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 compound %#llx"
105 " archive_id %d flags %#llx\n",
106 prefix, hal->hal_version, hal->hal_count,
107 hal->hal_compound_id, 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
113 " compound/cookie=%#llx/%#llx"
114 " action=%s extent=%#llx-%#llx gid=%#llx"
115 " datalen=%d data=[%s]\n",
117 PFID(&hai->hai_fid), PFID(&hai->hai_dfid),
118 hal->hal_compound_id, hai->hai_cookie,
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 *mti;
140 char fs_name[MTI_NAME_MAXLEN+1];
141 /* are we scanning the logs for housekeeping, or just looking
145 /* request to be send to agents */
146 int max_requests; /** vector size */
147 int request_cnt; /** used count */
148 struct hsm_scan_request *request;
151 struct hsm_thread_data {
152 struct mdt_thread_info *cdt_mti;
153 struct hsm_scan_request *request;
156 static int mdt_cdt_waiting_cb(const struct lu_env *env,
157 struct mdt_device *mdt,
158 struct llog_handle *llh,
159 struct llog_agent_req_rec *larr,
160 struct hsm_scan_data *hsd)
162 struct coordinator *cdt = &mdt->mdt_coordinator;
163 struct hsm_scan_request *request;
164 struct hsm_action_item *hai;
167 /* Are agents full? */
168 if (atomic_read(&cdt->cdt_request_count) >= cdt->cdt_max_requests)
171 /* first search whether the request is found in the list we
174 for (i = 0; i < hsd->request_cnt; i++) {
175 if (hsd->request[i].hal->hal_compound_id ==
176 larr->arr_compound_id) {
177 request = &hsd->request[i];
183 struct hsm_action_list *hal;
185 if (hsd->request_cnt == hsd->max_requests) {
186 if (!hsd->housekeeping) {
187 /* The request array is full, stop
188 * here. There might be more known
189 * requests that could be merged, but
190 * this avoid analyzing too many llogs
191 * for minor gains. */
192 RETURN(LLOG_PROC_BREAK);
194 /* Unknown request and no more room
195 * for a new request. Continue to scan
196 * to find other entries for already
197 * existing requests. */
202 request = &hsd->request[hsd->request_cnt];
204 /* allocates hai vector size just needs to be large
206 request->hal_sz = sizeof(*request->hal) +
207 cfs_size_round(MTI_NAME_MAXLEN + 1) +
208 2 * cfs_size_round(larr->arr_hai.hai_len);
209 OBD_ALLOC(hal, request->hal_sz);
213 hal->hal_version = HAL_VERSION;
214 strlcpy(hal->hal_fsname, hsd->fs_name, MTI_NAME_MAXLEN + 1);
215 hal->hal_compound_id = larr->arr_compound_id;
216 hal->hal_archive_id = larr->arr_archive_id;
217 hal->hal_flags = larr->arr_flags;
219 request->hal_used_sz = hal_size(hal);
222 hai = hai_first(hal);
224 /* request is known */
225 /* we check if record archive num is the same as the
226 * known request, if not we will serve it in multiple
227 * time because we do not know if the agent can serve
228 * multiple backend a use case is a compound made of
229 * multiple restore where the files are not archived
230 * in the same backend */
231 if (larr->arr_archive_id != request->hal->hal_archive_id)
234 if (request->hal_sz < request->hal_used_sz +
235 cfs_size_round(larr->arr_hai.hai_len)) {
236 /* Not enough room, need an extension */
240 sz = 2 * request->hal_sz;
241 OBD_ALLOC(hal_buffer, sz);
244 memcpy(hal_buffer, request->hal, request->hal_used_sz);
245 OBD_FREE(request->hal, request->hal_sz);
246 request->hal = hal_buffer;
247 request->hal_sz = sz;
250 hai = hai_first(request->hal);
251 for (i = 0; i < request->hal->hal_count; i++)
255 memcpy(hai, &larr->arr_hai, larr->arr_hai.hai_len);
256 hai->hai_cookie = larr->arr_hai.hai_cookie;
257 hai->hai_gid = larr->arr_hai.hai_gid;
259 request->hal_used_sz += cfs_size_round(hai->hai_len);
260 request->hal->hal_count++;
262 if (hai->hai_action != HSMA_CANCEL)
263 cdt_agent_record_hash_add(cdt, hai->hai_cookie,
264 llh->lgh_hdr->llh_cat_idx,
265 larr->arr_hdr.lrh_index);
270 static int mdt_cdt_started_cb(const struct lu_env *env,
271 struct mdt_device *mdt,
272 struct llog_handle *llh,
273 struct llog_agent_req_rec *larr,
274 struct hsm_scan_data *hsd)
276 struct coordinator *cdt = &mdt->mdt_coordinator;
277 struct hsm_action_item *hai = &larr->arr_hai;
278 struct cdt_agent_req *car;
279 time64_t now = ktime_get_real_seconds();
284 if (!hsd->housekeeping)
287 /* we search for a running request
288 * error may happen if coordinator crashes or stopped
289 * with running request
291 car = mdt_cdt_find_request(cdt, hai->hai_cookie);
293 last = larr->arr_req_change;
295 last = car->car_req_update;
298 /* test if request too long, if yes cancel it
299 * the same way the copy tool acknowledge a cancel request */
300 if (now <= last + cdt->cdt_active_req_timeout)
301 GOTO(out_car, rc = 0);
303 dump_llog_agent_req_rec("request timed out, start cleaning", larr);
306 car->car_req_update = now;
307 mdt_hsm_agent_update_statistics(cdt, 0, 1, 0, &car->car_uuid);
308 /* Remove car from memory list (LU-9075) */
309 mdt_cdt_remove_request(cdt, hai->hai_cookie);
312 /* Emit a changelog record for the failed action.*/
314 hsm_set_cl_error(&cl_flags, ECANCELED);
316 switch (hai->hai_action) {
318 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
321 hsm_set_cl_event(&cl_flags, HE_RESTORE);
324 hsm_set_cl_event(&cl_flags, HE_REMOVE);
327 hsm_set_cl_event(&cl_flags, HE_CANCEL);
330 /* Unknown record type, skip changelog. */
336 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
339 if (hai->hai_action == HSMA_RESTORE)
340 cdt_restore_handle_del(hsd->mti, cdt, &hai->hai_fid);
342 larr->arr_status = ARS_CANCELED;
343 larr->arr_req_change = now;
344 rc = llog_write(hsd->mti->mti_env, llh, &larr->arr_hdr,
345 larr->arr_hdr.lrh_index);
347 CERROR("%s: cannot update agent log: rc = %d\n",
348 mdt_obd_name(mdt), rc);
349 rc = LLOG_DEL_RECORD;
352 /* ct has completed a request, so a slot is available,
353 * signal the coordinator to find new work */
354 mdt_hsm_cdt_event(cdt);
357 mdt_cdt_put_request(car);
363 * llog_cat_process() callback, used to:
364 * - find waiting request and start action
365 * - purge canceled and done requests
366 * \param env [IN] environment
367 * \param llh [IN] llog handle
368 * \param hdr [IN] llog record
369 * \param data [IN/OUT] cb data = struct hsm_scan_data
371 * \retval -ve failure
373 static int mdt_coordinator_cb(const struct lu_env *env,
374 struct llog_handle *llh,
375 struct llog_rec_hdr *hdr,
378 struct llog_agent_req_rec *larr = (struct llog_agent_req_rec *)hdr;
379 struct hsm_scan_data *hsd = data;
380 struct mdt_device *mdt = hsd->mti->mti_mdt;
381 struct coordinator *cdt = &mdt->mdt_coordinator;
384 larr = (struct llog_agent_req_rec *)hdr;
385 dump_llog_agent_req_rec("mdt_coordinator_cb(): ", larr);
386 switch (larr->arr_status) {
388 RETURN(mdt_cdt_waiting_cb(env, mdt, llh, larr, hsd));
390 RETURN(mdt_cdt_started_cb(env, mdt, llh, larr, hsd));
392 if (!hsd->housekeeping)
