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 * (C) Copyright 2012 Commissariat a l'energie atomique et aux energies
25 * Copyright (c) 2016, Intel Corporation.
30 * lustre/mdt/mdt_hsm_cdt_agent.c
32 * Lustre HSM Coordinator
34 * Author: Jacques-Charles Lafoucriere <jacques-charles.lafoucriere@cea.fr>
35 * Author: Aurelien Degremont <aurelien.degremont@cea.fr>
38 #define DEBUG_SUBSYSTEM S_MDS
41 #include <obd_support.h>
42 #include <lustre_export.h>
43 #include <lprocfs_status.h>
44 #include <lustre_kernelcomm.h>
45 #include "mdt_internal.h"
52 * find a hsm_agent by uuid
53 * lock cdt_agent_lock needs to be held by caller
54 * \param cdt [IN] coordinator
55 * \param uuid [IN] agent UUID
56 * \retval hsm_agent pointer or NULL if not found
58 static struct hsm_agent *mdt_hsm_agent_lookup(struct coordinator *cdt,
59 const struct obd_uuid *uuid)
63 list_for_each_entry(ha, &cdt->cdt_agents, ha_list) {
64 if (obd_uuid_equals(&ha->ha_uuid, uuid))
71 * register a copy tool
72 * \param mti [IN] MDT context
73 * \param uuid [IN] client UUID to be registered
74 * \param count [IN] number of archives agent serves
75 * \param archive_id [IN] vector of archive number served by the copytool
79 int mdt_hsm_agent_register(struct mdt_thread_info *mti,
80 const struct obd_uuid *uuid,
81 int nr_archives, __u32 *archive_id)
83 struct coordinator *cdt = &mti->mti_mdt->mdt_coordinator;
84 struct hsm_agent *ha, *tmp;
88 /* no coordinator started, so we cannot serve requests */
89 if (cdt->cdt_state == CDT_STOPPED) {
90 LCONSOLE_WARN("HSM coordinator thread is not running - "
91 "denying agent registration.\n");
97 GOTO(out, rc = -ENOMEM);
100 ha->ha_archive_cnt = nr_archives;
101 if (ha->ha_archive_cnt != 0) {
104 sz = ha->ha_archive_cnt * sizeof(*ha->ha_archive_id);
105 OBD_ALLOC(ha->ha_archive_id, sz);
106 if (ha->ha_archive_id == NULL)
107 GOTO(out_free, rc = -ENOMEM);
108 memcpy(ha->ha_archive_id, archive_id, sz);
110 atomic_set(&ha->ha_requests, 0);
111 atomic_set(&ha->ha_success, 0);
112 atomic_set(&ha->ha_failure, 0);
114 down_write(&cdt->cdt_agent_lock);
115 tmp = mdt_hsm_agent_lookup(cdt, uuid);
117 LCONSOLE_WARN("HSM agent %s already registered\n",
119 up_write(&cdt->cdt_agent_lock);
120 GOTO(out_free, rc = -EEXIST);
123 list_add_tail(&ha->ha_list, &cdt->cdt_agents);
125 if (ha->ha_archive_cnt == 0)
126 CDEBUG(D_HSM, "agent %s registered for all archives\n",
127 obd_uuid2str(&ha->ha_uuid));
129 CDEBUG(D_HSM, "agent %s registered for %d archives\n",
130 obd_uuid2str(&ha->ha_uuid), ha->ha_archive_cnt);
132 up_write(&cdt->cdt_agent_lock);
137 if (ha != NULL && ha->ha_archive_id != NULL)
138 OBD_FREE(ha->ha_archive_id,
139 ha->ha_archive_cnt * sizeof(*ha->ha_archive_id));
147 * register a copy tool
148 * \param mti [IN] MDT context
149 * \param uuid [IN] uuid to be registered
150 * \param archive_mask [IN] bitmask of archive number served by the copytool
152 * \retval -ve failure
154 int mdt_hsm_agent_register_mask(struct mdt_thread_info *mti,
