Whamcloud - gitweb
LU-16097 quota: release preacquired quota when over limits
[fs/lustre-release.git] / lustre / quota / qmt_lock.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
20  *
21  * GPL HEADER END
22  */
23 /*
24  * Copyright (c) 2012, 2017, Intel Corporation.
25  * Use is subject to license terms.
26  *
27  * Author: Johann Lombardi <johann.lombardi@intel.com>
28  * Author: Niu    Yawei    <yawei.niu@intel.com>
29  */
30
31 #define DEBUG_SUBSYSTEM S_LQUOTA
32
33 #include <linux/kthread.h>
34
35 #include <lustre_dlm.h>
36 #include <lustre_swab.h>
37 #include <obd_class.h>
38
39 #include "qmt_internal.h"
40
41 /* intent policy function called from mdt_intent_opc() when the intent is of
42  * quota type */
43 int qmt_intent_policy(const struct lu_env *env, struct lu_device *ld,
44                       struct ptlrpc_request *req, struct ldlm_lock **lockp,
45                       int flags)
46 {
47         struct qmt_device       *qmt = lu2qmt_dev(ld);
48         struct ldlm_intent      *it;
49         struct quota_body       *reqbody;
50         struct quota_body       *repbody;
51         struct obd_uuid         *uuid;
52         struct lquota_lvb       *lvb;
53         struct ldlm_resource    *res = (*lockp)->l_resource;
54         struct ldlm_reply       *ldlm_rep;
55         int                      rc, lvb_len;
56         ENTRY;
57
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));
61
62         /* extract quota body and intent opc */
63         it = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
64         if (it == NULL)
65                 RETURN(err_serious(-EFAULT));
66
67         reqbody = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_BODY);
68         if (reqbody == NULL)
69                 RETURN(err_serious(-EFAULT));
70
71         /* prepare reply */
72         rc = req_capsule_server_pack(&req->rq_pill);
73         if (rc != 0) {
74                 CERROR("Can't pack response, rc %d\n", rc);
75                 RETURN(err_serious(rc));
76         }
77
78         repbody = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_BODY);
79         if (repbody == NULL)
80                 RETURN(err_serious(-EFAULT));
81
82         ldlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
83         if (ldlm_rep == NULL)
84                 RETURN(err_serious(-EFAULT));
85
86         uuid = &(*lockp)->l_export->exp_client_uuid;
87         switch (it->opc) {
88
89         case IT_QUOTA_DQACQ: {
90                 struct lquota_entry     *lqe;
91                 struct ldlm_lock        *lock;
92                 int idx, stype;
93
94                 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
95                         /* acquire on global lock? something is wrong ... */
96                         GOTO(out, rc = -EPROTO);
97
98                 /* verify global lock isn't stale */
99                 if (!lustre_handle_is_used(&reqbody->qb_glb_lockh))
100                         GOTO(out, rc = -ENOLCK);
101
102                 lock = ldlm_handle2lock(&reqbody->qb_glb_lockh);
103                 if (lock == NULL)
104                         GOTO(out, rc = -ENOLCK);
105                 LDLM_LOCK_PUT(lock);
106
107                 stype = qmt_uuid2idx(uuid, &idx);
108                 if (stype < 0)
109                         GOTO(out, rc = -EINVAL);
110
111                 /* TODO: it seems we don't need to get lqe from
112                  * lq_lvb_data anymore ... And do extra get
113                  * and put on it */
114                 lqe = res->lr_lvb_data;
115                 LASSERT(lqe != NULL);
116                 lqe_getref(lqe);
117
118                 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), stype,
119                                           lqe_qtype(lqe), &reqbody->qb_id,
120                                           NULL, idx);
121                 if (rc) {
122                         lqe_putref(lqe);
123                         GOTO(out, rc);
124                 }
125
126                 /* acquire quota space */
127                 rc = qmt_dqacq0(env, qmt, uuid,
128                                 reqbody->qb_flags, reqbody->qb_count,
129                                 reqbody->qb_usage, repbody,
130                                 qmt_dom(lqe_rtype(lqe), stype) ? -1 : idx);
131                 lqe_putref(lqe);
132                 qti_lqes_fini(env);
133                 if (rc)
134                         GOTO(out, rc);
135                 break;
136         }
137
138         case IT_QUOTA_CONN:
139                 /* new connection from slave */
140
141                 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0)
142                         /* connection on per-ID lock? something is wrong ... */
143                         GOTO(out, rc = -EPROTO);
144
145                 rc = qmt_pool_new_conn(env, qmt, &reqbody->qb_fid,
146                                        &repbody->qb_slv_fid,
147                                        &repbody->qb_slv_ver, uuid);
148                 if (rc)
149                         GOTO(out, rc);
150                 break;
151
152         default:
153                 CERROR("%s: invalid intent opcode: %llu\n", qmt->qmt_svname,
154                        it->opc);
155                 GOTO(out, rc = -EINVAL);
156         }
157
158         /* on success, pack lvb in reply */
159         lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB);
160         lvb_len = ldlm_lvbo_size(*lockp);
161         lvb_len = ldlm_lvbo_fill(*lockp, lvb, &lvb_len);
162         if (lvb_len < 0)
163                 GOTO(out, rc = lvb_len);
164
165         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len, RCL_SERVER);
166 out:
167         ldlm_rep->lock_policy_res2 = clear_serious(rc);
168         EXIT;
169         return ELDLM_OK;
170 }
171
172 /*
173  * Initialize quota LVB associated with quota indexes.
