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_progress_kernel pgs;
278 struct cdt_agent_req *car;
279 time64_t now = ktime_get_real_seconds();
283 if (!hsd->housekeeping)
286 /* we search for a running request
287 * error may happen if coordinator crashes or stopped
288 * with running request
290 car = mdt_cdt_find_request(cdt, larr->arr_hai.hai_cookie);
292 last = larr->arr_req_change;
294 last = car->car_req_update;
295 mdt_cdt_put_request(car);
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)
303 dump_llog_agent_req_rec("request timed out, start cleaning", larr);
304 /* a too old cancel request just needs to be removed
305 * this can happen, if copy tool does not support
306 * cancel for other requests, we have to remove the
307 * running request and notify the copytool */
308 pgs.hpk_fid = larr->arr_hai.hai_fid;
309 pgs.hpk_cookie = larr->arr_hai.hai_cookie;
310 pgs.hpk_extent = larr->arr_hai.hai_extent;
311 pgs.hpk_flags = HP_FLAG_COMPLETED;
312 pgs.hpk_errval = ENOSYS;
313 pgs.hpk_data_version = 0;
315 /* update request state, but do not record in llog, to
316 * avoid deadlock on cdt_llog_lock */
317 rc = mdt_hsm_update_request_state(hsd->mti, &pgs, 0);
319 CERROR("%s: cannot cleanup timed out request: "
320 DFID" for cookie %#llx action=%s\n",
322 PFID(&pgs.hpk_fid), pgs.hpk_cookie,
323 hsm_copytool_action2name(larr->arr_hai.hai_action));
326 /* The request no longer exists, forget
327 * about it, and do not send a cancel request
328 * to the client, for which an error will be
329 * sent back, leading to an endless cycle of
331 cdt_agent_record_hash_del(cdt, larr->arr_hai.hai_cookie);
332 RETURN(LLOG_DEL_RECORD);
335 /* XXX A cancel request cannot be cancelled. */
336 if (larr->arr_hai.hai_action == HSMA_CANCEL)
339 larr->arr_status = ARS_CANCELED;
340 larr->arr_req_change = now;
341 rc = llog_write(hsd->mti->mti_env, llh, &larr->arr_hdr,
342 larr->arr_hdr.lrh_index);
344 CERROR("%s: cannot update agent log: rc = %d\n",
345 mdt_obd_name(mdt), rc);
351 * llog_cat_process() callback, used to:
352 * - find waiting request and start action
353 * - purge canceled and done requests
354 * \param env [IN] environment
355 * \param llh [IN] llog handle
356 * \param hdr [IN] llog record
357 * \param data [IN/OUT] cb data = struct hsm_scan_data
359 * \retval -ve failure
361 static int mdt_coordinator_cb(const struct lu_env *env,
362 struct llog_handle *llh,
363 struct llog_rec_hdr *hdr,
366 struct llog_agent_req_rec *larr = (struct llog_agent_req_rec *)hdr;
367 struct hsm_scan_data *hsd = data;
368 struct mdt_device *mdt = hsd->mti->mti_mdt;
369 struct coordinator *cdt = &mdt->mdt_coordinator;
372 larr = (struct llog_agent_req_rec *)hdr;
373 dump_llog_agent_req_rec("mdt_coordinator_cb(): ", larr);
374 switch (larr->arr_status) {
376 RETURN(mdt_cdt_waiting_cb(env, mdt, llh, larr, hsd));
378 RETURN(mdt_cdt_started_cb(env, mdt, llh, larr, hsd));
380 if (!hsd->housekeeping)
383 if ((larr->arr_req_change + cdt->cdt_grace_delay) <
384 ktime_get_real_seconds()) {
385 cdt_agent_record_hash_del(cdt,
386 larr->arr_hai.hai_cookie);
387 RETURN(LLOG_DEL_RECORD);
395 * create /proc entries for coordinator
398 * \retval -ve failure
400 int hsm_cdt_procfs_init(struct mdt_device *mdt)
402 struct coordinator *cdt = &mdt->mdt_coordinator;
406 /* init /proc entries, failure is not critical */
407 cdt->cdt_proc_dir = lprocfs_register("hsm",
408 mdt2obd_dev(mdt)->obd_proc_entry,
409 lprocfs_mdt_hsm_vars, mdt);
410 if (IS_ERR(cdt->cdt_proc_dir)) {
411 rc = PTR_ERR(cdt->cdt_proc_dir);
412 CERROR("%s: Cannot create 'hsm' directory in mdt proc dir,"
413 " rc=%d\n", mdt_obd_name(mdt), rc);
414 cdt->cdt_proc_dir = NULL;
422 * remove /proc entries for coordinator
425 void hsm_cdt_procfs_fini(struct mdt_device *mdt)
427 struct coordinator *cdt = &mdt->mdt_coordinator;
429 if (cdt->cdt_proc_dir != NULL)
430 lprocfs_remove(&cdt->cdt_proc_dir);
434 * get vector of hsm cdt /proc vars
438 struct lprocfs_vars *hsm_cdt_get_proc_vars(void)
440 return lprocfs_mdt_hsm_vars;
443 /* Release the ressource used by the coordinator. Called when the
444 * coordinator is stopping. */
445 static void mdt_hsm_cdt_cleanup(struct mdt_device *mdt)
447 struct coordinator *cdt = &mdt->mdt_coordinator;
448 struct cdt_agent_req *car, *tmp1;
449 struct hsm_agent *ha, *tmp2;
450 struct cdt_restore_handle *crh, *tmp3;
451 struct mdt_thread_info *cdt_mti;
454 down_write(&cdt->cdt_request_lock);
455 list_for_each_entry_safe(car, tmp1, &cdt->cdt_request_list,
457 cfs_hash_del(cdt->cdt_request_cookie_hash,
458 &car->car_hai->hai_cookie,
459 &car->car_cookie_hash);
460 list_del(&car->car_request_list);
461 mdt_cdt_put_request(car);
463 up_write(&cdt->cdt_request_lock);
465 down_write(&cdt->cdt_agent_lock);
466 list_for_each_entry_safe(ha, tmp2, &cdt->cdt_agents, ha_list) {
467 list_del(&ha->ha_list);
470 up_write(&cdt->cdt_agent_lock);
472 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
473 mutex_lock(&cdt->cdt_restore_lock);
474 list_for_each_entry_safe(crh, tmp3, &cdt->cdt_restore_hdl, crh_list) {
475 list_del(&crh->crh_list);
476 /* give back layout lock */
477 mdt_object_unlock(cdt_mti, NULL, &crh->crh_lh, 1);
478 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
480 mutex_unlock(&cdt->cdt_restore_lock);
484 * Coordinator state transition table, indexed on enum cdt_states, taking
485 * from and to states. For instance since CDT_INIT to CDT_RUNNING is a
486 * valid transition, cdt_transition[CDT_INIT][CDT_RUNNING] is true.
488 static bool cdt_transition[CDT_STATES_COUNT][CDT_STATES_COUNT] = {
489 /* from -> to: stopped init running disable stopping */
490 /* stopped */ { true, true, false, false, false },
491 /* init */ { true, false, true, false, false },
492 /* running */ { false, false, true, true, true },
493 /* disable */ { false, false, true, true, true },
494 /* stopping */ { true, false, false, false, false }
498 * Change coordinator thread state
499 * Some combinations are not valid, so catch them here.
501 * Returns 0 on success, with old_state set if not NULL, or -EINVAL if
502 * the transition was not possible.