155 const struct obd_uuid *uuid, __u32 archive_mask)
157 int rc, i, nr_archives = 0;
158 __u32 *archive_id = NULL;
161 nr_archives = hweight32(archive_mask);
163 if (nr_archives != 0) {
164 OBD_ALLOC(archive_id, nr_archives * sizeof(*archive_id));
169 for (i = 0; i < sizeof(archive_mask) * 8; i++) {
170 if ((1 << i) & archive_mask) {
171 archive_id[nr_archives] = i + 1;
177 rc = mdt_hsm_agent_register(mti, uuid, nr_archives, archive_id);
179 if (archive_id != NULL)
180 OBD_FREE(archive_id, nr_archives * sizeof(*archive_id));
186 * unregister a copy tool
187 * \param mti [IN] MDT context
188 * \param uuid [IN] uuid to be unregistered
190 * \retval -ve failure
192 int mdt_hsm_agent_unregister(struct mdt_thread_info *mti,
193 const struct obd_uuid *uuid)
195 struct coordinator *cdt = &mti->mti_mdt->mdt_coordinator;
196 struct hsm_agent *ha;
200 /* no coordinator started, so we cannot serve requests */
201 if (cdt->cdt_state == CDT_STOPPED)
204 down_write(&cdt->cdt_agent_lock);
206 ha = mdt_hsm_agent_lookup(cdt, uuid);
208 list_del_init(&ha->ha_list);
210 up_write(&cdt->cdt_agent_lock);
213 GOTO(out, rc = -ENOENT);
215 if (ha->ha_archive_cnt != 0)
216 OBD_FREE(ha->ha_archive_id,
217 ha->ha_archive_cnt * sizeof(*ha->ha_archive_id));
220 rc = hsm_cancel_agent_requests(mti->mti_mdt, uuid);
223 CDEBUG(D_HSM, "agent %s unregistration: %d\n", obd_uuid2str(uuid), rc);
229 * update agent statistics
230 * \param mdt [IN] MDT device
231 * \param succ_rq [IN] number of success
232 * \param fail_rq [IN] number of failure
233 * \param new_rq [IN] number of new requests
234 * \param uuid [IN] agent uuid
235 * if all counters == 0, clear counters
237 * \retval -ve failure
239 int mdt_hsm_agent_update_statistics(struct coordinator *cdt,
240 int succ_rq, int fail_rq, int new_rq,
241 const struct obd_uuid *uuid)
243 struct hsm_agent *ha;
247 down_read(&cdt->cdt_agent_lock);
248 list_for_each_entry(ha, &cdt->cdt_agents, ha_list) {
249 if (obd_uuid_equals(&ha->ha_uuid, uuid)) {
250 if (succ_rq == 0 && fail_rq == 0 && new_rq == 0) {
251 atomic_set(&ha->ha_success, 0);
252 atomic_set(&ha->ha_failure, 0);
253 atomic_set(&ha->ha_requests, 0);
255 atomic_add(succ_rq, &ha->ha_success);
256 atomic_add(fail_rq, &ha->ha_failure);
257 atomic_add(new_rq, &ha->ha_requests);
258 atomic_sub(succ_rq, &ha->ha_requests);
259 atomic_sub(fail_rq, &ha->ha_requests);
267 up_read(&cdt->cdt_agent_lock);
272 * find the best agent
273 * \param cdt [IN] coordinator
274 * \param archive [IN] archive number
275 * \param uuid [OUT] agent who can serve archive
277 * \retval -ve failure
279 int mdt_hsm_find_best_agent(struct coordinator *cdt, __u32 archive,
280 struct obd_uuid *uuid)
282 int rc = -EAGAIN, i, load = -1;
283 struct hsm_agent *ha;
286 /* Choose an export to send a copytool req to */
287 down_read(&cdt->cdt_agent_lock);
288 list_for_each_entry(ha, &cdt->cdt_agents, ha_list) {
289 for (i = 0; (i < ha->ha_archive_cnt) &&
290 (ha->ha_archive_id[i] != archive); i++) {
