Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / quota / quota_context.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  lustre/quota/quota_context.c
5  *  Lustre Quota Context
6  *
7  *  Copyright (c) 2001-2005 Cluster File Systems, Inc.
8  *   Author: Niu YaWei <niu@clusterfs.com>
9  *
10  *   This file is part of Lustre, http://www.lustre.org.
11  *
12  *   No redistribution or use is permitted outside of Cluster File Systems, Inc.
13  *
14  */
15 #ifndef EXPORT_SYMTAB
16 # define EXPORT_SYMTAB
17 #endif
18
19 #define DEBUG_SUBSYSTEM S_MDS
20
21 #include <linux/version.h>
22 #include <linux/fs.h>
23 #include <asm/unistd.h>
24 #include <linux/slab.h>
25 #include <linux/quotaops.h>
26 #include <linux/module.h>
27 #include <linux/init.h>
28
29 #include <obd_class.h>
30 #include <lustre_quota.h>
31 #include <lustre_fsfilt.h>
32 #include "quota_internal.h"
33
34 unsigned long default_bunit_sz = 100 * 1024 * 1024;       /* 100M bytes */
35 unsigned long default_btune_ratio = 50;                   /* 50 percentage */
36 unsigned long default_iunit_sz = 5000;       /* 5000 inodes */
37 unsigned long default_itune_ratio = 50;      /* 50 percentage */
38
39 cfs_mem_cache_t *qunit_cachep = NULL;
40 struct list_head qunit_hash[NR_DQHASH];
41 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
42
43 struct lustre_qunit {
44         struct list_head lq_hash;               /* Hash list in memory */
45         atomic_t lq_refcnt;                     /* Use count */
46         struct lustre_quota_ctxt *lq_ctxt;      /* Quota context this applies to */
47         struct qunit_data lq_data;              /* See qunit_data */
48         unsigned int lq_opc;                    /* QUOTA_DQACQ, QUOTA_DQREL */
49         struct list_head lq_waiters;            /* All write threads waiting for this qunit */
50 };
51
52 int should_translate_quota (struct obd_import *imp)
53 {
54         ENTRY;
55
56         LASSERT(imp);
57         if ((imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64) && 
58             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT))
59                 RETURN(0);
60         else
61                 RETURN(1);
62 }
63
64 void qunit_cache_cleanup(void)
65 {
66         int i;
67         ENTRY;
68
69         spin_lock(&qunit_hash_lock);
70         for (i = 0; i < NR_DQHASH; i++)
71                 LASSERT(list_empty(qunit_hash + i));
72         spin_unlock(&qunit_hash_lock);
73
74         if (qunit_cachep) {
75                 int rc;
76                 rc = cfs_mem_cache_destroy(qunit_cachep);
77                 LASSERTF(rc == 0, "couldn't destory qunit_cache slab\n");
78                 qunit_cachep = NULL;
79         }
80         EXIT;
81 }
82
83 int qunit_cache_init(void)
84 {
85         int i;
86         ENTRY;
87
88         LASSERT(qunit_cachep == NULL);
89         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
90                                          sizeof(struct lustre_qunit),
91                                          0, 0);
92         if (!qunit_cachep)
93                 RETURN(-ENOMEM);
94
95         spin_lock(&qunit_hash_lock);
96         for (i = 0; i < NR_DQHASH; i++)
97                 INIT_LIST_HEAD(qunit_hash + i);
98         spin_unlock(&qunit_hash_lock);
99         RETURN(0);
100 }
101
102 static inline int
103 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
104              __attribute__((__const__));
105
106 static inline int
107 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
108 {
109         unsigned int id = qdata->qd_id;
110         unsigned int type = qdata->qd_flags & QUOTA_IS_GRP;
111
112         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
113         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
114         return tmp;
115 }
116
117 /* compute the remaining quota for certain gid or uid b=11693 */
118 int compute_remquota(struct obd_device *obd,
119                      struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
120 {
121         struct super_block *sb = qctxt->lqc_sb;
122         __u64 usage, limit;
123         struct obd_quotactl *qctl;
124         int ret = QUOTA_RET_OK;
125         __u32 qdata_type = qdata->qd_flags & QUOTA_IS_GRP;
126         ENTRY;
127
128         if (!sb_any_quota_enabled(sb))
129                 RETURN(QUOTA_RET_NOQUOTA);
130
131         /* ignore root user */
