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, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2012, Intel Corporation.
25 * Use is subject to license terms.
27 * Author: Johann Lombardi <johann.lombardi@intel.com>
28 * Author: Niu Yawei <yawei.niu@intel.com>
32 # define EXPORT_SYMTAB
35 #define DEBUG_SUBSYSTEM S_LQUOTA
37 #include <lustre_dlm.h>
38 #include <obd_class.h>
40 #include "qmt_internal.h"
42 /* intent policy function called from mdt_intent_opc() when the intent is of
44 int qmt_intent_policy(const struct lu_env *env, struct lu_device *ld,
45 struct ptlrpc_request *req, struct ldlm_lock **lockp,
48 struct qmt_device *qmt = lu2qmt_dev(ld);
49 struct ldlm_intent *it;
50 struct quota_body *reqbody;
51 struct quota_body *repbody;
52 struct obd_uuid *uuid;
53 struct lquota_lvb *lvb;
54 struct ldlm_resource *res = (*lockp)->l_resource;
58 req_capsule_extend(&req->rq_pill, &RQF_LDLM_INTENT_QUOTA);
59 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
60 ldlm_lvbo_size(*lockp));
62 /* extract quota body and intent opc */
63 it = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
65 RETURN(err_serious(-EFAULT));
67 reqbody = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_BODY);
69 RETURN(err_serious(-EFAULT));
72 rc = req_capsule_server_pack(&req->rq_pill);
74 CERROR("Can't pack response, rc %d\n", rc);
75 RETURN(err_serious(rc));
78 repbody = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_BODY);
80 RETURN(err_serious(-EFAULT));
82 uuid = &(*lockp)->l_export->exp_client_uuid;
85 case IT_QUOTA_DQACQ: {
86 struct lquota_entry *lqe;
87 struct ldlm_lock *lock;
89 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
90 /* acquire on global lock? something is wrong ... */
91 GOTO(out, rc = -EPROTO);
93 /* verify global lock isn't stale */
94 if (!lustre_handle_is_used(&reqbody->qb_glb_lockh))
95 GOTO(out, rc = -ENOLCK);
97 lock = ldlm_handle2lock(&reqbody->qb_glb_lockh);
99 GOTO(out, rc = -ENOLCK);
102 lqe = res->lr_lvb_data;
103 LASSERT(lqe != NULL);
106 /* acquire quota space */
107 rc = qmt_dqacq0(env, lqe, qmt, uuid, reqbody->qb_flags,
108 reqbody->qb_count, reqbody->qb_usage,
117 /* new connection from slave */
119 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0)
120 /* connection on per-ID lock? something is wrong ... */
121 GOTO(out, rc = -EPROTO);
123 rc = qmt_pool_new_conn(env, qmt, &reqbody->qb_fid,
124 &repbody->qb_slv_fid,
125 &repbody->qb_slv_ver, uuid);
131 CERROR("%s: invalid intent opcode: "LPU64"\n", qmt->qmt_svname,
133 GOTO(out, rc = err_serious(-EINVAL));
136 /* on success, pack lvb in reply */
137 lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB);
138 lvb_len = ldlm_lvbo_size(*lockp);
139 lvb_len = ldlm_lvbo_fill(*lockp, lvb, lvb_len);
140 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len, RCL_SERVER);
147 * Initialize quota LVB associated with quota indexes.
148 * Called with res->lr_lvb_sem held
150 int qmt_lvbo_init(struct lu_device *ld, struct ldlm_resource *res)
153 struct qmt_thread_info *qti;
154 struct qmt_device *qmt = lu2qmt_dev(ld);
155 int pool_id, pool_type, qtype;
159 LASSERT(res != NULL);
161 if (res->lr_type != LDLM_PLAIN)
164 if (res->lr_lvb_data ||
165 res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
172 /* initialize environment */
173 rc = lu_env_init(env, LCT_MD_THREAD);
180 /* extract global index FID and quota identifier */
181 fid_extract_quota_resid(&res->lr_name, &qti->qti_fid, &qti->qti_id);
183 /* sanity check the global index FID */
184 rc = lquota_extract_fid(&qti->qti_fid, &pool_id, &pool_type, &qtype);
186 CERROR("can't extract pool information from FID "DFID"\n",
187 PFID(&qti->qti_fid));
191 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
192 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
193 * we are thus dealing with an ID lock. */
194 struct lquota_entry *lqe;
196 /* Find the quota entry associated with the quota id */
197 lqe = qmt_pool_lqe_lookup(env, qmt, pool_id, pool_type, qtype,
200 GOTO(out, rc = PTR_ERR(lqe));
202 /* store reference to lqe in lr_lvb_data */
203 res->lr_lvb_data = lqe;
204 LQUOTA_DEBUG(lqe, "initialized res lvb");
206 struct dt_object *obj;
208 /* lookup global index */
209 obj = dt_locate(env, qmt->qmt_child, &qti->qti_fid);
211 GOTO(out, rc = PTR_ERR(obj));
212 if (!dt_object_exists(obj)) {
213 lu_object_put(env, &obj->do_lu);
214 GOTO(out, rc = -ENOENT);
217 /* store reference to global index object in lr_lvb_data */
218 res->lr_lvb_data = obj;
219 CDEBUG(D_QUOTA, DFID" initialized lvb\n", PFID(&qti->qti_fid));
222 res->lr_lvb_len = sizeof(struct lquota_lvb);
231 * Update LVB associated with the global quota index.