291 /* nothing to do, just skip unmatching records */
294 /* archive count == 0 means copy tool serves any backend */
295 if (ha->ha_archive_cnt != 0 && i == ha->ha_archive_cnt)
298 if (load == -1 || load > atomic_read(&ha->ha_requests)) {
299 load = atomic_read(&ha->ha_requests);
303 if (atomic_read(&ha->ha_requests) == 0)
306 up_read(&cdt->cdt_agent_lock);
311 int mdt_hsm_send_action_to_each_archive(struct mdt_thread_info *mti,
312 struct hsm_action_item *hai)
315 struct hsm_agent *ha;
316 __u32 archive_mask = 0;
317 struct coordinator *cdt = &mti->mti_mdt->mdt_coordinator;
319 /* return error by default in case all archive_ids have unregistered */
323 /* send action to all registered archive_ids */
324 down_read(&cdt->cdt_agent_lock);
325 list_for_each_entry(ha, &cdt->cdt_agents, ha_list) {
326 for (i = 0; (i < ha->ha_archive_cnt); i++) {
327 /* only send once for each archive_id */
328 if ((1 << ha->ha_archive_id[i]) & archive_mask)
330 archive_mask |= (1 << ha->ha_archive_id[i]);
332 /* XXX: instead of creating one request record per
333 * new action, it could make sense to gather
334 * all for the same archive_id as one compound
335 * request/id, like in mdt_hsm_add_actions() ?? */
336 compound_id = atomic_inc_return(&cdt->cdt_compound_id);
337 rc = mdt_agent_record_add(mti->mti_env, mti->mti_mdt,
339 ha->ha_archive_id[i], 0,
342 CERROR("%s: unable to add HSM remove request "
343 "for "DFID": rc=%d\n",
344 mdt_obd_name(mti->mti_mdt),
345 PFID(&hai->hai_fid), rc);
348 CDEBUG(D_HSM, "%s: added HSM remove request "
349 "for "DFID", archive_id=%d\n",
350 mdt_obd_name(mti->mti_mdt),
352 ha->ha_archive_id[i]);
355 /* early exit from loop due to error? */
356 if (i != ha->ha_archive_cnt)
359 up_read(&cdt->cdt_agent_lock);
365 * send a compound request to the agent
366 * \param mti [IN] context
367 * \param hal [IN] request (can be a kuc payload)
368 * \param purge [IN] purge mode (no record)
370 * \retval -ve failure
371 * This function supposes:
372 * - all actions are for the same archive number
373 * - in case of cancel, all cancel are for the same agent
374 * This implies that request split has to be done
375 * before when building the hal
377 int mdt_hsm_agent_send(struct mdt_thread_info *mti,
378 struct hsm_action_list *hal, bool purge)
380 struct obd_export *exp;
381 struct mdt_device *mdt = mti->mti_mdt;
382 struct coordinator *cdt = &mti->mti_mdt->mdt_coordinator;
383 struct hsm_action_list *buf = NULL;
384 struct hsm_action_item *hai;
385 struct obd_uuid uuid;
388 bool is_registered = false;
391 rc = mdt_hsm_find_best_agent(cdt, hal->hal_archive_id, &uuid);
392 if (rc && hal->hal_archive_id == 0) {
396 /* special case of remove requests with no archive_id specified,
397 * and no agent registered to serve all archives, then create a
398 * set of new requests, each to be sent to each registered
400 * Todo so, find all HSMA_REMOVE entries, and then :
401 * _ set completed status as SUCCESS (or FAIL?)
