYOUT);
1295 GOTO(out, rc = PTR_ERR(dobj));
1297 /* if copy tool closes the volatile before sending the final
1298 * progress through llapi_hsm_copy_end(), all the objects
1299 * are removed and mdd_swap_layout LBUG */
1300 if (!mdt_object_exists(dobj)) {
1301 CERROR("%s: Copytool has closed volatile file "DFID"\n",
1302 mdt_obd_name(mti->mti_mdt), PFID(dfid));
1303 GOTO(out_dobj, rc = -ENOENT);
1305 /* Since we only handle restores here, unconditionally use
1306 * SWAP_LAYOUTS_MDS_HSM flag to ensure original layout will
1307 * be preserved in case of failure during swap_layout and not
1308 * leave a file in an intermediate but incoherent state.
1309 * But need to setup HSM xattr of data FID before, reuse
1310 * mti and mh presets for FID in hsm_cdt_request_completed(),
1311 * only need to clear RELEASED and DIRTY.
1313 mh_common->mh_flags &= ~(HS_RELEASED | HS_DIRTY);
1314 rc = mdt_hsm_attr_set(mti, dobj, mh_common);
1316 rc = mo_swap_layouts(mti->mti_env,
1317 mdt_object_child(obj),
1318 mdt_object_child(dobj),
1319 SWAP_LAYOUTS_MDS_HSM);
1322 mdt_object_unlock_put(mti, dobj, dlh, 1);
1328 * update status of a completed request
1329 * \param mti [IN] context
1330 * \param pgs [IN] progress of the copy tool
1332 * \retval -ve failure
1334 static int hsm_cdt_request_completed(struct mdt_thread_info *mti,
1335 struct hsm_progress_kernel *pgs,
1336 const struct cdt_agent_req *car,
1337 enum agent_req_status *status)
1339 const struct lu_env *env = mti->mti_env;
1340 struct mdt_device *mdt = mti->mti_mdt;
1341 struct coordinator *cdt = &mdt->mdt_coordinator;
1342 struct mdt_object *obj = NULL;
1343 int cl_flags = 0, rc = 0;
1346 bool need_changelog = true;
1349 /* default is to retry */
1350 *status = ARS_WAITING;
1352 /* find object by FID, mdt_hsm_get_md_hsm() returns obj or err
1353 * if error/removed continue anyway to get correct reporting done */
1354 obj = mdt_hsm_get_md_hsm(mti, &car->car_hai->hai_fid, &mh);
1355 /* we will update MD HSM only if needed */
1356 is_mh_changed = false;
1358 /* no need to change mh->mh_arch_id
1359 * mdt_hsm_get_md_hsm() got it from disk and it is still valid
1361 if (pgs->hpk_errval != 0) {
1362 switch (pgs->hpk_errval) {
1364 /* the copy tool does not support cancel
1365 * so the cancel request is failed
1366 * As we cannot distinguish a cancel progress
1367 * from another action progress (they have the
1368 * same cookie), we suppose here the CT returns
1369 * ENOSYS only if does not support cancel
1371 /* this can also happen when cdt calls it to
1372 * for a timed out request */
1373 *status = ARS_FAILED;
1374 /* to have a cancel event in changelog */
1375 pgs->hpk_errval = ECANCELED;
1378 /* the request record has already been set to
1379 * ARS_CANCELED, this set the cancel request
1381 *status = ARS_SUCCEED;
1384 /* retry only if current policy or requested, and
1385 * object is not on error/removed */
1386 *status = (cdt->cdt_policy & CDT_NORETRY_ACTION ||
1387 !(pgs->hpk_flags & HP_FLAG_RETRY) ||
1388 IS_ERR(obj)) ? ARS_FAILED : ARS_WAITING;
1392 if (pgs->hpk_errval > CLF_HSM_MAXERROR) {
1393 CERROR("%s: Request %#llx on "DFID
1394 " failed, error code %d too large\n",
1396 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1398 hsm_set_cl_error(&cl_flags,
1399 CLF_HSM_ERROVERFLOW);
1402 hsm_set_cl_error(&cl_flags, pgs->hpk_errval);
1405 switch (car->car_hai->hai_action) {
