Whamcloud - gitweb
d5b40224d63505517ffd26d91f697a42e2cfd9af
[fs/lustre-release.git] / lustre / osp / osp_precreate.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, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, 2014, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/osp/osp_sync.c
37  *
38  * Lustre OST Proxy Device
39  *
40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41  * Author: Mikhail Pershin <mike.pershin@intel.com>
42  * Author: Di Wang <di.wang@intel.com>
43  */
44
45 #define DEBUG_SUBSYSTEM S_MDS
46
47 #include <linux/kthread.h>
48
49 #include <lustre_obdo.h>
50
51 #include "osp_internal.h"
52
53 /*
54  * there are two specific states to take care about:
55  *
56  * = import is disconnected =
57  *
58  * = import is inactive =
59  *   in this case osp_declare_object_create() returns an error
60  *
61  */
62
63 /*
64  **
65  * Check whether statfs data is expired
66  *
67  * OSP device caches statfs data for the target, the function checks
68  * whether the data is expired or not.
69  *
70  * \param[in] d         OSP device
71  *
72  * \retval              0 - not expired, 1 - expired
73  */
74 static inline int osp_statfs_need_update(struct osp_device *d)
75 {
76         return !cfs_time_before(cfs_time_current(),
77                                 d->opd_statfs_fresh_till);
78 }
79
80 /*
81  * OSP tries to maintain pool of available objects so that calls to create
82  * objects don't block most of time
83  *
84  * each time OSP gets connected to OST, we should start from precreation cleanup
85  */
86 static inline bool osp_precreate_running(struct osp_device *d)
87 {
88         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
89 }
90
91 static inline bool osp_precreate_stopped(struct osp_device *d)
92 {
93         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
94 }
95
96 static void osp_statfs_timer_cb(unsigned long _d)
97 {
98         struct osp_device *d = (struct osp_device *) _d;
99
100         LASSERT(d);
101         if (d->opd_pre != NULL && osp_precreate_running(d))
102                 wake_up(&d->opd_pre_waitq);
103 }
104
105 /**
106  * RPC interpret callback for OST_STATFS RPC
107  *
108  * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
109  * replied by the target. It's used to maintain statfs cache for the target.
110  * The function fills data from the reply if successful and schedules another
111  * update.
112  *
113  * \param[in] env       LU environment provided by the caller
114  * \param[in] req       RPC replied
115  * \param[in] aa        callback data
116  * \param[in] rc        RPC result
117  *
118  * \retval 0            on success
119  * \retval negative     negated errno on error
120  */
121 static int osp_statfs_interpret(const struct lu_env *env,
122                                 struct ptlrpc_request *req,
123                                 union ptlrpc_async_args *aa, int rc)
124 {
125         struct obd_import       *imp = req->rq_import;
126         struct obd_statfs       *msfs;
127         struct osp_device       *d;
128
129         ENTRY;
130
131         aa = ptlrpc_req_async_args(req);
132         d = aa->pointer_arg[0];
133         LASSERT(d);
134
135         if (rc != 0)
136                 GOTO(out, rc);
137
138         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
139         if (msfs == NULL)
140                 GOTO(out, rc = -EPROTO);
141
142         d->opd_statfs = *msfs;
143
144         osp_pre_update_status(d, rc);
145
146         /* schedule next update */
147         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
148         cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
149         d->opd_statfs_update_in_progress = 0;
150
151         CDEBUG(D_CACHE, "updated statfs %p\n", d);
152
153         RETURN(0);
154 out:
155         /* couldn't update statfs, try again as soon as possible */
156         if (d->opd_pre != NULL && osp_precreate_running(d))
157                 wake_up(&d->opd_pre_waitq);
158
159         if (req->rq_import_generation == imp->imp_generation)
160                 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
161                        d->opd_obd->obd_name, rc);
162         RETURN(rc);
163 }
164
165 /**
166  * Send OST_STATFS RPC
167  *
168  * Sends OST_STATFS RPC to refresh cached statfs data for the target.
169  * Also disables scheduled updates as times OSP may need to refresh
170  * statfs data before expiration. The function doesn't block, instead
171  * an interpretation callback osp_statfs_interpret() is used.