232 * This function is called from the DLM itself after a glimpse callback, in this
233 * case valid ptlrpc request is passed.
235 int qmt_lvbo_update(struct lu_device *ld, struct ldlm_resource *res,
236 struct ptlrpc_request *req, int increase_only)
239 struct qmt_thread_info *qti;
240 struct qmt_device *qmt = lu2qmt_dev(ld);
241 struct lquota_entry *lqe;
242 struct lquota_lvb *lvb;
243 struct ldlm_lock *lock;
244 struct obd_export *exp;
248 LASSERT(res != NULL);
253 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
254 /* no need to update lvb for global quota locks */
257 lvb = req_capsule_server_swab_get(&req->rq_pill, &RMF_DLM_LVB,
258 lustre_swab_lquota_lvb);
260 CERROR("%s: failed to extract lvb from request\n",
265 lqe = res->lr_lvb_data;
266 LASSERT(lqe != NULL);
269 LQUOTA_DEBUG(lqe, "releasing:"LPU64" may release:"LPU64,
270 lvb->lvb_id_rel, lvb->lvb_id_may_rel);
272 if (lvb->lvb_id_rel == 0) {
273 /* nothing to release */
274 if (lvb->lvb_id_may_rel != 0)
275 /* but might still release later ... */
276 lqe->lqe_may_rel += lvb->lvb_id_may_rel;
277 GOTO(out_lqe, rc = 0);
280 /* allocate environement */
283 GOTO(out_lqe, rc = -ENOMEM);
285 /* initialize environment */
286 rc = lu_env_init(env, LCT_MD_THREAD);
291 /* The request is a glimpse callback which was sent via the
292 * reverse import to the slave. What we care about here is the
293 * export associated with the slave and req->rq_export is
294 * definitely not what we are looking for (it is actually set to
296 * Therefore we extract the lock from the request argument
297 * and use lock->l_export. */
298 lock = ldlm_request_lock(req);
300 CERROR("%s: failed to get lock from request!\n",
302 GOTO(out_env_init, rc = PTR_ERR(lock));
305 exp = class_export_get(lock->l_export);
307 CERROR("%s: failed to get export from lock!\n",
309 GOTO(out_env_init, rc = -EFAULT);
312 /* release quota space */
313 rc = qmt_dqacq0(env, lqe, qmt, &exp->exp_client_uuid,
314 QUOTA_DQACQ_FL_REL, lvb->lvb_id_rel, 0, &qti->qti_body);
315 if (rc || qti->qti_body.qb_count != lvb->lvb_id_rel)
316 LQUOTA_ERROR(lqe, "failed to release quota space on glimpse "
317 LPU64"!="LPU64" rc:%d\n", qti->qti_body.qb_count,
318 lvb->lvb_id_rel, rc);
319 class_export_put(exp);
321 GOTO(out_env_init, rc);
333 * Report size of lvb to ldlm layer in order to allocate lvb buffer
334 * As far as quota locks are concerned, the size is static and is the same
335 * for both global and per-ID locks which shares the same lvb format.
