Whamcloud - gitweb
LU-10308 misc: update Intel copyright messages for 2017
[fs/lustre-release.git] / lustre / quota / qsd_handler.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 "qsd_internal.h"
34
35 /**
36  * helper function bumping lqe_pending_req if there is no quota request in
37  * flight for the lquota entry \a lqe. Otherwise, EBUSY is returned.
38  */
39 static inline int qsd_request_enter(struct lquota_entry *lqe)
40 {
41         /* is there already a quota request in flight? */
42         if (lqe->lqe_pending_req != 0) {
43                 LQUOTA_DEBUG(lqe, "already a request in flight");
44                 return -EBUSY;
45         }
46
47         if (lqe->lqe_pending_rel != 0) {
48                 LQUOTA_ERROR(lqe, "no request in flight with pending_rel=%llu",
49                              lqe->lqe_pending_rel);
50                 LBUG();
51         }
52
53         lqe->lqe_pending_req++;
54         return 0;
55 }
56
57 /**
58  * Companion of qsd_request_enter() dropping lqe_pending_req to 0.
59  */
60 static inline void qsd_request_exit(struct lquota_entry *lqe)
61 {
62         if (lqe->lqe_pending_req != 1) {
63                 LQUOTA_ERROR(lqe, "lqe_pending_req != 1!!!");
64                 LBUG();
65         }
66         lqe->lqe_pending_req--;
67         lqe->lqe_pending_rel = 0;
68         wake_up_all(&lqe->lqe_waiters);
69 }
70
71 /**
72  * Check whether a qsd instance is all set to send quota request to master.
73  * This includes checking whether:
74  * - the connection to master is set up and usable,
75  * - the qsd isn't stopping
76  * - reintegration has been successfully completed and all indexes are
77  *   up-to-date
78  *
79  * \param lqe - is the lquota entry for which we would like to send an quota
80  *              request
81  * \param lockh - is the remote handle of the global lock returned on success
82  *
83  * \retval 0 on success, appropriate error on failure
84  */
85 static int qsd_ready(struct lquota_entry *lqe, struct lustre_handle *lockh)
86 {
87         struct qsd_qtype_info   *qqi = lqe2qqi(lqe);
88         struct qsd_instance     *qsd = qqi->qqi_qsd;
89         struct obd_import       *imp = NULL;
90         struct ldlm_lock        *lock;
91         ENTRY;
92
93         read_lock(&qsd->qsd_lock);
94         /* is the qsd about to shut down? */
95         if (qsd->qsd_stopping) {
96                 read_unlock(&qsd->qsd_lock);
97                 LQUOTA_DEBUG(lqe, "dropping quota req since qsd is stopping");
98                 /* Target is about to shut down, client will retry */
99                 RETURN(-EINPROGRESS);
100         }
101
102         /* is the connection to the quota master ready? */
103         if (qsd->qsd_exp_valid)
104                 imp = class_exp2cliimp(qsd->qsd_exp);
105         if (imp == NULL || imp->imp_invalid) {
106                 read_unlock(&qsd->qsd_lock);
107                 LQUOTA_DEBUG(lqe, "connection to master not ready");
108                 RETURN(-ENOTCONN);
109         }
110
111         /* In most case, reintegration must have been triggered (when enable
112          * quota or on OST start), however, in rare race condition (enabling
113          * quota when starting OSTs), we might miss triggering reintegration
114          * for some qqi.
115          *
116          * If the previous reintegration failed for some reason, we'll
117          * re-trigger it here as well. */
118         if (!qqi->qqi_glb_uptodate || !qqi->qqi_slv_uptodate) {
119                 read_unlock(&qsd->qsd_lock);
120                 LQUOTA_DEBUG(lqe, "not up-to-date, dropping request and "
121                              "kicking off reintegration");
122                 qsd_start_reint_thread(qqi);
123                 RETURN(-EINPROGRESS);
124         }
125
126         /* Fill the remote global lock handle, master will check this handle
127          * to see if the slave is sending request with stale lock */
128         lustre_handle_copy(lockh, &qqi->qqi_lockh);
129         read_unlock(&qsd->qsd_lock);
130
131         if (!lustre_handle_is_used(lockh))
132                 RETURN(-ENOLCK);
133
134         lock = ldlm_handle2lock(lockh);
135         if (lock == NULL)
136                 RETURN(-ENOLCK);
137
138         /* return remote lock handle to be packed in quota request */
139         lustre_handle_copy(lockh, &lock->l_remote_handle);
140         LDLM_LOCK_PUT(lock);
141
142         RETURN(0);
143 }
144
145 /**
146  * Check whether any quota space adjustment (pre-acquire/release/report) is
147  * needed for a given quota ID. If a non-null \a qbody is passed, then the
148  * \a qbody structure (qb_count/flags/usage) is filled with appropriate data
149  * to be packed in the quota request.
150  *
151  * \param lqe   - is the lquota entry for which we would like to adjust quota
152  *                space.
153  * \param qbody - is the quota body to fill, if not NULL.
154  *
155  * \retval true  - space adjustment is required and \a qbody is filled, if not
156  *                 NULL
157  * \retval false - no space adjustment required
158  */
159 static bool qsd_calc_adjust(struct lquota_entry *lqe, struct quota_body *qbody)
160 {
161         __u64   usage, granted;
162         ENTRY;
163
164         usage   = lqe->lqe_usage;
165         usage  += lqe->lqe_pending_write + lqe->lqe_waiting_write;
166         granted = lqe->lqe_granted;
167
168         if (qbody != NULL)
169                 qbody->qb_flags = 0;
170
171         if (!lqe->lqe_enforced) {
172                 /* quota not enforced any more for this ID */
173                 if (granted != 0) {
174                         /* release all quota space unconditionally */
175                         LQUOTA_DEBUG(lqe, "not enforced, releasing all space");
176                         if (qbody != NULL) {
177                                 qbody->qb_count = granted;
178                                 qbody->qb_flags = QUOTA_DQACQ_FL_REL;
179                         }
180                         RETURN(true);
181                 }
182                 RETURN(false);
183         }
184
185         if (!lustre_handle_is_used(&lqe->lqe_lockh)) {
186                 /* No valid per-ID lock
187                  * When reporting quota (during reintegration or on setquota
188                  * glimpse), we should release granted space if usage is 0.