1407 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1410 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1413 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1416 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1417 CERROR("%s: Failed request %#llx on "DFID
1418 " cannot be a CANCEL\n",
1421 PFID(&pgs->hpk_fid));
1424 CERROR("%s: Failed request %#llx on "DFID
1425 " %d is an unknown action\n",
1427 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1428 car->car_hai->hai_action);
1433 *status = ARS_SUCCEED;
1434 switch (car->car_hai->hai_action) {
1436 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1437 /* set ARCHIVE keep EXIST and clear LOST and
1439 mh.mh_arch_ver = pgs->hpk_data_version;
1440 mh.mh_flags |= HS_ARCHIVED;
1441 mh.mh_flags &= ~(HS_LOST|HS_DIRTY);
1442 is_mh_changed = true;
1445 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1447 /* do not clear RELEASED and DIRTY here
1448 * this will occur in hsm_swap_layouts()
1451 /* Restoring has changed the file version on
1453 mh.mh_arch_ver = pgs->hpk_data_version;
1454 is_mh_changed = true;
1457 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1458 /* clear ARCHIVED EXISTS and LOST */
1459 mh.mh_flags &= ~(HS_ARCHIVED | HS_EXISTS | HS_LOST);
1460 is_mh_changed = true;
1463 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1464 CERROR("%s: Successful request %#llx on "DFID" cannot be a CANCEL\n",
1467 PFID(&pgs->hpk_fid));
1470 CERROR("%s: Successful request %#llx on "DFID" %d is an unknown action\n",
1472 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1473 car->car_hai->hai_action);
1479 /* rc != 0 means error when analysing action, it may come from
1480 * a crasy CT no need to manage DIRTY
1481 * and if mdt_hsm_get_md_hsm() has returned an error, mh has not been
1484 if (rc == 0 && !IS_ERR(obj))
1485 hsm_set_cl_flags(&cl_flags,
1486 mh.mh_flags & HS_DIRTY ? CLF_HSM_DIRTY : 0);
1488 /* unlock is done later, after layout lock management */
1489 if (is_mh_changed && !IS_ERR(obj))
1490 rc = mdt_hsm_attr_set(mti, obj, &mh);
1492 /* we give back layout lock only if restore was successful or
1493 * if no retry will be attempted and if object is still alive,
1494 * in other cases we just unlock the object */
1495 if (car->car_hai->hai_action == HSMA_RESTORE) {
1496 /* restore in data FID done, we swap the layouts
1497 * only if restore is successful */
1498 if (pgs->hpk_errval == 0 && !IS_ERR(obj)) {
1499 rc = hsm_swap_layouts(mti, obj, &car->car_hai->hai_dfid,
1502 if (cdt->cdt_policy & CDT_NORETRY_ACTION)
1503 *status = ARS_FAILED;
1504 pgs->hpk_errval = -rc;
1507 /* we have to retry, so keep layout lock */
1508 if (*status == ARS_WAITING)
1511 /* restore special case, need to create ChangeLog record
1512 * before to give back layout lock to avoid concurrent
1513 * file updater to post out of order ChangeLog */
1514 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1515 &car->car_hai->hai_fid);
1516 need_changelog = false;
1518 cdt_restore_handle_del(mti, cdt, &car->car_hai->hai_fid);
1524 /* always add a ChangeLog record */
1526 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1527 &car->car_hai->hai_fid);
1530 mdt_object_put(mti->mti_env, obj);
1536 * update status of a request
1537 * \param mti [IN] context
1538 * \param pgs [IN] progress of the copy tool
1540 * \retval -ve failure
1542 int mdt_hsm_update_request_state(struct mdt_thread_info *mti,
1543 struct hsm_progress_kernel *pgs)
1545 struct mdt_device *mdt = mti->mti_mdt;