337 int qmt_lvbo_size(struct lu_device *ld, struct ldlm_lock *lock)
339 return sizeof(struct lquota_lvb);
343 * Fill request buffer with quota lvb
345 int qmt_lvbo_fill(struct lu_device *ld, struct ldlm_lock *lock, void *lvb,
348 struct ldlm_resource *res = lock->l_resource;
349 struct lquota_lvb *qlvb = lvb;
352 LASSERT(res != NULL);
354 if (res->lr_type != LDLM_PLAIN || res->lr_lvb_data == NULL ||
355 res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
358 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
359 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
360 * we are thus dealing with an ID lock. */
361 struct lquota_entry *lqe = res->lr_lvb_data;
363 /* return current qunit value & edquot flags in lvb */
365 qlvb->lvb_id_qunit = lqe->lqe_qunit;
368 qlvb->lvb_flags = LQUOTA_FL_EDQUOT;
371 /* global quota lock */
374 struct dt_object *obj = res->lr_lvb_data;
380 /* initialize environment */
381 rc = lu_env_init(env, LCT_LOCAL);
387 /* return current version of global index */
388 qlvb->lvb_glb_ver = dt_version_get(env, obj);
394 RETURN(sizeof(struct lquota_lvb));
398 * Free lvb associated with a given ldlm resource
399 * we don't really allocate a lvb, lr_lvb_data just points to
400 * the appropriate backend structures.
402 int qmt_lvbo_free(struct lu_device *ld, struct ldlm_resource *res)
406 if (res->lr_lvb_data == NULL)
409 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
410 struct lquota_entry *lqe = res->lr_lvb_data;
412 /* release lqe reference */
415 struct dt_object *obj = res->lr_lvb_data;
423 /* initialize environment */
424 rc = lu_env_init(env, LCT_LOCAL);
430 /* release object reference */
431 lu_object_put(env, &obj->do_lu);
436 res->lr_lvb_data = NULL;
442 typedef int (*qmt_glimpse_cb_t)(const struct lu_env *, struct qmt_device *,
443 struct obd_uuid *, union ldlm_gl_desc *,
446 * Send glimpse callback to slaves holding a lock on resource \res.
447 * This is used to notify slaves of new quota settings or to claim quota space
450 * \param env - is the environment passed by the caller
451 * \param qmt - is the quota master target
452 * \param res - is the dlm resource associated with the quota object
453 * \param desc - is the glimpse descriptor to pack in glimpse callback
454 * \param cb - is the callback function called on every lock and determine
455 * whether a glimpse should be issued
456 * \param arg - is an opaq parameter passed to the callback function
458 static int qmt_glimpse_lock(const struct lu_env *env, struct qmt_device *qmt,
459 struct ldlm_resource *res, union ldlm_gl_desc *desc,
460 qmt_glimpse_cb_t cb, void *arg)
462 cfs_list_t *tmp, *pos;
463 CFS_LIST_HEAD(gl_list);
468 /* scan list of granted locks */
469 cfs_list_for_each(pos, &res->lr_granted) {
470 struct ldlm_glimpse_work *work;
471 struct ldlm_lock *lock;
472 struct obd_uuid *uuid;
474 lock = cfs_list_entry(pos, struct ldlm_lock, l_res_link);
475 LASSERT(lock->l_export);
476 uuid = &lock->l_export->exp_client_uuid;
479 rc = cb(env, qmt, uuid, desc, arg);
481 /* slave should not be notified */
484 /* something wrong happened, we still notify */
485 CERROR("%s: callback function failed to "
486 "determine whether slave %s should be "
487 "notified (%d)\n", qmt->qmt_svname,
488 obd_uuid2str(uuid), rc);
493 CERROR("%s: failed to notify %s\n", qmt->qmt_svname,
498 cfs_list_add_tail(&work->gl_list, &gl_list);
499 work->gl_lock = LDLM_LOCK_GET(lock);
501 work->gl_desc = desc;
506 if (cfs_list_empty(&gl_list)) {
507 CDEBUG(D_QUOTA, "%s: nobody to notify\n", qmt->qmt_svname);
511 /* issue glimpse callbacks to all connected slaves */
512 rc = ldlm_glimpse_locks(res, &gl_list);
514 cfs_list_for_each_safe(pos, tmp, &gl_list) {
515 struct ldlm_glimpse_work *work;
517 work = cfs_list_entry(pos, struct ldlm_glimpse_work, gl_list);
519 cfs_list_del(&work->gl_list);
520 CERROR("%s: failed to notify %s of new quota settings\n",
522 obd_uuid2str(&work->gl_lock->l_export->exp_client_uuid));
523 LDLM_LOCK_RELEASE(work->gl_lock);
531 * Send glimpse request to all global quota locks to push new quota setting to
534 * \param env - is the environment passed by the caller
535 * \param lqe - is the lquota entry which has new settings
536 * \param ver - is the version associated with the setting change