189                  * Otherwise, if the usage is less than granted, we need to
190                  * acquire the per-ID lock to make sure the unused grant can be
191                  * reclaimed by per-ID lock glimpse. */
192                 if (usage == 0) {
193                         /* no on-disk usage and no outstanding activity, release
194                          * space */
195                         if (granted != 0) {
196                                 LQUOTA_DEBUG(lqe, "no usage, releasing all "
197                                              "space");
198                                 if (qbody != NULL) {
199                                         qbody->qb_count = granted;
200                                         qbody->qb_flags = QUOTA_DQACQ_FL_REL;
201                                 }
202                                 RETURN(true);
203                         }
204                         LQUOTA_DEBUG(lqe, "no usage + no granted, nothing to "
205                                      "do");
206                         RETURN(false);
207                 }
208
209                 if (lqe->lqe_usage < lqe->lqe_granted) {
210                         /* holding quota space w/o any lock, enqueue per-ID lock
211                          * again */
212                         LQUOTA_DEBUG(lqe, "(re)acquiring per-ID lock");
213                         if (qbody != NULL) {
214                                 qbody->qb_count = 0;
215                                 qbody->qb_flags = QUOTA_DQACQ_FL_ACQ;
216                         }
217                         RETURN(true);
218                 }
219
220                 if (lqe->lqe_usage > lqe->lqe_granted) {
221                         /* quota overrun, report usage */
222                         LQUOTA_DEBUG(lqe, "overrun, reporting usage");
223                         if (qbody != NULL) {
224                                 qbody->qb_usage = lqe->lqe_usage;
225                                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
226                         }
227                         RETURN(true);
228                 }
229                 LQUOTA_DEBUG(lqe, "granted matches usage, nothing to do");
230                 RETURN(false);
231         }
232
233         /* valid per-ID lock
234          * Apply good old quota qunit adjustment logic which has been around
235          * since lustre 1.4:
236          * 1. release spare quota space? */
237         if (granted > usage + lqe->lqe_qunit) {
238                 /* pre-release quota space */
239                 if (qbody == NULL)
240                         RETURN(true);
241                 qbody->qb_count = granted - usage;
242                 /* if usage == 0, release all granted space */
243                 if (usage) {
244                         /* try to keep one qunit of quota space */
245                         qbody->qb_count -= lqe->lqe_qunit;
246                         /* but don't release less than qtune to avoid releasing
247                          * space too often */
248                         if (qbody->qb_count < lqe->lqe_qtune)
249                                 qbody->qb_count = lqe->lqe_qtune;
250                 }
251                 qbody->qb_flags = QUOTA_DQACQ_FL_REL;
252                 RETURN(true);
253         }
254
255         /* 2. Any quota overrun? */
256         if (lqe->lqe_usage > lqe->lqe_granted) {
257                 /* we overconsumed quota space, we report usage in request so
258                  * that master can adjust it unconditionally */
259                 if (qbody == NULL)
260                         RETURN(true);
261                 qbody->qb_usage = lqe->lqe_usage;
262                 granted         = lqe->lqe_usage;
263                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
264         }
265
266         /* 3. Time to pre-acquire? */
267         if (!lqe->lqe_edquot && !lqe->lqe_nopreacq && usage > 0 &&
268             lqe->lqe_qunit != 0 && granted < usage + lqe->lqe_qtune) {
269                 /* To pre-acquire quota space, we report how much spare quota
270                  * space the slave currently owns, then the master will grant us
271                  * back how much we can pretend given the current state of
272                  * affairs */
273                 if (qbody == NULL)
274                         RETURN(true);
275                 if (granted <= usage)
276                         qbody->qb_count = 0;
277                 else
278                         qbody->qb_count = granted - usage;
279                 qbody->qb_flags |= QUOTA_DQACQ_FL_PREACQ;
280                 RETURN(true);
281         }
282
283         if (qbody != NULL)
284                 RETURN(qbody->qb_flags != 0);
285         else
286                 RETURN(false);
287 }
288
289 /**
290  * Helper function returning true when quota space need to be adjusted (some
291  * unused space should be free or pre-acquire) and false otherwise.