395 if ((larr->arr_req_change + cdt->cdt_grace_delay) <
396 ktime_get_real_seconds()) {
397 cdt_agent_record_hash_del(cdt,
398 larr->arr_hai.hai_cookie);
399 RETURN(LLOG_DEL_RECORD);
407 * create /proc entries for coordinator
410 * \retval -ve failure
412 int hsm_cdt_procfs_init(struct mdt_device *mdt)
414 struct coordinator *cdt = &mdt->mdt_coordinator;
418 /* init /proc entries, failure is not critical */
419 cdt->cdt_proc_dir = lprocfs_register("hsm",
420 mdt2obd_dev(mdt)->obd_proc_entry,
421 lprocfs_mdt_hsm_vars, mdt);
422 if (IS_ERR(cdt->cdt_proc_dir)) {
423 rc = PTR_ERR(cdt->cdt_proc_dir);
424 CERROR("%s: Cannot create 'hsm' directory in mdt proc dir,"
425 " rc=%d\n", mdt_obd_name(mdt), rc);
426 cdt->cdt_proc_dir = NULL;
434 * remove /proc entries for coordinator
437 void hsm_cdt_procfs_fini(struct mdt_device *mdt)
439 struct coordinator *cdt = &mdt->mdt_coordinator;
441 if (cdt->cdt_proc_dir != NULL)
442 lprocfs_remove(&cdt->cdt_proc_dir);
446 * get vector of hsm cdt /proc vars
450 struct lprocfs_vars *hsm_cdt_get_proc_vars(void)
452 return lprocfs_mdt_hsm_vars;
455 /* Release the ressource used by the coordinator. Called when the
456 * coordinator is stopping. */
457 static void mdt_hsm_cdt_cleanup(struct mdt_device *mdt)
459 struct coordinator *cdt = &mdt->mdt_coordinator;
460 struct cdt_agent_req *car, *tmp1;
461 struct hsm_agent *ha, *tmp2;
462 struct cdt_restore_handle *crh, *tmp3;
463 struct mdt_thread_info *cdt_mti;
466 down_write(&cdt->cdt_request_lock);
467 list_for_each_entry_safe(car, tmp1, &cdt->cdt_request_list,
469 cfs_hash_del(cdt->cdt_request_cookie_hash,
470 &car->car_hai->hai_cookie,
471 &car->car_cookie_hash);
472 list_del(&car->car_request_list);
473 mdt_cdt_put_request(car);
475 up_write(&cdt->cdt_request_lock);
477 down_write(&cdt->cdt_agent_lock);
478 list_for_each_entry_safe(ha, tmp2, &cdt->cdt_agents, ha_list) {
479 list_del(&ha->ha_list);
482 up_write(&cdt->cdt_agent_lock);
484 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
485 mutex_lock(&cdt->cdt_restore_lock);
486 list_for_each_entry_safe(crh, tmp3, &cdt->cdt_restore_handle_list,
488 list_del(&crh->crh_list);
489 /* give back layout lock */
490 mdt_object_unlock(cdt_mti, NULL, &crh->crh_lh, 1);
491 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
493 mutex_unlock(&cdt->cdt_restore_lock);
497 * Coordinator state transition table, indexed on enum cdt_states, taking
498 * from and to states. For instance since CDT_INIT to CDT_RUNNING is a
499 * valid transition, cdt_transition[CDT_INIT][CDT_RUNNING] is true.
501 static bool cdt_transition[CDT_STATES_COUNT][CDT_STATES_COUNT] = {
502 /* from -> to: stopped init running disable stopping */
503 /* stopped */ { true, true, false, false, false },
504 /* init */ { true, false, true, false, false },
505 /* running */ { false, false, true, true, true },
506 /* disable */ { false, false, true, true, true },
507 /* stopping */ { true, false, false, false, false }
511 * Change coordinator thread state
512 * Some combinations are not valid, so catch them here.
514 * Returns 0 on success, with old_state set if not NULL, or -EINVAL if
515 * the transition was not possible.
517 static int set_cdt_state(struct coordinator *cdt, enum cdt_states new_state,
518 enum cdt_states *old_state)
521 enum cdt_states state;
523 spin_lock(&cdt->cdt_state_lock);
525 state = cdt->cdt_state;
527 if (cdt_transition[state][new_state]) {
528 cdt->cdt_state = new_state;
529 spin_unlock(&cdt->cdt_state_lock);
534 spin_unlock(&cdt->cdt_state_lock);
536 "unexpected coordinator transition, from=%s, to=%s\n",
537 cdt_mdt_state2str(state), cdt_mdt_state2str(new_state));
546 * \param data [IN] obd device
548 * \retval -ve failure
550 static int mdt_coordinator(void *data)
552 struct hsm_thread_data *thread_data = data;
553 struct mdt_thread_info *mti = thread_data->cdt_mti;
554 struct mdt_device *mdt = mti->mti_mdt;
555 struct coordinator *cdt = &mdt->mdt_coordinator;
556 struct hsm_scan_data hsd = { NULL };
557 time64_t last_housekeeping = 0;
562 /* set up hsd->request and max_requests */
563 hsd.max_requests = cdt->cdt_max_requests;
564 request_sz = hsd.max_requests * sizeof(*hsd.request);
565 hsd.request = thread_data->request;
567 CDEBUG(D_HSM, "%s: coordinator thread starting, pid=%d\n",
568 mdt_obd_name(mdt), current_pid());
571 obd_uuid2fsname(hsd.fs_name, mdt_obd_name(mdt), MTI_NAME_MAXLEN);
573 set_cdt_state(cdt, CDT_RUNNING, NULL);
575 /* Inform mdt_hsm_cdt_start(). */
576 wake_up_all(&cdt->cdt_waitq);
583 struct hsm_record_update *updates;
585 /* Limit execution of the expensive requests traversal
586 * to at most one second. This prevents repeatedly
587 * locking/unlocking the catalog for each request
588 * and preventing other HSM operations from happening
590 wait_event_interruptible_timeout(cdt->cdt_waitq,
591 kthread_should_stop() ||
592 cdt->cdt_wakeup_coordinator,
593 cfs_time_seconds(1));
595 cdt->cdt_wakeup_coordinator = false;
596 CDEBUG(D_HSM, "coordinator resumes\n");
598 if (kthread_should_stop()) {
599 CDEBUG(D_HSM, "Coordinator stops\n");
604 /* if coordinator is suspended continue to wait */
605 if (cdt->cdt_state == CDT_DISABLE) {
606 CDEBUG(D_HSM, "disable state, coordinator sleeps\n");
610 /* If no event, and no housekeeping to do, continue to
612 if (last_housekeeping + cdt->cdt_loop_period <=
613 ktime_get_real_seconds()) {
614 last_housekeeping = ktime_get_real_seconds();
615 hsd.housekeeping = true;
616 } else if (cdt->cdt_event) {
617 hsd.housekeeping = false;
622 cdt->cdt_event = false;
624 CDEBUG(D_HSM, "coordinator starts reading llog\n");
626 if (hsd.max_requests != cdt->cdt_max_requests) {
627 /* cdt_max_requests has changed,
628 * we need to allocate a new buffer
630 struct hsm_scan_request *tmp = NULL;
631 int max_requests = cdt->cdt_max_requests;
632 OBD_ALLOC_LARGE(tmp, max_requests *
633 sizeof(struct hsm_scan_request));
635 CERROR("Failed to resize request buffer, "
636 "keeping it at %d\n",
638 cdt->cdt_max_requests = hsd.max_requests;
640 OBD_FREE_LARGE(hsd.request, request_sz);
641 hsd.max_requests = max_requests;
642 request_sz = hsd.max_requests *
643 sizeof(struct hsm_scan_request);
650 rc = cdt_llog_process(mti->mti_env, mdt, mdt_coordinator_cb,
655 CDEBUG(D_HSM, "found %d requests to send\n", hsd.request_cnt);
657 if (list_empty(&cdt->cdt_agents)) {
658 CDEBUG(D_HSM, "no agent available, "
659 "coordinator sleeps\n");
663 /* Compute how many HAI we have in all the requests */
665 for (i = 0; i < hsd.request_cnt; i++) {
666 const struct hsm_scan_request *request =
669 updates_cnt += request->hal->hal_count;
672 /* Allocate a temporary array to store the cookies to
673 * update, and their status. */
674 updates_sz = updates_cnt * sizeof(*updates);
675 OBD_ALLOC(updates, updates_sz);