1546 struct coordinator *cdt = &mdt->mdt_coordinator;
1547 struct cdt_agent_req *car;
1551 /* no coordinator started, so we cannot serve requests */
1552 if (cdt->cdt_state == CDT_STOPPED)
1555 /* first do sanity checks */
1556 car = mdt_cdt_update_request(cdt, pgs);
1558 CERROR("%s: Cannot find running request for cookie %#llx"
1561 pgs->hpk_cookie, PFID(&pgs->hpk_fid));
1563 RETURN(PTR_ERR(car));
1566 CDEBUG(D_HSM, "Progress received for fid="DFID" cookie=%#llx"
1567 " action=%s flags=%d err=%d fid="DFID" dfid="DFID"\n",
1568 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1569 hsm_copytool_action2name(car->car_hai->hai_action),
1570 pgs->hpk_flags, pgs->hpk_errval,
1571 PFID(&car->car_hai->hai_fid),
1572 PFID(&car->car_hai->hai_dfid));
1574 /* progress is done on FID or data FID depending of the action and
1575 * of the copy progress */
1576 /* for restore progress is used to send back the data FID to cdt */
1577 if (car->car_hai->hai_action == HSMA_RESTORE &&
1578 lu_fid_eq(&car->car_hai->hai_fid, &car->car_hai->hai_dfid))
1579 car->car_hai->hai_dfid = pgs->hpk_fid;
1581 if ((car->car_hai->hai_action == HSMA_RESTORE ||
1582 car->car_hai->hai_action == HSMA_ARCHIVE) &&
1583 (!lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_dfid) &&
1584 !lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_fid))) {
1585 CERROR("%s: Progress on "DFID" for cookie %#llx"
1586 " does not match request FID "DFID" nor data FID "
1589 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1590 PFID(&car->car_hai->hai_fid),
1591 PFID(&car->car_hai->hai_dfid));
1592 GOTO(out, rc = -EINVAL);
1595 if (pgs->hpk_errval != 0 && !(pgs->hpk_flags & HP_FLAG_COMPLETED)) {
1596 CERROR("%s: Progress on "DFID" for cookie %#llx action=%s"
1597 " is not coherent (err=%d and not completed"
1600 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1601 hsm_copytool_action2name(car->car_hai->hai_action),
1602 pgs->hpk_errval, pgs->hpk_flags);
1603 GOTO(out, rc = -EINVAL);
1606 /* now progress is valid */
1608 /* we use a root like ucred */
1609 hsm_init_ucred(mdt_ucred(mti));
1611 if (pgs->hpk_flags & HP_FLAG_COMPLETED) {
1612 enum agent_req_status status;
1613 struct hsm_record_update update;
1616 rc = hsm_cdt_request_completed(mti, pgs, car, &status);
1618 CDEBUG(D_HSM, "updating record: fid="DFID" cookie=%#llx action=%s "
1620 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1621 hsm_copytool_action2name(car->car_hai->hai_action),
1622 agent_req_status2name(status));
1624 /* update record first (LU-9075) */
1625 update.cookie = pgs->hpk_cookie;
1626 update.status = status;
1628 rc1 = mdt_agent_record_update(mti->mti_env, mdt,
1631 CERROR("%s: mdt_agent_record_update() failed,"
1632 " rc=%d, cannot update status to %s"
1633 " for cookie %#llx\n",
1634 mdt_obd_name(mdt), rc1,
1635 agent_req_status2name(status),
1637 rc = (rc != 0 ? rc : rc1);
1639 /* then remove request from memory list (LU-9075) */
1640 mdt_cdt_remove_request(cdt, pgs->hpk_cookie);
1642 /* ct has completed a request, so a slot is available,
1643 * signal the coordinator to find new work */
1644 mdt_hsm_cdt_event(cdt);
1646 /* if copytool send a progress on a canceled request
1647 * we inform copytool it should stop
1649 if (car->car_canceled == 1)
1655 /* remove ref got from mdt_cdt_update_request() */
1656 mdt_cdt_put_request(car);
1663 * data passed to llog_cat_process() callback
1664 * to cancel requests