504 static int set_cdt_state(struct coordinator *cdt, enum cdt_states new_state,
505 enum cdt_states *old_state)
508 enum cdt_states state;
510 spin_lock(&cdt->cdt_state_lock);
512 state = cdt->cdt_state;
514 if (cdt_transition[state][new_state]) {
515 cdt->cdt_state = new_state;
516 spin_unlock(&cdt->cdt_state_lock);
521 spin_unlock(&cdt->cdt_state_lock);
523 "unexpected coordinator transition, from=%s, to=%s\n",
524 cdt_mdt_state2str(state), cdt_mdt_state2str(new_state));
533 * \param data [IN] obd device
535 * \retval -ve failure
537 static int mdt_coordinator(void *data)
539 struct hsm_thread_data *thread_data = data;
540 struct mdt_thread_info *mti = thread_data->cdt_mti;
541 struct mdt_device *mdt = mti->mti_mdt;
542 struct coordinator *cdt = &mdt->mdt_coordinator;
543 struct hsm_scan_data hsd = { NULL };
544 time64_t last_housekeeping = 0;
549 /* set up hsd->request and max_requests */
550 hsd.max_requests = cdt->cdt_max_requests;
551 request_sz = hsd.max_requests * sizeof(*hsd.request);
552 hsd.request = thread_data->request;
554 CDEBUG(D_HSM, "%s: coordinator thread starting, pid=%d\n",
555 mdt_obd_name(mdt), current_pid());
558 obd_uuid2fsname(hsd.fs_name, mdt_obd_name(mdt), MTI_NAME_MAXLEN);
560 set_cdt_state(cdt, CDT_RUNNING, NULL);
562 /* Inform mdt_hsm_cdt_start(). */
563 wake_up_all(&cdt->cdt_waitq);
570 struct hsm_record_update *updates;
572 /* Limit execution of the expensive requests traversal
573 * to at most one second. This prevents repeatedly
574 * locking/unlocking the catalog for each request
575 * and preventing other HSM operations from happening
577 wait_event_interruptible_timeout(cdt->cdt_waitq,
578 kthread_should_stop() ||
579 cdt->cdt_wakeup_coordinator,
580 cfs_time_seconds(1));
582 cdt->cdt_wakeup_coordinator = false;
583 CDEBUG(D_HSM, "coordinator resumes\n");
585 if (kthread_should_stop()) {
586 CDEBUG(D_HSM, "Coordinator stops\n");
591 /* if coordinator is suspended continue to wait */
592 if (cdt->cdt_state == CDT_DISABLE) {
593 CDEBUG(D_HSM, "disable state, coordinator sleeps\n");
597 /* If no event, and no housekeeping to do, continue to
599 if (last_housekeeping + cdt->cdt_loop_period <=
600 ktime_get_real_seconds()) {
601 last_housekeeping = ktime_get_real_seconds();
602 hsd.housekeeping = true;
603 } else if (cdt->cdt_event) {
604 hsd.housekeeping = false;
609 cdt->cdt_event = false;
611 CDEBUG(D_HSM, "coordinator starts reading llog\n");
613 if (hsd.max_requests != cdt->cdt_max_requests) {
614 /* cdt_max_requests has changed,
615 * we need to allocate a new buffer
617 struct hsm_scan_request *tmp = NULL;
618 int max_requests = cdt->cdt_max_requests;
619 OBD_ALLOC_LARGE(tmp, max_requests *
620 sizeof(struct hsm_scan_request));
622 CERROR("Failed to resize request buffer, "
623 "keeping it at %d\n",
625 cdt->cdt_max_requests = hsd.max_requests;
627 OBD_FREE_LARGE(hsd.request, request_sz);
628 hsd.max_requests = max_requests;
629 request_sz = hsd.max_requests *
630 sizeof(struct hsm_scan_request);
637 rc = cdt_llog_process(mti->mti_env, mdt, mdt_coordinator_cb,
642 CDEBUG(D_HSM, "found %d requests to send\n", hsd.request_cnt);
644 if (list_empty(&cdt->cdt_agents)) {
645 CDEBUG(D_HSM, "no agent available, "
646 "coordinator sleeps\n");
650 /* Compute how many HAI we have in all the requests */
652 for (i = 0; i < hsd.request_cnt; i++) {
653 const struct hsm_scan_request *request =
656 updates_cnt += request->hal->hal_count;
659 /* Allocate a temporary array to store the cookies to
660 * update, and their status. */
661 updates_sz = updates_cnt * sizeof(*updates);
662 OBD_ALLOC(updates, updates_sz);
663 if (updates == NULL) {
664 CERROR("%s: Cannot allocate memory (%d o) "
666 mdt_obd_name(mdt), updates_sz, updates_cnt);
670 /* here hsd contains a list of requests to be started */
671 for (i = 0; i < hsd.request_cnt; i++) {
672 struct hsm_scan_request *request = &hsd.request[i];
673 struct hsm_action_list *hal = request->hal;
674 struct hsm_action_item *hai;
677 /* still room for work ? */
678 if (atomic_read(&cdt->cdt_request_count) >=
679 cdt->cdt_max_requests)
682 rc = mdt_hsm_agent_send(mti, hal, 0);
683 /* if failure, we suppose it is temporary
684 * if the copy tool failed to do the request
685 * it has to use hsm_progress
688 /* set up cookie vector to set records status
689 * after copy tools start or failed
691 hai = hai_first(hal);
692 for (j = 0; j < hal->hal_count; j++) {
693 updates[update_idx].cookie = hai->hai_cookie;
694 updates[update_idx].status =
695 (rc ? ARS_WAITING : ARS_STARTED);
702 rc = mdt_agent_record_update(mti->mti_env, mdt,
703 updates, update_idx);
705 CERROR("%s: mdt_agent_record_update() failed, "
706 "rc=%d, cannot update records "
708 mdt_obd_name(mdt), rc, update_idx);
711 OBD_FREE(updates, updates_sz);
714 /* free hal allocated by callback */
715 for (i = 0; i < hsd.request_cnt; i++) {
716 struct hsm_scan_request *request = &hsd.request[i];
718 OBD_FREE(request->hal, request->hal_sz);
723 OBD_FREE_LARGE(hsd.request, request_sz);
725 mdt_hsm_cdt_cleanup(mdt);
728 CERROR("%s: coordinator thread exiting, process=%d, rc=%d\n",
729 mdt_obd_name(mdt), current_pid(), rc);
731 CDEBUG(D_HSM, "%s: coordinator thread exiting, process=%d,"
733 mdt_obd_name(mdt), current_pid());
739 * lookup a restore handle by FID
740 * caller needs to hold cdt_restore_lock
741 * \param cdt [IN] coordinator
742 * \param fid [IN] FID
743 * \retval cdt_restore_handle found
744 * \retval NULL not found
746 struct cdt_restore_handle *mdt_hsm_restore_hdl_find(struct coordinator *cdt,
747 const struct lu_fid *fid)
749 struct cdt_restore_handle *crh;
752 list_for_each_entry(crh, &cdt->cdt_restore_hdl, crh_list) {
753 if (lu_fid_eq(&crh->crh_fid, fid))
760 * data passed to llog_cat_process() callback
761 * to scan requests and take actions
763 struct hsm_restore_data {
764 struct mdt_thread_info *hrd_mti;
768 * llog_cat_process() callback, used to:
769 * - find restore request and allocate the restore handle