174  * Called with res->lr_lvb_sem held
175  */
176 int qmt_lvbo_init(struct lu_device *ld, struct ldlm_resource *res)
177 {
178         struct lu_env           *env;
179         struct qmt_thread_info  *qti;
180         struct qmt_device       *qmt = lu2qmt_dev(ld);
181         int                      pool_type, qtype;
182         int                      rc;
183         ENTRY;
184
185         LASSERT(res != NULL);
186
187         if (res->lr_type != LDLM_PLAIN)
188                 RETURN(-ENOTSUPP);
189
190         if (res->lr_lvb_data ||
191             res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
192                 RETURN(0);
193
194         env = lu_env_find();
195         LASSERT(env);
196         qti = qmt_info(env);
197
198         /* extract global index FID and quota identifier */
199         fid_extract_from_quota_res(&qti->qti_fid, &qti->qti_id, &res->lr_name);
200
201         /* sanity check the global index FID */
202         rc = lquota_extract_fid(&qti->qti_fid, &pool_type, &qtype);
203         if (rc) {
204                 CERROR("can't extract glb index information from FID "DFID"\n",
205                        PFID(&qti->qti_fid));
206                 GOTO(out, rc);
207         }
208
209         if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
210                 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
211                  * we are thus dealing with an ID lock. */
212                 struct qmt_pool_info    *pool;
213                 struct lquota_entry     *lqe;
214                 struct lqe_glbl_data    *lgd;
215
216                 pool = qmt_pool_lookup_glb(env, qmt, pool_type);
217                 if (IS_ERR(pool))
218                         GOTO(out, rc = -ENOMEM);
219
220                 /* Find the quota entry associated with the quota id */
221                 lqe = qmt_pool_lqe_lookup(env, qmt, pool_type, qtype,
222                                           &qti->qti_id, NULL);
223                 if (IS_ERR(lqe)) {
224                         qpi_putref(env, pool);
225                         GOTO(out, rc = PTR_ERR(lqe));
226                 }
227
228                 /* TODO: need something like qmt_extend_lqe_gd that has
229                  * to be calledeach time when qpi_slv_nr is incremented */
230                 lgd = qmt_alloc_lqe_gd(pool, qtype);
231                 if (!lgd) {
232                         lqe_putref(lqe);
233                         qpi_putref(env, pool);
234                         GOTO(out, rc = -ENOMEM);
235                 }
236
237                 qmt_setup_lqe_gd(env, qmt, lqe, lgd, pool_type);
238
239                 /* store reference to lqe in lr_lvb_data */
240                 res->lr_lvb_data = lqe;
241                 qpi_putref(env, pool);
242                 LQUOTA_DEBUG(lqe, "initialized res lvb");
243         } else {
244                 struct dt_object        *obj;
245
246                 /* lookup global index */
247                 obj = dt_locate(env, qmt->qmt_child, &qti->qti_fid);
248                 if (IS_ERR(obj))
249                         GOTO(out, rc = PTR_ERR(obj));
250                 if (!dt_object_exists(obj)) {
251                         dt_object_put(env, obj);
252                         GOTO(out, rc = -ENOENT);
253                 }
254
255                 /* store reference to global index object in lr_lvb_data */
256                 res->lr_lvb_data = obj;
257                 CDEBUG(D_QUOTA, DFID" initialized lvb\n", PFID(&qti->qti_fid));
258         }
259
260         res->lr_lvb_len = sizeof(struct lquota_lvb);
261         EXIT;
262 out:
263         return rc;
264 }
265
266 /* clear lge_qunit/edquot_nu flags -
267  * slave recieved new qunit and edquot.
268  *
269  * \retval      true if revoke is needed - qunit
270  *              for this slave reaches least_qunit
271  */
272 static bool qmt_clear_lgeg_arr_nu(struct lquota_entry *lqe, int stype, int idx)
273 {
274         unsigned long least_qunit = lqe2qpi(lqe)->qpi_least_qunit;
275         struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
276
277         /* There is no array to store lge for the case of DOM.
278          * Ignore it until MDT pools will be ready. */
279         if (!qmt_dom(lqe_rtype(lqe), stype)) {
280                 lqe->lqe_glbl_data->lqeg_arr[idx].lge_qunit_nu = 0;
281                 lqe->lqe_glbl_data->lqeg_arr[idx].lge_edquot_nu = 0;
282
283                 /* We shouldn't call revoke for DOM case, it will be updated
284                  * at qmt_id_lock_glimpse. */
285                 return (lgd->lqeg_arr[idx].lge_qunit == least_qunit);
286         }
287
288         return false;
289 }
290
291 static bool qmt_set_revoke(struct lu_env *env, struct lquota_entry *lqe,
292                           int stype, int idx)
293 {
294         unsigned long least_qunit = lqe2qpi(lqe)->qpi_least_qunit;
295         struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
296         bool notify = false;
297
298         if (lgd->lqeg_arr[idx].lge_qunit == least_qunit) {
299                 int i;
300
301                 qti_lqes_write_lock(env);
302                 for (i = 0; i < qti_lqes_cnt(env); i++) {
303                         LQUOTA_DEBUG(qti_lqes(env)[i],
304                                      "idx %d lge_qunit %llu least_qunit %lu\n",
305                                      idx, lgd->lqeg_arr[idx].lge_qunit,
306                                      least_qunit);
307                         if (qti_lqes(env)[i]->lqe_qunit == least_qunit) {
308                                 qti_lqes(env)[i]->lqe_revoke_time =
309                                                         ktime_get_seconds();
310                                 notify |= qmt_adjust_edquot(qti_lqes(env)[i],
311                                                   ktime_get_real_seconds());
312                         }
313                 }
314                 qti_lqes_write_unlock(env);
315         }
316         return notify;
317 }
318
319 /*
320  * Update LVB associated with the global quota index.