1666 struct hsm_cancel_all_data {
1667 struct mdt_device *mdt;
1671 * llog_cat_process() callback, used to:
1672 * - purge all requests
1673 * \param env [IN] environment
1674 * \param llh [IN] llog handle
1675 * \param hdr [IN] llog record
1676 * \param data [IN] cb data = struct hsm_cancel_all_data
1678 * \retval -ve failure
1680 static int mdt_cancel_all_cb(const struct lu_env *env,
1681 struct llog_handle *llh,
1682 struct llog_rec_hdr *hdr, void *data)
1684 struct llog_agent_req_rec *larr;
1685 struct hsm_cancel_all_data *hcad;
1689 larr = (struct llog_agent_req_rec *)hdr;
1691 if (larr->arr_status == ARS_WAITING ||
1692 larr->arr_status == ARS_STARTED) {
1693 larr->arr_status = ARS_CANCELED;
1694 larr->arr_req_change = ktime_get_real_seconds();
1695 rc = llog_write(env, llh, hdr, hdr->lrh_index);
1702 * cancel all actions
1703 * \param obd [IN] MDT device
1705 static int hsm_cancel_all_actions(struct mdt_device *mdt)
1708 struct lu_context session;
1709 struct mdt_thread_info *mti;
1710 struct coordinator *cdt = &mdt->mdt_coordinator;
1711 struct cdt_agent_req *car;
1712 struct hsm_action_list *hal = NULL;
1713 struct hsm_action_item *hai;
1714 struct hsm_cancel_all_data hcad;
1715 int hal_sz = 0, hal_len, rc;
1716 enum cdt_states old_state;
1719 rc = lu_env_init(&env, LCT_MD_THREAD);
1723 /* for mdt_ucred(), lu_ucred stored in lu_ucred_key */
1724 rc = lu_context_init(&session, LCT_SERVER_SESSION);
1728 lu_context_enter(&session);
1729 env.le_ses = &session;
1731 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
1732 LASSERT(mti != NULL);
1734 mti->mti_env = &env;
1737 hsm_init_ucred(mdt_ucred(mti));
1739 mutex_lock(&cdt->cdt_state_lock);
1740 old_state = cdt->cdt_state;
1742 /* disable coordinator */
1743 rc = set_cdt_state_locked(cdt, CDT_DISABLE);
1745 GOTO(out_cdt_state_unlock, rc);
1747 /* send cancel to all running requests */
1748 down_read(&cdt->cdt_request_lock);
1749 list_for_each_entry(car, &cdt->cdt_request_list, car_request_list) {
1750 mdt_cdt_get_request(car);
1751 /* request is not yet removed from list, it will be done
1752 * when copytool will return progress
1755 if (car->car_hai->hai_action == HSMA_CANCEL) {
1756 mdt_cdt_put_request(car);
1761 hal_len = sizeof(*hal) + cfs_size_round(MTI_NAME_MAXLEN + 1) +
1762 cfs_size_round(car->car_hai->hai_len);
1764 if (hal_len > hal_sz && hal_sz > 0) {
1765 /* not enough room, free old buffer */
1766 OBD_FREE(hal, hal_sz);
1770 /* empty buffer, allocate one */
1773 OBD_ALLOC(hal, hal_sz);
1775 mdt_cdt_put_request(car);
1776 up_read(&cdt->cdt_request_lock);
1777 GOTO(out_cdt_state, rc = -ENOMEM);
1781 hal->hal_version = HAL_VERSION;
1782 obd_uuid2fsname(hal->hal_fsname, mdt_obd_name(mdt),
1784 hal->hal_fsname[MTI_NAME_MAXLEN] = '\0';
1785 hal->hal_archive_id = car->car_archive_id;
1786 hal->hal_flags = car->car_flags;
1789 hai = hai_first(hal);
1790 memcpy(hai, car->car_hai, car->car_hai->hai_len);
1791 hai->hai_action = HSMA_CANCEL;
1794 /* it is possible to safely call mdt_hsm_agent_send()
1795 * (ie without a deadlock on cdt_request_lock), because the
1796 * write lock is taken only if we are not in purge mode
1797 * (mdt_hsm_agent_send() does not call mdt_cdt_add_request()
1798 * nor mdt_cdt_remove_request())
1800 /* no conflict with cdt thread because cdt is disable and we
1801 * have the request lock */
1802 mdt_hsm_agent_send(mti, hal, 1);
1804 mdt_cdt_put_request(car);
1806 up_read(&cdt->cdt_request_lock);