770 * \param env [IN] environment
771 * \param llh [IN] llog handle
772 * \param hdr [IN] llog record
773 * \param data [IN/OUT] cb data = struct hsm_restore_data
775 * \retval -ve failure
777 static int hsm_restore_cb(const struct lu_env *env,
778 struct llog_handle *llh,
779 struct llog_rec_hdr *hdr, void *data)
781 struct llog_agent_req_rec *larr;
782 struct hsm_restore_data *hrd;
783 struct cdt_restore_handle *crh;
784 struct hsm_action_item *hai;
785 struct mdt_thread_info *mti;
786 struct coordinator *cdt;
787 struct mdt_object *child;
793 cdt = &mti->mti_mdt->mdt_coordinator;
795 larr = (struct llog_agent_req_rec *)hdr;
796 hai = &larr->arr_hai;
797 if (hai->hai_cookie > cdt->cdt_last_cookie)
798 /* update the cookie to avoid collision */
799 cdt->cdt_last_cookie = hai->hai_cookie + 1;
801 if (hai->hai_action != HSMA_RESTORE ||
802 agent_req_in_final_state(larr->arr_status))
805 /* restore request not in a final state */
807 /* force replay of restore requests left in started state from previous
808 * CDT context, to be canceled later if finally found to be incompatible
809 * when being re-started */
810 if (larr->arr_status == ARS_STARTED) {
811 larr->arr_status = ARS_WAITING;
812 larr->arr_req_change = ktime_get_real_seconds();
813 rc = llog_write(env, llh, hdr, hdr->lrh_index);
818 OBD_SLAB_ALLOC_PTR(crh, mdt_hsm_cdt_kmem);
822 crh->crh_fid = hai->hai_fid;
823 /* in V1 all file is restored
824 crh->extent.start = hai->hai_extent.offset;
825 crh->extent.end = hai->hai_extent.offset + hai->hai_extent.length;
827 crh->crh_extent.start = 0;
828 crh->crh_extent.end = hai->hai_extent.length;
829 /* get the layout lock */
830 mdt_lock_reg_init(&crh->crh_lh, LCK_EX);
831 child = mdt_object_find_lock(mti, &crh->crh_fid, &crh->crh_lh,
832 MDS_INODELOCK_LAYOUT);
834 GOTO(out, rc = PTR_ERR(child));
837 /* we choose to not keep a reference
838 * on the object during the restore time which can be very long */
839 mdt_object_put(mti->mti_env, child);
841 mutex_lock(&cdt->cdt_restore_lock);
842 list_add_tail(&crh->crh_list, &cdt->cdt_restore_hdl);
843 mutex_unlock(&cdt->cdt_restore_lock);
850 * restore coordinator state at startup
851 * the goal is to take a layout lock for each registered restore request
852 * \param mti [IN] context
854 static int mdt_hsm_pending_restore(struct mdt_thread_info *mti)
856 struct hsm_restore_data hrd;
862 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, hsm_restore_cb, &hrd,
868 static int hsm_init_ucred(struct lu_ucred *uc)
872 uc->uc_valid = UCRED_OLD;
881 uc->uc_suppgids[0] = -1;
882 uc->uc_suppgids[1] = -1;
883 uc->uc_cap = CFS_CAP_FS_MASK;
886 uc->uc_identity = NULL;
892 * initialize coordinator struct
893 * \param mdt [IN] device
895 * \retval -ve failure
897 int mdt_hsm_cdt_init(struct mdt_device *mdt)
899 struct coordinator *cdt = &mdt->mdt_coordinator;
900 struct mdt_thread_info *cdt_mti = NULL;
904 init_waitqueue_head(&cdt->cdt_waitq);
905 init_rwsem(&cdt->cdt_llog_lock);
906 init_rwsem(&cdt->cdt_agent_lock);
907 init_rwsem(&cdt->cdt_request_lock);
908 mutex_init(&cdt->cdt_restore_lock);
909 spin_lock_init(&cdt->cdt_state_lock);
910 set_cdt_state(cdt, CDT_STOPPED, NULL);
912 INIT_LIST_HEAD(&cdt->cdt_request_list);
913 INIT_LIST_HEAD(&cdt->cdt_agents);
914 INIT_LIST_HEAD(&cdt->cdt_restore_hdl);
916 cdt->cdt_request_cookie_hash = cfs_hash_create("REQUEST_COOKIE_HASH",
923 &cdt_request_cookie_hash_ops,
925 if (cdt->cdt_request_cookie_hash == NULL)
928 cdt->cdt_agent_record_hash = cfs_hash_create("AGENT_RECORD_HASH",
935 &cdt_agent_record_hash_ops,
937 if (cdt->cdt_agent_record_hash == NULL)
938 GOTO(out_request_cookie_hash, rc = -ENOMEM);
940 rc = lu_env_init(&cdt->cdt_env, LCT_MD_THREAD);
942 GOTO(out_agent_record_hash, rc);
944 /* for mdt_ucred(), lu_ucred stored in lu_ucred_key */
945 rc = lu_context_init(&cdt->cdt_session, LCT_SERVER_SESSION);
949 lu_context_enter(&cdt->cdt_session);
950 cdt->cdt_env.le_ses = &cdt->cdt_session;
952 cdt_mti = lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
953 LASSERT(cdt_mti != NULL);
955 cdt_mti->mti_env = &cdt->cdt_env;
956 cdt_mti->mti_mdt = mdt;
958 hsm_init_ucred(mdt_ucred(cdt_mti));
960 /* default values for /proc tunnables
961 * can be override by MGS conf */
962 cdt->cdt_default_archive_id = 1;
963 cdt->cdt_grace_delay = 60;
964 cdt->cdt_loop_period = 10;
965 cdt->cdt_max_requests = 3;
966 cdt->cdt_policy = CDT_DEFAULT_POLICY;
967 cdt->cdt_active_req_timeout = 3600;
969 /* Initialize cdt_compound_id here to allow its usage for
970 * delayed requests from RAoLU policy */
971 atomic_set(&cdt->cdt_compound_id, ktime_get_real_seconds());
973 /* by default do not remove archives on last unlink */
974 cdt->cdt_remove_archive_on_last_unlink = false;
979 lu_env_fini(&cdt->cdt_env);
980 out_agent_record_hash:
981 cfs_hash_putref(cdt->cdt_agent_record_hash);
982 cdt->cdt_agent_record_hash = NULL;
983 out_request_cookie_hash:
984 cfs_hash_putref(cdt->cdt_request_cookie_hash);
985 cdt->cdt_request_cookie_hash = NULL;
991 * free a coordinator thread
992 * \param mdt [IN] device
994 int mdt_hsm_cdt_fini(struct mdt_device *mdt)
996 struct coordinator *cdt = &mdt->mdt_coordinator;
999 lu_context_exit(cdt->cdt_env.le_ses);
1000 lu_context_fini(cdt->cdt_env.le_ses);
1002 lu_env_fini(&cdt->cdt_env);
1004 cfs_hash_putref(cdt->cdt_agent_record_hash);
1005 cdt->cdt_agent_record_hash = NULL;
1007 cfs_hash_putref(cdt->cdt_request_cookie_hash);
1008 cdt->cdt_request_cookie_hash = NULL;
1014 * start a coordinator thread
1015 * \param mdt [IN] device
1017 * \retval -ve failure
1019 static int mdt_hsm_cdt_start(struct mdt_device *mdt)
1021 struct coordinator *cdt = &mdt->mdt_coordinator;
1024 struct task_struct *task;
1026 struct hsm_thread_data thread_data;
1029 /* functions defined but not yet used
1030 * this avoid compilation warning
1032 ptr = dump_requests;
1034 rc = set_cdt_state(cdt, CDT_INIT, NULL);
1036 CERROR("%s: Coordinator already started or stopping\n",