321  * This function is called from the DLM itself after a glimpse callback, in this
322  * case valid ptlrpc request is passed.
323  */
324 int qmt_lvbo_update(struct lu_device *ld, struct ldlm_resource *res,
325                     struct ptlrpc_request *req, int increase_only)
326 {
327         struct lu_env           *env;
328         struct qmt_thread_info  *qti;
329         struct qmt_device       *qmt = lu2qmt_dev(ld);
330         struct lquota_entry     *lqe;
331         struct lquota_lvb       *lvb;
332         struct ldlm_lock        *lock;
333         struct obd_export       *exp;
334         bool                     need_revoke;
335         int                      rc = 0, idx, stype;
336         ENTRY;
337
338         LASSERT(res != NULL);
339
340         if (req == NULL)
341                 RETURN(0);
342
343         if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
344                 /* no need to update lvb for global quota locks */
345                 RETURN(0);
346
347         lvb = req_capsule_server_swab_get(&req->rq_pill, &RMF_DLM_LVB,
348                                           lustre_swab_lquota_lvb);
349         if (lvb == NULL) {
350                 CERROR("%s: failed to extract lvb from request\n",
351                        qmt->qmt_svname);
352                 RETURN(-EFAULT);
353         }
354
355         lqe = res->lr_lvb_data;
356         LASSERT(lqe != NULL);
357         lqe_getref(lqe);
358
359         /* allocate environement */
360         env = lu_env_find();
361         LASSERT(env);
362         qti = qmt_info(env);
363
364         /* The request is a glimpse callback which was sent via the
365          * reverse import to the slave. What we care about here is the
366          * export associated with the slave and req->rq_export is
367          * definitely not what we are looking for (it is actually set to
368          * NULL here).
369          * Therefore we extract the lock from the request argument
370          * and use lock->l_export. */
371         lock = ldlm_request_lock(req);
372         if (IS_ERR(lock)) {
373                 CERROR("%s: failed to get lock from request!\n",
374                        qmt->qmt_svname);
375                 GOTO(out, rc = PTR_ERR(lock));
376         }
377
378         exp = class_export_get(lock->l_export);
379         if (exp == NULL) {
380                 CERROR("%s: failed to get export from lock!\n",
381                        qmt->qmt_svname);
382                 GOTO(out, rc = -EFAULT);
383         }
384
385         stype = qmt_uuid2idx(&exp->exp_client_uuid, &idx);
386         if (stype < 0)
387                 GOTO(out_exp, rc = stype);
388
389         need_revoke = qmt_clear_lgeg_arr_nu(lqe, stype, idx);
390         if (lvb->lvb_id_rel == 0) {
391                 /* nothing to release */
392                 if (lvb->lvb_id_may_rel != 0)
393                         /* but might still release later ... */
394                         lqe->lqe_may_rel += lvb->lvb_id_may_rel;
395         }
396
397         if (!need_revoke && lvb->lvb_id_rel == 0)
398                 GOTO(out_exp, rc = 0);
399
400         rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), stype,
401                                   lqe_qtype(lqe), &lqe->lqe_id, NULL, idx);
402         if (rc)
403                 GOTO(out_exp, rc);
404
405         if (need_revoke && qmt_set_revoke(env, lqe, stype, idx)) {
406                 mutex_lock(&lqe->lqe_glbl_data_lock);
407                 if (lqe->lqe_glbl_data) {
408                         qmt_seed_glbe_edquot(env, lqe->lqe_glbl_data);
409                         qmt_id_lock_notify(qmt, lqe);
410                 }
411                 mutex_unlock(&lqe->lqe_glbl_data_lock);
412         }
413
414         if (lvb->lvb_id_rel) {
415                 LQUOTA_DEBUG(lqe, "releasing:%llu may release:%llu",
416                              lvb->lvb_id_rel, lvb->lvb_id_may_rel);
417
418                 /* release quota space */
419                 rc = qmt_dqacq0(env, qmt, &exp->exp_client_uuid,
420                                 QUOTA_DQACQ_FL_REL, lvb->lvb_id_rel,
421                                 0, &qti->qti_body,
422                                 qmt_dom(lqe_rtype(lqe), stype) ? -1 : idx);
423                 if (rc || qti->qti_body.qb_count != lvb->lvb_id_rel)
424                         LQUOTA_ERROR(lqe,
425                                      "failed to release quota space on glimpse %llu!=%llu : rc = %d\n",
426                                      qti->qti_body.qb_count,
427                                      lvb->lvb_id_rel, rc);
428         }
429         qti_lqes_fini(env);
430         if (rc)
431                 GOTO(out_exp, rc);
432         EXIT;
433 out_exp:
434         class_export_put(exp);
435 out:
436         lqe_putref(lqe);
437         return rc;
438 }
439
440 /*
441  * Report size of lvb to ldlm layer in order to allocate lvb buffer
442  * As far as quota locks are concerned, the size is static and is the same
443  * for both global and per-ID locks which shares the same lvb format.