172  *
173  * \param[in] d         OSP device
174  */
175 static int osp_statfs_update(struct osp_device *d)
176 {
177         struct ptlrpc_request   *req;
178         struct obd_import       *imp;
179         union ptlrpc_async_args *aa;
180         int                      rc;
181
182         ENTRY;
183
184         CDEBUG(D_CACHE, "going to update statfs\n");
185
186         imp = d->opd_obd->u.cli.cl_import;
187         LASSERT(imp);
188
189         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
190         if (req == NULL)
191                 RETURN(-ENOMEM);
192
193         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
194         if (rc) {
195                 ptlrpc_request_free(req);
196                 RETURN(rc);
197         }
198         ptlrpc_request_set_replen(req);
199         req->rq_request_portal = OST_CREATE_PORTAL;
200         ptlrpc_at_set_req_timeout(req);
201
202         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
203         aa = ptlrpc_req_async_args(req);
204         aa->pointer_arg[0] = d;
205
206         /*
207          * no updates till reply
208          */
209         cfs_timer_disarm(&d->opd_statfs_timer);
210         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
211         d->opd_statfs_update_in_progress = 1;
212
213         ptlrpcd_add_req(req);
214
215         RETURN(0);
216 }
217
218 /**
219  * Schedule an immediate update for statfs data
220  *
221  * If cached statfs data claim no free space, but OSP has got a request to
222  * destroy an object (so release some space probably), then we may need to
223  * refresh cached statfs data sooner than planned. The function checks there
224  * is no statfs update going and schedules immediate update if so.
225  * XXX: there might be a case where removed object(s) do not add free space (empty
226  * object). If the number of such deletions is high, then we can start to update
227  * statfs too often causing a RPC storm. some throttling is needed...
228  *
229  * \param[in] d         OSP device where statfs data needs to be refreshed
230  */
231 void osp_statfs_need_now(struct osp_device *d)
232 {
233         if (!d->opd_statfs_update_in_progress) {
234                 /*
235                  * if current status is -ENOSPC (lack of free space on OST)
236                  * then we should poll OST immediately once object destroy
237                  * is replied
238                  */
239                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
240                 cfs_timer_disarm(&d->opd_statfs_timer);
241                 wake_up(&d->opd_pre_waitq);
242         }
243 }
244
245 /**
246  * Return number of precreated objects
247  *
248  * A simple helper to calculate the number of precreated objects on the device.
249  *
250  * \param[in] env       LU environment provided by the caller
251  * \param[in] osp       OSP device
252  *
253  * \retval              the number of the precreated objects
254  */
255 static inline int osp_objs_precreated(const struct lu_env *env,
256                                       struct osp_device *osp)
257 {
258         return osp_fid_diff(&osp->opd_pre_last_created_fid,
259                             &osp->opd_pre_used_fid);
260 }
261
262 /**
263  * Check pool of precreated objects is nearly empty
264  *
265  * We should not wait till the pool of the precreated objects is exhausted,
266  * because then there will be a long period of OSP being unavailable for the
267  * new creations due to lenghty precreate RPC. Instead we ask for another
268  * precreation ahead and hopefully have it ready before the current pool is
269  * empty. Notice this function relies on an external locking.
270  *
271  * \param[in] env       LU environment provided by the caller
272  * \param[in] d         OSP device
273  *
274  * \retval              0 - current pool is good enough, 1 - time to precreate
275  */
276 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
277                                                   struct osp_device *d)
278 {
279         int window = osp_objs_precreated(env, d);
280
281         /* don't consider new precreation till OST is healty and
282          * has free space */
283         return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
284                 (d->opd_pre_status == 0));
285 }
286
287 /**
288  * Check pool of precreated objects
289  *
290  * This is protected version of osp_precreate_near_empty_nolock(), check that
291  * for the details.
292  *
293  * \param[in] env       LU environment provided by the caller
294  * \param[in] d         OSP device
295  *
296  * \retval              0 - current pool is good enough, 1 - time to precreate
297  */
298 static inline int osp_precreate_near_empty(const struct lu_env *env,
299                                            struct osp_device *d)
300 {
301         int rc;
302
303         /* XXX: do we really need locking here? */
304         spin_lock(&d->opd_pre_lock);
305         rc = osp_precreate_near_empty_nolock(env, d);
306         spin_unlock(&d->opd_pre_lock);
307         return rc;
308 }
309
310 /**
311  * Check given sequence is empty
312  *
313  * Returns a binary result whether the given sequence has some IDs left
314  * or not. Find the details in osp_fid_end_seq(). This is a lock protected
315  * version of that function.