132         if (qdata->qd_id == 0 && qdata_type == USRQUOTA)
133                 RETURN(QUOTA_RET_NOLIMIT);
134
135         OBD_ALLOC_PTR(qctl);
136         if (qctl == NULL) 
137                 RETURN(-ENOMEM);
138
139         /* get fs quota usage & limit */
140         qctl->qc_cmd = Q_GETQUOTA;
141         qctl->qc_id = qdata->qd_id;
142         qctl->qc_type = qdata_type;
143         ret = fsfilt_quotactl(obd, sb, qctl);
144         if (ret) {
145                 if (ret == -ESRCH)      /* no limit */
146                         ret = QUOTA_RET_NOLIMIT;
147                 else
148                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)", 
149                                ret);
150                 GOTO(out, ret);
151         }
152
153         usage = qctl->qc_dqblk.dqb_curspace;
154         limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
155         if (!limit){            /* no limit */
156                 ret = QUOTA_RET_NOLIMIT;
157                 GOTO(out, ret);
158         }
159
160         if (limit >= usage)
161                 qdata->qd_count = limit - usage;
162         else
163                 qdata->qd_count = 0;
164         EXIT;
165 out:
166         OBD_FREE_PTR(qctl);
167         return ret;
168 }
169
170 /* caller must hold qunit_hash_lock */
171 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
172                                               struct lustre_quota_ctxt *qctxt,
173                                               struct qunit_data *qdata)
174 {
175         struct lustre_qunit *qunit = NULL;
176         struct qunit_data *tmp;
177
178         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
179         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
180                 tmp = &qunit->lq_data;
181                 if (qunit->lq_ctxt == qctxt &&
182                     qdata->qd_id == tmp->qd_id && qdata->qd_flags == tmp->qd_flags)
183                         return qunit;
184         }
185         return NULL;
186 }
187
188 /* check_cur_qunit - check the current usage of qunit.
189  * @qctxt: quota context
190  * @qdata: the type of quota unit to be checked
191  *
192  * return: 1 - need acquire qunit;
193  *         2 - need release qunit;
194  *         0 - need do nothing.
195  *       < 0 - error.
196  */
197 static int
198 check_cur_qunit(struct obd_device *obd,
199                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
200 {
201         struct super_block *sb = qctxt->lqc_sb;
202         unsigned long qunit_sz, tune_sz;
203         __u64 usage, limit;
204         struct obd_quotactl *qctl;
205         int ret = 0;
206         __u32 qdata_type = qdata->qd_flags & QUOTA_IS_GRP;
207         __u32 is_blk = (qdata->qd_flags & QUOTA_IS_BLOCK) >> 1;
208         ENTRY;
209
210         if (!sb_any_quota_enabled(sb))
211                 RETURN(0);
212
213         OBD_ALLOC_PTR(qctl);
214         if (qctl == NULL)
215                 RETURN(-ENOMEM);
216
217         /* get fs quota usage & limit */
218         qctl->qc_cmd = Q_GETQUOTA;
219         qctl->qc_id = qdata->qd_id;
220         qctl->qc_type = qdata_type;
221         ret = fsfilt_quotactl(obd, sb, qctl);
222         if (ret) {
223                 if (ret == -ESRCH)      /* no limit */
224                         ret = 0;
225                 else
226                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
227                 GOTO(out, ret);
228         }
229
230         if (is_blk) {
231                 usage = qctl->qc_dqblk.dqb_curspace;
232                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
233                 qunit_sz = qctxt->lqc_bunit_sz;
234                 tune_sz = qctxt->lqc_btune_sz;
235
236                 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
237         } else {
238                 usage = qctl->qc_dqblk.dqb_curinodes;
239                 limit = qctl->qc_dqblk.dqb_ihardlimit;
240                 qunit_sz = qctxt->lqc_iunit_sz;
241                 tune_sz = qctxt->lqc_itune_sz;
242         }
243
244         /* ignore the no quota limit case */
245         if (!limit)
246                 GOTO(out, ret = 0);
247
248         /* we don't count the MIN_QLIMIT */
249         if ((limit == MIN_QLIMIT && !is_blk) ||
250             (toqb(limit) == MIN_QLIMIT && is_blk))
251                 limit = 0;
252