444  */
445 int qmt_lvbo_size(struct lu_device *ld, struct ldlm_lock *lock)
446 {
447         return sizeof(struct lquota_lvb);
448 }
449
450 /*
451  * Fill request buffer with quota lvb
452  */
453 int qmt_lvbo_fill(struct lu_device *ld, struct ldlm_lock *lock, void *lvb,
454                   int lvblen)
455 {
456         struct ldlm_resource *res = lock->l_resource;
457         struct lquota_lvb *qlvb = lvb;
458         struct lu_env *env;
459         int rc;
460         ENTRY;
461
462         LASSERT(res != NULL);
463         rc = 0;
464
465         if (res->lr_type != LDLM_PLAIN || res->lr_lvb_data == NULL ||
466             res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
467                 RETURN(-EINVAL);
468
469         env = lu_env_find();
470         LASSERT(env);
471
472         if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
473                 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
474                  * we are thus dealing with an ID lock. */
475                 struct lquota_entry *lqe = res->lr_lvb_data;
476                 struct qmt_device *qmt;
477                 struct obd_uuid *uuid;
478                 int idx;
479
480                 uuid = &(lock)->l_export->exp_client_uuid;
481                 rc = qmt_uuid2idx(uuid, &idx);
482                 if (rc < 0)
483                         RETURN(rc);
484                 qmt = lu2qmt_dev(ld);
485                 /* return current qunit value & edquot flags in lvb */
486                 lqe_getref(lqe);
487                 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc,
488                                           lqe_qtype(lqe), &lqe->lqe_id,
489                                           NULL, idx);
490                 if (!rc) {
491                         qlvb->lvb_id_qunit = qti_lqes_min_qunit(env);
492                         qlvb->lvb_flags = 0;
493                         if (qti_lqes_edquot(env))
494                                 qlvb->lvb_flags = LQUOTA_FL_EDQUOT;
495                         qti_lqes_fini(env);
496                 }
497                 CDEBUG(D_QUOTA, "uuid %s lqe_id %lu, edquot %llu qunit %llu\n",
498                        (char *)uuid, (unsigned long)lqe->lqe_id.qid_uid,
499                        qlvb->lvb_flags, qlvb->lvb_id_qunit);
500                 lqe_putref(lqe);
501         } else {
502                 /* global quota lock */
503                 struct dt_object *obj = res->lr_lvb_data;
504
505                 /* return current version of global index */
506                 qlvb->lvb_glb_ver = dt_version_get(env, obj);
507         }
508
509         RETURN(rc = rc ?: sizeof(struct lquota_lvb));
510 }
511
512 /*
513  * Free lvb associated with a given ldlm resource
514  * we don't really allocate a lvb, lr_lvb_data just points to
515  * the appropriate backend structures.
516  */
517 int qmt_lvbo_free(struct lu_device *ld, struct ldlm_resource *res)
518 {
519         ENTRY;
520
521         if (res->lr_lvb_data == NULL)
522                 RETURN(0);
523
524         if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
525                 struct lquota_entry *lqe = res->lr_lvb_data;
526                 struct lqe_glbl_data *lgd;
527
528                 mutex_lock(&lqe->lqe_glbl_data_lock);
529                 lgd = lqe->lqe_glbl_data;
530                 lqe->lqe_glbl_data = NULL;
531                 mutex_unlock(&lqe->lqe_glbl_data_lock);
532                 qmt_free_lqe_gd(lgd);
533
534                 /* release lqe reference */
535                 lqe_putref(lqe);
536         } else {
537                 struct dt_object *obj = res->lr_lvb_data;
538                 /* release object reference */
539                 dt_object_put(lu_env_find(), obj);
540         }
541
542         res->lr_lvb_data = NULL;
543         res->lr_lvb_len  = 0;
544
545         RETURN(0);
546 }
547
548 typedef int (*qmt_glimpse_cb_t)(struct ldlm_lock *, struct lquota_entry *);
549
550 struct qmt_gl_lock_array {
551         unsigned long             q_max;
552         unsigned long             q_cnt;
553         struct ldlm_lock        **q_locks;
554 };
555
556 static void qmt_free_lock_array(struct qmt_gl_lock_array *array)
557 {
558         int i;
559
560         if (array->q_max == 0) {
561                 LASSERT(array->q_locks == NULL);
562                 return;
563         }
564
565         for (i = 0; i < array->q_cnt; i++) {
566                 LASSERT(array->q_locks[i]);
567                 LDLM_LOCK_RELEASE(array->q_locks[i]);
568                 array->q_locks[i] = NULL;
569         }
570         array->q_cnt = 0;
571         OBD_FREE_PTR_ARRAY(array->q_locks, array->q_max);
572         array->q_locks = NULL;
573         array->q_max = 0;
574 }
575
576 static int qmt_alloc_lock_array(struct ldlm_resource *res,
577                                 struct qmt_gl_lock_array *array,
578                                 qmt_glimpse_cb_t cb, void *arg)
579 {
580         struct lquota_entry *lqe = arg;
581         struct list_head *pos;
582         unsigned long count = 0;
583         int fail_cnt = 0;
584         ENTRY;
585
586         LASSERT(!array->q_max && !array->q_cnt && !array->q_locks);
587 again:
588         if (cb)
589                 mutex_lock(&lqe->lqe_glbl_data_lock);
590         lock_res(res);