1041 CLASSERT(1 << (CDT_POLICY_SHIFT_COUNT - 1) == CDT_POLICY_LAST);
1042 cdt->cdt_policy = CDT_DEFAULT_POLICY;
1044 /* just need to be larger than previous one */
1045 /* cdt_last_cookie is protected by cdt_llog_lock */
1046 cdt->cdt_last_cookie = ktime_get_real_seconds();
1047 atomic_set(&cdt->cdt_request_count, 0);
1048 atomic_set(&cdt->cdt_archive_count, 0);
1049 atomic_set(&cdt->cdt_restore_count, 0);
1050 atomic_set(&cdt->cdt_remove_count, 0);
1051 cdt->cdt_user_request_mask = (1UL << HSMA_RESTORE);
1052 cdt->cdt_group_request_mask = (1UL << HSMA_RESTORE);
1053 cdt->cdt_other_request_mask = (1UL << HSMA_RESTORE);
1055 /* to avoid deadlock when start is made through /proc
1056 * /proc entries are created by the coordinator thread */
1058 /* set up list of started restore requests */
1059 thread_data.cdt_mti =
1060 lu_context_key_get(&cdt->cdt_env.le_ctx, &mdt_thread_key);
1061 rc = mdt_hsm_pending_restore(thread_data.cdt_mti);
1063 CERROR("%s: cannot take the layout locks needed"
1064 " for registered restore: %d\n",
1065 mdt_obd_name(mdt), rc);
1067 if (mdt->mdt_bottom->dd_rdonly)
1070 /* Allocate the initial hsd.request[] vector*/
1071 request_sz = cdt->cdt_max_requests * sizeof(struct hsm_scan_request);
1072 OBD_ALLOC_LARGE(thread_data.request, request_sz);
1073 if (!thread_data.request) {
1074 set_cdt_state(cdt, CDT_STOPPED, NULL);
1078 task = kthread_run(mdt_coordinator, &thread_data, "hsm_cdtr");
1081 set_cdt_state(cdt, CDT_STOPPED, NULL);
1082 OBD_FREE(thread_data.request, request_sz);
1083 CERROR("%s: error starting coordinator thread: %d\n",
1084 mdt_obd_name(mdt), rc);
1086 cdt->cdt_task = task;
1087 wait_event(cdt->cdt_waitq,
1088 cdt->cdt_state != CDT_INIT);
1089 CDEBUG(D_HSM, "%s: coordinator thread started\n",
1098 * stop a coordinator thread
1099 * \param mdt [IN] device
1101 int mdt_hsm_cdt_stop(struct mdt_device *mdt)
1103 struct coordinator *cdt = &mdt->mdt_coordinator;
1107 /* stop coordinator thread */
1108 rc = set_cdt_state(cdt, CDT_STOPPING, NULL);
1110 kthread_stop(cdt->cdt_task);
1111 cdt->cdt_task = NULL;
1112 set_cdt_state(cdt, CDT_STOPPED, NULL);
1118 static int mdt_hsm_set_exists(struct mdt_thread_info *mti,
1119 const struct lu_fid *fid,
1122 struct mdt_object *obj;
1126 obj = mdt_hsm_get_md_hsm(mti, fid, &mh);
1128 GOTO(out, rc = PTR_ERR(obj));
1130 if (mh.mh_flags & HS_EXISTS &&
1131 mh.mh_arch_id == archive_id)
1132 GOTO(out_obj, rc = 0);
1134 mh.mh_flags |= HS_EXISTS;
1135 mh.mh_arch_id = archive_id;
1136 rc = mdt_hsm_attr_set(mti, obj, &mh);
1139 mdt_object_put(mti->mti_env, obj);
1145 * register all requests from an hal in the memory list
1146 * \param mti [IN] context
1147 * \param hal [IN] request
1148 * \param uuid [OUT] in case of CANCEL, the uuid of the agent
1149 * which is running the CT
1151 * \retval -ve failure
1153 int mdt_hsm_add_hal(struct mdt_thread_info *mti,
1154 struct hsm_action_list *hal, struct obd_uuid *uuid)
1156 struct mdt_device *mdt = mti->mti_mdt;
1157 struct coordinator *cdt = &mdt->mdt_coordinator;
1158 struct hsm_action_item *hai;
1162 /* register request in memory list */
1163 hai = hai_first(hal);
1164 for (i = 0; i < hal->hal_count; i++, hai = hai_next(hai)) {
1165 struct cdt_agent_req *car;
1167 /* in case of a cancel request, we first mark the ondisk
1168 * record of the request we want to stop as canceled
1169 * this does not change the cancel record
1170 * it will be done when updating the request status
1172 if (hai->hai_action == HSMA_CANCEL) {
1173 struct hsm_record_update update = {
1174 .cookie = hai->hai_cookie,
1175 .status = ARS_CANCELED,
1178 rc = mdt_agent_record_update(mti->mti_env, mti->mti_mdt,
1181 CERROR("%s: mdt_agent_record_update() failed, "
1182 "rc=%d, cannot update status to %s "
1183 "for cookie %#llx\n",
1184 mdt_obd_name(mdt), rc,
1185 agent_req_status2name(ARS_CANCELED),
1190 /* find the running request to set it canceled */
1191 car = mdt_cdt_find_request(cdt, hai->hai_cookie);
1193 car->car_canceled = 1;
1194 /* uuid has to be changed to the one running the
1195 * request to cancel */
1196 *uuid = car->car_uuid;
1197 mdt_cdt_put_request(car);
1199 /* no need to memorize cancel request
1200 * this also avoid a deadlock when we receive
1201 * a purge all requests command
1206 if (hai->hai_action == HSMA_ARCHIVE) {
1207 rc = mdt_hsm_set_exists(mti, &hai->hai_fid,
1208 hal->hal_archive_id);
1215 car = mdt_cdt_alloc_request(hal->hal_compound_id,
1216 hal->hal_archive_id, hal->hal_flags,
1219 GOTO(out, rc = PTR_ERR(car));
1221 rc = mdt_cdt_add_request(cdt, car);
1223 mdt_cdt_free_request(car);
1230 * swap layouts between 2 fids
1231 * \param mti [IN] context
1234 * \param mh_common [IN] MD HSM
1236 static int hsm_swap_layouts(struct mdt_thread_info *mti,
1237 struct mdt_object *obj, const struct lu_fid *dfid,
1238 struct md_hsm *mh_common)
1240 struct mdt_object *dobj;
1241 struct mdt_lock_handle *dlh;
1245 if (!mdt_object_exists(obj))
1246 GOTO(out, rc = -ENOENT);
1248 /* we already have layout lock on obj so take only
1250 dlh = &mti->mti_lh[MDT_LH_OLD];
1251 mdt_lock_reg_init(dlh, LCK_EX);
1252 dobj = mdt_object_find_lock(mti, dfid, dlh, MDS_INODELOCK_LAYOUT);
1254 GOTO(out, rc = PTR_ERR(dobj));
1256 /* if copy tool closes the volatile before sending the final
1257 * progress through llapi_hsm_copy_end(), all the objects
1258 * are removed and mdd_swap_layout LBUG */
1259 if (!mdt_object_exists(dobj)) {
1260 CERROR("%s: Copytool has closed volatile file "DFID"\n",
1261 mdt_obd_name(mti->mti_mdt), PFID(dfid));
1262 GOTO(out_dobj, rc = -ENOENT);
1264 /* Since we only handle restores here, unconditionally use
1265 * SWAP_LAYOUTS_MDS_HSM flag to ensure original layout will
1266 * be preserved in case of failure during swap_layout and not
1267 * leave a file in an intermediate but incoherent state.
1268 * But need to setup HSM xattr of data FID before, reuse
1269 * mti and mh presets for FID in hsm_cdt_request_completed(),
1270 * only need to clear RELEASED and DIRTY.