676 if (updates == NULL) {
677 CERROR("%s: Cannot allocate memory (%d o) "
679 mdt_obd_name(mdt), updates_sz, updates_cnt);
683 /* here hsd contains a list of requests to be started */
684 for (i = 0; i < hsd.request_cnt; i++) {
685 struct hsm_scan_request *request = &hsd.request[i];
686 struct hsm_action_list *hal = request->hal;
687 struct hsm_action_item *hai;
690 /* still room for work ? */
691 if (atomic_read(&cdt->cdt_request_count) >=
692 cdt->cdt_max_requests)
695 rc = mdt_hsm_agent_send(mti, hal, 0);
696 /* if failure, we suppose it is temporary
697 * if the copy tool failed to do the request
698 * it has to use hsm_progress
701 /* set up cookie vector to set records status
702 * after copy tools start or failed
704 hai = hai_first(hal);
705 for (j = 0; j < hal->hal_count; j++) {
706 updates[update_idx].cookie = hai->hai_cookie;
707 updates[update_idx].status =
708 (rc ? ARS_WAITING : ARS_STARTED);
715 rc = mdt_agent_record_update(mti->mti_env, mdt,
716 updates, update_idx);
718 CERROR("%s: mdt_agent_record_update() failed, "
719 "rc=%d, cannot update records "
721 mdt_obd_name(mdt), rc, update_idx);
724 OBD_FREE(updates, updates_sz);
727 /* free hal allocated by callback */
728 for (i = 0; i < hsd.request_cnt; i++) {
729 struct hsm_scan_request *request = &hsd.request[i];
731 OBD_FREE(request->hal, request->hal_sz);
736 OBD_FREE_LARGE(hsd.request, request_sz);
738 mdt_hsm_cdt_cleanup(mdt);
741 CERROR("%s: coordinator thread exiting, process=%d, rc=%d\n",
742 mdt_obd_name(mdt), current_pid(), rc);
744 CDEBUG(D_HSM, "%s: coordinator thread exiting, process=%d,"
746 mdt_obd_name(mdt), current_pid());
751 int cdt_restore_handle_add(struct mdt_thread_info *mti, struct coordinator *cdt,
752 const struct lu_fid *fid,
753 const struct hsm_extent *he)
755 struct cdt_restore_handle *crh;
756 struct mdt_object *obj;
760 OBD_SLAB_ALLOC_PTR(crh, mdt_hsm_cdt_kmem);
765 /* in V1 all file is restored
766 * crh->extent.start = he->offset;
767 * crh->extent.end = he->offset + he->length;
769 crh->crh_extent.start = 0;
770 crh->crh_extent.end = he->length;
771 /* get the layout lock */
772 mdt_lock_reg_init(&crh->crh_lh, LCK_EX);
773 obj = mdt_object_find_lock(mti, &crh->crh_fid, &crh->crh_lh,
774 MDS_INODELOCK_LAYOUT);
776 GOTO(out_crh, rc = PTR_ERR(obj));
778 /* We do not keep a reference on the object during the restore
779 * which can be very long. */
780 mdt_object_put(mti->mti_env, obj);
782 mutex_lock(&cdt->cdt_restore_lock);
783 if (unlikely(cdt->cdt_state == CDT_STOPPED ||
784 cdt->cdt_state == CDT_STOPPING)) {
785 mutex_unlock(&cdt->cdt_restore_lock);
786 GOTO(out_lh, rc = -EAGAIN);
789 list_add_tail(&crh->crh_list, &cdt->cdt_restore_handle_list);
790 mutex_unlock(&cdt->cdt_restore_lock);
794 mdt_object_unlock(mti, NULL, &crh->crh_lh, 1);
796 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
802 * lookup a restore handle by FID
803 * caller needs to hold cdt_restore_lock
804 * \param cdt [IN] coordinator
805 * \param fid [IN] FID
806 * \retval cdt_restore_handle found
807 * \retval NULL not found
809 struct cdt_restore_handle *cdt_restore_handle_find(struct coordinator *cdt,
810 const struct lu_fid *fid)
812 struct cdt_restore_handle *crh;
815 list_for_each_entry(crh, &cdt->cdt_restore_handle_list, crh_list) {
816 if (lu_fid_eq(&crh->crh_fid, fid))
823 void cdt_restore_handle_del(struct mdt_thread_info *mti,
824 struct coordinator *cdt, const struct lu_fid *fid)
826 struct cdt_restore_handle *crh;
828 /* give back layout lock */
829 mutex_lock(&cdt->cdt_restore_lock);
830 crh = cdt_restore_handle_find(cdt, fid);
832 list_del(&crh->crh_list);
833 mutex_unlock(&cdt->cdt_restore_lock);
838 /* XXX We pass a NULL object since the restore handle does not
839 * keep a reference on the object being restored. */
840 mdt_object_unlock(mti, NULL, &crh->crh_lh, 1);
841 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
845 * data passed to llog_cat_process() callback
846 * to scan requests and take actions
848 struct hsm_restore_data {
849 struct mdt_thread_info *hrd_mti;
853 * llog_cat_process() callback, used to:
854 * - find restore request and allocate the restore handle
855 * \param env [IN] environment
856 * \param llh [IN] llog handle
857 * \param hdr [IN] llog record
858 * \param data [IN/OUT] cb data = struct hsm_restore_data
860 * \retval -ve failure
862 static int hsm_restore_cb(const struct lu_env *env,
863 struct llog_handle *llh,
864 struct llog_rec_hdr *hdr, void *data)
866 struct llog_agent_req_rec *larr;
867 struct hsm_restore_data *hrd;
868 struct hsm_action_item *hai;
869 struct mdt_thread_info *mti;
870 struct coordinator *cdt;
876 cdt = &mti->mti_mdt->mdt_coordinator;
878 larr = (struct llog_agent_req_rec *)hdr;
879 hai = &larr->arr_hai;
880 if (hai->hai_cookie > cdt->cdt_last_cookie)
881 /* update the cookie to avoid collision */
882 cdt->cdt_last_cookie = hai->hai_cookie + 1;
884 if (hai->hai_action != HSMA_RESTORE ||
885 agent_req_in_final_state(larr->arr_status))
888 /* restore request not in a final state */
890 /* force replay of restore requests left in started state from previous
891 * CDT context, to be canceled later if finally found to be incompatible
892 * when being re-started */
893 if (larr->arr_status == ARS_STARTED) {
894 larr->arr_status = ARS_WAITING;
895 larr->arr_req_change = ktime_get_real_seconds();
896 rc = llog_write(env, llh, hdr, hdr->lrh_index);
901 rc = cdt_restore_handle_add(mti, cdt, &hai->hai_fid, &hai->hai_extent);
907 * restore coordinator state at startup
908 * the goal is to take a layout lock for each registered restore request
909 * \param mti [IN] context
911 static int mdt_hsm_pending_restore(struct mdt_thread_info *mti)
913 struct hsm_restore_data hrd;
919 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, hsm_restore_cb, &hrd,
925 static int hsm_init_ucred(struct lu_ucred *uc)
929 uc->uc_valid = UCRED_OLD;
938 uc->uc_suppgids[0] = -1;
939 uc->uc_suppgids[1] = -1;
940 uc->uc_cap = CFS_CAP_FS_MASK;
943 uc->uc_identity = NULL;
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 spin_lock_init(&cdt->cdt_state_lock);
967 set_cdt_state(cdt, CDT_STOPPED, NULL);
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 /* Initialize cdt_compound_id here to allow its usage for
1027 * delayed requests from RAoLU policy */
1028 atomic_set(&cdt->cdt_compound_id, ktime_get_real_seconds());
1030 /* by default do not remove archives on last unlink */
1031 cdt->cdt_remove_archive_on_last_unlink = false;
1036 lu_env_fini(&cdt->cdt_env);
1037 out_agent_record_hash:
1038 cfs_hash_putref(cdt->cdt_agent_record_hash);
1039 cdt->cdt_agent_record_hash = NULL;
1040 out_request_cookie_hash:
1041 cfs_hash_putref(cdt->cdt_request_cookie_hash);
1042 cdt->cdt_request_cookie_hash = NULL;