253         LASSERT(qdata->qd_count == 0);
254         if (limit <= usage + tune_sz) {
255                 while (qdata->qd_count + limit <= usage + tune_sz)
256                         qdata->qd_count += qunit_sz;
257                 ret = 1;
258         } else if (limit > usage + qunit_sz + tune_sz) {
259                 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz)
260                         qdata->qd_count += qunit_sz;
261                 ret = 2;
262         }
263         LASSERT(ret == 0 || qdata->qd_count);
264         EXIT;
265 out:
266         OBD_FREE_PTR(qctl);
267         return ret;
268 }
269
270 /* caller must hold qunit_hash_lock */
271 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
272                                             struct qunit_data *qdata)
273 {
274         unsigned int hashent = qunit_hashfn(qctxt, qdata);
275         struct lustre_qunit *qunit;
276         ENTRY;
277
278         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
279         qunit = find_qunit(hashent, qctxt, qdata);
280         RETURN(qunit);
281 }
282
283 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
284                                         struct qunit_data *qdata, int opc)
285 {
286         struct lustre_qunit *qunit = NULL;
287         ENTRY;
288
289         OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
290         if (qunit == NULL)
291                 RETURN(NULL);
292
293         INIT_LIST_HEAD(&qunit->lq_hash);
294         INIT_LIST_HEAD(&qunit->lq_waiters);
295         atomic_set(&qunit->lq_refcnt, 1);
296         qunit->lq_ctxt = qctxt;
297         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
298         qunit->lq_opc = opc;
299
300         RETURN(qunit);
301 }
302
303 static inline void free_qunit(struct lustre_qunit *qunit)
304 {
305         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
306 }
307
308 static inline void qunit_get(struct lustre_qunit *qunit)
309 {
310         atomic_inc(&qunit->lq_refcnt);
311 }
312
313 static void qunit_put(struct lustre_qunit *qunit)
314 {
315         LASSERT(atomic_read(&qunit->lq_refcnt));
316         if (atomic_dec_and_test(&qunit->lq_refcnt))
317                 free_qunit(qunit);
318 }
319
320 static void
321 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
322 {
323         struct list_head *head;
324
325         LASSERT(list_empty(&qunit->lq_hash));
326         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
327         list_add(&qunit->lq_hash, head);
328 }
329
330 static void remove_qunit_nolock(struct lustre_qunit *qunit)
331 {
332         LASSERT(!list_empty(&qunit->lq_hash));
333         list_del_init(&qunit->lq_hash);
334 }
335
336 struct qunit_waiter {
337         struct list_head qw_entry;
338         cfs_waitq_t      qw_waitq;
339         int qw_rc;
340 };
341
342 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
343                                  (limit = count) : (limit += count)
344
345
346 /* FIXME check if this mds is the master of specified id */
347 static int
348 is_master(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
349           unsigned int id, int type)
350 {
351         return qctxt->lqc_handler ? 1 : 0;
352 }
353
354 static int
355 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
356                struct qunit_data *qdata, int opc, int wait);
357
358 static int split_before_schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
359                                        struct qunit_data *qdata, int opc, int wait)
360 {
361         int rc = 0, ret;
362         struct qunit_data tmp_qdata;
363         ENTRY;
364
365         LASSERT(qdata);
366         if (qctxt->lqc_import)
367                 while (should_translate_quota(qctxt->lqc_import) &&
368                        qdata->qd_count > MAX_QUOTA_COUNT32) {
369
370                         tmp_qdata = *qdata;
371                         tmp_qdata.qd_count = MAX_QUOTA_COUNT32;
372                         qdata->qd_count -= tmp_qdata.qd_count;
373                         ret = schedule_dqacq(obd, qctxt, &tmp_qdata, opc, wait);
374                         if (!rc)
375                                 rc = ret;
376                 }
377
378         if (qdata->qd_count){
379                 ret = schedule_dqacq(obd, qctxt, qdata, opc, wait);
380                 if (!rc)
381                         rc = ret;