591         /* scan list of granted locks */
592         list_for_each(pos, &res->lr_granted) {
593                 struct ldlm_lock *lock;
594                 int rc;
595
596                 lock = list_entry(pos, struct ldlm_lock, l_res_link);
597                 LASSERT(lock->l_export);
598
599                 if (cb != NULL) {
600                         rc = cb(lock, arg);
601                         /* slave should not be notified */
602                         if (rc == 0)
603                                 continue;
604                 }
605
606                 count++;
607                 if (array->q_max != 0 && array->q_cnt < array->q_max) {
608                         array->q_locks[array->q_cnt] = LDLM_LOCK_GET(lock);
609                         array->q_cnt++;
610                 }
611         }
612         unlock_res(res);
613         if (cb)
614                 mutex_unlock(&lqe->lqe_glbl_data_lock);
615
616         if (count > array->q_max) {
617                 qmt_free_lock_array(array);
618                 if (++fail_cnt > 5)
619                         RETURN(-EAGAIN);
620                 /*
621                  * allocate more slots in case of more qualified locks are
622                  * found during next loop
623                  */
624                 array->q_max = count + count / 2 + 10;
625                 count = 0;
626                 LASSERT(array->q_locks == NULL && array->q_cnt == 0);
627                 OBD_ALLOC_PTR_ARRAY(array->q_locks, array->q_max);
628                 if (array->q_locks == NULL) {
629                         array->q_max = 0;
630                         RETURN(-ENOMEM);
631                 }
632
633                 goto again;
634         }
635         RETURN(0);
636 }
637
638 static void qmt_setup_id_desc(struct ldlm_lock *lock, union ldlm_gl_desc *desc,
639                               struct lquota_entry *lqe)
640 {
641         struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
642         int idx, stype;
643         __u64 qunit;
644         bool edquot;
645
646         stype = qmt_uuid2idx(uuid, &idx);
647         LASSERT(stype >= 0);
648
649         /* DOM case - set global lqe settings */
650         if (qmt_dom(lqe_rtype(lqe), stype)) {
651                 edquot = lqe->lqe_edquot;
652                 qunit = lqe->lqe_qunit;
653         } else {
654                 struct lqe_glbl_data *lgd;
655
656                 mutex_lock(&lqe->lqe_glbl_data_lock);
657                 lgd = lqe->lqe_glbl_data;
658                 if (lgd) {
659                         edquot = lgd->lqeg_arr[idx].lge_edquot;
660                         qunit = lgd->lqeg_arr[idx].lge_qunit;
661                 } else {
662                         edquot = lqe->lqe_edquot;
663                         qunit = lqe->lqe_qunit;
664                 }
665                 mutex_unlock(&lqe->lqe_glbl_data_lock);
666         }
667
668         /* fill glimpse descriptor with lqe settings */
669         desc->lquota_desc.gl_flags = edquot ? LQUOTA_FL_EDQUOT : 0;
670         desc->lquota_desc.gl_qunit = qunit;
671         CDEBUG(D_QUOTA, "setup desc: stype %d idx %d, edquot %llu qunit %llu\n",
672                          stype, idx, desc->lquota_desc.gl_flags,
673                          desc->lquota_desc.gl_qunit);
674 }
675
676 /*
677  * Send glimpse callback to slaves holding a lock on resource \res.
678  * This is used to notify slaves of new quota settings or to claim quota space
679  * back.
680  *
681  * \param env  - is the environment passed by the caller
682  * \param qmt  - is the quota master target
683  * \param res  - is the dlm resource associated with the quota object
684  * \param desc - is the glimpse descriptor to pack in glimpse callback
685  * \param cb   - is the callback function called on every lock and determine
686  *               whether a glimpse should be issued
687  * \param arg  - is an opaq parameter passed to the callback function
688  */
689 static int qmt_glimpse_lock(const struct lu_env *env, struct qmt_device *qmt,
690                             struct ldlm_resource *res, union ldlm_gl_desc *desc,
691                             qmt_glimpse_cb_t cb, struct lquota_entry *lqe)
692 {
693         union ldlm_gl_desc *descs = NULL;
694         struct list_head *tmp, *pos;
695         LIST_HEAD(gl_list);
696         struct qmt_gl_lock_array locks;
697         unsigned long i, locks_count;
698         int rc = 0;
699         ENTRY;
700
701         memset(&locks, 0, sizeof(locks));
702         rc = qmt_alloc_lock_array(res, &locks, cb, lqe);
703         if (rc) {
704                 CERROR("%s: failed to allocate glimpse lock array (%d)\n",
705                        qmt->qmt_svname, rc);
706                 RETURN(rc);
707         }
708         if (!locks.q_cnt) {
709                 CDEBUG(D_QUOTA, "%s: no granted locks to send glimpse\n",
710                        qmt->qmt_svname);
711                 RETURN(0);
712         }
713         CDEBUG(D_QUOTA, "found granted locks %lu\n", locks.q_cnt);
714         locks_count = locks.q_cnt;
715
716         /* Use one desc for all works, when called from qmt_glb_lock_notify */
717         if (cb && locks.q_cnt > 1) {
718                 /* TODO: think about to store this preallocated descs
719                  * in lqe_global in lqeg_arr as a part of lqe_glbl_entry.