1809 OBD_FREE(hal, hal_sz);
1811 /* cancel all on-disk records */
1814 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, mdt_cancel_all_cb,
1815 &hcad, 0, 0, WRITE);
1817 /* Enable coordinator, unless the coordinator was stopping. */
1818 set_cdt_state_locked(cdt, old_state);
1819 out_cdt_state_unlock:
1820 mutex_unlock(&cdt->cdt_state_lock);
1822 lu_context_exit(&session);
1823 lu_context_fini(&session);
1831 * check if a request is compatible with file status
1832 * \param hai [IN] request description
1833 * \param archive_id [IN] request archive id
1834 * \param rq_flags [IN] request flags
1835 * \param hsm [IN] file HSM metadata
1838 bool mdt_hsm_is_action_compat(const struct hsm_action_item *hai,
1839 u32 archive_id, u64 rq_flags,
1840 const struct md_hsm *hsm)
1842 int is_compat = false;
1846 hsm_flags = hsm->mh_flags;
1847 switch (hai->hai_action) {
1849 if (!(hsm_flags & HS_NOARCHIVE) &&
1850 (hsm_flags & HS_DIRTY || !(hsm_flags & HS_ARCHIVED)))
1853 if (hsm_flags & HS_EXISTS &&
1855 archive_id != hsm->mh_arch_id)
1860 if (!(hsm_flags & HS_DIRTY) && (hsm_flags & HS_RELEASED) &&
1861 hsm_flags & HS_ARCHIVED && !(hsm_flags & HS_LOST))
1865 if (!(hsm_flags & HS_RELEASED) &&
1866 (hsm_flags & (HS_ARCHIVED | HS_EXISTS)))
1873 CDEBUG(D_HSM, "fid="DFID" action=%s flags=%#llx"
1874 " extent=%#llx-%#llx hsm_flags=%.8X %s\n",
1875 PFID(&hai->hai_fid),
1876 hsm_copytool_action2name(hai->hai_action), rq_flags,
1877 hai->hai_extent.offset, hai->hai_extent.length,
1879 (is_compat ? "compatible" : "uncompatible"));
1885 * /proc interface used to get/set HSM behaviour (cdt->cdt_policy)
1887 static const struct {
1891 } hsm_policy_names[] = {
1892 { CDT_NONBLOCKING_RESTORE, "NonBlockingRestore", "NBR"},
1893 { CDT_NORETRY_ACTION, "NoRetryAction", "NRA"},
1898 * convert a policy name to a bit
1899 * \param name [IN] policy name
1901 * \retval policy bit
1903 static __u64 hsm_policy_str2bit(const char *name)
1907 for (i = 0; hsm_policy_names[i].bit != 0; i++)
1908 if (strcmp(hsm_policy_names[i].nickname, name) == 0 ||
1909 strcmp(hsm_policy_names[i].name, name) == 0)
1910 return hsm_policy_names[i].bit;
1915 * convert a policy bit field to a string
1916 * \param mask [IN] policy bit field
1917 * \param hexa [IN] print mask before bit names
1918 * \param buffer [OUT] string
1919 * \param count [IN] size of buffer
1921 static void hsm_policy_bit2str(struct seq_file *m, const __u64 mask,
1929 seq_printf(m, "(%#llx) ", mask);
1931 for (i = 0; i < CDT_POLICY_SHIFT_COUNT; i++) {
1934 for (j = 0; hsm_policy_names[j].bit != 0; j++) {
1935 if (hsm_policy_names[j].bit == bit)
1939 seq_printf(m, "[%s] ", hsm_policy_names[j].name);
1941 seq_printf(m, "%s ", hsm_policy_names[j].name);
1943 /* remove last ' ' */
1948 /* methods to read/write HSM policy flags */
1949 static int mdt_hsm_policy_seq_show(struct seq_file *m, void *data)
1951 struct mdt_device *mdt = m->private;
1952 struct coordinator *cdt = &mdt->mdt_coordinator;
1955 hsm_policy_bit2str(m, cdt->cdt_policy, false);
1960 mdt_hsm_policy_seq_write(struct file *file, const char __user *buffer,
1961 size_t count, loff_t *off)
1963 struct seq_file *m = file->private_data;
1964 struct mdt_device *mdt = m->private;
1965 struct coordinator *cdt = &mdt->mdt_coordinator;
1966 char *start, *token, sign;
1969 __u64 add_mask, remove_mask, set_mask;
1973 if (count + 1 > PAGE_SIZE)
1976 OBD_ALLOC(buf, count + 1);
1980 if (copy_from_user(buf, buffer, count))