382         }
383
384         RETURN(rc);
385 }
386
387 static int
388 dqacq_completion(struct obd_device *obd,
389                  struct lustre_quota_ctxt *qctxt,
390                  struct qunit_data *qdata, int rc, int opc)
391 {
392         struct lustre_qunit *qunit = NULL;
393         struct super_block *sb = qctxt->lqc_sb;
394         unsigned long qunit_sz;
395         struct qunit_waiter *qw, *tmp;
396         int err = 0;
397         __u32 qdata_type = qdata->qd_flags & QUOTA_IS_GRP;
398         __u32 is_blk = (qdata->qd_flags & QUOTA_IS_BLOCK) >> 1;
399         __u64 qd_tmp = qdata->qd_count;
400         unsigned long div_r;
401         ENTRY;
402
403         LASSERT(qdata);
404         qunit_sz = is_blk ? qctxt->lqc_bunit_sz : qctxt->lqc_iunit_sz;
405         div_r = do_div(qd_tmp, qunit_sz);
406         LASSERT(!div_r);
407
408         /* update local operational quota file */
409         if (rc == 0) {
410                 __u32 count = QUSG(qdata->qd_count, is_blk);
411                 struct obd_quotactl *qctl;
412                 __u64 *hardlimit;
413
414                 OBD_ALLOC_PTR(qctl);
415                 if (qctl == NULL)
416                         GOTO(out, err = -ENOMEM);
417
418                 /* acq/rel qunit for specified uid/gid is serialized,
419                  * so there is no race between get fs quota limit and
420                  * set fs quota limit */
421                 qctl->qc_cmd = Q_GETQUOTA;
422                 qctl->qc_id = qdata->qd_id;
423                 qctl->qc_type = qdata_type;
424                 err = fsfilt_quotactl(obd, sb, qctl);
425                 if (err) {
426                         CERROR("error get quota fs limit! (rc:%d)\n", err);
427                         GOTO(out_mem, err);
428                 }
429
430                 if (is_blk) {
431                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
432                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
433                 } else {
434                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
435                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
436                 }
437
438                 switch (opc) {
439                 case QUOTA_DQACQ:
440                         INC_QLIMIT(*hardlimit, count);
441                         break;
442                 case QUOTA_DQREL:
443                         LASSERT(count < *hardlimit);
444                         *hardlimit -= count;
445                         break;
446                 default:
447                         LBUG();
448                 }
449
450                 /* clear quota limit */
451                 if (count == 0)
452                         *hardlimit = 0;
453
454                 qctl->qc_cmd = Q_SETQUOTA;
455                 err = fsfilt_quotactl(obd, sb, qctl);
456                 if (err)
457                         CERROR("error set quota fs limit! (rc:%d)\n", err);
458
459                 QDATA_DEBUG(qdata, "%s completion\n",
460                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
461 out_mem:
462                 OBD_FREE_PTR(qctl);
463         } else if (rc == -EDQUOT) {
464                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
465         } else if (rc == -EBUSY) {
466                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
467         } else {
468                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
469         }
470 out:
471         /* remove the qunit from hash */
472         spin_lock(&qunit_hash_lock);
473
474         qunit = dqacq_in_flight(qctxt, qdata);
475         /* this qunit has been removed by qctxt_cleanup() */
476         if (!qunit) {
477                 spin_unlock(&qunit_hash_lock);
478                 RETURN(err);
479         }
480
481         LASSERT(opc == qunit->lq_opc);
482         remove_qunit_nolock(qunit);
483
484         /* wake up all waiters */
485         list_for_each_entry_safe(qw, tmp, &qunit->lq_waiters, qw_entry) {
486                 list_del_init(&qw->qw_entry);
487                 qw->qw_rc = rc;
488                 wake_up(&qw->qw_waitq);
489         }
490
491         spin_unlock(&qunit_hash_lock);
492
493         qunit_put(qunit);
494
495         /* don't reschedule in such cases:
496          *   - acq/rel failure, but not for quota recovery.