720                  * The benefit is that we don't need to allocate/free
721                  * and setup this descs each time. But the drawback is
722                  * memory use (sizeof ldlm_gl_desc * OST_COUNT * user_number).
723                  * for examfple it could be 88 * 256 * 10 000 about 225 MB. */
724                 OBD_ALLOC(descs,
725                           sizeof(struct ldlm_gl_lquota_desc) * locks.q_cnt);
726                 if (!descs) {
727                         CERROR("%s: alloc glimpse lock array failed: rc = %d\n",
728                                qmt->qmt_svname, rc);
729                         qmt_free_lock_array(&locks);
730                         RETURN(-ENOMEM);
731                 }
732         }
733
734         for (i = locks.q_cnt; i > 0; i--) {
735                 struct ldlm_glimpse_work *work;
736
737                 OBD_ALLOC_PTR(work);
738                 if (work == NULL) {
739                         CERROR("%s: failed to notify a lock.\n",
740                                qmt->qmt_svname);
741                         continue;
742                 }
743
744                 if (cb) {
745                         if (descs)
746                                 desc = &descs[i - 1];
747                         qmt_setup_id_desc(locks.q_locks[i - 1], desc, lqe);
748                         work->gl_interpret_data = lqe;
749                 }
750
751                 list_add_tail(&work->gl_list, &gl_list);
752                 work->gl_lock  = locks.q_locks[i - 1];
753                 work->gl_flags = 0;
754                 work->gl_desc  = desc;
755
756                 locks.q_locks[i - 1] = NULL;
757                 locks.q_cnt--;
758         }
759
760         qmt_free_lock_array(&locks);
761
762         if (list_empty(&gl_list)) {
763                 CDEBUG(D_QUOTA, "%s: nobody to notify\n", qmt->qmt_svname);
764                 GOTO(out, rc = 0);
765         }
766
767         /* issue glimpse callbacks to all connected slaves */
768         rc = ldlm_glimpse_locks(res, &gl_list);
769
770         list_for_each_safe(pos, tmp, &gl_list) {
771                 struct ldlm_glimpse_work *work;
772
773                 work = list_entry(pos, struct ldlm_glimpse_work, gl_list);
774
775                 list_del(&work->gl_list);
776                 CERROR("%s: failed to notify %s of new quota settings\n",
777                        qmt->qmt_svname,
778                        obd_uuid2str(&work->gl_lock->l_export->exp_client_uuid));
779                 LDLM_LOCK_RELEASE(work->gl_lock);
780                 OBD_FREE_PTR(work);
781         }
782 out:
783         if (descs)
784                 OBD_FREE(descs,
785                          sizeof(struct ldlm_gl_lquota_desc) * locks_count);
786
787         RETURN(rc);
788 }
789
790 /*
791  * Send glimpse request to all global quota locks to push new quota setting to
792  * slaves.
793  *
794  * \param env - is the environment passed by the caller
795  * \param lqe - is the lquota entry which has new settings
796  * \param ver - is the version associated with the setting change
797  */
798 void qmt_glb_lock_notify(const struct lu_env *env, struct lquota_entry *lqe,
799                          __u64 ver)
800 {
801         struct qmt_thread_info  *qti = qmt_info(env);
802         struct qmt_pool_info    *pool = lqe2qpi(lqe);
803         struct ldlm_resource    *res = NULL;
804         ENTRY;
805
806         lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
807
808         /* send glimpse callback to notify slaves of new quota settings */
809         qti->qti_gl_desc.lquota_desc.gl_id        = lqe->lqe_id;
810         qti->qti_gl_desc.lquota_desc.gl_flags     = 0;
811         if (lqe->lqe_is_default) {
812                 qti->qti_gl_desc.lquota_desc.gl_hardlimit = 0;
813                 qti->qti_gl_desc.lquota_desc.gl_softlimit = 0;
814                 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
815                                                         LQUOTA_FLAG_DEFAULT);
816
817         } else if (lqe->lqe_is_deleted) {
818                 qti->qti_gl_desc.lquota_desc.gl_hardlimit = 0;
819                 qti->qti_gl_desc.lquota_desc.gl_softlimit = 0;
820                 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
821                                                         LQUOTA_FLAG_DELETED);
822         } else if (lqe->lqe_is_reset) {
823                 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
824                 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
825                 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
826                                                         LQUOTA_FLAG_RESET);
827         } else if (lqe->lqe_granted > lqe->lqe_hardlimit) {
828                 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
829                 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
830                 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
831                                                         LQUOTA_FLAG_REVOKE);
832         } else {
833                 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
834                 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
835                 qti->qti_gl_desc.lquota_desc.gl_time = lqe->lqe_gracetime;
836         }
837         qti->qti_gl_desc.lquota_desc.gl_ver       = ver;
838
839         /* look up ldlm resource associated with global index */
840         fid_build_reg_res_name(&qti->qti_fid, &qti->qti_resid);
841         res = ldlm_resource_get(pool->qpi_qmt->qmt_ns, &qti->qti_resid,
842                                 LDLM_PLAIN, 0);
843         if (IS_ERR(res)) {
844                 /* this might happen if no slaves have enqueued global quota
845                  * locks yet */
846                 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource associated "
847                              "with "DFID, PFID(&qti->qti_fid));
848                 RETURN_EXIT;
849         }
850
851         qmt_glimpse_lock(env, pool->qpi_qmt, res, &qti->qti_gl_desc,
852                          NULL, NULL);
853         ldlm_resource_putref(res);
854         EXIT;
855 }
856
857 /* Callback function used to select locks that should be glimpsed when
858  * broadcasting the new qunit value */
859 static int qmt_id_lock_cb(struct ldlm_lock *lock, struct lquota_entry *lqe)
860 {
861         struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
862         struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
863         int idx;
864         int stype = qmt_uuid2idx(uuid, &idx);
865
866         LASSERT(stype == QMT_STYPE_OST || stype == QMT_STYPE_MDT);
867
868         /* Quota pools support only OSTs, despite MDTs also could be registered
869          * as LQUOTA_RES_DT devices(DOM). */
870         if (qmt_dom(lqe_rtype(lqe), stype))
871                 return 1;
872         else
873                 return lgd ? lgd->lqeg_arr[idx].lge_edquot_nu ||
874                        lgd->lqeg_arr[idx].lge_qunit_nu : 0;
875 }
876
877
878 /*
879  * Send glimpse request on per-ID lock to push new qunit value to slave.