316  *
317  * \param[in] env       LU environment provided by the caller
318  * \param[in] osp       OSP device
319  *
320  * \retval              0 - current sequence has no IDs, 1 - otherwise
321  */
322 static inline int osp_create_end_seq(const struct lu_env *env,
323                                      struct osp_device *osp)
324 {
325         struct lu_fid *fid = &osp->opd_pre_used_fid;
326         int rc;
327
328         spin_lock(&osp->opd_pre_lock);
329         rc = osp_fid_end_seq(env, fid);
330         spin_unlock(&osp->opd_pre_lock);
331         return rc;
332 }
333
334 /**
335  * Write FID into into last_oid/last_seq file
336  *
337  * The function stores the sequence and the in-sequence id into two dedicated
338  * files. The sync argument can be used to request synchronous commit, so the
339  * function won't return until the updates are committed.
340  *
341  * \param[in] env       LU environment provided by the caller
342  * \param[in] osp       OSP device
343  * \param[in] fid       fid where sequence/id is taken
344  * \param[in] sync      update mode: 0 - asynchronously, 1 - synchronously
345  *
346  * \retval 0            on success
347  * \retval negative     negated errno on error
348  **/
349 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
350                                  struct lu_fid *fid, int sync)
351 {
352         struct osp_thread_info  *oti = osp_env_info(env);
353         struct lu_buf      *lb_oid = &oti->osi_lb;
354         struct lu_buf      *lb_oseq = &oti->osi_lb2;
355         loff_t             oid_off;
356         loff_t             oseq_off;
357         struct thandle    *th;
358         int                   rc;
359         ENTRY;
360
361         /* Note: through f_oid is only 32 bits, it will also write 64 bits
362          * for oid to keep compatibility with the previous version. */
363         lb_oid->lb_buf = &fid->f_oid;
364         lb_oid->lb_len = sizeof(u64);
365         oid_off = sizeof(u64) * osp->opd_index;
366
367         lb_oseq->lb_buf = &fid->f_seq;
368         lb_oseq->lb_len = sizeof(u64);
369         oseq_off = sizeof(u64) * osp->opd_index;
370
371         th = dt_trans_create(env, osp->opd_storage);
372         if (IS_ERR(th))
373                 RETURN(PTR_ERR(th));
374
375         th->th_sync |= sync;
376         rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
377                                      lb_oid, oid_off, th);
378         if (rc != 0)
379                 GOTO(out, rc);
380
381         rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
382                                      lb_oseq, oseq_off, th);
383         if (rc != 0)
384                 GOTO(out, rc);
385
386         rc = dt_trans_start_local(env, osp->opd_storage, th);
387         if (rc != 0)
388                 GOTO(out, rc);
389
390         rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
391                              &oid_off, th);
392         if (rc != 0) {
393                 CERROR("%s: can not write to last seq file: rc = %d\n",
394                         osp->opd_obd->obd_name, rc);
395                 GOTO(out, rc);
396         }
397         rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
398                              &oseq_off, th);
399         if (rc) {
400                 CERROR("%s: can not write to last seq file: rc = %d\n",
401                         osp->opd_obd->obd_name, rc);
402                 GOTO(out, rc);
403         }
404 out:
405         dt_trans_stop(env, osp->opd_storage, th);
406         RETURN(rc);
407 }
408
409 /**
410  * Switch to another sequence
411  *
412  * When a current sequence has no available IDs left, OSP has to switch to
413  * another new sequence. OSP requests it using the regular FLDB protocol
414  * and stores synchronously before that is used in precreated. This is needed
415  * to basically have the sequences referenced (not orphaned), otherwise it's
416  * possible that OST has some objects precreated and the clients have data
417  * written to it, but after MDT failover nobody refers those objects and OSP
418  * has no idea that the sequence need cleanup to be done.