497          *   - local dqacq/dqrel.
498          *   - local disk io failure.
499          */
500         if (err || (rc && rc != -EBUSY) || 
501             is_master(obd, qctxt, qdata->qd_id, qdata_type))
502                 RETURN(err);
503
504         /* reschedule another dqacq/dqrel if needed */
505         qdata->qd_count = 0;
506         rc = check_cur_qunit(obd, qctxt, qdata);
507         if (rc > 0) {
508                 int opc;
509                 opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
510                 rc = split_before_schedule_dqacq(obd, qctxt, qdata, opc, 0);
511                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc);
512         }
513         RETURN(err);
514 }
515
516 struct dqacq_async_args {
517         struct lustre_quota_ctxt *aa_ctxt;
518         struct lustre_qunit *aa_qunit;
519 };
520
521 static int dqacq_interpret(struct ptlrpc_request *req, void *data, int rc)
522 {
523         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
524         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
525         struct lustre_qunit *qunit = aa->aa_qunit;
526         struct obd_device *obd = req->rq_import->imp_obd;
527         struct qunit_data *qdata = NULL;
528         struct qunit_data_old *qdata_old = NULL;
529         ENTRY;
530
531         LASSERT(req);
532         LASSERT(req->rq_import);
533         if ((req->rq_import->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)  &&
534             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
535                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
536                 qdata = lustre_swab_reqbuf(req, REPLY_REC_OFF, sizeof(*qdata), lustre_swab_qdata);
537         } else {
538                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
539                 qdata_old = lustre_swab_reqbuf(req, REPLY_REC_OFF, sizeof(struct qunit_data_old),
540                                                lustre_swab_qdata_old);
541                 qdata = lustre_quota_old_to_new(qdata_old);
542         }
543         if (qdata == NULL) {
544                 DEBUG_REQ(D_ERROR, req, "error unpacking qunit_data");
545                 RETURN(-EPROTO);
546         }
547
548         LASSERT(qdata->qd_id == qunit->lq_data.qd_id &&
549                 (qdata->qd_flags & QUOTA_IS_GRP) == (qunit->lq_data.qd_flags & QUOTA_IS_GRP) &&
550                 (qdata->qd_count == qunit->lq_data.qd_count ||
551                  qdata->qd_count == 0));
552
553         QDATA_DEBUG(qdata, "%s interpret rc(%d).\n",
554                     lustre_msg_get_opc(req->rq_reqmsg) == QUOTA_DQACQ ?