880  *
881  * \param env  - is the environment passed by the caller
882  * \param qmt  - is the quota master target device
883  * \param lqe  - is the lquota entry with the new qunit value
884  * \param uuid - is the uuid of the slave acquiring space, if any
885  */
886 static void qmt_id_lock_glimpse(const struct lu_env *env,
887                                 struct qmt_device *qmt,
888                                 struct lquota_entry *lqe, struct obd_uuid *uuid)
889 {
890         struct qmt_thread_info  *qti = qmt_info(env);
891         struct qmt_pool_info    *pool = lqe2qpi(lqe);
892         struct ldlm_resource    *res = NULL;
893         ENTRY;
894
895         if (!lqe->lqe_enforced)
896                 RETURN_EXIT;
897
898         lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
899         fid_build_quota_res_name(&qti->qti_fid, &lqe->lqe_id, &qti->qti_resid);
900         res = ldlm_resource_get(qmt->qmt_ns, &qti->qti_resid, LDLM_PLAIN, 0);
901         if (IS_ERR(res)) {
902                 /* this might legitimately happens if slaves haven't had the
903                  * opportunity to enqueue quota lock yet. */
904                 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource for per-ID "
905                              "lock "DFID, PFID(&qti->qti_fid));
906                 lqe_write_lock(lqe);
907                 if (lqe->lqe_revoke_time == 0 &&
908                     lqe->lqe_qunit == pool->qpi_least_qunit)
909                         lqe->lqe_revoke_time = ktime_get_seconds();
910                 lqe_write_unlock(lqe);
911                 RETURN_EXIT;
912         }
913
914         lqe_write_lock(lqe);
915         /*
916          * It is possible to add an lqe in a 2nd time while the same lqe
917          * from the 1st time is still sending glimpse
918          */
919         if (lqe->lqe_gl)
920                 GOTO(out, 0);
921         /* The purpose of glimpse callback on per-ID lock is twofold:
922          * - notify slaves of new qunit value and hope they will release some
923          *   spare quota space in return
924          * - notify slaves that master ran out of quota space and there is no
925          *   need to send acquire request any more until further notice */
926
927         /* TODO: it is not clear how to implement below case for all lqes
928          * from where slaves will be notified in qmt_glimpse_lock. Because
929          * here we have just global lqe with an array of OSTs that should
930          * be notified. Theoretically we can find all lqes that includes
931          * these OSTs, but it is not trivial. So I would propose to move
932          * this case to another place ... */
933         if (lqe->lqe_revoke_time == 0 &&
934             lqe->lqe_qunit == pool->qpi_least_qunit)
935                 /* reset lqe_may_rel, it will be updated on glimpse callback
936                  * replies if needed */
937                 lqe->lqe_may_rel = 0;
938
939         lqe->lqe_gl = true;
940         lqe_write_unlock(lqe);
941
942         /* issue glimpse callback to slaves */
943         if (lqe->lqe_glbl_data)
944                 qmt_glimpse_lock(env, qmt, res, &qti->qti_gl_desc,
945                                  qmt_id_lock_cb, lqe);
946
947         lqe_write_lock(lqe);
948         if (lqe->lqe_revoke_time == 0 &&
949             lqe->lqe_qunit == pool->qpi_least_qunit) {
950                 lqe->lqe_revoke_time = ktime_get_seconds();
951                 qmt_adjust_edquot(lqe, ktime_get_real_seconds());
952         }
953         LASSERT(lqe->lqe_gl);
954         lqe->lqe_gl = false;
955 out:
956         lqe_write_unlock(lqe);
957         ldlm_resource_putref(res);
958         EXIT;
959 }
960
961 /*
962  * Schedule a glimpse request on per-ID locks to push new qunit value or
963  * edquot flag to quota slaves.