1272 mh_common->mh_flags &= ~(HS_RELEASED | HS_DIRTY);
1273 rc = mdt_hsm_attr_set(mti, dobj, mh_common);
1275 rc = mo_swap_layouts(mti->mti_env,
1276 mdt_object_child(obj),
1277 mdt_object_child(dobj),
1278 SWAP_LAYOUTS_MDS_HSM);
1281 mdt_object_unlock_put(mti, dobj, dlh, 1);
1287 * update status of a completed request
1288 * \param mti [IN] context
1289 * \param pgs [IN] progress of the copy tool
1290 * \param update_record [IN] update llog record
1292 * \retval -ve failure
1294 static int hsm_cdt_request_completed(struct mdt_thread_info *mti,
1295 struct hsm_progress_kernel *pgs,
1296 const struct cdt_agent_req *car,
1297 enum agent_req_status *status)
1299 const struct lu_env *env = mti->mti_env;
1300 struct mdt_device *mdt = mti->mti_mdt;
1301 struct coordinator *cdt = &mdt->mdt_coordinator;
1302 struct mdt_object *obj = NULL;
1303 int cl_flags = 0, rc = 0;
1306 bool need_changelog = true;
1309 /* default is to retry */
1310 *status = ARS_WAITING;
1312 /* find object by FID, mdt_hsm_get_md_hsm() returns obj or err
1313 * if error/removed continue anyway to get correct reporting done */
1314 obj = mdt_hsm_get_md_hsm(mti, &car->car_hai->hai_fid, &mh);
1315 /* we will update MD HSM only if needed */
1316 is_mh_changed = false;
1318 /* no need to change mh->mh_arch_id
1319 * mdt_hsm_get_md_hsm() got it from disk and it is still valid
1321 if (pgs->hpk_errval != 0) {
1322 switch (pgs->hpk_errval) {
1324 /* the copy tool does not support cancel
1325 * so the cancel request is failed
1326 * As we cannot distinguish a cancel progress
1327 * from another action progress (they have the
1328 * same cookie), we suppose here the CT returns
1329 * ENOSYS only if does not support cancel
1331 /* this can also happen when cdt calls it to
1332 * for a timed out request */
1333 *status = ARS_FAILED;
1334 /* to have a cancel event in changelog */
1335 pgs->hpk_errval = ECANCELED;
1338 /* the request record has already been set to
1339 * ARS_CANCELED, this set the cancel request
1341 *status = ARS_SUCCEED;
1344 /* retry only if current policy or requested, and
1345 * object is not on error/removed */
1346 *status = (cdt->cdt_policy & CDT_NORETRY_ACTION ||
1347 !(pgs->hpk_flags & HP_FLAG_RETRY) ||
1348 IS_ERR(obj)) ? ARS_FAILED : ARS_WAITING;
1352 if (pgs->hpk_errval > CLF_HSM_MAXERROR) {
1353 CERROR("%s: Request %#llx on "DFID
1354 " failed, error code %d too large\n",
1356 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1358 hsm_set_cl_error(&cl_flags,
1359 CLF_HSM_ERROVERFLOW);
1362 hsm_set_cl_error(&cl_flags, pgs->hpk_errval);
1365 switch (car->car_hai->hai_action) {
1367 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1370 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1373 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1376 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1377 CERROR("%s: Failed request %#llx on "DFID
1378 " cannot be a CANCEL\n",
1381 PFID(&pgs->hpk_fid));
1384 CERROR("%s: Failed request %#llx on "DFID
1385 " %d is an unknown action\n",
1387 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1388 car->car_hai->hai_action);
1393 *status = ARS_SUCCEED;
1394 switch (car->car_hai->hai_action) {
1396 hsm_set_cl_event(&cl_flags, HE_ARCHIVE);
1397 /* set ARCHIVE keep EXIST and clear LOST and
1399 mh.mh_arch_ver = pgs->hpk_data_version;
1400 mh.mh_flags |= HS_ARCHIVED;
1401 mh.mh_flags &= ~(HS_LOST|HS_DIRTY);
1402 is_mh_changed = true;
1405 hsm_set_cl_event(&cl_flags, HE_RESTORE);
1407 /* do not clear RELEASED and DIRTY here
1408 * this will occur in hsm_swap_layouts()
1411 /* Restoring has changed the file version on
1413 mh.mh_arch_ver = pgs->hpk_data_version;
1414 is_mh_changed = true;
1417 hsm_set_cl_event(&cl_flags, HE_REMOVE);
1418 /* clear ARCHIVED EXISTS and LOST */
1419 mh.mh_flags &= ~(HS_ARCHIVED | HS_EXISTS | HS_LOST);
1420 is_mh_changed = true;
1423 hsm_set_cl_event(&cl_flags, HE_CANCEL);
1424 CERROR("%s: Successful request %#llx on "DFID" cannot be a CANCEL\n",
1427 PFID(&pgs->hpk_fid));
1430 CERROR("%s: Successful request %#llx on "DFID" %d is an unknown action\n",
1432 pgs->hpk_cookie, PFID(&pgs->hpk_fid),
1433 car->car_hai->hai_action);
1439 /* rc != 0 means error when analysing action, it may come from
1440 * a crasy CT no need to manage DIRTY
1441 * and if mdt_hsm_get_md_hsm() has returned an error, mh has not been
1444 if (rc == 0 && !IS_ERR(obj))
1445 hsm_set_cl_flags(&cl_flags,
1446 mh.mh_flags & HS_DIRTY ? CLF_HSM_DIRTY : 0);
1448 /* unlock is done later, after layout lock management */
1449 if (is_mh_changed && !IS_ERR(obj))
1450 rc = mdt_hsm_attr_set(mti, obj, &mh);
1452 /* we give back layout lock only if restore was successful or
1453 * if no retry will be attempted and if object is still alive,
1454 * in other cases we just unlock the object */
1455 if (car->car_hai->hai_action == HSMA_RESTORE) {
1456 struct cdt_restore_handle *crh;
1458 /* restore in data FID done, we swap the layouts
1459 * only if restore is successful */
1460 if (pgs->hpk_errval == 0 && !IS_ERR(obj)) {
1461 rc = hsm_swap_layouts(mti, obj, &car->car_hai->hai_dfid,
1464 if (cdt->cdt_policy & CDT_NORETRY_ACTION)
1465 *status = ARS_FAILED;
1466 pgs->hpk_errval = -rc;
1469 /* we have to retry, so keep layout lock */
1470 if (*status == ARS_WAITING)
1473 /* restore special case, need to create ChangeLog record
1474 * before to give back layout lock to avoid concurrent
1475 * file updater to post out of order ChangeLog */
1476 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1477 &car->car_hai->hai_fid);
1478 need_changelog = false;
1480 /* give back layout lock */
1481 mutex_lock(&cdt->cdt_restore_lock);
1482 crh = mdt_hsm_restore_hdl_find(cdt, &car->car_hai->hai_fid);
1484 list_del(&crh->crh_list);
1485 mutex_unlock(&cdt->cdt_restore_lock);
1486 /* Just give back layout lock, we keep the reference
1487 * which is given back later with the lock for HSM
1489 * XXX obj may be invalid so we do not pass it. */
1491 mdt_object_unlock(mti, NULL, &crh->crh_lh, 1);
1494 OBD_SLAB_FREE_PTR(crh, mdt_hsm_cdt_kmem);
1500 /* always add a ChangeLog record */
1502 mo_changelog(env, CL_HSM, cl_flags, mdt->mdt_child,
1503 &car->car_hai->hai_fid);
1506 mdt_object_put(mti->mti_env, obj);
1512 * update status of a request
1513 * \param mti [IN] context
1514 * \param pgs [IN] progress of the copy tool
1515 * \param update_record [IN] update llog record
1517 * \retval -ve failure
1519 int mdt_hsm_update_request_state(struct mdt_thread_info *mti,
1520 struct hsm_progress_kernel *pgs,
1521 const int update_record)
1523 struct mdt_device *mdt = mti->mti_mdt;
1524 struct coordinator *cdt = &mdt->mdt_coordinator;
1525 struct cdt_agent_req *car;
1529 /* no coordinator started, so we cannot serve requests */