555                     "DQACQ" : "DQREL", rc);
556
557         rc = dqacq_completion(obd, qctxt, qdata, rc,
558                               lustre_msg_get_opc(req->rq_reqmsg));
559
560         RETURN(rc);
561 }
562
563 static int got_qunit(struct qunit_waiter *waiter)
564 {
565         int rc = 0;
566         ENTRY;
567         spin_lock(&qunit_hash_lock);
568         rc = list_empty(&waiter->qw_entry);
569         spin_unlock(&qunit_hash_lock);
570         RETURN(rc);
571 }
572
573 static int
574 schedule_dqacq(struct obd_device *obd,
575                struct lustre_quota_ctxt *qctxt,
576                struct qunit_data *qdata, int opc, int wait)
577 {
578         struct lustre_qunit *qunit, *empty;
579         struct qunit_waiter qw;
580         struct l_wait_info lwi = { 0 };
581         struct ptlrpc_request *req;
582         struct qunit_data *reqdata;
583         struct dqacq_async_args *aa;
584         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*reqdata) };
585         int rc = 0;
586         ENTRY;
587
588         INIT_LIST_HEAD(&qw.qw_entry);
589         init_waitqueue_head(&qw.qw_waitq);
590         qw.qw_rc = 0;
591
592         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
593                 RETURN(-ENOMEM);
594
595         spin_lock(&qunit_hash_lock);
596
597         qunit = dqacq_in_flight(qctxt, qdata);
598         if (qunit) {
599                 if (wait)
600                         list_add_tail(&qw.qw_entry, &qunit->lq_waiters);
601                 spin_unlock(&qunit_hash_lock);
602
603                 free_qunit(empty);
604                 goto wait_completion;
605         }
606         qunit = empty;
607         insert_qunit_nolock(qctxt, qunit);
608         if (wait)
609                 list_add_tail(&qw.qw_entry, &qunit->lq_waiters);
610         spin_unlock(&qunit_hash_lock);
611
612         LASSERT(qunit);
613
614         /* master is going to dqacq/dqrel from itself */
615         if (is_master(obd, qctxt, qdata->qd_id, qdata->qd_flags & QUOTA_IS_GRP)) {
616                 int rc2;
617                 QDATA_DEBUG(qdata, "local %s.\n",
618                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
619                 rc = qctxt->lqc_handler(obd, qdata, opc);
620                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
621                 RETURN((rc && rc != -EDQUOT) ? rc : rc2);
622         }
623
624         /* build dqacq/dqrel request */
625         LASSERT(qctxt->lqc_import);
626         req = ptlrpc_prep_req(qctxt->lqc_import, LUSTRE_MDS_VERSION, opc, 2,
627                               size, NULL);
628         if (!req) {
629                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
630                 RETURN(-ENOMEM);
631         }
632
633         LASSERT(!should_translate_quota(qctxt->lqc_import) || 
634                 qdata->qd_count <= MAX_QUOTA_COUNT32);
635         if (should_translate_quota(qctxt->lqc_import))
636         {
637                 struct qunit_data_old *reqdata_old, *tmp;
638                         
639                 reqdata_old = lustre_msg_buf(req->rq_reqmsg, REPLY_REC_OFF, 
640                                              sizeof(*reqdata_old));
641                 tmp = lustre_quota_new_to_old(qdata);
642                 *reqdata_old = *tmp;
643                 size[1] = sizeof(*reqdata_old);
644                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
645         } else {
646                 reqdata = lustre_msg_buf(req->rq_reqmsg, REPLY_REC_OFF,
647                                          sizeof(*reqdata));
648                 *reqdata = *qdata;
649                 size[1] = sizeof(*reqdata);
650                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
651         }
652         ptlrpc_req_set_repsize(req, 2, size);
653
654         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
655         aa = (struct dqacq_async_args *)&req->rq_async_args;
656         aa->aa_ctxt = qctxt;
657         aa->aa_qunit = qunit;
658
659         req->rq_interpret_reply = dqacq_interpret;
660         ptlrpcd_add_req(req);
661
662         QDATA_DEBUG(qdata, "%s scheduled.\n",
663                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
664 wait_completion:
665         if (wait && qunit) {
666                 struct qunit_data *p = &qunit->lq_data;