292  */
293 static inline bool qsd_adjust_needed(struct lquota_entry *lqe)
294 {
295         return qsd_calc_adjust(lqe, NULL);
296 }
297
298 /**
299  * Callback function called when an acquire/release request sent to the master
300  * is completed
301  */
302 static void qsd_req_completion(const struct lu_env *env,
303                                struct qsd_qtype_info *qqi,
304                                struct quota_body *reqbody,
305                                struct quota_body *repbody,
306                                struct lustre_handle *lockh,
307                                struct lquota_lvb *lvb,
308                                void *arg, int ret)
309 {
310         struct lquota_entry     *lqe = (struct lquota_entry *)arg;
311         struct qsd_thread_info  *qti;
312         int                      rc;
313         bool                     adjust = false, cancel = false;
314         ENTRY;
315
316         LASSERT(qqi != NULL && lqe != NULL);
317
318         /* environment passed by ptlrpcd is mostly used by CLIO and hasn't the
319          * DT tags set. */
320         rc = lu_env_refill_by_tags((struct lu_env *)env, LCT_DT_THREAD, 0);
321         if (rc) {
322                 LQUOTA_ERROR(lqe, "failed to refill environmnent %d", rc);
323                 lqe_write_lock(lqe);
324                 /* can't afford to adjust quota space with no suitable lu_env */
325                 GOTO(out_noadjust, rc);
326         }
327         qti = qsd_info(env);
328
329         lqe_write_lock(lqe);
330         LQUOTA_DEBUG(lqe, "DQACQ returned %d, flags:0x%x", ret,
331                      reqbody->qb_flags);
332
333         /* despite -EDQUOT & -EINPROGRESS errors, the master might still
334          * grant us back quota space to adjust quota overrun */
335         if (ret != 0 && ret != -EDQUOT && ret != -EINPROGRESS) {
336                 if (ret != -ETIMEDOUT && ret != -ENOTCONN &&
337                    ret != -ESHUTDOWN && ret != -EAGAIN)
338                         /* print errors only if return code is unexpected */
339                         LQUOTA_ERROR(lqe, "DQACQ failed with %d, flags:0x%x",
340                                      ret, reqbody->qb_flags);
341                 GOTO(out, ret);
342         }
343
344         /* Set the lqe_lockh */
345         if (lustre_handle_is_used(lockh) &&
346             !lustre_handle_equal(lockh, &lqe->lqe_lockh))
347                 lustre_handle_copy(&lqe->lqe_lockh, lockh);
348
349         /* If the replied qb_count is zero, it means master didn't process
350          * the DQACQ since the limit for this ID has been removed, so we
351          * should not update quota entry & slave index copy neither. */
352         if (repbody != NULL && repbody->qb_count != 0) {
353                 LQUOTA_DEBUG(lqe, "DQACQ qb_count:%llu", repbody->qb_count);
354
355                 if (req_is_rel(reqbody->qb_flags)) {
356                         if (lqe->lqe_granted < repbody->qb_count) {
357                                 LQUOTA_ERROR(lqe, "can't release more space "
358                                              "than owned %llu<%llu",
359                                              lqe->lqe_granted,
360                                              repbody->qb_count);
361                                 lqe->lqe_granted = 0;
362                         } else {
363                                 lqe->lqe_granted -= repbody->qb_count;
364                         }
365                         /* Cancel the per-ID lock initiatively when there
366                          * isn't any usage & grant, which can avoid master
367                          * sending glimpse unnecessarily to this slave on
368                          * quota revoking */
369                         if (!lqe->lqe_pending_write && !lqe->lqe_granted &&
370                             !lqe->lqe_waiting_write && !lqe->lqe_usage)
371                                 cancel = true;
372                 } else {
373                         lqe->lqe_granted += repbody->qb_count;
374                 }
375                 qti->qti_rec.lqr_slv_rec.qsr_granted = lqe->lqe_granted;
376                 lqe_write_unlock(lqe);
377
378                 /* Update the slave index file in the dedicated thread. So far,
379                  * We don't update the version of slave index copy on DQACQ.
380                  * No locking is necessary since nobody can change
381                  * lqe->lqe_granted while lqe->lqe_pending_req > 0 */
382                 qsd_upd_schedule(qqi, lqe, &lqe->lqe_id, &qti->qti_rec, 0,
383                                  false);
384                 lqe_write_lock(lqe);
385         }
386
387         /* extract information from lvb */
388         if (ret == 0 && lvb != NULL) {
389                 if (lvb->lvb_id_qunit != 0)
390                         qsd_set_qunit(lqe, lvb->lvb_id_qunit);
391                 qsd_set_edquot(lqe, !!(lvb->lvb_flags & LQUOTA_FL_EDQUOT));
392         } else if (repbody != NULL && repbody->qb_qunit != 0) {
393                 qsd_set_qunit(lqe, repbody->qb_qunit);
394         }
395
396         /* turn off pre-acquire if it failed with -EDQUOT. This is done to avoid
397          * flooding the master with acquire request. Pre-acquire will be turned
398          * on again as soon as qunit is modified */
399         if (req_is_preacq(reqbody->qb_flags) && ret == -EDQUOT)
400                 lqe->lqe_nopreacq = true;
401 out:
402         adjust = qsd_adjust_needed(lqe);
403         if (reqbody && req_is_acq(reqbody->qb_flags) && ret != -EDQUOT) {
404                 lqe->lqe_acq_rc = ret;
405                 lqe->lqe_acq_time = ktime_get_seconds();
406         }
407 out_noadjust:
408         qsd_request_exit(lqe);
409         lqe_write_unlock(lqe);
410
411         /* release reference on per-ID lock */
412         if (lustre_handle_is_used(lockh))
413                 ldlm_lock_decref(lockh, qsd_id_einfo.ei_mode);
414
415         if (cancel) {
416                 qsd_adjust_schedule(lqe, false, true);
417         } else if (adjust) {
418                 if (!ret || ret == -EDQUOT)
419                         qsd_adjust_schedule(lqe, false, false);
420                 else
421                         qsd_adjust_schedule(lqe, true, false);
422         }
423         lqe_putref(lqe);
424
425         if (lvb)
426                 OBD_FREE_PTR(lvb);
427         EXIT;
428 }
429
430 /**
431  * Try to consume local quota space.