402 * _ create a new LLOG record for each archive_id
403 * presently being served by any CT
405 hai = hai_first(hal);
406 for (i = 0; i < hal->hal_count; i++,
407 hai = hai_next(hai)) {
408 struct hsm_record_update update;
410 /* only removes are concerned */
411 if (hai->hai_action != HSMA_REMOVE) {
412 /* count if other actions than HSMA_REMOVE,
413 * to return original error/rc */
418 /* send remove request to all registered archive_ids */
419 rc2 = mdt_hsm_send_action_to_each_archive(mti, hai);
423 /* only update original request as SUCCEED if it has
424 * been successfully broadcasted to all available
426 * XXX: this should only cause duplicates to be sent,
427 * unless a method to record already successfully
428 * reached archive_ids is implemented */
430 update.cookie = hai->hai_cookie;
431 update.status = ARS_SUCCEED;
432 rc2 = mdt_agent_record_update(mti->mti_env, mdt,
435 CERROR("%s: mdt_agent_record_update() "
436 "failed, cannot update "
437 "status to %s for cookie "
440 agent_req_status2name(ARS_SUCCEED),
441 hai->hai_cookie, rc2);
445 /* only remove requests with archive_id=0 */
452 CERROR("%s: Cannot find agent for archive %d: rc = %d\n",
453 mdt_obd_name(mdt), hal->hal_archive_id, rc);
457 CDEBUG(D_HSM, "Agent %s selected for archive %d\n", obd_uuid2str(&uuid),
458 hal->hal_archive_id);
461 buf = kuc_alloc(len, KUC_TRANSPORT_HSM, HMT_ACTION_LIST);
463 RETURN(PTR_ERR(buf));
464 memcpy(buf, hal, len);
466 /* Check if request is still valid (cf file hsm flags) */
467 fail_request = false;
468 hai = hai_first(hal);
469 for (i = 0; i < hal->hal_count; i++, hai = hai_next(hai)) {
470 struct mdt_object *obj;
473 if (hai->hai_action == HSMA_CANCEL)
476 obj = mdt_hsm_get_md_hsm(mti, &hai->hai_fid, &hsm);
478 mdt_object_put(mti->mti_env, obj);
479 } else if (PTR_ERR(obj) == -ENOENT) {
480 struct hsm_record_update update = {
481 .cookie = hai->hai_cookie,
482 .status = ARS_FAILED,
485 if (hai->hai_action == HSMA_REMOVE)
489 rc = mdt_agent_record_update(mti->mti_env, mdt,
492 CERROR("%s: mdt_agent_record_update() failed, "
493 "cannot update status to %s for cookie "
496 agent_req_status2name(ARS_FAILED),
497 hai->hai_cookie, rc);
503 GOTO(out_buf, rc = PTR_ERR(obj));
506 if (!mdt_hsm_is_action_compat(hai, hal->hal_archive_id,
507 hal->hal_flags, &hsm)) {
508 struct hsm_record_update update = {
509 .cookie = hai->hai_cookie,
510 .status = ARS_FAILED,
513 /* incompatible request, we abort the request */
514 /* next time coordinator will wake up, it will
515 * make the same compound with valid only
518 rc = mdt_agent_record_update(mti->mti_env, mdt,
521 CERROR("%s: mdt_agent_record_update() failed, "
522 "cannot update status to %s for cookie "
525 agent_req_status2name(ARS_FAILED),
526 hai->hai_cookie, rc);
530 /* if restore and record status updated, give
531 * back granted layout lock */
532 if (hai->hai_action == HSMA_RESTORE) {
533 struct cdt_restore_handle *crh = NULL;
535 mutex_lock(&cdt->cdt_restore_lock);
536 crh = mdt_hsm_restore_hdl_find(cdt,
539 list_del(&crh->crh_list);
540 mutex_unlock(&cdt->cdt_restore_lock);
542 mdt_object_unlock(mti, NULL,
544 OBD_SLAB_FREE_PTR(crh,
551 /* we found incompatible requests, so the compound cannot be send
552 * as is. Bad records have been invalidated in llog.
553 * Valid one will be reschedule next time coordinator will wake up
554 * So no need the rebuild a full valid compound request now
557 GOTO(out_buf, rc = 0);
559 /* Cancel memory registration is useless for purge
560 * non registration avoid a deadlock :
561 * in case of failure we have to take the write lock
562 * to remove entry which conflict with the read loack needed
566 /* set is_registered even if failure because we may have
567 * partial work done */
568 is_registered = true;
569 rc = mdt_hsm_add_hal(mti, hal, &uuid);
574 /* Uses the ldlm reverse import; this rpc will be seen by
575 * the ldlm_callback_handler. Note this sends a request RPC
576 * from a server (MDT) to a client (MDC), backwards of normal comms.