1530 if (cdt->cdt_state == CDT_STOPPED)
1533 /* first do sanity checks */
1534 car = mdt_cdt_update_request(cdt, pgs);
1536 CERROR("%s: Cannot find running request for cookie %#llx"
1539 pgs->hpk_cookie, PFID(&pgs->hpk_fid));
1541 RETURN(PTR_ERR(car));
1544 CDEBUG(D_HSM, "Progress received for fid="DFID" cookie=%#llx"
1545 " action=%s flags=%d err=%d fid="DFID" dfid="DFID"\n",
1546 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1547 hsm_copytool_action2name(car->car_hai->hai_action),
1548 pgs->hpk_flags, pgs->hpk_errval,
1549 PFID(&car->car_hai->hai_fid),
1550 PFID(&car->car_hai->hai_dfid));
1552 /* progress is done on FID or data FID depending of the action and
1553 * of the copy progress */
1554 /* for restore progress is used to send back the data FID to cdt */
1555 if (car->car_hai->hai_action == HSMA_RESTORE &&
1556 lu_fid_eq(&car->car_hai->hai_fid, &car->car_hai->hai_dfid))
1557 car->car_hai->hai_dfid = pgs->hpk_fid;
1559 if ((car->car_hai->hai_action == HSMA_RESTORE ||
1560 car->car_hai->hai_action == HSMA_ARCHIVE) &&
1561 (!lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_dfid) &&
1562 !lu_fid_eq(&pgs->hpk_fid, &car->car_hai->hai_fid))) {
1563 CERROR("%s: Progress on "DFID" for cookie %#llx"
1564 " does not match request FID "DFID" nor data FID "
1567 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1568 PFID(&car->car_hai->hai_fid),
1569 PFID(&car->car_hai->hai_dfid));
1570 GOTO(out, rc = -EINVAL);
1573 if (pgs->hpk_errval != 0 && !(pgs->hpk_flags & HP_FLAG_COMPLETED)) {
1574 CERROR("%s: Progress on "DFID" for cookie %#llx action=%s"
1575 " is not coherent (err=%d and not completed"
1578 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1579 hsm_copytool_action2name(car->car_hai->hai_action),
1580 pgs->hpk_errval, pgs->hpk_flags);
1581 GOTO(out, rc = -EINVAL);
1584 /* now progress is valid */
1586 /* we use a root like ucred */
1587 hsm_init_ucred(mdt_ucred(mti));
1589 if (pgs->hpk_flags & HP_FLAG_COMPLETED) {
1590 enum agent_req_status status;
1592 rc = hsm_cdt_request_completed(mti, pgs, car, &status);
1594 CDEBUG(D_HSM, "%s record: fid="DFID" cookie=%#llx action=%s "
1596 update_record ? "Updating" : "Not updating",
1597 PFID(&pgs->hpk_fid), pgs->hpk_cookie,
1598 hsm_copytool_action2name(car->car_hai->hai_action),
1599 agent_req_status2name(status));
1601 /* update record first (LU-9075) */
1602 if (update_record) {
1604 struct hsm_record_update update = {
1605 .cookie = pgs->hpk_cookie,
1609 rc1 = mdt_agent_record_update(mti->mti_env, mdt,
1612 CERROR("%s: mdt_agent_record_update() failed,"
1613 " rc=%d, cannot update status to %s"
1614 " for cookie %#llx\n",
1615 mdt_obd_name(mdt), rc1,
1616 agent_req_status2name(status),
1618 rc = (rc != 0 ? rc : rc1);
1621 /* then remove request from memory list (LU-9075) */
1622 mdt_cdt_remove_request(cdt, pgs->hpk_cookie);
1624 /* ct has completed a request, so a slot is available,
1625 * signal the coordinator to find new work */
1626 mdt_hsm_cdt_event(cdt);
1628 /* if copytool send a progress on a canceled request
1629 * we inform copytool it should stop
1631 if (car->car_canceled == 1)
1637 /* remove ref got from mdt_cdt_update_request() */
1638 mdt_cdt_put_request(car);
1645 * data passed to llog_cat_process() callback
1646 * to cancel requests
1648 struct hsm_cancel_all_data {
1649 struct mdt_device *mdt;
1653 * llog_cat_process() callback, used to:
1654 * - purge all requests
1655 * \param env [IN] environment
1656 * \param llh [IN] llog handle
1657 * \param hdr [IN] llog record
1658 * \param data [IN] cb data = struct hsm_cancel_all_data
1660 * \retval -ve failure
1662 static int mdt_cancel_all_cb(const struct lu_env *env,
1663 struct llog_handle *llh,
1664 struct llog_rec_hdr *hdr, void *data)
1666 struct llog_agent_req_rec *larr;
1667 struct hsm_cancel_all_data *hcad;
1671 larr = (struct llog_agent_req_rec *)hdr;
1673 if (larr->arr_status == ARS_WAITING ||
1674 larr->arr_status == ARS_STARTED) {
1675 larr->arr_status = ARS_CANCELED;
1676 larr->arr_req_change = ktime_get_real_seconds();
1677 rc = llog_write(env, llh, hdr, hdr->lrh_index);
1684 * cancel all actions
1685 * \param obd [IN] MDT device
1687 static int hsm_cancel_all_actions(struct mdt_device *mdt)
1690 struct lu_context session;
1691 struct mdt_thread_info *mti;
1692 struct coordinator *cdt = &mdt->mdt_coordinator;
1693 struct cdt_agent_req *car;
1694 struct hsm_action_list *hal = NULL;
1695 struct hsm_action_item *hai;
1696 struct hsm_cancel_all_data hcad;
1697 int hal_sz = 0, hal_len, rc;
1698 enum cdt_states old_state;
1701 rc = lu_env_init(&env, LCT_MD_THREAD);
1705 /* for mdt_ucred(), lu_ucred stored in lu_ucred_key */
1706 rc = lu_context_init(&session, LCT_SERVER_SESSION);
1710 lu_context_enter(&session);
1711 env.le_ses = &session;
1713 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
1714 LASSERT(mti != NULL);
1716 mti->mti_env = &env;
1719 hsm_init_ucred(mdt_ucred(mti));
1721 /* disable coordinator */
1722 rc = set_cdt_state(cdt, CDT_DISABLE, &old_state);
1726 /* send cancel to all running requests */
1727 down_read(&cdt->cdt_request_lock);
1728 list_for_each_entry(car, &cdt->cdt_request_list, car_request_list) {
1729 mdt_cdt_get_request(car);
1730 /* request is not yet removed from list, it will be done
1731 * when copytool will return progress
1734 if (car->car_hai->hai_action == HSMA_CANCEL) {
1735 mdt_cdt_put_request(car);
1740 hal_len = sizeof(*hal) + cfs_size_round(MTI_NAME_MAXLEN + 1) +
1741 cfs_size_round(car->car_hai->hai_len);
1743 if (hal_len > hal_sz && hal_sz > 0) {
1744 /* not enough room, free old buffer */
1745 OBD_FREE(hal, hal_sz);
1749 /* empty buffer, allocate one */
1752 OBD_ALLOC(hal, hal_sz);
1754 mdt_cdt_put_request(car);
1755 up_read(&cdt->cdt_request_lock);
1756 GOTO(out_cdt_state, rc = -ENOMEM);
1760 hal->hal_version = HAL_VERSION;
1761 obd_uuid2fsname(hal->hal_fsname, mdt_obd_name(mdt),
1763 hal->hal_fsname[MTI_NAME_MAXLEN] = '\0';
1764 hal->hal_compound_id = car->car_compound_id;
1765 hal->hal_archive_id = car->car_archive_id;
1766 hal->hal_flags = car->car_flags;
1769 hai = hai_first(hal);
1770 memcpy(hai, car->car_hai, car->car_hai->hai_len);
1771 hai->hai_action = HSMA_CANCEL;
1774 /* it is possible to safely call mdt_hsm_agent_send()
1775 * (ie without a deadlock on cdt_request_lock), because the
1776 * write lock is taken only if we are not in purge mode
1777 * (mdt_hsm_agent_send() does not call mdt_cdt_add_request()
1778 * nor mdt_cdt_remove_request())
1780 /* no conflict with cdt thread because cdt is disable and we
1781 * have the request lock */
1782 mdt_hsm_agent_send(mti, hal, 1);
1784 mdt_cdt_put_request(car);
1786 up_read(&cdt->cdt_request_lock);