432  *
433  * \param lqe   - is the qid entry to be processed
434  * \param space - is the amount of quota space needed to complete the operation
435  *
436  * \retval 0       - success
437  * \retval -EDQUOT - out of quota
438  * \retval -EAGAIN - need to acquire space from master
439  */
440 static int qsd_acquire_local(struct lquota_entry *lqe, __u64 space)
441 {
442         __u64   usage;
443         int     rc;
444         ENTRY;
445
446         if (!lqe->lqe_enforced)
447                 /* not enforced any more, we are good */
448                 RETURN(-ESRCH);
449
450         lqe_write_lock(lqe);
451         /* use latest usage */
452         usage = lqe->lqe_usage;
453         /* take pending write into account */
454         usage += lqe->lqe_pending_write;
455
456         if (space + usage <= lqe->lqe_granted - lqe->lqe_pending_rel) {
457                 /* Yay! we got enough space */
458                 lqe->lqe_pending_write += space;
459                 lqe->lqe_waiting_write -= space;
460                 rc = 0;
461         /* lqe_edquot flag is used to avoid flooding dqacq requests when
462          * the user is over quota, however, the lqe_edquot could be stale
463          * sometimes due to the race reply of dqacq vs. id lock glimpse
464          * (see LU-4505), so we revalidate it every 5 seconds. */
465         } else if (lqe->lqe_edquot &&
466                    (lqe->lqe_edquot_time > ktime_get_seconds() - 5)) {
467                 rc = -EDQUOT;
468         }else {
469                 rc = -EAGAIN;
470         }
471         lqe_write_unlock(lqe);
472
473         RETURN(rc);
474 }
475
476 /**
477  * Compute how much quota space should be acquire from the master based
478  * on how much is currently granted to this slave and pending/waiting
479  * operations.
480  *
481  * \param lqe - is the lquota entry for which we would like to adjust quota
482  *              space.
483  * \param qbody - is the quota body of the acquire request to fill
484  *
485  * \retval true  - space acquisition is needed and qbody is filled
486  * \retval false - no space acquisition required
487  */
488 static inline bool qsd_calc_acquire(struct lquota_entry *lqe,
489                                     struct quota_body *qbody)
490 {
491         __u64   usage, granted;
492
493         usage   = lqe->lqe_usage;
494         usage  += lqe->lqe_pending_write + lqe->lqe_waiting_write;
495         granted = lqe->lqe_granted;
496
497         qbody->qb_flags = 0;
498
499         /* if we overconsumed quota space, we report usage in request so that
500          * master can adjust it unconditionally */
501         if (lqe->lqe_usage > lqe->lqe_granted) {
502                 qbody->qb_usage = lqe->lqe_usage;
503                 qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
504                 granted = lqe->lqe_usage;
505         }
506
507         /* acquire as much as needed, but not more */
508         if (usage > granted) {
509                 qbody->qb_count  = usage - granted;
510                 qbody->qb_flags |= QUOTA_DQACQ_FL_ACQ;
511         }
512
513         return qbody->qb_flags != 0;
514 }
515
516 /**
517  * Acquire quota space from master.
518  * There are at most 1 in-flight dqacq/dqrel.
519  *
520  * \param env    - the environment passed by the caller
521  * \param lqe    - is the qid entry to be processed
522  *
523  * \retval 0            - success
524  * \retval -EDQUOT      - out of quota
525  * \retval -EINPROGRESS - inform client to retry write/create
526  * \retval -EBUSY       - already a quota request in flight
527  * \retval -ve          - other appropriate errors
528  */
529 static int qsd_acquire_remote(const struct lu_env *env,
530                               struct lquota_entry *lqe)
531 {
532         struct qsd_thread_info  *qti = qsd_info(env);
533         struct quota_body       *qbody = &qti->qti_body;
534         struct qsd_instance     *qsd;
535         struct qsd_qtype_info   *qqi;
536         int                      rc;
537         ENTRY;
538
539         memset(qbody, 0, sizeof(*qbody));
540         rc = qsd_ready(lqe, &qbody->qb_glb_lockh);
541         if (rc)
542                 RETURN(rc);
543
544         qqi = lqe2qqi(lqe);
545         qsd = qqi->qqi_qsd;
546
547         lqe_write_lock(lqe);
548
549         /* is quota really enforced for this id? */
550         if (!lqe->lqe_enforced) {
551                 lqe_write_unlock(lqe);
552                 LQUOTA_DEBUG(lqe, "quota not enforced any more");
553                 RETURN(0);
554         }
555
556         /* fill qb_count & qb_flags */
557         if (!qsd_calc_acquire(lqe, qbody)) {
558                 lqe_write_unlock(lqe);
559                 LQUOTA_DEBUG(lqe, "No acquire required");
560                 RETURN(0);
561         }
562
563         /* check whether an acquire request completed recently */
564         if (lqe->lqe_acq_rc != 0 &&
565             lqe->lqe_acq_time > ktime_get_seconds() - 1) {
566                 lqe_write_unlock(lqe);
567                 LQUOTA_DEBUG(lqe, "using cached return code %d", lqe->lqe_acq_rc);
568                 RETURN(lqe->lqe_acq_rc);
569         }
570
571         /* only 1 quota request in flight for a given ID is allowed */
572         rc = qsd_request_enter(lqe);
573         if (rc) {
574                 lqe_write_unlock(lqe);
575                 RETURN(rc);
576         }
577
578         lustre_handle_copy(&qti->qti_lockh, &lqe->lqe_lockh);
579         lqe_write_unlock(lqe);
580
581         /* hold a refcount until completion */
582         lqe_getref(lqe);
583
584         /* fill other quota body fields */
585         qbody->qb_fid = qqi->qqi_fid;
586         qbody->qb_id  = lqe->lqe_id;
587
588         /* check whether we already own a valid lock for this ID */
589         rc = qsd_id_lock_match(&qti->qti_lockh, &qbody->qb_lockh);
590         if (rc) {
591                 struct lquota_lvb *lvb;
592
593                 OBD_ALLOC_PTR(lvb);
594                 if (lvb == NULL) {
595                         rc = -ENOMEM;
596                         qsd_req_completion(env, qqi, qbody, NULL,
597                                            &qti->qti_lockh, NULL, lqe, rc);
598                         RETURN(rc);
599                 }
600                 /* no lock found, should use intent */
601                 rc = qsd_intent_lock(env, qsd->qsd_exp, qbody, true,
602                                      IT_QUOTA_DQACQ, qsd_req_completion,
603                                      qqi, lvb, (void *)lqe);
604         } else {
605                 /* lock found, should use regular dqacq */
606                 rc = qsd_send_dqacq(env, qsd->qsd_exp, qbody, true,
607                                     qsd_req_completion, qqi, &qti->qti_lockh,
608                                     lqe);
609         }
610
611         /* the completion function will be called by qsd_send_dqacq or
612          * qsd_intent_lock */
613         RETURN(rc);
614 }
615
616 /**
617  * Acquire \a space of quota space in order to complete an operation.