419  * While this is very expensive operation, it's supposed to happen very very
420  * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
421  *
422  * \param[in] env       LU environment provided by the caller
423  * \param[in] osp       OSP device
424  *
425  * \retval 0            on success
426  * \retval negative     negated errno on error
427  */
428 static int osp_precreate_rollover_new_seq(struct lu_env *env,
429                                           struct osp_device *osp)
430 {
431         struct lu_fid   *fid = &osp_env_info(env)->osi_fid;
432         struct lu_fid   *last_fid = &osp->opd_last_used_fid;
433         int             rc;
434         ENTRY;
435
436         rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
437         if (rc != 0) {
438                 CERROR("%s: alloc fid error: rc = %d\n",
439                        osp->opd_obd->obd_name, rc);
440                 RETURN(rc);
441         }
442
443         fid->f_oid = 1;
444         fid->f_ver = 0;
445         LASSERTF(fid_seq(fid) != fid_seq(last_fid),
446                  "fid "DFID", last_fid "DFID"\n", PFID(fid),
447                  PFID(last_fid));
448
449         rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
450         if (rc != 0) {
451                 CERROR("%s: Can not update oid/seq file: rc = %d\n",
452                        osp->opd_obd->obd_name, rc);
453                 RETURN(rc);
454         }
455
456         LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
457                       osp->opd_obd->obd_name, fid_seq(last_fid),
458                       fid_seq(fid));
459         /* Update last_xxx to the new seq */
460         spin_lock(&osp->opd_pre_lock);
461         osp->opd_last_used_fid = *fid;
462         osp->opd_gap_start_fid = *fid;
463         osp->opd_pre_used_fid = *fid;
464         osp->opd_pre_last_created_fid = *fid;
465         spin_unlock(&osp->opd_pre_lock);
466
467         RETURN(rc);
468 }
469
470 /**
471  * Find IDs available in current sequence
472  *
473  * The function calculates the highest possible ID and the number of IDs
474  * available in the current sequence OSP is using. The number is limited
475  * artifically by the caller (grow param) and the number of IDs available
476  * in the sequence by nature. The function doesn't require an external
477  * locking.
478  *
479  * \param[in] env       LU environment provided by the caller
480  * \param[in] osp       OSP device
481  * \param[in] fid       FID the caller wants to start with
482  * \param[in] grow      how many the caller wants
483  * \param[out] fid      the highest calculated FID
484  * \param[out] grow     the number of available IDs calculated
485  *
486  * \retval              0 on success, 1 - the sequence is empty
487  */
488 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
489                               struct lu_fid *fid, int *grow)
490 {
491         struct osp_thread_info  *osi = osp_env_info(env);
492         __u64                   end;
493         int                     i = 0;
494
495         if (fid_is_idif(fid)) {
496                 struct lu_fid   *last_fid;
497                 struct ost_id   *oi = &osi->osi_oi;
498
499                 spin_lock(&osp->opd_pre_lock);
500                 last_fid = &osp->opd_pre_last_created_fid;
501                 fid_to_ostid(last_fid, oi);
502                 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
503                 *grow = end - ostid_id(oi);
504                 ostid_set_id(oi, ostid_id(oi) + *grow);
505                 spin_unlock(&osp->opd_pre_lock);
506
507                 if (*grow == 0)
508                         return 1;
509
510                 ostid_to_fid(fid, oi, osp->opd_index);
511                 return 0;
512         }
513
514         spin_lock(&osp->opd_pre_lock);
515         *fid = osp->opd_pre_last_created_fid;
516         end = fid->f_oid;
517         end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
518         *grow = end - fid->f_oid;
519         fid->f_oid += end - fid->f_oid;
520         spin_unlock(&osp->opd_pre_lock);
521
522         CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
523                *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
524
525         return *grow > 0 ? 0 : 1;
526 }
527
528 /**
529  * Prepare and send precreate RPC
530  *
531  * The function finds how many objects should be precreated.  Then allocates,
532  * prepares and schedules precreate RPC synchronously. Upon reply the function
533  * wake ups the threads waiting for the new objects on this target. If the
534  * target wasn't able to create all the objects requested, then the next
535  * precreate will be asking less objects (i.e. slow precreate down).