1048 * free a coordinator thread
1049 * \param mdt [IN] device
1051 int mdt_hsm_cdt_fini(struct mdt_device *mdt)
1053 struct coordinator *cdt = &mdt->mdt_coordinator;
1056 lu_context_exit(cdt->cdt_env.le_ses);
1057 lu_context_fini(cdt->cdt_env.le_ses);
1059 lu_env_fini(&cdt->cdt_env);
1061 cfs_hash_putref(cdt->cdt_agent_record_hash);
1062 cdt->cdt_agent_record_hash = NULL;
1064 cfs_hash_putref(cdt->cdt_request_cookie_hash);
1065 cdt->cdt_request_cookie_hash = NULL;
1071 * start a coordinator thread
1072 * \param mdt [IN] device
1074 * \retval -ve failure
1076 static int mdt_hsm_cdt_start(struct mdt_device *mdt)
1078 struct coordinator *cdt = &mdt->mdt_coordinator;
1081 struct task_struct *task;
1083 struct hsm_thread_data thread_data;
1086 /* functions defined but not yet used
1087 * this avoid compilation warning
1089 ptr = dump_requests;
1091 rc = set_cdt_state(cdt, CDT_INIT, NULL);
1093 CERROR("%s: Coordinator already started or stopping\n",
1098 CLASSERT(1 << (CDT_POLICY_SHIFT_COUNT - 1) == CDT_POLICY_LAST);
1099 cdt->cdt_policy = CDT_DEFAULT_POLICY;
1101 /* just need to be larger than previous one */
1102 /* cdt_last_cookie is protected by cdt_llog_lock */
1103 cdt->cdt_last_cookie = ktime_get_real_seconds();
1104 atomic_set(&cdt->cdt_request_count, 0);
1105 atomic_set(&cdt->cdt_archive_count, 0);
1106 atomic_set(&cdt->cdt_restore_count, 0);
1107 atomic_set(&cdt->cdt_remove_count, 0);
1108 cdt->cdt_user_request_mask = (1UL << HSMA_RESTORE);
1109 cdt->cdt_group_request_mask = (1UL << HSMA_RESTORE);
1110 cdt->cdt_other_request_mask = (1UL << HSMA_RESTORE);
1112 /* to avoid deadlock when start is made through /proc
1113 * /proc entries are created by the coordinator thread */
1115 /* set up list of started restore requests */
1116 thread_data.cdt_mti =
1117 lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
1118 rc = mdt_hsm_pending_restore(thread_data.cdt_mti);
1120 CERROR("%s: cannot take the layout locks needed"
1121 " for registered restore: %d\n",
1122 mdt_obd_name(mdt), rc);
1124 if (mdt->mdt_bottom->dd_rdonly)
1127 /* Allocate the initial hsd.request[] vector*/
1128 request_sz = cdt->cdt_max_requests * sizeof(struct hsm_scan_request);
1129 OBD_ALLOC_LARGE(thread_data.request, request_sz);
1130 if (!thread_data.request) {
1131 set_cdt_state(cdt, CDT_STOPPED, NULL);
1135 task = kthread_run(mdt_coordinator, &thread_data, "hsm_cdtr");
1138 set_cdt_state(cdt, CDT_STOPPED, NULL);
1139 OBD_FREE(thread_data.request, request_sz);
1140 CERROR("%s: error starting coordinator thread: %d\n",
1141 mdt_obd_name(mdt), rc);
1143 cdt->cdt_task = task;
1144 wait_event(cdt->cdt_waitq,
1145 cdt->cdt_state != CDT_INIT);
1146 CDEBUG(D_HSM, "%s: coordinator thread started\n",
1155 * stop a coordinator thread
1156 * \param mdt [IN] device
1158 int mdt_hsm_cdt_stop(struct mdt_device *mdt)
1160 struct coordinator *cdt = &mdt->mdt_coordinator;
1164 /* stop coordinator thread */
1165 rc = set_cdt_state(cdt, CDT_STOPPING, NULL);
1167 kthread_stop(cdt->cdt_task);
1168 cdt->cdt_task = NULL;
1169 set_cdt_state(cdt, CDT_STOPPED, NULL);
1175 static int mdt_hsm_set_exists(struct mdt_thread_info *mti,
1176 const struct lu_fid *fid,
1179 struct mdt_object *obj;
1183 obj = mdt_hsm_get_md_hsm(mti, fid, &mh);
1185 GOTO(out, rc = PTR_ERR(obj));
1187 if (mh.mh_flags & HS_EXISTS &&
1188 mh.mh_arch_id == archive_id)
1189 GOTO(out_obj, rc = 0);
1191 mh.mh_flags |= HS_EXISTS;
1192 mh.mh_arch_id = archive_id;
1193 rc = mdt_hsm_attr_set(mti, obj, &mh);
1196 mdt_object_put(mti->mti_env, obj);
1202 * register all requests from an hal in the memory list
1203 * \param mti [IN] context
1204 * \param hal [IN] request
1205 * \param uuid [OUT] in case of CANCEL, the uuid of the agent
1206 * which is running the CT
1208 * \retval -ve failure
1210 int mdt_hsm_add_hal(struct mdt_thread_info *mti,
1211 struct hsm_action_list *hal, struct obd_uuid *uuid)
1213 struct mdt_device *mdt = mti->mti_mdt;
1214 struct coordinator *cdt = &mdt->mdt_coordinator;
1215 struct hsm_action_item *hai;
1219 /* register request in memory list */
1220 hai = hai_first(hal);
1221 for (i = 0; i < hal->hal_count; i++, hai = hai_next(hai)) {
1222 struct cdt_agent_req *car;
1224 /* in case of a cancel request, we first mark the ondisk
1225 * record of the request we want to stop as canceled
1226 * this does not change the cancel record
1227 * it will be done when updating the request status
1229 if (hai->hai_action == HSMA_CANCEL) {
1230 struct hsm_record_update update = {
1231 .cookie = hai->hai_cookie,
1232 .status = ARS_CANCELED,
1235 rc = mdt_agent_record_update(mti->mti_env, mti->mti_mdt,
1238 CERROR("%s: mdt_agent_record_update() failed, "
1239 "rc=%d, cannot update status to %s "
1240 "for cookie %#llx\n",
1241 mdt_obd_name(mdt), rc,
1242 agent_req_status2name(ARS_CANCELED),
1247 /* find the running request to set it canceled */
1248 car = mdt_cdt_find_request(cdt, hai->hai_cookie);
1250 car->car_canceled = 1;
1251 /* uuid has to be changed to the one running the
1252 * request to cancel */
1253 *uuid = car->car_uuid;
1254 mdt_cdt_put_request(car);
1256 /* no need to memorize cancel request
1257 * this also avoid a deadlock when we receive
1258 * a purge all requests command
1263 if (hai->hai_action == HSMA_ARCHIVE) {
1264 rc = mdt_hsm_set_exists(mti, &hai->hai_fid,
1265 hal->hal_archive_id);
1272 car = mdt_cdt_alloc_request(hal->hal_compound_id,
1273 hal->hal_archive_id, hal->hal_flags,
1276 GOTO(out, rc = PTR_ERR(car));
1278 rc = mdt_cdt_add_request(cdt, car);
1280 mdt_cdt_free_request(car);
1287 * swap layouts between 2 fids
1288 * \param mti [IN] context
1291 * \param mh_common [IN] MD HSM
1293 static int hsm_swap_layouts(struct mdt_thread_info *mti,
1294 struct mdt_object *obj, const struct lu_fid *dfid,
1295 struct md_hsm *mh_common)
1297 struct mdt_object *dobj;
1298 struct mdt_lock_handle *dlh;
1302 if (!mdt_object_exists(obj))
1303 GOTO(out, rc = -ENOENT);
1305 /* we already have layout lock on obj so take only
1307 dlh = &mti->mti_lh[MDT_LH_OLD];
1308 mdt_lock_reg_init(dlh, LCK_EX);
1309 dobj = mdt_object_find_lock(mti, dfid, dlh, MDS_INODELOCK_LAYOUT);
1311 GOTO(out, rc = PTR_ERR(dobj));
1313 /* if copy tool closes the volatile before sending the final
1314 * progress through llapi_hsm_copy_end(), all the objects
1315 * are removed and mdd_swap_layout LBUG */
1316 if (!mdt_object_exists(dobj)) {
1317 CERROR("%s: Copytool has closed volatile file "DFID"\n",
1318 mdt_obd_name(mti->mti_mdt), PFID(dfid));
1319 GOTO(out_dobj, rc = -ENOENT);
1321 /* Since we only handle restores here, unconditionally use
1322 * SWAP_LAYOUTS_MDS_HSM flag to ensure original layout will
1323 * be preserved in case of failure during swap_layout and not
1324 * leave a file in an intermediate but incoherent state.