538 void qmt_glb_lock_notify(const struct lu_env *env, struct lquota_entry *lqe,
541 struct qmt_thread_info *qti = qmt_info(env);
542 struct qmt_pool_info *pool = lqe2qpi(lqe);
543 struct ldlm_resource *res = NULL;
547 lquota_generate_fid(&qti->qti_fid, pool->qpi_key & 0x0000ffff,
548 pool->qpi_key >> 16, lqe->lqe_site->lqs_qtype);
550 /* send glimpse callback to notify slaves of new quota settings */
551 qti->qti_gl_desc.lquota_desc.gl_id = lqe->lqe_id;
552 qti->qti_gl_desc.lquota_desc.gl_flags = 0;
553 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
554 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
555 qti->qti_gl_desc.lquota_desc.gl_time = lqe->lqe_gracetime;
556 qti->qti_gl_desc.lquota_desc.gl_ver = ver;
558 /* look up ldlm resource associated with global index */
559 fid_build_reg_res_name(&qti->qti_fid, &qti->qti_resid);
560 res = ldlm_resource_get(pool->qpi_qmt->qmt_ns, NULL, &qti->qti_resid,
563 /* this might happen if no slaves have enqueued global quota
565 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource associated "
566 "with "DFID, PFID(&qti->qti_fid));
570 rc = qmt_glimpse_lock(env, pool->qpi_qmt, res, &qti->qti_gl_desc,
572 ldlm_resource_putref(res);
576 /* Callback function used to select locks that should be glimpsed when
577 * broadcasting the new qunit value */
578 static int qmt_id_lock_cb(const struct lu_env *env, struct qmt_device *qmt,
579 struct obd_uuid *uuid, union ldlm_gl_desc *desc,
582 struct obd_uuid *slv_uuid = arg;
584 if (slv_uuid != NULL && obd_uuid_equals(uuid, slv_uuid))
590 * Send glimpse request on per-ID lock to push new qunit value to slave.
592 * \param env - is the environment passed by the caller
593 * \param qmt - is the quota master target device
594 * \param lqe - is the lquota entry with the new qunit value
595 * \param uuid - is the uuid of the slave acquiring space, if any
597 static void qmt_id_lock_glimpse(const struct lu_env *env,
598 struct qmt_device *qmt,
599 struct lquota_entry *lqe, struct obd_uuid *uuid)
601 struct qmt_thread_info *qti = qmt_info(env);
602 struct qmt_pool_info *pool = lqe2qpi(lqe);
603 struct ldlm_resource *res = NULL;
607 if (!lqe->lqe_enforced)
610 lquota_generate_fid(&qti->qti_fid, pool->qpi_key & 0x0000ffff,
611 pool->qpi_key >> 16, lqe->lqe_site->lqs_qtype);
612 fid_build_quota_resid(&qti->qti_fid, &lqe->lqe_id, &qti->qti_resid);
613 res = ldlm_resource_get(qmt->qmt_ns, NULL, &qti->qti_resid, LDLM_PLAIN,
616 /* this might legitimately happens if slaves haven't had the
617 * opportunity to enqueue quota lock yet. */
618 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource for per-ID "
619 "lock "DFID, PFID(&qti->qti_fid));
621 if (lqe->lqe_revoke_time == 0 &&
622 lqe->lqe_qunit == pool->qpi_least_qunit)
623 lqe->lqe_revoke_time = cfs_time_current_64();
624 lqe_write_unlock(lqe);
629 /* The purpose of glimpse callback on per-ID lock is twofold:
630 * - notify slaves of new qunit value and hope they will release some
631 * spare quota space in return
632 * - notify slaves that master ran out of quota space and there is no
633 * need to send acquire request any more until further notice */
635 /* fill glimpse descriptor with lqe settings */
637 qti->qti_gl_desc.lquota_desc.gl_flags = LQUOTA_FL_EDQUOT;
639 qti->qti_gl_desc.lquota_desc.gl_flags = 0;
640 qti->qti_gl_desc.lquota_desc.gl_qunit = lqe->lqe_qunit;
642 if (lqe->lqe_revoke_time == 0 &&
643 qti->qti_gl_desc.lquota_desc.gl_qunit == pool->qpi_least_qunit)
644 /* reset lqe_may_rel, it will be updated on glimpse callback
645 * replies if needed */
646 lqe->lqe_may_rel = 0;
648 /* The rebalance thread is the only thread which can issue glimpses */
649 LASSERT(!lqe->lqe_gl);
651 lqe_write_unlock(lqe);
653 /* issue glimpse callback to slaves */
654 rc = qmt_glimpse_lock(env, qmt, res, &qti->qti_gl_desc,
655 uuid ? qmt_id_lock_cb : NULL, (void *)uuid);
658 if (lqe->lqe_revoke_time == 0 &&
659 qti->qti_gl_desc.lquota_desc.gl_qunit == pool->qpi_least_qunit &&
660 lqe->lqe_qunit == pool->qpi_least_qunit) {
661 lqe->lqe_revoke_time = cfs_time_current_64();
662 qmt_adjust_edquot(lqe, cfs_time_current_sec());
664 LASSERT(lqe->lqe_gl);
666 lqe_write_unlock(lqe);
668 ldlm_resource_putref(res);
673 * Schedule a glimpse request on per-ID locks to push new qunit value or
674 * edquot flag to quota slaves.