1981 GOTO(out, rc = -EFAULT);
1986 CDEBUG(D_HSM, "%s: receive new policy: '%s'\n", mdt_obd_name(mdt),
1989 add_mask = remove_mask = set_mask = 0;
1991 token = strsep(&start, "\n ");
1997 if (sign == '-' || sign == '+')
2000 policy = hsm_policy_str2bit(token);
2002 CWARN("%s: '%s' is unknown, "
2003 "supported policies are:\n", mdt_obd_name(mdt),
2005 hsm_policy_bit2str(m, 0, false);
2006 GOTO(out, rc = -EINVAL);
2010 remove_mask |= policy;
2020 } while (start != NULL);
2022 CDEBUG(D_HSM, "%s: new policy: rm=%#llx add=%#llx set=%#llx\n",
2023 mdt_obd_name(mdt), remove_mask, add_mask, set_mask);
2025 /* if no sign in all string, it is a clear and set
2026 * if some sign found, all unsigned are converted
2028 * P1 P2 = set to P1 and P2
2029 * P1 -P2 = add P1 clear P2 same as +P1 -P2
2031 if (remove_mask == 0 && add_mask == 0) {
2032 cdt->cdt_policy = set_mask;
2034 cdt->cdt_policy |= set_mask | add_mask;
2035 cdt->cdt_policy &= ~remove_mask;
2038 GOTO(out, rc = count);
2041 OBD_FREE(buf, count + 1);
2044 LPROC_SEQ_FOPS(mdt_hsm_policy);
2046 #define GENERATE_PROC_METHOD(VAR) \
2047 static int mdt_hsm_##VAR##_seq_show(struct seq_file *m, void *data) \
2049 struct mdt_device *mdt = m->private; \
2050 struct coordinator *cdt = &mdt->mdt_coordinator; \
2053 seq_printf(m, "%llu\n", (__u64)cdt->VAR); \
2057 mdt_hsm_##VAR##_seq_write(struct file *file, const char __user *buffer, \
2058 size_t count, loff_t *off) \
2061 struct seq_file *m = file->private_data; \
2062 struct mdt_device *mdt = m->private; \
2063 struct coordinator *cdt = &mdt->mdt_coordinator; \
2068 rc = kstrtouint_from_user(buffer, count, 0, &val); \
2079 GENERATE_PROC_METHOD(cdt_loop_period)
2080 GENERATE_PROC_METHOD(cdt_grace_delay)
2081 GENERATE_PROC_METHOD(cdt_active_req_timeout)
2082 GENERATE_PROC_METHOD(cdt_max_requests)
2083 GENERATE_PROC_METHOD(cdt_default_archive_id)
2086 * procfs write method for MDT/hsm_control
2087 * proc entry is in mdt directory so data is mdt obd_device pointer
2089 #define CDT_ENABLE_CMD "enabled"
2090 #define CDT_STOP_CMD "shutdown"
2091 #define CDT_DISABLE_CMD "disabled"
2092 #define CDT_PURGE_CMD "purge"
2093 #define CDT_HELP_CMD "help"
2094 #define CDT_MAX_CMD_LEN 10
2097 mdt_hsm_cdt_control_seq_write(struct file *file, const char __user *buffer,
2098 size_t count, loff_t *off)
2100 struct seq_file *m = file->private_data;
2101 struct obd_device *obd = m->private;
2102 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2103 struct coordinator *cdt = &(mdt->mdt_coordinator);
2105 char kernbuf[CDT_MAX_CMD_LEN];
2108 if (count == 0 || count >= sizeof(kernbuf))
2111 if (copy_from_user(kernbuf, buffer, count))
2115 if (kernbuf[count - 1] == '\n')
2116 kernbuf[count - 1] = 0;
2119 if (strcmp(kernbuf, CDT_ENABLE_CMD) == 0) {
2120 if (cdt->cdt_state == CDT_DISABLE) {
2121 rc = set_cdt_state(cdt, CDT_RUNNING);
2122 mdt_hsm_cdt_event(cdt);
2123 wake_up(&cdt->cdt_waitq);
2125 rc = mdt_hsm_cdt_start(mdt);
2127 } else if (strcmp(kernbuf, CDT_STOP_CMD) == 0) {
2128 if ((cdt->cdt_state == CDT_STOPPING) ||
2129 (cdt->cdt_state == CDT_STOPPED)) {
2130 CERROR("%s: Coordinator already stopped\n",
2134 rc = mdt_hsm_cdt_stop(mdt);
2136 } else if (strcmp(kernbuf, CDT_DISABLE_CMD) == 0) {
2137 if ((cdt->cdt_state == CDT_STOPPING) ||
2138 (cdt->cdt_state == CDT_STOPPED)) {
2139 CERROR("%s: Coordinator is stopped\n",
2143 rc = set_cdt_state(cdt, CDT_DISABLE);