536  *
537  * \param[in] env       LU environment provided by the caller
538  * \param[in] d         OSP device
539  *
540  * \retval 0            on success
541  * \retval negative     negated errno on error
542  **/
543 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
544 {
545         struct osp_thread_info  *oti = osp_env_info(env);
546         struct ptlrpc_request   *req;
547         struct obd_import       *imp;
548         struct ost_body         *body;
549         int                      rc, grow, diff;
550         struct lu_fid           *fid = &oti->osi_fid;
551         ENTRY;
552
553         /* don't precreate new objects till OST healthy and has free space */
554         if (unlikely(d->opd_pre_status)) {
555                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
556                        d->opd_obd->obd_name, d->opd_pre_status);
557                 RETURN(0);
558         }
559
560         /*
561          * if not connection/initialization is compeleted, ignore
562          */
563         imp = d->opd_obd->u.cli.cl_import;
564         LASSERT(imp);
565
566         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
567         if (req == NULL)
568                 RETURN(-ENOMEM);
569         req->rq_request_portal = OST_CREATE_PORTAL;
570         /* we should not resend create request - anyway we will have delorphan
571          * and kill these objects */
572         req->rq_no_delay = req->rq_no_resend = 1;
573
574         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
575         if (rc) {
576                 ptlrpc_request_free(req);
577                 RETURN(rc);
578         }
579
580         spin_lock(&d->opd_pre_lock);
581         if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
582                 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
583         grow = d->opd_pre_create_count;
584         spin_unlock(&d->opd_pre_lock);
585
586         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
587         LASSERT(body);
588
589         *fid = d->opd_pre_last_created_fid;
590         rc = osp_precreate_fids(env, d, fid, &grow);
591         if (rc == 1) {
592                 /* Current seq has been used up*/
593                 if (!osp_is_fid_client(d)) {
594                         osp_pre_update_status(d, -ENOSPC);
595                         rc = -ENOSPC;
596                 }
597                 wake_up(&d->opd_pre_waitq);
598                 GOTO(out_req, rc);
599         }
600
601         if (!osp_is_fid_client(d)) {
602                 /* Non-FID client will always send seq 0 because of
603                  * compatiblity */
604                 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
605                 fid->f_seq = 0;
606         }
607
608         fid_to_ostid(fid, &body->oa.o_oi);
609         body->oa.o_valid = OBD_MD_FLGROUP;
610
611         ptlrpc_request_set_replen(req);
612
613         rc = ptlrpc_queue_wait(req);
614         if (rc) {
615                 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
616                        rc);
617                 GOTO(out_req, rc);
618         }
619         LASSERT(req->rq_transno == 0);
620
621         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
622         if (body == NULL)
623                 GOTO(out_req, rc = -EPROTO);
624
625         ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
626         if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
627                 CERROR("%s: precreate fid "DFID" < local used fid "DFID
628                        ": rc = %d\n", d->opd_obd->obd_name,
629                        PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
630                 GOTO(out_req, rc = -ESTALE);
631         }
632
633         diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
634
635         spin_lock(&d->opd_pre_lock);
636         if (diff < grow) {
637                 /* the OST has not managed to create all the
638                  * objects we asked for */
639                 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
640                 d->opd_pre_create_slow = 1;
641         } else {
642                 /* the OST is able to keep up with the work,
643                  * we could consider increasing create_count
644                  * next time if needed */
645                 d->opd_pre_create_slow = 0;
646         }
647
648         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
649         fid_to_ostid(fid, &body->oa.o_oi);
650
651         d->opd_pre_last_created_fid = *fid;
652         spin_unlock(&d->opd_pre_lock);
653
654         CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
655                d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
656                PFID(&d->opd_pre_last_created_fid));
657 out_req:
658         /* now we can wakeup all users awaiting for objects */
659         osp_pre_update_status(d, rc);
660         wake_up(&d->opd_pre_user_waitq);
661
662         ptlrpc_req_finished(req);
663         RETURN(rc);
664 }
665
666 /**
667  * Get last precreated object from target (OST)
668  *
669  * Sends synchronous RPC to the target (OST) to learn the last precreated
670  * object. This later is used to remove all unused objects (cleanup orphan
671  * procedure). Also, the next object after one we got will be used as a
672  * starting point for the new precreates.