1325 * But need to setup HSM xattr of data FID before, reuse
1326 * mti and mh presets for FID in hsm_cdt_request_completed(),
1327 * only need to clear RELEASED and DIRTY.
1329 mh_common->mh_flags &= ~(HS_RELEASED | HS_DIRTY);
1330 rc = mdt_hsm_attr_set(mti, dobj, mh_common);
1332 rc = mo_swap_layouts(mti->mti_env,
1333 mdt_object_child(obj),
1334 mdt_object_child(dobj),
1335 SWAP_LAYOUTS_MDS_HSM);
1338 mdt_object_unlock_put(mti, dobj, dlh, 1);
1344 * update status of a completed request
1345 * \param mti [IN] context
1346 * \param pgs [IN] progress of the copy tool
1348 * \retval -ve failure
1350 static int hsm_cdt_request_completed(struct mdt_thread_info *mti,
1351 struct hsm_progress_kernel *pgs,
1352 const struct cdt_agent_req *car,
1353 enum agent_req_status *status)
1355 const struct lu_env *env = mti->mti_env;
1356 struct mdt_device *mdt = mti->mti_mdt;
1357 struct coordinator *cdt = &mdt->mdt_coordinator;
1358 struct mdt_object *obj = NULL;
1359 int cl_flags = 0, rc = 0;
1362 bool need_changelog = true;
1365 /* default is to retry */
1366 *status = ARS_WAITING;
1368 /* find object by FID, mdt_hsm_get_md_hsm() returns obj or err
1369 * if error/removed continue anyway to get correct reporting done */
1370 obj = mdt_hsm_get_md_hsm(mti, &car->car_hai->hai_fid, &mh);
1371 /* we will update MD HSM only if needed */
1372 is_mh_changed = false;
1374 /* no need to change mh->mh_arch_id
1375 * mdt_hsm_get_md_hsm() got it from disk and it is still valid
1377 if (pgs->hpk_errval != 0) {
1378 switch (pgs->hpk_errval) {
1380 /* the copy tool does not support cancel
1381 * so the cancel request is failed
1382 * As we cannot distinguish a cancel progress
1383 * from another action progress (they have the
1384 * same cookie), we suppose here the CT returns
1385 * ENOSYS only if does not support cancel
1387 /* this can also happen when cdt calls it to
1388 * for a timed out request */
1389 *status = ARS_FAILED;
1390 /* to have a cancel event in changelog */
1391 pgs->hpk_errval = ECANCELED;
1394 /* the request record has already been set to
1395 * ARS_CANCELED, this set the cancel request
1397 *status = ARS_SUCCEED;
1400 /* retry only if current policy or requested, and
1401 * object is not on error/removed */
1402 *status = (cdt->cdt_policy & CDT_NORETRY_ACTION ||
1403 !(pgs->hpk_flags & HP_FLAG_RETRY) ||
1404 IS_ERR(obj)) ? ARS_FAILED : ARS_WAITING;
1408 if (pgs->hpk_errval > CLF_HSM_MAXERROR) {
1409 CERROR("%s: Request %#llx on "DFID
1410 " failed, error code %d too large\n",
1412 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1414 hsm_set_cl_error(&cl_flags,
1415 CLF_HSM_ERROVERFLOW);
1418 hsm_set_cl_error(&cl_flags, pgs->hpk_errval);
1421 switch (car->car_hai->hai_action) {
1423 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1426 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1429 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1432 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1433 CERROR("%s: Failed request %#llx on "DFID
1434 " cannot be a CANCEL\n",
1437 PFID(&pgs->hpk_fid));
1440 CERROR("%s: Failed request %#llx on "DFID
1441 " %d is an unknown action\n",
1443 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1444 car->car_hai->hai_action);
1449 *status = ARS_SUCCEED;
1450 switch (car->car_hai->hai_action) {
1452 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1453 /* set ARCHIVE keep EXIST and clear LOST and
1455 mh.mh_arch_ver = pgs->hpk_data_version;
1456 mh.mh_flags |= HS_ARCHIVED;
1457 mh.mh_flags &= ~(HS_LOST|HS_DIRTY);
1458 is_mh_changed = true;
1461 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1463 /* do not clear RELEASED and DIRTY here
1464 * this will occur in hsm_swap_layouts()
1467 /* Restoring has changed the file version on
1469 mh.mh_arch_ver = pgs->hpk_data_version;
1470 is_mh_changed = true;
1473 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1474 /* clear ARCHIVED EXISTS and LOST */
1475 mh.mh_flags &= ~(HS_ARCHIVED | HS_EXISTS | HS_LOST);
1476 is_mh_changed = true;
1479 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1480 CERROR("%s: Successful request %#llx on "DFID" cannot be a CANCEL\n",
1483 PFID(&pgs->hpk_fid));
1486 CERROR("%s: Successful request %#llx on "DFID" %d is an unknown action\n",
1488 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1489 car->car_hai->hai_action);
1495 /* rc != 0 means error when analysing action, it may come from
1496 * a crasy CT no need to manage DIRTY
1497 * and if mdt_hsm_get_md_hsm() has returned an error, mh has not been
1500 if (rc == 0 && !IS_ERR(obj))
1501 hsm_set_cl_flags(&cl_flags,
1502 mh.mh_flags & HS_DIRTY ? CLF_HSM_DIRTY : 0);
1504 /* unlock is done later, after layout lock management */
1505 if (is_mh_changed && !IS_ERR(obj))
1506 rc = mdt_hsm_attr_set(mti, obj, &mh);
1508 /* we give back layout lock only if restore was successful or
1509 * if no retry will be attempted and if object is still alive,
1510 * in other cases we just unlock the object */
1511 if (car->car_hai->hai_action == HSMA_RESTORE) {
1512 /* restore in data FID done, we swap the layouts
1513 * only if restore is successful */
1514 if (pgs->hpk_errval == 0 && !IS_ERR(obj)) {
1515 rc = hsm_swap_layouts(mti, obj, &car->car_hai->hai_dfid,
1518 if (cdt->cdt_policy & CDT_NORETRY_ACTION)
1519 *status = ARS_FAILED;
1520 pgs->hpk_errval = -rc;
1523 /* we have to retry, so keep layout lock */
1524 if (*status == ARS_WAITING)
1527 /* restore special case, need to create ChangeLog record
1528 * before to give back layout lock to avoid concurrent
1529 * file updater to post out of order ChangeLog */
1530 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1531 &car->car_hai->hai_fid);
1532 need_changelog = false;
1534 cdt_restore_handle_del(mti, cdt, &car->car_hai->hai_fid);
1540 /* always add a ChangeLog record */
1542 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1543 &car->car_hai->hai_fid);
1546 mdt_object_put(mti->mti_env, obj);
1552 * update status of a request
1553 * \param mti [IN] context
1554 * \param pgs [IN] progress of the copy tool
1556 * \retval -ve failure
1558 int mdt_hsm_update_request_state(struct mdt_thread_info *mti,
1559 struct hsm_progress_kernel *pgs)
1561 struct mdt_device *mdt = mti->mti_mdt;
1562 struct coordinator *cdt = &mdt->mdt_coordinator;
1563 struct cdt_agent_req *car;
1567 /* no coordinator started, so we cannot serve requests */
1568 if (cdt->cdt_state == CDT_STOPPED)
1571 /* first do sanity checks */
1572 car = mdt_cdt_update_request(cdt, pgs);
1574 CERROR("%s: Cannot find running request for cookie %#llx"
1577 pgs->hpk_cookie, PFID(&pgs->hpk_fid));