2145 } else if (strcmp(kernbuf, CDT_PURGE_CMD) == 0) {
2146 rc = hsm_cancel_all_actions(mdt);
2147 } else if (strcmp(kernbuf, CDT_HELP_CMD) == 0) {
2155 CERROR("%s: Valid coordinator control commands are: "
2156 "%s %s %s %s %s\n", mdt_obd_name(mdt),
2157 CDT_ENABLE_CMD, CDT_STOP_CMD, CDT_DISABLE_CMD,
2158 CDT_PURGE_CMD, CDT_HELP_CMD);
2166 int mdt_hsm_cdt_control_seq_show(struct seq_file *m, void *data)
2168 struct obd_device *obd = m->private;
2169 struct coordinator *cdt;
2172 cdt = &(mdt_dev(obd->obd_lu_dev)->mdt_coordinator);
2174 seq_printf(m, "%s\n", cdt_mdt_state2str(cdt->cdt_state));
2180 mdt_hsm_request_mask_show(struct seq_file *m, __u64 mask)
2186 for (i = 0; i < 8 * sizeof(mask); i++) {
2187 if (mask & (1UL << i)) {
2188 seq_printf(m, "%s%s", first ? "" : " ",
2189 hsm_copytool_action2name(i));
2199 mdt_hsm_user_request_mask_seq_show(struct seq_file *m, void *data)
2201 struct mdt_device *mdt = m->private;
2202 struct coordinator *cdt = &mdt->mdt_coordinator;
2204 return mdt_hsm_request_mask_show(m, cdt->cdt_user_request_mask);
2208 mdt_hsm_group_request_mask_seq_show(struct seq_file *m, void *data)
2210 struct mdt_device *mdt = m->private;
2211 struct coordinator *cdt = &mdt->mdt_coordinator;
2213 return mdt_hsm_request_mask_show(m, cdt->cdt_group_request_mask);
2217 mdt_hsm_other_request_mask_seq_show(struct seq_file *m, void *data)
2219 struct mdt_device *mdt = m->private;
2220 struct coordinator *cdt = &mdt->mdt_coordinator;
2222 return mdt_hsm_request_mask_show(m, cdt->cdt_other_request_mask);
2225 static inline enum hsm_copytool_action
2226 hsm_copytool_name2action(const char *name)
2228 if (strcasecmp(name, "NOOP") == 0)
2230 else if (strcasecmp(name, "ARCHIVE") == 0)
2231 return HSMA_ARCHIVE;
2232 else if (strcasecmp(name, "RESTORE") == 0)
2233 return HSMA_RESTORE;
2234 else if (strcasecmp(name, "REMOVE") == 0)
2236 else if (strcasecmp(name, "CANCEL") == 0)
2243 mdt_write_hsm_request_mask(struct file *file, const char __user *user_buf,
2244 size_t user_count, __u64 *mask)
2246 char *buf, *pos, *name;
2252 if (!(user_count < 4096))
2255 buf_size = user_count + 1;
2257 OBD_ALLOC(buf, buf_size);
2261 if (copy_from_user(buf, user_buf, buf_size - 1))
2262 GOTO(out, rc = -EFAULT);
2264 buf[buf_size - 1] = '\0';
2267 while ((name = strsep(&pos, " \t\v\n")) != NULL) {
2273 action = hsm_copytool_name2action(name);
2275 GOTO(out, rc = -EINVAL);
2277 new_mask |= (1UL << action);
2283 OBD_FREE(buf, buf_size);
2289 mdt_hsm_user_request_mask_seq_write(struct file *file, const char __user *buf,
2290 size_t count, loff_t *off)
2292 struct seq_file *m = file->private_data;
2293 struct mdt_device *mdt = m->private;
2294 struct coordinator *cdt = &mdt->mdt_coordinator;
2296 return mdt_write_hsm_request_mask(file, buf, count,
2297 &cdt->cdt_user_request_mask);
2301 mdt_hsm_group_request_mask_seq_write(struct file *file, const char __user *buf,
2302 size_t count, loff_t *off)
2304 struct seq_file *m = file->private_data;
2305 struct mdt_device *mdt = m->private;
2306 struct coordinator *cdt = &mdt->mdt_coordinator;
2308 return mdt_write_hsm_request_mask(file, buf, count,
2309 &cdt->cdt_group_request_mask);
2313 mdt_hsm_other_request_mask_seq_write(struct file *file, const char __user *buf,
2314 size_t count, loff_t *off)
2316 struct seq_file *m = file->private_data;
2317 struct mdt_device *mdt = m->private;
2318 struct coordinator *cdt = &mdt->mdt_coordinator;
2320 return mdt_write_hsm_request_mask(file, buf, count,