676 * \param qmt - is the quota master target device
677 * \param lqe - is the lquota entry with the new qunit value
679 void qmt_id_lock_notify(struct qmt_device *qmt, struct lquota_entry *lqe)
685 spin_lock(&qmt->qmt_reba_lock);
686 if (!qmt->qmt_stopping && cfs_list_empty(&lqe->lqe_link)) {
687 cfs_list_add_tail(&lqe->lqe_link, &qmt->qmt_reba_list);
690 spin_unlock(&qmt->qmt_reba_lock);
693 cfs_waitq_signal(&qmt->qmt_reba_thread.t_ctl_waitq);
700 * The rebalance thread is in charge of sending glimpse callbacks on per-ID
701 * quota locks owned by slaves in order to notify them of:
702 * - a qunit shrink in which case slaves might release quota space back in
704 * - set/clear edquot flag used to cache the "quota exhausted" state of the
705 * master. When the flag is set, slaves know that there is no need to
706 * try to acquire quota from the master since this latter has already
707 * distributed all the space.
709 static int qmt_reba_thread(void *arg)
711 struct qmt_device *qmt = (struct qmt_device *)arg;
712 struct ptlrpc_thread *thread = &qmt->qmt_reba_thread;
713 struct l_wait_info lwi = { 0 };
715 struct lquota_entry *lqe, *tmp;
716 char pname[MTI_NAME_MAXLEN];
724 rc = lu_env_init(env, LCT_MD_THREAD);
726 CERROR("%s: failed to init env.", qmt->qmt_svname);
731 snprintf(pname, MTI_NAME_MAXLEN, "qmt_reba_%s", qmt->qmt_svname);
732 cfs_daemonize(pname);
734 thread_set_flags(thread, SVC_RUNNING);
735 cfs_waitq_signal(&thread->t_ctl_waitq);
738 l_wait_event(thread->t_ctl_waitq,
739 !cfs_list_empty(&qmt->qmt_reba_list) ||
740 !thread_is_running(thread), &lwi);
742 spin_lock(&qmt->qmt_reba_lock);
743 cfs_list_for_each_entry_safe(lqe, tmp, &qmt->qmt_reba_list,
745 cfs_list_del_init(&lqe->lqe_link);
746 spin_unlock(&qmt->qmt_reba_lock);
748 if (thread_is_running(thread))
749 qmt_id_lock_glimpse(env, qmt, lqe, NULL);
752 spin_lock(&qmt->qmt_reba_lock);
754 spin_unlock(&qmt->qmt_reba_lock);
756 if (!thread_is_running(thread))
761 thread_set_flags(thread, SVC_STOPPED);
762 cfs_waitq_signal(&thread->t_ctl_waitq);
767 * Start rebalance thread. Called when the QMT is being setup
769 int qmt_start_reba_thread(struct qmt_device *qmt)
771 struct ptlrpc_thread *thread = &qmt->qmt_reba_thread;
772 struct l_wait_info lwi = { 0 };
776 rc = cfs_create_thread(qmt_reba_thread, (void *)qmt, 0);
778 CERROR("%s: failed to start rebalance thread (%d)\n",
779 qmt->qmt_svname, rc);
780 thread_set_flags(thread, SVC_STOPPED);
784 l_wait_event(thread->t_ctl_waitq,
785 thread_is_running(thread) || thread_is_stopped(thread),
792 * Stop rebalance thread. Called when the QMT is about to shutdown.
794 void qmt_stop_reba_thread(struct qmt_device *qmt)
796 struct ptlrpc_thread *thread = &qmt->qmt_reba_thread;
798 if (!thread_is_stopped(thread)) {
799 struct l_wait_info lwi = { 0 };
801 thread_set_flags(thread, SVC_STOPPING);
802 cfs_waitq_signal(&thread->t_ctl_waitq);
804 l_wait_event(thread->t_ctl_waitq, thread_is_stopped(thread),
807 LASSERT(cfs_list_empty(&qmt->qmt_reba_list));