1579 RETURN(PTR_ERR(car));
1582 CDEBUG(D_HSM, "Progress received for fid="DFID" cookie=%#llx"
1583 " action=%s flags=%d err=%d fid="DFID" dfid="DFID"\n",
1584 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1585 hsm_copytool_action2name(car->car_hai->hai_action),
1586 pgs->hpk_flags, pgs->hpk_errval,
1587 PFID(&car->car_hai->hai_fid),
1588 PFID(&car->car_hai->hai_dfid));
1590 /* progress is done on FID or data FID depending of the action and
1591 * of the copy progress */
1592 /* for restore progress is used to send back the data FID to cdt */
1593 if (car->car_hai->hai_action == HSMA_RESTORE &&
1594 lu_fid_eq(&car->car_hai->hai_fid, &car->car_hai->hai_dfid))
1595 car->car_hai->hai_dfid = pgs->hpk_fid;
1597 if ((car->car_hai->hai_action == HSMA_RESTORE ||
1598 car->car_hai->hai_action == HSMA_ARCHIVE) &&
1599 (!lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_dfid) &&
1600 !lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_fid))) {
1601 CERROR("%s: Progress on "DFID" for cookie %#llx"
1602 " does not match request FID "DFID" nor data FID "
1605 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1606 PFID(&car->car_hai->hai_fid),
1607 PFID(&car->car_hai->hai_dfid));
1608 GOTO(out, rc = -EINVAL);
1611 if (pgs->hpk_errval != 0 && !(pgs->hpk_flags & HP_FLAG_COMPLETED)) {
1612 CERROR("%s: Progress on "DFID" for cookie %#llx action=%s"
1613 " is not coherent (err=%d and not completed"
1616 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1617 hsm_copytool_action2name(car->car_hai->hai_action),
1618 pgs->hpk_errval, pgs->hpk_flags);
1619 GOTO(out, rc = -EINVAL);
1622 /* now progress is valid */
1624 /* we use a root like ucred */
1625 hsm_init_ucred(mdt_ucred(mti));
1627 if (pgs->hpk_flags & HP_FLAG_COMPLETED) {
1628 enum agent_req_status status;
1629 struct hsm_record_update update;
1632 rc = hsm_cdt_request_completed(mti, pgs, car, &status);
1634 CDEBUG(D_HSM, "updating record: fid="DFID" cookie=%#llx action=%s "
1636 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1637 hsm_copytool_action2name(car->car_hai->hai_action),
1638 agent_req_status2name(status));
1640 /* update record first (LU-9075) */
1641 update.cookie = pgs->hpk_cookie;
1642 update.status = status;
1644 rc1 = mdt_agent_record_update(mti->mti_env, mdt,
1647 CERROR("%s: mdt_agent_record_update() failed,"
1648 " rc=%d, cannot update status to %s"
1649 " for cookie %#llx\n",
1650 mdt_obd_name(mdt), rc1,
1651 agent_req_status2name(status),
1653 rc = (rc != 0 ? rc : rc1);
1655 /* then remove request from memory list (LU-9075) */
1656 mdt_cdt_remove_request(cdt, pgs->hpk_cookie);
1658 /* ct has completed a request, so a slot is available,
1659 * signal the coordinator to find new work */
1660 mdt_hsm_cdt_event(cdt);
1662 /* if copytool send a progress on a canceled request
1663 * we inform copytool it should stop
1665 if (car->car_canceled == 1)
1671 /* remove ref got from mdt_cdt_update_request() */
1672 mdt_cdt_put_request(car);
1679 * data passed to llog_cat_process() callback
1680 * to cancel requests
1682 struct hsm_cancel_all_data {
1683 struct mdt_device *mdt;
1687 * llog_cat_process() callback, used to:
1688 * - purge all requests
1689 * \param env [IN] environment
1690 * \param llh [IN] llog handle
1691 * \param hdr [IN] llog record
1692 * \param data [IN] cb data = struct hsm_cancel_all_data
1694 * \retval -ve failure
1696 static int mdt_cancel_all_cb(const struct lu_env *env,
1697 struct llog_handle *llh,
1698 struct llog_rec_hdr *hdr, void *data)
1700 struct llog_agent_req_rec *larr;
1701 struct hsm_cancel_all_data *hcad;
1705 larr = (struct llog_agent_req_rec *)hdr;
1707 if (larr->arr_status == ARS_WAITING ||
1708 larr->arr_status == ARS_STARTED) {
1709 larr->arr_status = ARS_CANCELED;
1710 larr->arr_req_change = ktime_get_real_seconds();
1711 rc = llog_write(env, llh, hdr, hdr->lrh_index);
1718 * cancel all actions
1719 * \param obd [IN] MDT device
1721 static int hsm_cancel_all_actions(struct mdt_device *mdt)
1724 struct lu_context session;
1725 struct mdt_thread_info *mti;
1726 struct coordinator *cdt = &mdt->mdt_coordinator;
1727 struct cdt_agent_req *car;
1728 struct hsm_action_list *hal = NULL;
1729 struct hsm_action_item *hai;
1730 struct hsm_cancel_all_data hcad;
1731 int hal_sz = 0, hal_len, rc;
1732 enum cdt_states old_state;
1735 rc = lu_env_init(&env, LCT_MD_THREAD);
1739 /* for mdt_ucred(), lu_ucred stored in lu_ucred_key */
1740 rc = lu_context_init(&session, LCT_SERVER_SESSION);
1744 lu_context_enter(&session);
1745 env.le_ses = &session;
1747 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
1748 LASSERT(mti != NULL);
1750 mti->mti_env = &env;
1753 hsm_init_ucred(mdt_ucred(mti));
1755 /* disable coordinator */
1756 rc = set_cdt_state(cdt, CDT_DISABLE, &old_state);
1760 /* send cancel to all running requests */
1761 down_read(&cdt->cdt_request_lock);
1762 list_for_each_entry(car, &cdt->cdt_request_list, car_request_list) {
1763 mdt_cdt_get_request(car);
1764 /* request is not yet removed from list, it will be done
1765 * when copytool will return progress
1768 if (car->car_hai->hai_action == HSMA_CANCEL) {
1769 mdt_cdt_put_request(car);
1774 hal_len = sizeof(*hal) + cfs_size_round(MTI_NAME_MAXLEN + 1) +
1775 cfs_size_round(car->car_hai->hai_len);
1777 if (hal_len > hal_sz && hal_sz > 0) {
1778 /* not enough room, free old buffer */
1779 OBD_FREE(hal, hal_sz);
1783 /* empty buffer, allocate one */
1786 OBD_ALLOC(hal, hal_sz);
1788 mdt_cdt_put_request(car);
1789 up_read(&cdt->cdt_request_lock);
1790 GOTO(out_cdt_state, rc = -ENOMEM);
1794 hal->hal_version = HAL_VERSION;
1795 obd_uuid2fsname(hal->hal_fsname, mdt_obd_name(mdt),
1797 hal->hal_fsname[MTI_NAME_MAXLEN] = '\0';
1798 hal->hal_compound_id = car->car_compound_id;
1799 hal->hal_archive_id = car->car_archive_id;
1800 hal->hal_flags = car->car_flags;
1803 hai = hai_first(hal);
1804 memcpy(hai, car->car_hai, car->car_hai->hai_len);
1805 hai->hai_action = HSMA_CANCEL;
1808 /* it is possible to safely call mdt_hsm_agent_send()
1809 * (ie without a deadlock on cdt_request_lock), because the
1810 * write lock is taken only if we are not in purge mode
1811 * (mdt_hsm_agent_send() does not call mdt_cdt_add_request()
1812 * nor mdt_cdt_remove_request())
1814 /* no conflict with cdt thread because cdt is disable and we
1815 * have the request lock */
1816 mdt_hsm_agent_send(mti, hal, 1);
1818 mdt_cdt_put_request(car);
1820 up_read(&cdt->cdt_request_lock);
1823 OBD_FREE(hal, hal_sz);
1825 /* cancel all on-disk records */
1828 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, mdt_cancel_all_cb,
1829 &hcad, 0, 0, WRITE);