618  * Try to consume local quota space first and send acquire request to quota
619  * master if required.
620  *
621  * \param env   - the environment passed by the caller
622  * \param lqe   - is the qid entry to be processed
623  * \param space - is the amount of quota required for the operation
624  * \param ret   - is the return code (-EDQUOT, -EINPROGRESS, ...)
625  *
626  * \retval true  - exit from l_wait_event and real return value in \a ret
627  * \retval false - continue waiting
628  */
629 static bool qsd_acquire(const struct lu_env *env, struct lquota_entry *lqe,
630                         long long space, int *ret)
631 {
632         int rc = 0, count;
633         ENTRY;
634
635         for (count = 0; rc == 0; count++) {
636                 LQUOTA_DEBUG(lqe, "acquiring:%lld count=%d", space, count);
637
638                 if (lqe2qqi(lqe)->qqi_qsd->qsd_stopping) {
639                         rc = -EINPROGRESS;
640                         break;
641                 }
642
643                 /* refresh disk usage */
644                 rc = qsd_refresh_usage(env, lqe);
645                 if (rc)
646                         break;
647
648                 /* try to consume local quota space first */
649                 rc = qsd_acquire_local(lqe, space);
650                 if (rc != -EAGAIN)
651                         /* rc == 0, Wouhou! enough local quota space
652                          * rc < 0, something bad happened */
653                          break;
654
655                 /* need to acquire more quota space from master */
656                 rc = qsd_acquire_remote(env, lqe);
657         }
658
659         if (rc == -EBUSY)
660                 /* already a request in flight, continue waiting */
661                 RETURN(false);
662         *ret = rc;
663         RETURN(true); /* exit from l_wait_event */
664 }
665
666 /**
667  * Quota enforcement handler. If local quota can satisfy this operation,
668  * return success, otherwise, acquire more quota from master.
669  * (for write operation, if master isn't available at this moment, return
670  * -EINPROGRESS to inform client to retry the write)
671  *
672  * \param env   - the environment passed by the caller
673  * \param qsd   - is the qsd instance associated with the device in charge
674  *                of the operation.
675  * \param qid   - is the qid information attached in the transaction handle
676  * \param space - is the space required by the operation
677  * \param flags - if the operation is write, return caller no user/group
678  *                and sync commit flags
679  *
680  * \retval 0            - success
681  * \retval -EDQUOT      - out of quota
682  * \retval -EINPROGRESS - inform client to retry write
683  * \retval -ve          - other appropriate errors
684  */
685 static int qsd_op_begin0(const struct lu_env *env, struct qsd_qtype_info *qqi,
686                          struct lquota_id_info *qid, long long space,
687                          int *flags)
688 {
689         struct lquota_entry *lqe;
690         struct l_wait_info lwi;
691         int qtype_flag = 0;
692         int rc, ret = -EINPROGRESS;
693         ENTRY;
694
695         if (qid->lqi_qentry != NULL) {
696                 /* we already had to deal with this id for this transaction */
697                 lqe = qid->lqi_qentry;
698                 if (!lqe->lqe_enforced)
699                         RETURN(0);
700         } else {
701                 /* look up lquota entry associated with qid */
702                 lqe = lqe_locate(env, qqi->qqi_site, &qid->lqi_id);
703                 if (IS_ERR(lqe))
704                         RETURN(PTR_ERR(lqe));
705                 if (!lqe->lqe_enforced) {
706                         lqe_putref(lqe);
707                         RETURN(0);
708                 }
709                 qid->lqi_qentry = lqe;
710                 /* lqe will be released in qsd_op_end() */
711         }
712
713         if (space <= 0) {
714                 /* when space is negative or null, we don't need to consume
715                  * quota space. That said, we still want to perform space
716                  * adjustments in qsd_op_end, so we return here, but with
717                  * a reference on the lqe */
718                 if (flags != NULL) {
719                         rc = qsd_refresh_usage(env, lqe);
720                         GOTO(out_flags, rc);
721                 }
722                 RETURN(0);
723         }
724
725         LQUOTA_DEBUG(lqe, "op_begin space:%lld", space);
726
727         lqe_write_lock(lqe);
728         lqe->lqe_waiting_write += space;
729         lqe_write_unlock(lqe);
730
731         /* acquire quota space for the operation, cap overall wait time to
732          * prevent a service thread from being stuck for too long */
733         lwi = LWI_TIMEOUT(cfs_time_seconds(qsd_wait_timeout(qqi->qqi_qsd)),
734                           NULL, NULL);
735         rc = l_wait_event(lqe->lqe_waiters, qsd_acquire(env, lqe, space, &ret),
736                           &lwi);
737
738         if (rc == 0 && ret == 0) {
739                 qid->lqi_space += space;
740         } else {
741                 if (rc == 0)
742                         rc = ret;
743
744                 LQUOTA_DEBUG(lqe, "acquire quota failed:%d", rc);
745
746                 lqe_write_lock(lqe);
747                 lqe->lqe_waiting_write -= space;
748
749                 if (flags && lqe->lqe_pending_write != 0)
750                         /* Inform OSD layer that there are pending writes.