673  *
674  * \param[in] env       LU environment provided by the caller
675  * \param[in] d         OSP device
676  *
677  * \retval 0            on success
678  * \retval negative     negated errno on error
679  **/
680 static int osp_get_lastfid_from_ost(const struct lu_env *env,
681                                     struct osp_device *d)
682 {
683         struct ptlrpc_request   *req = NULL;
684         struct obd_import       *imp;
685         struct lu_fid           *last_fid;
686         char                    *tmp;
687         int                     rc;
688         ENTRY;
689
690         imp = d->opd_obd->u.cli.cl_import;
691         LASSERT(imp);
692
693         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
694         if (req == NULL)
695                 RETURN(-ENOMEM);
696
697         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
698                              sizeof(KEY_LAST_FID));
699
700         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
701         if (rc) {
702                 ptlrpc_request_free(req);
703                 RETURN(rc);
704         }
705
706         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
707         memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
708
709         req->rq_no_delay = req->rq_no_resend = 1;
710         last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
711         fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
712
713         ptlrpc_request_set_replen(req);
714
715         rc = ptlrpc_queue_wait(req);
716         if (rc) {
717                 /* bad-bad OST.. let sysadm sort this out */
718                 if (rc == -ENOTSUPP) {
719                         CERROR("%s: server does not support FID: rc = %d\n",
720                                d->opd_obd->obd_name, -ENOTSUPP);
721                 }
722                 ptlrpc_set_import_active(imp, 0);
723                 GOTO(out, rc);
724         }
725
726         last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
727         if (last_fid == NULL) {
728                 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
729                 GOTO(out, rc = -EPROTO);
730         }
731
732         if (!fid_is_sane(last_fid)) {
733                 CERROR("%s: Got insane last_fid "DFID"\n",
734                        d->opd_obd->obd_name, PFID(last_fid));
735                 GOTO(out, rc = -EPROTO);
736         }
737
738         /* Only update the last used fid, if the OST has objects for
739          * this sequence, i.e. fid_oid > 0 */
740         if (fid_oid(last_fid) > 0)
741                 d->opd_last_used_fid = *last_fid;
742
743         CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
744                PFID(last_fid));
745
746 out:
747         ptlrpc_req_finished(req);
748         RETURN(rc);
749 }
750
751 /**
752  * Cleanup orphans on OST
753  *
754  * This function is called in a contex of a dedicated thread handling
755  * all the precreation suff. The function waits till local recovery
756  * is complete, then identify all the unreferenced objects (orphans)
757  * using the highest ID referenced by a local and the highest object
758  * precreated by the target. The found range is a subject to removal
759  * using specially flagged RPC. During this process OSP is marked
760  * unavailable for new objects.
761  *
762  * \param[in] env       LU environment provided by the caller
763  * \param[in] d         OSP device
764  *
765  * \retval 0            on success
766  * \retval negative     negated errno on error
767  */
768 static int osp_precreate_cleanup_orphans(struct lu_env *env,
769                                          struct osp_device *d)
770 {
771         struct osp_thread_info  *osi = osp_env_info(env);
772         struct lu_fid           *last_fid = &osi->osi_fid;
773         struct ptlrpc_request   *req = NULL;
774         struct obd_import       *imp;
775         struct ost_body         *body;
776         struct l_wait_info       lwi = { 0 };
777         int                      update_status = 0;
778         int                      rc;
779         int                      diff;
780
781         ENTRY;
782
783         /*
784          * wait for local recovery to finish, so we can cleanup orphans
785          * orphans are all objects since "last used" (assigned), but
786          * there might be objects reserved and in some cases they won't
787          * be used. we can't cleanup them till we're sure they won't be
788          * used. also can't we allow new reservations because they may
789          * end up getting orphans being cleaned up below. so we block
790          * new reservations and wait till all reserved objects either
791          * user or released.
792          */
793         spin_lock(&d->opd_pre_lock);
794         d->opd_pre_recovering = 1;
795         spin_unlock(&d->opd_pre_lock);
796         /*
797          * The locking above makes sure the opd_pre_reserved check below will
798          * catch all osp_precreate_reserve() calls who find
799          * "!opd_pre_recovering".