1831 /* Enable coordinator, unless the coordinator was stopping. */
1832 set_cdt_state(cdt, old_state, NULL);
1833 lu_context_exit(&session);
1834 lu_context_fini(&session);
1842 * check if a request is compatible with file status
1843 * \param hai [IN] request description
1844 * \param archive_id [IN] request archive id
1845 * \param rq_flags [IN] request flags
1846 * \param hsm [IN] file HSM metadata
1849 bool mdt_hsm_is_action_compat(const struct hsm_action_item *hai,
1850 u32 archive_id, u64 rq_flags,
1851 const struct md_hsm *hsm)
1853 int is_compat = false;
1857 hsm_flags = hsm->mh_flags;
1858 switch (hai->hai_action) {
1860 if (!(hsm_flags & HS_NOARCHIVE) &&
1861 (hsm_flags & HS_DIRTY || !(hsm_flags & HS_ARCHIVED)))
1864 if (hsm_flags & HS_EXISTS &&
1866 archive_id != hsm->mh_arch_id)
1871 if (!(hsm_flags & HS_DIRTY) && (hsm_flags & HS_RELEASED) &&
1872 hsm_flags & HS_ARCHIVED && !(hsm_flags & HS_LOST))
1876 if (!(hsm_flags & HS_RELEASED) &&
1877 (hsm_flags & (HS_ARCHIVED | HS_EXISTS)))
1884 CDEBUG(D_HSM, "fid="DFID" action=%s flags=%#llx"
1885 " extent=%#llx-%#llx hsm_flags=%.8X %s\n",
1886 PFID(&hai->hai_fid),
1887 hsm_copytool_action2name(hai->hai_action), rq_flags,
1888 hai->hai_extent.offset, hai->hai_extent.length,
1890 (is_compat ? "compatible" : "uncompatible"));
1896 * /proc interface used to get/set HSM behaviour (cdt->cdt_policy)
1898 static const struct {
1902 } hsm_policy_names[] = {
1903 { CDT_NONBLOCKING_RESTORE, "NonBlockingRestore", "NBR"},
1904 { CDT_NORETRY_ACTION, "NoRetryAction", "NRA"},
1909 * convert a policy name to a bit
1910 * \param name [IN] policy name
1912 * \retval policy bit
1914 static __u64 hsm_policy_str2bit(const char *name)
1918 for (i = 0; hsm_policy_names[i].bit != 0; i++)
1919 if (strcmp(hsm_policy_names[i].nickname, name) == 0 ||
1920 strcmp(hsm_policy_names[i].name, name) == 0)
1921 return hsm_policy_names[i].bit;
1926 * convert a policy bit field to a string
1927 * \param mask [IN] policy bit field
1928 * \param hexa [IN] print mask before bit names
1929 * \param buffer [OUT] string
1930 * \param count [IN] size of buffer
1932 static void hsm_policy_bit2str(struct seq_file *m, const __u64 mask,
1940 seq_printf(m, "(%#llx) ", mask);
1942 for (i = 0; i < CDT_POLICY_SHIFT_COUNT; i++) {
1945 for (j = 0; hsm_policy_names[j].bit != 0; j++) {
1946 if (hsm_policy_names[j].bit == bit)
1950 seq_printf(m, "[%s] ", hsm_policy_names[j].name);
1952 seq_printf(m, "%s ", hsm_policy_names[j].name);
1954 /* remove last ' ' */
1959 /* methods to read/write HSM policy flags */
1960 static int mdt_hsm_policy_seq_show(struct seq_file *m, void *data)
1962 struct mdt_device *mdt = m->private;
1963 struct coordinator *cdt = &mdt->mdt_coordinator;
1966 hsm_policy_bit2str(m, cdt->cdt_policy, false);
1971 mdt_hsm_policy_seq_write(struct file *file, const char __user *buffer,
1972 size_t count, loff_t *off)
1974 struct seq_file *m = file->private_data;
1975 struct mdt_device *mdt = m->private;
1976 struct coordinator *cdt = &mdt->mdt_coordinator;
1977 char *start, *token, sign;
1980 __u64 add_mask, remove_mask, set_mask;
1984 if (count + 1 > PAGE_SIZE)
1987 OBD_ALLOC(buf, count + 1);
1991 if (copy_from_user(buf, buffer, count))
1992 GOTO(out, rc = -EFAULT);
1997 CDEBUG(D_HSM, "%s: receive new policy: '%s'\n", mdt_obd_name(mdt),
2000 add_mask = remove_mask = set_mask = 0;
2002 token = strsep(&start, "\n ");
2008 if (sign == '-' || sign == '+')
2011 policy = hsm_policy_str2bit(token);
2013 CWARN("%s: '%s' is unknown, "
2014 "supported policies are:\n", mdt_obd_name(mdt),
2016 hsm_policy_bit2str(m, 0, false);
2017 GOTO(out, rc = -EINVAL);
2021 remove_mask |= policy;
2031 } while (start != NULL);
2033 CDEBUG(D_HSM, "%s: new policy: rm=%#llx add=%#llx set=%#llx\n",
2034 mdt_obd_name(mdt), remove_mask, add_mask, set_mask);
2036 /* if no sign in all string, it is a clear and set
2037 * if some sign found, all unsigned are converted
2039 * P1 P2 = set to P1 and P2
2040 * P1 -P2 = add P1 clear P2 same as +P1 -P2
2042 if (remove_mask == 0 && add_mask == 0) {
2043 cdt->cdt_policy = set_mask;
2045 cdt->cdt_policy |= set_mask | add_mask;
2046 cdt->cdt_policy &= ~remove_mask;
2049 GOTO(out, rc = count);
2052 OBD_FREE(buf, count + 1);
2055 LPROC_SEQ_FOPS(mdt_hsm_policy);
2057 #define GENERATE_PROC_METHOD(VAR) \
2058 static int mdt_hsm_##VAR##_seq_show(struct seq_file *m, void *data) \
2060 struct mdt_device *mdt = m->private; \
2061 struct coordinator *cdt = &mdt->mdt_coordinator; \
2064 seq_printf(m, "%llu\n", (__u64)cdt->VAR); \
2068 mdt_hsm_##VAR##_seq_write(struct file *file, const char __user *buffer, \
2069 size_t count, loff_t *off) \
2072 struct seq_file *m = file->private_data; \
2073 struct mdt_device *mdt = m->private; \
2074 struct coordinator *cdt = &mdt->mdt_coordinator; \
2079 rc = lprocfs_str_to_s64(buffer, count, &val); \
2082 if (val > 0 && val < INT_MAX) { \
2089 GENERATE_PROC_METHOD(cdt_loop_period)
2090 GENERATE_PROC_METHOD(cdt_grace_delay)
2091 GENERATE_PROC_METHOD(cdt_active_req_timeout)
2092 GENERATE_PROC_METHOD(cdt_max_requests)
2093 GENERATE_PROC_METHOD(cdt_default_archive_id)
2096 * procfs write method for MDT/hsm_control
2097 * proc entry is in mdt directory so data is mdt obd_device pointer
2099 #define CDT_ENABLE_CMD "enabled"
2100 #define CDT_STOP_CMD "shutdown"
2101 #define CDT_DISABLE_CMD "disabled"
2102 #define CDT_PURGE_CMD "purge"
2103 #define CDT_HELP_CMD "help"
2104 #define CDT_MAX_CMD_LEN 10
2107 mdt_hsm_cdt_control_seq_write(struct file *file, const char __user *buffer,
2108 size_t count, loff_t *off)
2110 struct seq_file *m = file->private_data;
2111 struct obd_device *obd = m->private;
2112 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2113 struct coordinator *cdt = &(mdt->mdt_coordinator);
2115 char kernbuf[CDT_MAX_CMD_LEN];
2118 if (count == 0 || count >= sizeof(kernbuf))
2121 if (copy_from_user(kernbuf, buffer, count))
2125 if (kernbuf[count - 1] == '\n')
2126 kernbuf[count - 1] = 0;
2129 if (strcmp(kernbuf, CDT_ENABLE_CMD) == 0) {
2130 if (cdt->cdt_state == CDT_DISABLE) {
2131 rc = set_cdt_state(cdt, CDT_RUNNING, NULL);
2132 mdt_hsm_cdt_event(cdt);
2133 wake_up(&cdt->cdt_waitq);
2135 rc = mdt_hsm_cdt_start(mdt);
2137 } else if (strcmp(kernbuf, CDT_STOP_CMD) == 0) {
2138 if ((cdt->cdt_state == CDT_STOPPING) ||
2139 (cdt->cdt_state == CDT_STOPPED)) {
2140 CERROR("%s: Coordinator already stopped\n",
2144 rc = mdt_hsm_cdt_stop(mdt);
2146 } else if (strcmp(kernbuf, CDT_DISABLE_CMD) == 0) {
2147 if ((cdt->cdt_state == CDT_STOPPING) ||
2148 (cdt->cdt_state == CDT_STOPPED)) {
2149 CERROR("%s: Coordinator is stopped\n",
2153 rc = set_cdt_state(cdt, CDT_DISABLE, NULL);