667                 QDATA_DEBUG(p, "wait for dqacq.\n");
668
669                 l_wait_event(qw.qw_waitq, got_qunit(&qw), &lwi);
670                 if (qw.qw_rc == 0)
671                         rc = -EAGAIN;
672
673                 CDEBUG(D_QUOTA, "wait dqacq done. (rc:%d)\n", qw.qw_rc);
674         }
675         RETURN(rc);
676 }
677
678 int
679 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
680                    uid_t uid, gid_t gid, __u32 isblk, int wait)
681 {
682         int ret, rc = 0, i = USRQUOTA;
683         __u32 id[MAXQUOTAS] = { uid, gid };
684         struct qunit_data qdata[MAXQUOTAS];
685         ENTRY;
686
687         CLASSERT(MAXQUOTAS < 4);
688         if (!sb_any_quota_enabled(qctxt->lqc_sb))
689                 RETURN(0);
690
691         for (i = 0; i < MAXQUOTAS; i++) {
692                 qdata[i].qd_id = id[i];
693                 qdata[i].qd_flags = 0;
694                 qdata[i].qd_flags |= i;
695                 qdata[i].qd_flags |= isblk ? QUOTA_IS_BLOCK : 0;        
696                 qdata[i].qd_count = 0;
697
698                 ret = check_cur_qunit(obd, qctxt, &qdata[i]);
699                 if (ret > 0) {
700                         int opc;
701                         /* need acquire or release */
702                         opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
703                         ret = split_before_schedule_dqacq(obd, qctxt, &qdata[i], 
704                                                           opc, wait);
705                         if (!rc)
706                                 rc = ret;
707                 }
708         }
709
710         RETURN(rc);
711 }
712
713 int
714 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
715                          unsigned short type, int isblk)
716 {
717         struct lustre_qunit *qunit = NULL;
718         struct qunit_waiter qw;
719         struct qunit_data qdata;
720         struct l_wait_info lwi = { 0 };
721         ENTRY;
722
723         INIT_LIST_HEAD(&qw.qw_entry);
724         init_waitqueue_head(&qw.qw_waitq);
725         qw.qw_rc = 0;
726
727         qdata.qd_id = id;
728         qdata.qd_flags = 0;
729         qdata.qd_flags |= type;
730         qdata.qd_flags |= isblk ? QUOTA_IS_BLOCK : 0;
731         qdata.qd_count = 0;
732
733         spin_lock(&qunit_hash_lock);
734
735         qunit = dqacq_in_flight(qctxt, &qdata);
736         if (qunit)
737                 list_add_tail(&qw.qw_entry, &qunit->lq_waiters);
738
739         spin_unlock(&qunit_hash_lock);
740
741         if (qunit) {
742                 struct qunit_data *p = &qdata;
743                 QDATA_DEBUG(p, "wait for dqacq completion.\n");
744                 l_wait_event(qw.qw_waitq, got_qunit(&qw), &lwi);
745                 QDATA_DEBUG(p, "wait dqacq done. (rc:%d)\n", qw.qw_rc);
746         }
747         RETURN(0);
748 }
749
750 int
751 qctxt_init(struct lustre_quota_ctxt *qctxt, struct super_block *sb,
752            dqacq_handler_t handler)
753 {
754         int rc = 0;
755         ENTRY;
756
757         rc = ptlrpcd_addref();
758         if (rc)
759                 RETURN(rc);
760
761         qctxt->lqc_handler = handler;
762         qctxt->lqc_sb = sb;
763         qctxt->lqc_import = NULL;
764         qctxt->lqc_recovery = 0;
765         qctxt->lqc_atype = 0;
766         qctxt->lqc_status= 0;
767         qctxt->lqc_bunit_sz = default_bunit_sz;
768         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
769         qctxt->lqc_iunit_sz = default_iunit_sz;
770         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
771
772         RETURN(0);
773 }
774
775 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
776 {
777         struct lustre_qunit *qunit, *tmp;
778         struct qunit_waiter *qw, *tmp2;
779         int i;
780         ENTRY;
781
782         spin_lock(&qunit_hash_lock);
783
784         for (i = 0; i < NR_DQHASH; i++) {
785                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
786                         if (qunit->lq_ctxt != qctxt)
787                                 continue;
788
789                         remove_qunit_nolock(qunit);
790                         /* wake up all waiters */
791                         list_for_each_entry_safe(qw, tmp2, &qunit->lq_waiters,
792                                                  qw_entry) {