751                          * It might want to retry after a sync if appropriate */
752                          *flags |= QUOTA_FL_SYNC;
753                 lqe_write_unlock(lqe);
754
755                 /* convert recoverable error into -EINPROGRESS, client will
756                  * retry */
757                 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ENOLCK ||
758                     rc == -EAGAIN || rc == -EINTR) {
759                         rc = -EINPROGRESS;
760                 } else if (rc == -ESRCH) {
761                         rc = 0;
762                         LQUOTA_ERROR(lqe, "ID isn't enforced on master, it "
763                                      "probably due to a legeal race, if this "
764                                      "message is showing up constantly, there "
765                                      "could be some inconsistence between "
766                                      "master & slave, and quota reintegration "
767                                      "needs be re-triggered.");
768                 }
769         }
770
771         if (flags != NULL) {
772 out_flags:
773                 LASSERT(qid->lqi_is_blk);
774                 if (rc != 0) {
775                         *flags |= lquota_over_fl(qqi->qqi_qtype);
776                 } else {
777                         __u64   usage;
778
779                         lqe_read_lock(lqe);
780                         usage  = lqe->lqe_usage;
781                         usage += lqe->lqe_pending_write;
782                         usage += lqe->lqe_waiting_write;
783                         usage += qqi->qqi_qsd->qsd_sync_threshold;
784
785                         qtype_flag = lquota_over_fl(qqi->qqi_qtype);
786                         /* if we should notify client to start sync write */
787                         if (usage >= lqe->lqe_granted - lqe->lqe_pending_rel)
788                                 *flags |= qtype_flag;
789                         else
790                                 *flags &= ~qtype_flag;
791                         lqe_read_unlock(lqe);
792                 }
793         }
794         RETURN(rc);
795 }
796
797 /**
798  * helper function comparing two lquota_id_info structures
799  */
800 static inline bool qid_equal(struct lquota_id_info *q1,
801                              struct lquota_id_info *q2)
802 {
803         if (q1->lqi_type != q2->lqi_type)
804                 return false;
805         return (q1->lqi_id.qid_uid == q2->lqi_id.qid_uid) ? true : false;
806 }
807
808 /**
809  * Enforce quota, it's called in the declaration of each operation.
810  * qsd_op_end() will then be called later once all the operations have been
811  * completed in order to release/adjust the quota space.
812  *
813  * \param env   - the environment passed by the caller
814  * \param qsd   - is the qsd instance associated with the device in charge of
815  *                the operation.
816  * \param trans - is the quota transaction information
817  * \param qi    - qid & space required by current operation
818  * \param flags - if the operation is write, return caller no user/group and
819  *                sync commit flags
820  *
821  * \retval 0            - success
822  * \retval -EDQUOT      - out of quota
823  * \retval -EINPROGRESS - inform client to retry write
824  * \retval -ve          - other appropriate errors
825  */
826 int qsd_op_begin(const struct lu_env *env, struct qsd_instance *qsd,
827                  struct lquota_trans *trans, struct lquota_id_info *qi,
828                  int *flags)
829 {
830         int     i, rc;
831         bool    found = false;
832         ENTRY;
833
834         if (unlikely(qsd == NULL))
835                 RETURN(0);
836
837         if (qsd->qsd_dev->dd_rdonly)
838                 RETURN(0);
839
840         /* We don't enforce quota until the qsd_instance is started */
841         read_lock(&qsd->qsd_lock);
842         if (!qsd->qsd_started) {
843                 read_unlock(&qsd->qsd_lock);
844                 RETURN(0);
845         }
846         read_unlock(&qsd->qsd_lock);
847
848         /* ignore block quota on MDTs, ignore inode quota on OSTs */
849         if ((!qsd->qsd_is_md && !qi->lqi_is_blk) ||
850             (qsd->qsd_is_md && qi->lqi_is_blk))
851                 RETURN(0);
852
853         /* ignore quota enforcement request when:
854          *    - quota isn't enforced for this quota type
855          * or - the user/group is root
856          * or - quota accounting isn't enabled */
857         if (!qsd_type_enabled(qsd, qi->lqi_type) || qi->lqi_id.qid_uid == 0 ||
858             (qsd->qsd_type_array[qi->lqi_type])->qqi_acct_failed)
859                 RETURN(0);
860
861         LASSERTF(trans->lqt_id_cnt <= QUOTA_MAX_TRANSIDS, "id_cnt=%d\n",
862                  trans->lqt_id_cnt);
863         /* check whether we already allocated a slot for this id */
864         for (i = 0; i < trans->lqt_id_cnt; i++) {
865                 if (qid_equal(qi, &trans->lqt_ids[i])) {
866                         found = true;
867                         /* make sure we are not mixing inodes & blocks */
868                         LASSERT(trans->lqt_ids[i].lqi_is_blk == qi->lqi_is_blk);
869                         break;
870                 }
871         }
872
873         if (!found) {
874                 if (unlikely(i >= QUOTA_MAX_TRANSIDS)) {
875                         CERROR("%s: more than %d qids enforced for a "
876                                "transaction?\n", qsd->qsd_svname, i);
877                         RETURN(-EINVAL);
878                 }
879
880                 /* fill new slot */
881                 trans->lqt_ids[i].lqi_id     = qi->lqi_id;
882                 trans->lqt_ids[i].lqi_type   = qi->lqi_type;
883                 trans->lqt_ids[i].lqi_is_blk = qi->lqi_is_blk;
884                 trans->lqt_id_cnt++;
885         }
886
887         /* manage quota enforcement for this ID */
888         rc = qsd_op_begin0(env, qsd->qsd_type_array[qi->lqi_type],
889                            &trans->lqt_ids[i], qi->lqi_space, flags);
890         RETURN(rc);
891 }
892 EXPORT_SYMBOL(qsd_op_begin);
893
894 /**
895  * Adjust quota space (by acquiring or releasing) hold by the quota slave.