964  *
965  * \param qmt  - is the quota master target device
966  * \param lqe  - is the lquota entry with the new qunit value
967  */
968 void qmt_id_lock_notify(struct qmt_device *qmt, struct lquota_entry *lqe)
969 {
970         bool    added = false;
971         ENTRY;
972
973         LASSERT(lqe->lqe_is_global);
974         lqe_getref(lqe);
975         spin_lock(&qmt->qmt_reba_lock);
976         if (!qmt->qmt_stopping && list_empty(&lqe->lqe_link)) {
977                 list_add_tail(&lqe->lqe_link, &qmt->qmt_reba_list);
978                 added = true;
979                 if (qmt->qmt_reba_task)
980                         wake_up_process(qmt->qmt_reba_task);
981         }
982         spin_unlock(&qmt->qmt_reba_lock);
983
984         if (!added)
985                 lqe_putref(lqe);
986         EXIT;
987 }
988
989 struct qmt_reba_args {
990         struct qmt_device       *qra_dev;
991         struct lu_env            qra_env;
992         struct completion       *qra_started;
993 };
994
995 #ifndef TASK_IDLE
996 #define TASK_IDLE TASK_INTERRUPTIBLE
997 #endif
998
999 /*
1000  * The rebalance thread is in charge of sending glimpse callbacks on per-ID
1001  * quota locks owned by slaves in order to notify them of:
1002  * - a qunit shrink in which case slaves might release quota space back in
1003  *   glimpse reply.
1004  * - set/clear edquot flag used to cache the "quota exhausted" state of the
1005  *   master. When the flag is set, slaves know that there is no need to
1006  *   try to acquire quota from the master since this latter has already
1007  *   distributed all the space.
1008  */
1009 static int qmt_reba_thread(void *_args)
1010 {
1011         struct qmt_reba_args    *args = _args;
1012         struct qmt_device       *qmt = args->qra_dev;
1013         struct lu_env           *env = &args->qra_env;
1014         struct lquota_entry     *lqe, *tmp;
1015         ENTRY;
1016
1017         complete(args->qra_started);
1018         while (({set_current_state(TASK_IDLE);
1019                  !kthread_should_stop(); })) {
1020
1021                 spin_lock(&qmt->qmt_reba_lock);
1022                 list_for_each_entry_safe(lqe, tmp, &qmt->qmt_reba_list,
1023                                          lqe_link) {
1024                         __set_current_state(TASK_RUNNING);
1025                         list_del_init(&lqe->lqe_link);
1026                         spin_unlock(&qmt->qmt_reba_lock);
1027
1028                         /* lqe_ref == 1 means we hold the last ref,
1029                          * so no need to send glimpse callbacks.
1030                          */
1031                         if (!kthread_should_stop() &&
1032                             atomic_read(&lqe->lqe_ref) > 1)
1033                                 qmt_id_lock_glimpse(env, qmt, lqe, NULL);
1034
1035                         lqe_putref(lqe);
1036                         spin_lock(&qmt->qmt_reba_lock);
1037                 }
1038                 spin_unlock(&qmt->qmt_reba_lock);
1039                 schedule();
1040         }
1041         __set_current_state(TASK_RUNNING);
1042
1043         lu_env_remove(env);
1044         lu_env_fini(env);
1045         OBD_FREE_PTR(args);
1046         RETURN(0);
1047 }
1048
1049 /*
1050  * Start rebalance thread. Called when the QMT is being setup
1051  */
1052 int qmt_start_reba_thread(struct qmt_device *qmt)
1053 {
1054         struct task_struct *task;
1055         struct qmt_reba_args *args;
1056         DECLARE_COMPLETION_ONSTACK(started);
1057         int rc;
1058         ENTRY;
1059
1060         OBD_ALLOC_PTR(args);
1061         if (args == NULL)
1062                 RETURN(-ENOMEM);
1063         args->qra_dev = qmt;
1064         args->qra_started = &started;
1065
1066         rc = lu_env_init(&args->qra_env, LCT_MD_THREAD);
1067         if (rc) {
1068                 CERROR("%s: failed to init env.\n", qmt->qmt_svname);
1069                 GOTO(out_env, rc);
1070         }
1071
1072         task = kthread_create(qmt_reba_thread, args,
1073                               "qmt_reba_%s", qmt->qmt_svname);
1074         if (IS_ERR(task)) {
1075                 CERROR("%s: failed to start rebalance thread (%ld)\n",
1076                        qmt->qmt_svname, PTR_ERR(task));
1077                 GOTO(out_env_fini, rc = PTR_ERR(task));
1078         }
1079
1080         rc = lu_env_add_task(&args->qra_env, task);
1081         if (rc) {
1082                 kthread_stop(task);
1083                 GOTO(out_env_fini, rc);
1084         }
1085         qmt->qmt_reba_task = task;
1086         wake_up_process(task);
1087         wait_for_completion(&started);
1088
1089         RETURN(0);
1090 out_env_fini:
1091         lu_env_fini(&args->qra_env);
1092 out_env:
1093         OBD_FREE_PTR(args);
1094         RETURN(rc);
1095 }
1096
1097 /*
1098  * Stop rebalance thread. Called when the QMT is about to shutdown.
1099  */
1100 void qmt_stop_reba_thread(struct qmt_device *qmt)
1101 {
1102         struct task_struct *task;
1103
1104         spin_lock(&qmt->qmt_reba_lock);
1105         task = qmt->qmt_reba_task;
1106         qmt->qmt_reba_task = NULL;
1107         spin_unlock(&qmt->qmt_reba_lock);
1108
1109         if (task)
1110                 kthread_stop(task);
1111
1112         LASSERT(list_empty(&qmt->qmt_reba_list));
1113 }