793                                 list_del_init(&qw->qw_entry);
794                                 qw->qw_rc = 0;
795                                 wake_up(&qw->qw_waitq);
796                         }
797                         qunit_put(qunit);
798                 }
799         }
800
801         spin_unlock(&qunit_hash_lock);
802
803         ptlrpcd_decref();
804
805         EXIT;
806 }
807
808 struct qslave_recov_thread_data {
809         struct obd_device *obd;
810         struct lustre_quota_ctxt *qctxt;
811         struct completion comp;
812 };
813
814 /* FIXME only recovery block quota by now */
815 static int qslave_recovery_main(void *arg)
816 {
817         struct qslave_recov_thread_data *data = arg;
818         struct obd_device *obd = data->obd;
819         struct lustre_quota_ctxt *qctxt = data->qctxt;
820         unsigned int type;
821         int rc = 0;
822         ENTRY;
823
824         ptlrpc_daemonize("qslave_recovd");
825
826         complete(&data->comp);
827
828         if (qctxt->lqc_recovery)
829                 RETURN(0);
830         qctxt->lqc_recovery = 1;
831
832         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
833                 struct qunit_data qdata;
834                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
835                 struct list_head id_list;
836                 struct dquot_id *dqid, *tmp;
837                 int ret;
838
839                 LOCK_DQONOFF_MUTEX(dqopt);
840                 if (!sb_has_quota_enabled(qctxt->lqc_sb, type)) {
841                         UNLOCK_DQONOFF_MUTEX(dqopt);
842                         break;
843                 }
844
845                 LASSERT(dqopt->files[type] != NULL);
846                 INIT_LIST_HEAD(&id_list);
847 #ifndef KERNEL_SUPPORTS_QUOTA_READ 
848                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
849 #else
850                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
851 #endif
852                 UNLOCK_DQONOFF_MUTEX(dqopt);
853                 if (rc)
854                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
855
856                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
857                         list_del_init(&dqid->di_link);
858                         /* skip slave recovery on itself */
859                         if (is_master(obd, qctxt, dqid->di_id, type))
860                                 goto free;
861                         if (rc && rc != -EBUSY)
862                                 goto free;
863
864                         qdata.qd_id = dqid->di_id;
865                         qdata.qd_flags = 0;
866                         qdata.qd_flags |= type;
867                         qdata.qd_flags |= QUOTA_IS_BLOCK;
868                         qdata.qd_count = 0;
869
870                         ret = check_cur_qunit(obd, qctxt, &qdata);
871                         if (ret > 0) {
872                                 int opc;
873                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
874                                 rc = split_before_schedule_dqacq(obd, qctxt, &qdata, opc, 0);
875                         } else
876                                 rc = 0;
877
878                         if (rc)
879                                 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
880                                        "qslave recovery failed! (id:%d type:%d "
881                                        " rc:%d)\n", dqid->di_id, type, rc);
882 free:
883                         kfree(dqid);
884                 }
885         }
886
887         qctxt->lqc_recovery = 0;
888         RETURN(rc);
889 }
890
891 void
892 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
893 {
894         struct qslave_recov_thread_data data;
895         int rc;
896         ENTRY;
897
898         if (!sb_any_quota_enabled(qctxt->lqc_sb))
899                 goto exit;
900
901         data.obd = obd;
902         data.qctxt = qctxt;
903         init_completion(&data.comp);
904
905         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
906         if (rc < 0) {
907                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
908                 goto exit;
909         }
910         wait_for_completion(&data.comp);
911 exit:
912         EXIT;
913 }
914