896  * This function is called after each quota request completion and during
897  * reintegration in order to report usage or re-acquire quota locks.
898  * Space adjustment is aborted if there is already a quota request in flight
899  * for this ID.
900  *
901  * \param env    - the environment passed by the caller
902  * \param lqe    - is the qid entry to be processed
903  *
904  * \retval 0 on success, appropriate errors on failure
905  */
906 int qsd_adjust(const struct lu_env *env, struct lquota_entry *lqe)
907 {
908         struct qsd_thread_info  *qti = qsd_info(env);
909         struct quota_body       *qbody = &qti->qti_body;
910         struct qsd_instance     *qsd;
911         struct qsd_qtype_info   *qqi;
912         int                      rc;
913         bool                     intent = false;
914         ENTRY;
915
916         memset(qbody, 0, sizeof(*qbody));
917         rc = qsd_ready(lqe, &qbody->qb_glb_lockh);
918         if (rc) {
919                 /* add to adjust list again to trigger adjustment later when
920                  * slave is ready */
921                 LQUOTA_DEBUG(lqe, "delaying adjustment since qsd isn't ready");
922                 qsd_adjust_schedule(lqe, true, false);
923                 RETURN(0);
924         }
925
926         qqi = lqe2qqi(lqe);
927         qsd = qqi->qqi_qsd;
928
929         if (qsd->qsd_dev->dd_rdonly)
930                 RETURN(0);
931
932         lqe_write_lock(lqe);
933
934         /* fill qb_count & qb_flags */
935         if (!qsd_calc_adjust(lqe, qbody)) {
936                 lqe_write_unlock(lqe);
937                 LQUOTA_DEBUG(lqe, "no adjustment required");
938                 RETURN(0);
939         }
940
941         /* only 1 quota request in flight for a given ID is allowed */
942         rc = qsd_request_enter(lqe);
943         if (rc) {
944                 /* already a request in flight, space adjustment will be run
945                  * again on request completion */
946                 lqe_write_unlock(lqe);
947                 RETURN(0);
948         }
949
950         if (req_is_rel(qbody->qb_flags))
951                 lqe->lqe_pending_rel = qbody->qb_count;
952         lustre_handle_copy(&qti->qti_lockh, &lqe->lqe_lockh);
953         lqe_write_unlock(lqe);
954
955         /* hold a refcount until completion */
956         lqe_getref(lqe);
957
958         /* fill other quota body fields */
959         qbody->qb_fid = qqi->qqi_fid;
960         qbody->qb_id  = lqe->lqe_id;
961
962         if (req_is_acq(qbody->qb_flags) || req_is_preacq(qbody->qb_flags)) {
963                 /* check whether we own a valid lock for this ID */
964                 rc = qsd_id_lock_match(&qti->qti_lockh, &qbody->qb_lockh);
965                 if (rc) {
966                         memset(&qti->qti_lockh, 0, sizeof(qti->qti_lockh));
967                         if (req_is_preacq(qbody->qb_flags)) {
968                                 if (req_has_rep(qbody->qb_flags))
969                                         /* still want to report usage */
970                                         qbody->qb_flags = QUOTA_DQACQ_FL_REPORT;
971                                 else
972                                         /* no pre-acquire if no per-ID lock */
973                                         GOTO(out, rc = -ENOLCK);
974                         } else {
975                                 /* no lock found, should use intent */
976                                 intent = true;
977                         }
978                 } else if (req_is_acq(qbody->qb_flags) &&
979                            qbody->qb_count == 0) {
980                         /* found cached lock, no need to acquire */
981                         GOTO(out, rc = 0);
982                 }
983         } else {
984                 /* release and report don't need a per-ID lock */
985                 memset(&qti->qti_lockh, 0, sizeof(qti->qti_lockh));
986         }
987
988         if (!intent) {
989                 rc = qsd_send_dqacq(env, qsd->qsd_exp, qbody, false,
990                                     qsd_req_completion, qqi, &qti->qti_lockh,
991                                     lqe);
992         } else {
993                 struct lquota_lvb *lvb;
994
995                 OBD_ALLOC_PTR(lvb);
996                 if (lvb == NULL)
997                         GOTO(out, rc = -ENOMEM);
998
999                 rc = qsd_intent_lock(env, qsd->qsd_exp, qbody, false,
1000                                      IT_QUOTA_DQACQ, qsd_req_completion,
1001                                      qqi, lvb, (void *)lqe);
1002         }
1003         /* the completion function will be called by qsd_send_dqacq or
1004          * qsd_intent_lock */
1005         RETURN(rc);
1006 out:
1007         qsd_req_completion(env, qqi, qbody, NULL, &qti->qti_lockh, NULL, lqe,
1008                            rc);
1009         return rc;
1010 }
1011
1012 /**
1013  * Post quota operation, pre-acquire/release quota from master.
1014  *
1015  * \param  env  - the environment passed by the caller
1016  * \param  qsd  - is the qsd instance attached to the OSD device which
1017  *                is handling the operation.