578 exp = cfs_hash_lookup(mdt2obd_dev(mdt)->obd_uuid_hash, &uuid);
579 if (exp == NULL || exp->exp_disconnected) {
581 class_export_put(exp);
582 /* This should clean up agents on evicted exports */
584 CERROR("%s: agent uuid (%s) not found, unregistering:"
586 mdt_obd_name(mdt), obd_uuid2str(&uuid), rc);
587 mdt_hsm_agent_unregister(mti, &uuid);
591 /* send request to agent */
592 rc = do_set_info_async(exp->exp_imp_reverse, LDLM_SET_INFO,
594 sizeof(KEY_HSM_COPYTOOL_SEND),
595 KEY_HSM_COPYTOOL_SEND,
596 kuc_len(len), kuc_ptr(buf), NULL);
599 CERROR("%s: cannot send request to agent '%s': rc = %d\n",
600 mdt_obd_name(mdt), obd_uuid2str(&uuid), rc);
602 class_export_put(exp);
605 CDEBUG(D_HSM, "Lost connection to agent '%s', unregistering\n",
606 obd_uuid2str(&uuid));
607 mdt_hsm_agent_unregister(mti, &uuid);
611 if (rc != 0 && is_registered) {
612 /* in case of error, we have to unregister requests */
613 hai = hai_first(hal);
614 for (i = 0; i < hal->hal_count; i++, hai = hai_next(hai)) {
615 if (hai->hai_action == HSMA_CANCEL)
617 mdt_cdt_remove_request(cdt, hai->hai_cookie);
628 * update status of a request
630 * \param pgs [IN] progress of the copy tool
632 * \retval -ve failure
634 int mdt_hsm_coordinator_update(struct mdt_thread_info *mti,
635 struct hsm_progress_kernel *pgs)
640 /* ask to coordinator to update request state and
641 * to record on disk the result */
642 rc = mdt_hsm_update_request_state(mti, pgs, 1);
647 * seq_file method called to start access to /proc file
649 static void *mdt_hsm_agent_proc_start(struct seq_file *s, loff_t *off)
651 struct mdt_device *mdt = s->private;
652 struct coordinator *cdt = &mdt->mdt_coordinator;
653 struct list_head *pos;
657 down_read(&cdt->cdt_agent_lock);
659 if (list_empty(&cdt->cdt_agents))
663 RETURN(SEQ_START_TOKEN);
666 list_for_each(pos, &cdt->cdt_agents) {
676 * seq_file method called to get next item
677 * just returns NULL at eof
679 static void *mdt_hsm_agent_proc_next(struct seq_file *s, void *v, loff_t *p)
681 struct mdt_device *mdt = s->private;
682 struct coordinator *cdt = &mdt->mdt_coordinator;
683 struct list_head *pos = v;
686 if (pos == SEQ_START_TOKEN)
687 pos = cdt->cdt_agents.next;
692 if (pos != &cdt->cdt_agents)
700 static int mdt_hsm_agent_proc_show(struct seq_file *s, void *v)
702 struct list_head *pos = v;
703 struct hsm_agent *ha;
707 if (pos == SEQ_START_TOKEN)
710 ha = list_entry(pos, struct hsm_agent, ha_list);
711 seq_printf(s, "uuid=%s archive_id=", ha->ha_uuid.uuid);
712 if (ha->ha_archive_cnt == 0) {
713 seq_printf(s, "ANY");
715 seq_printf(s, "%d", ha->ha_archive_id[0]);
716 for (i = 1; i < ha->ha_archive_cnt; i++)
717 seq_printf(s, ",%d", ha->ha_archive_id[i]);
720 seq_printf(s, " requests=[current:%d ok:%d errors:%d]\n",
721 atomic_read(&ha->ha_requests),
722 atomic_read(&ha->ha_success),
723 atomic_read(&ha->ha_failure));
728 * seq_file method called to stop access to /proc file
730 static void mdt_hsm_agent_proc_stop(struct seq_file *s, void *v)
732 struct mdt_device *mdt = s->private;
733 struct coordinator *cdt = &mdt->mdt_coordinator;
735 up_read(&cdt->cdt_agent_lock);
738 /* hsm agent list proc functions */
739 static const struct seq_operations mdt_hsm_agent_proc_ops = {
740 .start = mdt_hsm_agent_proc_start,
741 .next = mdt_hsm_agent_proc_next,
742 .show = mdt_hsm_agent_proc_show,
743 .stop = mdt_hsm_agent_proc_stop,
747 * public function called at open of /proc file to get
750 static int lprocfs_open_hsm_agent(struct inode *inode, struct file *file)
756 rc = seq_open(file, &mdt_hsm_agent_proc_ops);
760 s = file->private_data;
761 s->private = PDE_DATA(inode);
766 /* methods to access hsm agent list */
767 const struct file_operations mdt_hsm_agent_fops = {
768 .owner = THIS_MODULE,
769 .open = lprocfs_open_hsm_agent,
772 .release = lprocfs_seq_release,