1789 OBD_FREE(hal, hal_sz);
1791 /* cancel all on-disk records */
1794 rc = cdt_llog_process(mti->mti_env, mti->mti_mdt, mdt_cancel_all_cb,
1795 &hcad, 0, 0, WRITE);
1797 /* Enable coordinator, unless the coordinator was stopping. */
1798 set_cdt_state(cdt, old_state, NULL);
1799 lu_context_exit(&session);
1800 lu_context_fini(&session);
1808 * check if a request is compatible with file status
1809 * \param hai [IN] request description
1810 * \param archive_id [IN] request archive id
1811 * \param rq_flags [IN] request flags
1812 * \param hsm [IN] file HSM metadata
1815 bool mdt_hsm_is_action_compat(const struct hsm_action_item *hai,
1816 u32 archive_id, u64 rq_flags,
1817 const struct md_hsm *hsm)
1819 int is_compat = false;
1823 hsm_flags = hsm->mh_flags;
1824 switch (hai->hai_action) {
1826 if (!(hsm_flags & HS_NOARCHIVE) &&
1827 (hsm_flags & HS_DIRTY || !(hsm_flags & HS_ARCHIVED)))
1830 if (hsm_flags & HS_EXISTS &&
1832 archive_id != hsm->mh_arch_id)
1837 if (!(hsm_flags & HS_DIRTY) && (hsm_flags & HS_RELEASED) &&
1838 hsm_flags & HS_ARCHIVED && !(hsm_flags & HS_LOST))
1842 if (!(hsm_flags & HS_RELEASED) &&
1843 (hsm_flags & (HS_ARCHIVED | HS_EXISTS)))
1850 CDEBUG(D_HSM, "fid="DFID" action=%s flags=%#llx"
1851 " extent=%#llx-%#llx hsm_flags=%.8X %s\n",
1852 PFID(&hai->hai_fid),
1853 hsm_copytool_action2name(hai->hai_action), rq_flags,
1854 hai->hai_extent.offset, hai->hai_extent.length,
1856 (is_compat ? "compatible" : "uncompatible"));
1862 * /proc interface used to get/set HSM behaviour (cdt->cdt_policy)
1864 static const struct {
1868 } hsm_policy_names[] = {
1869 { CDT_NONBLOCKING_RESTORE, "NonBlockingRestore", "NBR"},
1870 { CDT_NORETRY_ACTION, "NoRetryAction", "NRA"},
1875 * convert a policy name to a bit
1876 * \param name [IN] policy name
1878 * \retval policy bit
1880 static __u64 hsm_policy_str2bit(const char *name)
1884 for (i = 0; hsm_policy_names[i].bit != 0; i++)
1885 if (strcmp(hsm_policy_names[i].nickname, name) == 0 ||
1886 strcmp(hsm_policy_names[i].name, name) == 0)
1887 return hsm_policy_names[i].bit;
1892 * convert a policy bit field to a string
1893 * \param mask [IN] policy bit field
1894 * \param hexa [IN] print mask before bit names
1895 * \param buffer [OUT] string
1896 * \param count [IN] size of buffer
1898 static void hsm_policy_bit2str(struct seq_file *m, const __u64 mask,
1906 seq_printf(m, "(%#llx) ", mask);
1908 for (i = 0; i < CDT_POLICY_SHIFT_COUNT; i++) {
1911 for (j = 0; hsm_policy_names[j].bit != 0; j++) {
1912 if (hsm_policy_names[j].bit == bit)
1916 seq_printf(m, "[%s] ", hsm_policy_names[j].name);
1918 seq_printf(m, "%s ", hsm_policy_names[j].name);
1920 /* remove last ' ' */
1925 /* methods to read/write HSM policy flags */
1926 static int mdt_hsm_policy_seq_show(struct seq_file *m, void *data)
1928 struct mdt_device *mdt = m->private;
1929 struct coordinator *cdt = &mdt->mdt_coordinator;
1932 hsm_policy_bit2str(m, cdt->cdt_policy, false);
1937 mdt_hsm_policy_seq_write(struct file *file, const char __user *buffer,
1938 size_t count, loff_t *off)
1940 struct seq_file *m = file->private_data;
1941 struct mdt_device *mdt = m->private;
1942 struct coordinator *cdt = &mdt->mdt_coordinator;
1943 char *start, *token, sign;
1946 __u64 add_mask, remove_mask, set_mask;
1950 if (count + 1 > PAGE_SIZE)
1953 OBD_ALLOC(buf, count + 1);
1957 if (copy_from_user(buf, buffer, count))
1958 GOTO(out, rc = -EFAULT);
1963 CDEBUG(D_HSM, "%s: receive new policy: '%s'\n", mdt_obd_name(mdt),
1966 add_mask = remove_mask = set_mask = 0;
1968 token = strsep(&start, "\n ");
1974 if (sign == '-' || sign == '+')
1977 policy = hsm_policy_str2bit(token);
1979 CWARN("%s: '%s' is unknown, "
1980 "supported policies are:\n", mdt_obd_name(mdt),
1982 hsm_policy_bit2str(m, 0, false);
1983 GOTO(out, rc = -EINVAL);
1987 remove_mask |= policy;
1997 } while (start != NULL);
1999 CDEBUG(D_HSM, "%s: new policy: rm=%#llx add=%#llx set=%#llx\n",
2000 mdt_obd_name(mdt), remove_mask, add_mask, set_mask);
2002 /* if no sign in all string, it is a clear and set
2003 * if some sign found, all unsigned are converted
2005 * P1 P2 = set to P1 and P2
2006 * P1 -P2 = add P1 clear P2 same as +P1 -P2
2008 if (remove_mask == 0 && add_mask == 0) {
2009 cdt->cdt_policy = set_mask;
2011 cdt->cdt_policy |= set_mask | add_mask;
2012 cdt->cdt_policy &= ~remove_mask;
2015 GOTO(out, rc = count);
2018 OBD_FREE(buf, count + 1);
2021 LPROC_SEQ_FOPS(mdt_hsm_policy);
2023 #define GENERATE_PROC_METHOD(VAR) \
2024 static int mdt_hsm_##VAR##_seq_show(struct seq_file *m, void *data) \
2026 struct mdt_device *mdt = m->private; \
2027 struct coordinator *cdt = &mdt->mdt_coordinator; \
2030 seq_printf(m, "%llu\n", (__u64)cdt->VAR); \
2034 mdt_hsm_##VAR##_seq_write(struct file *file, const char __user *buffer, \
2035 size_t count, loff_t *off) \
2038 struct seq_file *m = file->private_data; \
2039 struct mdt_device *mdt = m->private; \
2040 struct coordinator *cdt = &mdt->mdt_coordinator; \
2045 rc = lprocfs_str_to_s64(buffer, count, &val); \
2048 if (val > 0 && val < INT_MAX) { \
2055 GENERATE_PROC_METHOD(cdt_loop_period)
2056 GENERATE_PROC_METHOD(cdt_grace_delay)
2057 GENERATE_PROC_METHOD(cdt_active_req_timeout)
2058 GENERATE_PROC_METHOD(cdt_max_requests)
2059 GENERATE_PROC_METHOD(cdt_default_archive_id)
2062 * procfs write method for MDT/hsm_control
2063 * proc entry is in mdt directory so data is mdt obd_device pointer
2065 #define CDT_ENABLE_CMD "enabled"
2066 #define CDT_STOP_CMD "shutdown"
2067 #define CDT_DISABLE_CMD "disabled"
2068 #define CDT_PURGE_CMD "purge"
2069 #define CDT_HELP_CMD "help"
2070 #define CDT_MAX_CMD_LEN 10
2073 mdt_hsm_cdt_control_seq_write(struct file *file, const char __user *buffer,
2074 size_t count, loff_t *off)
2076 struct seq_file *m = file->private_data;
2077 struct obd_device *obd = m->private;
2078 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2079 struct coordinator *cdt = &(mdt->mdt_coordinator);
2081 char kernbuf[CDT_MAX_CMD_LEN];
2084 if (count == 0 || count >= sizeof(kernbuf))
2087 if (copy_from_user(kernbuf, buffer, count))
2091 if (kernbuf[count - 1] == '\n')
2092 kernbuf[count - 1] = 0;
2095 if (strcmp(kernbuf, CDT_ENABLE_CMD) == 0) {
2096 if (cdt->cdt_state == CDT_DISABLE) {
2097 rc = set_cdt_state(cdt, CDT_RUNNING, NULL);
2098 mdt_hsm_cdt_event(cdt);
2099 wake_up(&cdt->cdt_waitq);
2101 rc = mdt_hsm_cdt_start(mdt);
2103 } else if (strcmp(kernbuf, CDT_STOP_CMD) == 0) {
2104 if ((cdt->cdt_state == CDT_STOPPING) ||
2105 (cdt->cdt_state == CDT_STOPPED)) {
2106 CERROR("%s: Coordinator already stopped\n",
2110 rc = mdt_hsm_cdt_stop(mdt);
2112 } else if (strcmp(kernbuf, CDT_DISABLE_CMD) == 0) {
2113 if ((cdt->cdt_state == CDT_STOPPING) ||
2114 (cdt->cdt_state == CDT_STOPPED)) {