1018  * \param  qqi  - is the qsd_qtype_info structure associated with the quota ID
1019  *                subject to the operation
1020  * \param  qid  - stores information related to his ID for the operation
1021  *                which has just completed
1022  *
1023  * \retval 0    - success
1024  * \retval -ve  - failure
1025  */
1026 static void qsd_op_end0(const struct lu_env *env, struct qsd_qtype_info *qqi,
1027                         struct lquota_id_info *qid)
1028 {
1029         struct lquota_entry     *lqe;
1030         bool                     adjust;
1031         ENTRY;
1032
1033         lqe = qid->lqi_qentry;
1034         if (lqe == NULL)
1035                 RETURN_EXIT;
1036         qid->lqi_qentry = NULL;
1037
1038         /* refresh cached usage if a suitable environment is passed */
1039         if (env != NULL)
1040                 qsd_refresh_usage(env, lqe);
1041
1042         lqe_write_lock(lqe);
1043         if (qid->lqi_space > 0)
1044                 lqe->lqe_pending_write -= qid->lqi_space;
1045         if (env != NULL)
1046                 adjust = qsd_adjust_needed(lqe);
1047         else
1048                 adjust = true;
1049         lqe_write_unlock(lqe);
1050
1051         if (adjust) {
1052                 /* pre-acquire/release quota space is needed */
1053                 if (env != NULL)
1054                         qsd_adjust(env, lqe);
1055                 else
1056                         /* no suitable environment, handle adjustment in
1057                          * separate thread context */
1058                         qsd_adjust_schedule(lqe, false, false);
1059         }
1060         lqe_putref(lqe);
1061         EXIT;
1062 }
1063
1064 /**
1065  * Post quota operation. It's called after each operation transaction stopped.
1066  *
1067  * \param  env   - the environment passed by the caller
1068  * \param  qsd   - is the qsd instance associated with device which is handling
1069  *                 the operation.
1070  * \param  qids  - all qids information attached in the transaction handle
1071  * \param  count - is the number of qid entries in the qids array.
1072  *
1073  * \retval 0     - success
1074  * \retval -ve   - failure
1075  */
1076 void qsd_op_end(const struct lu_env *env, struct qsd_instance *qsd,
1077                 struct lquota_trans *trans)
1078 {
1079         int i;
1080         ENTRY;
1081
1082         if (unlikely(qsd == NULL))
1083                 RETURN_EXIT;
1084
1085         if (qsd->qsd_dev->dd_rdonly)
1086                 RETURN_EXIT;
1087
1088         /* We don't enforce quota until the qsd_instance is started */
1089         read_lock(&qsd->qsd_lock);
1090         if (!qsd->qsd_started) {
1091                 read_unlock(&qsd->qsd_lock);
1092                 RETURN_EXIT;
1093         }
1094         read_unlock(&qsd->qsd_lock);
1095
1096         LASSERT(trans != NULL);
1097
1098         for (i = 0; i < trans->lqt_id_cnt; i++) {
1099                 struct qsd_qtype_info *qqi;
1100
1101                 if (trans->lqt_ids[i].lqi_qentry == NULL)
1102                         continue;
1103
1104                 qqi = qsd->qsd_type_array[trans->lqt_ids[i].lqi_type];
1105                 qsd_op_end0(env, qqi, &trans->lqt_ids[i]);
1106         }
1107
1108         /* reset id_count to 0 so that a second accidental call to qsd_op_end()
1109          * does not result in failure */
1110         trans->lqt_id_cnt = 0;
1111         EXIT;
1112 }
1113 EXPORT_SYMBOL(qsd_op_end);
1114
1115 /**
1116  * Trigger pre-acquire/release if necessary.
1117  * It's only used by ldiskfs osd so far. When unlink a file in ldiskfs, the
1118  * quota accounting isn't updated when the transaction stopped. Instead, it'll
1119  * be updated on the final iput, so qsd_op_adjust() will be called then (in
1120  * osd_object_delete()) to trigger quota release if necessary.
1121  *
1122  * \param env - the environment passed by the caller
1123  * \param qsd - is the qsd instance associated with the device in charge
1124  *              of the operation.
1125  * \param qid - is the lquota ID of the user/group for which to trigger
1126  *              quota space adjustment
1127  * \param qtype - is the quota type (USRQUOTA or GRPQUOTA)
1128  */
1129 void qsd_op_adjust(const struct lu_env *env, struct qsd_instance *qsd,
1130                    union lquota_id *qid, int qtype)
1131 {
1132         struct lquota_entry    *lqe;
1133         struct qsd_qtype_info  *qqi;
1134         bool                    adjust;
1135         ENTRY;
1136
1137         if (unlikely(qsd == NULL))
1138                 RETURN_EXIT;
1139
1140         /* We don't enforce quota until the qsd_instance is started */
1141         read_lock(&qsd->qsd_lock);
1142         if (!qsd->qsd_started) {
1143                 read_unlock(&qsd->qsd_lock);
1144                 RETURN_EXIT;
1145         }
1146         read_unlock(&qsd->qsd_lock);
1147
1148         qqi = qsd->qsd_type_array[qtype];
1149         LASSERT(qqi);
1150
1151         if (!qsd_type_enabled(qsd, qtype) || qqi->qqi_acct_obj == NULL ||
1152             qid->qid_uid == 0)
1153                 RETURN_EXIT;
1154
1155         read_lock(&qsd->qsd_lock);
1156         if (!qsd->qsd_started) {
1157                 read_unlock(&qsd->qsd_lock);
1158                 RETURN_EXIT;
1159         }
1160         read_unlock(&qsd->qsd_lock);
1161
1162         lqe = lqe_locate(env, qqi->qqi_site, qid);
1163         if (IS_ERR(lqe)) {
1164                 CERROR("%s: fail to locate lqe for id:%llu, type:%d\n",
1165                        qsd->qsd_svname, qid->qid_uid, qtype);
1166                 RETURN_EXIT;
1167         }
1168
1169         qsd_refresh_usage(env, lqe);
1170
1171         lqe_read_lock(lqe);
1172         adjust = qsd_adjust_needed(lqe);
1173         lqe_read_unlock(lqe);
1174
1175         if (adjust)
1176                 qsd_adjust(env, lqe);
1177
1178         lqe_putref(lqe);
1179         EXIT;
1180 }
1181 EXPORT_SYMBOL(qsd_op_adjust);