2321 &cdt->cdt_other_request_mask);
2324 static int mdt_hsm_cdt_raolu_seq_show(struct seq_file *m, void *data)
2326 struct mdt_device *mdt = m->private;
2327 struct coordinator *cdt = &mdt->mdt_coordinator;
2330 seq_printf(m, "%d\n", (int)cdt->cdt_remove_archive_on_last_unlink);
2335 mdt_hsm_cdt_raolu_seq_write(struct file *file, const char __user *buffer,
2336 size_t count, loff_t *off)
2339 struct seq_file *m = file->private_data;
2340 struct mdt_device *mdt = m->private;
2341 struct coordinator *cdt = &mdt->mdt_coordinator;
2346 rc = kstrtobool_from_user(buffer, count, &val);
2350 cdt->cdt_remove_archive_on_last_unlink = val;
2354 LPROC_SEQ_FOPS(mdt_hsm_cdt_loop_period);
2355 LPROC_SEQ_FOPS(mdt_hsm_cdt_grace_delay);
2356 LPROC_SEQ_FOPS(mdt_hsm_cdt_active_req_timeout);
2357 LPROC_SEQ_FOPS(mdt_hsm_cdt_max_requests);
2358 LPROC_SEQ_FOPS(mdt_hsm_cdt_default_archive_id);
2359 LPROC_SEQ_FOPS(mdt_hsm_user_request_mask);
2360 LPROC_SEQ_FOPS(mdt_hsm_group_request_mask);
2361 LPROC_SEQ_FOPS(mdt_hsm_other_request_mask);
2362 LPROC_SEQ_FOPS(mdt_hsm_cdt_raolu);
2364 /* Read-only proc files for request counters */
2365 static int mdt_hsm_cdt_archive_count_seq_show(struct seq_file *m, void *data)
2367 struct mdt_device *mdt = m->private;
2368 struct coordinator *cdt = &mdt->mdt_coordinator;
2371 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_archive_count));
2375 static int mdt_hsm_cdt_restore_count_seq_show(struct seq_file *m, void *data)
2377 struct mdt_device *mdt = m->private;
2378 struct coordinator *cdt = &mdt->mdt_coordinator;
2381 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_restore_count));
2385 static int mdt_hsm_cdt_remove_count_seq_show(struct seq_file *m, void *data)
2387 struct mdt_device *mdt = m->private;
2388 struct coordinator *cdt = &mdt->mdt_coordinator;
2391 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_remove_count));
2395 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_archive_count);
2396 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_restore_count);
2397 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_remove_count);
2399 static struct lprocfs_vars lprocfs_mdt_hsm_vars[] = {
2401 .fops = &mdt_hsm_agent_fops },
2402 { .name = "actions",
2403 .fops = &mdt_hsm_actions_fops,
2404 .proc_mode = 0444 },
2405 { .name = "default_archive_id",
2406 .fops = &mdt_hsm_cdt_default_archive_id_fops },
2407 { .name = "grace_delay",
2408 .fops = &mdt_hsm_cdt_grace_delay_fops },
2409 { .name = "loop_period",
2410 .fops = &mdt_hsm_cdt_loop_period_fops },
2411 { .name = "max_requests",
2412 .fops = &mdt_hsm_cdt_max_requests_fops },
2414 .fops = &mdt_hsm_policy_fops },
2415 { .name = "active_request_timeout",
2416 .fops = &mdt_hsm_cdt_active_req_timeout_fops },
2417 { .name = "active_requests",
2418 .fops = &mdt_hsm_active_requests_fops },
2419 { .name = "user_request_mask",
2420 .fops = &mdt_hsm_user_request_mask_fops, },
2421 { .name = "group_request_mask",
2422 .fops = &mdt_hsm_group_request_mask_fops, },
2423 { .name = "other_request_mask",
2424 .fops = &mdt_hsm_other_request_mask_fops, },
2425 { .name = "remove_archive_on_last_unlink",
2426 .fops = &mdt_hsm_cdt_raolu_fops, },
2427 { .name = "archive_count",
2428 .fops = &mdt_hsm_cdt_archive_count_fops, },
2429 { .name = "restore_count",
2430 .fops = &mdt_hsm_cdt_restore_count_fops, },
2431 { .name = "remove_count",
2432 .fops = &mdt_hsm_cdt_remove_count_fops, },