2115 CERROR("%s: Coordinator is stopped\n",
2119 rc = set_cdt_state(cdt, CDT_DISABLE, NULL);
2121 } else if (strcmp(kernbuf, CDT_PURGE_CMD) == 0) {
2122 rc = hsm_cancel_all_actions(mdt);
2123 } else if (strcmp(kernbuf, CDT_HELP_CMD) == 0) {
2131 CERROR("%s: Valid coordinator control commands are: "
2132 "%s %s %s %s %s\n", mdt_obd_name(mdt),
2133 CDT_ENABLE_CMD, CDT_STOP_CMD, CDT_DISABLE_CMD,
2134 CDT_PURGE_CMD, CDT_HELP_CMD);
2142 int mdt_hsm_cdt_control_seq_show(struct seq_file *m, void *data)
2144 struct obd_device *obd = m->private;
2145 struct coordinator *cdt;
2148 cdt = &(mdt_dev(obd->obd_lu_dev)->mdt_coordinator);
2150 seq_printf(m, "%s\n", cdt_mdt_state2str(cdt->cdt_state));
2156 mdt_hsm_request_mask_show(struct seq_file *m, __u64 mask)
2162 for (i = 0; i < 8 * sizeof(mask); i++) {
2163 if (mask & (1UL << i)) {
2164 seq_printf(m, "%s%s", first ? "" : " ",
2165 hsm_copytool_action2name(i));
2175 mdt_hsm_user_request_mask_seq_show(struct seq_file *m, void *data)
2177 struct mdt_device *mdt = m->private;
2178 struct coordinator *cdt = &mdt->mdt_coordinator;
2180 return mdt_hsm_request_mask_show(m, cdt->cdt_user_request_mask);
2184 mdt_hsm_group_request_mask_seq_show(struct seq_file *m, void *data)
2186 struct mdt_device *mdt = m->private;
2187 struct coordinator *cdt = &mdt->mdt_coordinator;
2189 return mdt_hsm_request_mask_show(m, cdt->cdt_group_request_mask);
2193 mdt_hsm_other_request_mask_seq_show(struct seq_file *m, void *data)
2195 struct mdt_device *mdt = m->private;
2196 struct coordinator *cdt = &mdt->mdt_coordinator;
2198 return mdt_hsm_request_mask_show(m, cdt->cdt_other_request_mask);
2201 static inline enum hsm_copytool_action
2202 hsm_copytool_name2action(const char *name)
2204 if (strcasecmp(name, "NOOP") == 0)
2206 else if (strcasecmp(name, "ARCHIVE") == 0)
2207 return HSMA_ARCHIVE;
2208 else if (strcasecmp(name, "RESTORE") == 0)
2209 return HSMA_RESTORE;
2210 else if (strcasecmp(name, "REMOVE") == 0)
2212 else if (strcasecmp(name, "CANCEL") == 0)
2219 mdt_write_hsm_request_mask(struct file *file, const char __user *user_buf,
2220 size_t user_count, __u64 *mask)
2222 char *buf, *pos, *name;
2228 if (!(user_count < 4096))
2231 buf_size = user_count + 1;
2233 OBD_ALLOC(buf, buf_size);
2237 if (copy_from_user(buf, user_buf, buf_size - 1))
2238 GOTO(out, rc = -EFAULT);
2240 buf[buf_size - 1] = '\0';
2243 while ((name = strsep(&pos, " \t\v\n")) != NULL) {
2249 action = hsm_copytool_name2action(name);
2251 GOTO(out, rc = -EINVAL);
2253 new_mask |= (1UL << action);
2259 OBD_FREE(buf, buf_size);
2265 mdt_hsm_user_request_mask_seq_write(struct file *file, const char __user *buf,
2266 size_t count, loff_t *off)
2268 struct seq_file *m = file->private_data;
2269 struct mdt_device *mdt = m->private;
2270 struct coordinator *cdt = &mdt->mdt_coordinator;
2272 return mdt_write_hsm_request_mask(file, buf, count,
2273 &cdt->cdt_user_request_mask);
2277 mdt_hsm_group_request_mask_seq_write(struct file *file, const char __user *buf,
2278 size_t count, loff_t *off)
2280 struct seq_file *m = file->private_data;
2281 struct mdt_device *mdt = m->private;
2282 struct coordinator *cdt = &mdt->mdt_coordinator;
2284 return mdt_write_hsm_request_mask(file, buf, count,
2285 &cdt->cdt_group_request_mask);
2289 mdt_hsm_other_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_other_request_mask);
2300 static int mdt_hsm_cdt_raolu_seq_show(struct seq_file *m, void *data)
2302 struct mdt_device *mdt = m->private;
2303 struct coordinator *cdt = &mdt->mdt_coordinator;
2306 seq_printf(m, "%d\n", (int)cdt->cdt_remove_archive_on_last_unlink);
2311 mdt_hsm_cdt_raolu_seq_write(struct file *file, const char __user *buffer,
2312 size_t count, loff_t *off)
2315 struct seq_file *m = file->private_data;
2316 struct mdt_device *mdt = m->private;
2317 struct coordinator *cdt = &mdt->mdt_coordinator;
2322 rc = lprocfs_str_to_s64(buffer, count, &val);
2326 cdt->cdt_remove_archive_on_last_unlink = val;
2330 LPROC_SEQ_FOPS(mdt_hsm_cdt_loop_period);
2331 LPROC_SEQ_FOPS(mdt_hsm_cdt_grace_delay);
2332 LPROC_SEQ_FOPS(mdt_hsm_cdt_active_req_timeout);
2333 LPROC_SEQ_FOPS(mdt_hsm_cdt_max_requests);
2334 LPROC_SEQ_FOPS(mdt_hsm_cdt_default_archive_id);
2335 LPROC_SEQ_FOPS(mdt_hsm_user_request_mask);
2336 LPROC_SEQ_FOPS(mdt_hsm_group_request_mask);
2337 LPROC_SEQ_FOPS(mdt_hsm_other_request_mask);
2338 LPROC_SEQ_FOPS(mdt_hsm_cdt_raolu);
2340 /* Read-only proc files for request counters */
2341 static int mdt_hsm_cdt_archive_count_seq_show(struct seq_file *m, void *data)
2343 struct mdt_device *mdt = m->private;
2344 struct coordinator *cdt = &mdt->mdt_coordinator;
2347 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_archive_count));
2351 static int mdt_hsm_cdt_restore_count_seq_show(struct seq_file *m, void *data)
2353 struct mdt_device *mdt = m->private;
2354 struct coordinator *cdt = &mdt->mdt_coordinator;
2357 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_restore_count));
2361 static int mdt_hsm_cdt_remove_count_seq_show(struct seq_file *m, void *data)
2363 struct mdt_device *mdt = m->private;
2364 struct coordinator *cdt = &mdt->mdt_coordinator;
2367 seq_printf(m, "%d\n", atomic_read(&cdt->cdt_remove_count));
2371 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_archive_count);
2372 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_restore_count);
2373 LPROC_SEQ_FOPS_RO(mdt_hsm_cdt_remove_count);
2375 static struct lprocfs_vars lprocfs_mdt_hsm_vars[] = {
2377 .fops = &mdt_hsm_agent_fops },
2378 { .name = "actions",
2379 .fops = &mdt_hsm_actions_fops,
2380 .proc_mode = 0444 },
2381 { .name = "default_archive_id",
2382 .fops = &mdt_hsm_cdt_default_archive_id_fops },
2383 { .name = "grace_delay",
2384 .fops = &mdt_hsm_cdt_grace_delay_fops },
2385 { .name = "loop_period",
2386 .fops = &mdt_hsm_cdt_loop_period_fops },
2387 { .name = "max_requests",
2388 .fops = &mdt_hsm_cdt_max_requests_fops },
2390 .fops = &mdt_hsm_policy_fops },
2391 { .name = "active_request_timeout",
2392 .fops = &mdt_hsm_cdt_active_req_timeout_fops },
2393 { .name = "active_requests",
2394 .fops = &mdt_hsm_active_requests_fops },
2395 { .name = "user_request_mask",
2396 .fops = &mdt_hsm_user_request_mask_fops, },
2397 { .name = "group_request_mask",
2398 .fops = &mdt_hsm_group_request_mask_fops, },
2399 { .name = "other_request_mask",
2400 .fops = &mdt_hsm_other_request_mask_fops, },
2401 { .name = "remove_archive_on_last_unlink",
2402 .fops = &mdt_hsm_cdt_raolu_fops, },
2403 { .name = "archive_count",
2404 .fops = &mdt_hsm_cdt_archive_count_fops, },
2405 { .name = "restore_count",
2406 .fops = &mdt_hsm_cdt_restore_count_fops, },
2407 { .name = "remove_count",
2408 .fops = &mdt_hsm_cdt_remove_count_fops, },