2155 } else if (strcmp(kernbuf, CDT_PURGE_CMD) == 0) {
2156 rc = hsm_cancel_all_actions(mdt);
2157 } else if (strcmp(kernbuf, CDT_HELP_CMD) == 0) {
2165 CERROR("%s: Valid coordinator control commands are: "
2166 "%s %s %s %s %s\n", mdt_obd_name(mdt),
2167 CDT_ENABLE_CMD, CDT_STOP_CMD, CDT_DISABLE_CMD,
2168 CDT_PURGE_CMD, CDT_HELP_CMD);
2176 int mdt_hsm_cdt_control_seq_show(struct seq_file *m, void *data)
2178 struct obd_device *obd = m->private;
2179 struct coordinator *cdt;
2182 cdt = &(mdt_dev(obd->obd_lu_dev)->mdt_coordinator);
2184 seq_printf(m, "%s\n", cdt_mdt_state2str(cdt->cdt_state));
2190 mdt_hsm_request_mask_show(struct seq_file *m, __u64 mask)
2196 for (i = 0; i < 8 * sizeof(mask); i++) {
2197 if (mask & (1UL << i)) {
2198 seq_printf(m, "%s%s", first ? "" : " ",
2199 hsm_copytool_action2name(i));
2209 mdt_hsm_user_request_mask_seq_show(struct seq_file *m, void *data)
2211 struct mdt_device *mdt = m->private;
2212 struct coordinator *cdt = &mdt->mdt_coordinator;
2214 return mdt_hsm_request_mask_show(m, cdt->cdt_user_request_mask);
2218 mdt_hsm_group_request_mask_seq_show(struct seq_file *m, void *data)
2220 struct mdt_device *mdt = m->private;
2221 struct coordinator *cdt = &mdt->mdt_coordinator;
2223 return mdt_hsm_request_mask_show(m, cdt->cdt_group_request_mask);
2227 mdt_hsm_other_request_mask_seq_show(struct seq_file *m, void *data)
2229 struct mdt_device *mdt = m->private;
2230 struct coordinator *cdt = &mdt->mdt_coordinator;
2232 return mdt_hsm_request_mask_show(m, cdt->cdt_other_request_mask);
2235 static inline enum hsm_copytool_action
2236 hsm_copytool_name2action(const char *name)
2238 if (strcasecmp(name, "NOOP") == 0)
2240 else if (strcasecmp(name, "ARCHIVE") == 0)
2241 return HSMA_ARCHIVE;
2242 else if (strcasecmp(name, "RESTORE") == 0)
2243 return HSMA_RESTORE;
2244 else if (strcasecmp(name, "REMOVE") == 0)
2246 else if (strcasecmp(name, "CANCEL") == 0)
2253 mdt_write_hsm_request_mask(struct file *file, const char __user *user_buf,
2254 size_t user_count, __u64 *mask)
2256 char *buf, *pos, *name;
2262 if (!(user_count < 4096))
2265 buf_size = user_count + 1;
2267 OBD_ALLOC(buf, buf_size);
2271 if (copy_from_user(buf, user_buf, buf_size - 1))
2272 GOTO(out, rc = -EFAULT);
2274 buf[buf_size - 1] = '\0';
2277 while ((name = strsep(&pos, " \t\v\n")) != NULL) {
2283 action = hsm_copytool_name2action(name);
2285 GOTO(out, rc = -EINVAL);
2287 new_mask |= (1UL << action);
2293 OBD_FREE(buf, buf_size);
2299 mdt_hsm_user_request_mask_seq_write(struct file *file, const char __user *buf,
2300 size_t count, loff_t *off)
2302 struct seq_file *m = file->private_data;
2303 struct mdt_device *mdt = m->private;
2304 struct coordinator *cdt = &mdt->mdt_coordinator;
2306 return mdt_write_hsm_request_mask(file, buf, count,
2307 &cdt->cdt_user_request_mask);
2311 mdt_hsm_group_request_mask_seq_write(struct file *file, const char __user *buf,
2312 size_t count, loff_t *off)
2314 struct seq_file *m = file->private_data;
2315 struct mdt_device *mdt = m->private;
2316 struct coordinator *cdt = &mdt->mdt_coordinator;
2318 return mdt_write_hsm_request_mask(file, buf, count,
2319 &cdt->cdt_group_request_mask);
2323 mdt_hsm_other_request_mask_seq_write(struct file *file, const char __user *buf,
2324 size_t count, loff_t *off)
2326 struct seq_file *m = file->private_data;
2327 struct mdt_device *mdt = m->private;
2328 struct coordinator *cdt = &mdt->mdt_coordinator;
2330 return mdt_write_hsm_request_mask(file, buf, count,
2331 &cdt->cdt_other_request_mask);
2334 static int mdt_hsm_cdt_raolu_seq_show(struct seq_file *m, void *data)
2336 struct mdt_device *mdt = m->private;
2337 struct coordinator *cdt = &mdt->mdt_coordinator;
2340 seq_printf(m, "%d\n", (int)cdt->cdt_remove_archive_on_last_unlink);
2345 mdt_hsm_cdt_raolu_seq_write(struct file *file, const char __user *buffer,
2346 size_t count, loff_t *off)
2349 struct seq_file *m = file->private_data;
2350 struct mdt_device *mdt = m->private;
2351 struct coordinator *cdt = &mdt->mdt_coordinator;
2356 rc = lprocfs_str_to_s64(buffer, count, &val);
2360 cdt->cdt_remove_archive_on_last_unlink = val;
2364 LPROC_SEQ_FOPS(mdt_hsm_cdt_loop_period);
2365 LPROC_SEQ_FOPS(mdt_hsm_cdt_grace_delay);
2366 LPROC_SEQ_FOPS(mdt_hsm_cdt_active_req_timeout);
2367 LPROC_SEQ_FOPS(mdt_hsm_cdt_max_requests);
2368 LPROC_SEQ_FOPS(mdt_hsm_cdt_default_archive_id);
2369 LPROC_SEQ_FOPS(mdt_hsm_user_request_mask);
2370 LPROC_SEQ_FOPS(mdt_hsm_group_request_mask);
2371 LPROC_SEQ_FOPS(mdt_hsm_other_request_mask);
2372 LPROC_SEQ_FOPS(mdt_hsm_cdt_raolu);
2374 /* Read-only proc files for request counters */
2375 static int mdt_hsm_cdt_archive_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_archive_count));
2385 static int mdt_hsm_cdt_restore_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_restore_count));
2395 static int mdt_hsm_cdt_remove_count_seq_show(struct seq_file *m, void *data)
2397 struct mdt_device *mdt = m->private;
2398 struct coordinator *cdt = &mdt->mdt_coordinator;
2401 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_remove_count));
2405 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_archive_count);
2406 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_restore_count);
2407 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_remove_count);
2409 static struct lprocfs_vars lprocfs_mdt_hsm_vars[] = {
2411 .fops = &mdt_hsm_agent_fops },
2412 { .name = "actions",
2413 .fops = &mdt_hsm_actions_fops,
2414 .proc_mode = 0444 },
2415 { .name = "default_archive_id",
2416 .fops = &mdt_hsm_cdt_default_archive_id_fops },
2417 { .name = "grace_delay",
2418 .fops = &mdt_hsm_cdt_grace_delay_fops },
2419 { .name = "loop_period",
2420 .fops = &mdt_hsm_cdt_loop_period_fops },
2421 { .name = "max_requests",
2422 .fops = &mdt_hsm_cdt_max_requests_fops },
2424 .fops = &mdt_hsm_policy_fops },
2425 { .name = "active_request_timeout",
2426 .fops = &mdt_hsm_cdt_active_req_timeout_fops },
2427 { .name = "active_requests",
2428 .fops = &mdt_hsm_active_requests_fops },
2429 { .name = "user_request_mask",
2430 .fops = &mdt_hsm_user_request_mask_fops, },
2431 { .name = "group_request_mask",
2432 .fops = &mdt_hsm_group_request_mask_fops, },
2433 { .name = "other_request_mask",
2434 .fops = &mdt_hsm_other_request_mask_fops, },
2435 { .name = "remove_archive_on_last_unlink",
2436 .fops = &mdt_hsm_cdt_raolu_fops, },
2437 { .name = "archive_count",
2438 .fops = &mdt_hsm_cdt_archive_count_fops, },
2439 { .name = "restore_count",
2440 .fops = &mdt_hsm_cdt_restore_count_fops, },
2441 { .name = "remove_count",
2442 .fops = &mdt_hsm_cdt_remove_count_fops, },