800          */
801         l_wait_event(d->opd_pre_waitq,
802                      (!d->opd_pre_reserved && d->opd_recovery_completed) ||
803                      !osp_precreate_running(d) || d->opd_got_disconnected,
804                      &lwi);
805         if (!osp_precreate_running(d) || d->opd_got_disconnected)
806                 GOTO(out, rc = -EAGAIN);
807
808         CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
809                d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
810
811         *last_fid = d->opd_last_used_fid;
812         /* The OSP should already get the valid seq now */
813         LASSERT(!fid_is_zero(last_fid));
814         if (fid_oid(&d->opd_last_used_fid) < 2) {
815                 /* lastfid looks strange... ask OST */
816                 rc = osp_get_lastfid_from_ost(env, d);
817                 if (rc)
818                         GOTO(out, rc);
819         }
820
821         imp = d->opd_obd->u.cli.cl_import;
822         LASSERT(imp);
823
824         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
825         if (req == NULL)
826                 GOTO(out, rc = -ENOMEM);
827
828         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
829         if (rc) {
830                 ptlrpc_request_free(req);
831                 req = NULL;
832                 GOTO(out, rc);
833         }
834
835         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
836         if (body == NULL)
837                 GOTO(out, rc = -EPROTO);
838
839         body->oa.o_flags = OBD_FL_DELORPHAN;
840         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
841
842         fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
843
844         ptlrpc_request_set_replen(req);
845
846         /* Don't resend the delorphan req */
847         req->rq_no_resend = req->rq_no_delay = 1;
848
849         rc = ptlrpc_queue_wait(req);
850         if (rc) {
851                 update_status = 1;
852                 GOTO(out, rc);
853         }
854
855         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
856         if (body == NULL)
857                 GOTO(out, rc = -EPROTO);
858
859         /*
860          * OST provides us with id new pool starts from in body->oa.o_id
861          */
862         ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
863
864         spin_lock(&d->opd_pre_lock);
865         diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
866         if (diff > 0) {
867                 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
868                 d->opd_pre_last_created_fid = d->opd_last_used_fid;
869         } else {
870                 d->opd_pre_create_count = OST_MIN_PRECREATE;
871                 d->opd_pre_last_created_fid = *last_fid;
872         }
873         /*
874          * This empties the pre-creation pool and effectively blocks any new
875          * reservations.
876          */
877         LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
878                 LUSTRE_DATA_SEQ_MAX_WIDTH);
879         d->opd_pre_used_fid = d->opd_pre_last_created_fid;
880         d->opd_pre_create_slow = 0;
881         spin_unlock(&d->opd_pre_lock);
882
883         CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
884                "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
885                PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
886 out:
887         if (req)
888                 ptlrpc_req_finished(req);
889
890         spin_lock(&d->opd_pre_lock);
891         d->opd_pre_recovering = 0;
892         spin_unlock(&d->opd_pre_lock);
893
894         /*
895          * If rc is zero, the pre-creation window should have been emptied.
896          * Since waking up the herd would be useless without pre-created
897          * objects, we defer the signal to osp_precreate_send() in that case.
898          */
899         if (rc != 0) {
900                 if (update_status) {
901                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
902                                d->opd_obd->obd_name, rc);
903                         /* we can't proceed from here, OST seem to
904                          * be in a bad shape, better to wait for
905                          * a new instance of the server and repeat
906                          * from the beginning. notify possible waiters
907                          * this OSP isn't quite functional yet */
908                         osp_pre_update_status(d, rc);
909                 } else {
910                         wake_up(&d->opd_pre_user_waitq);
911                 }
912         }
913
914         RETURN(rc);
915 }
916
917 /**
918  * Update precreate status using statfs data
919  *
920  * The function decides whether this OSP should be used for new objects.
921  * IOW, whether this OST is used up or has some free space. Cached statfs
922  * data is used to make this decision. If the latest result of statfs
923  * request (rc argument) is not success, then just mark OSP unavailable
924  * right away.
925
926  * Add a bit of hysteresis so this flag isn't continually flapping,
927  * and ensure that new files don't get extremely fragmented due to
928  * only a small amount of available space in the filesystem.
929  * We want to set the NOSPC flag when there is less than ~0.1% free
930  * and clear it when there is at least ~0.2% free space, so:
931  *                   avail < ~0.1% max          max = avail + used
932  *            1025 * avail < avail + used       used = blocks - free
933  *            1024 * avail < used
934  *            1024 * avail < blocks - free
935  *                   avail < ((blocks - free) >> 10)
936  *
937  * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
938  * lose that amount of space so in those cases we report no space left
939  * if their is less than 1 GB left.
940  * the function updates current precreation status used: functional or not
941  *
942  * \param[in] d         OSP device
943  * \param[in] rc        new precreate status for device \a d
944  *
945  * \retval 0            on success
946  * \retval negative     negated errno on error
947  */
948 void osp_pre_update_status(struct osp_device *d, int rc)
949 {
950         struct obd_statfs       *msfs = &d->opd_statfs;
951         int                      old = d->opd_pre_status;
952         __u64                    used;
953
954         d->opd_pre_status = rc;
955         if (rc)
956                 goto out;
957
958         if (likely(msfs->os_type)) {
959                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
960                                     1 << 30);
961                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
962                         d->opd_pre_status = -ENOSPC;
963                         if (old != -ENOSPC)
964                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
965                                        LPU64" free, "LPU64" used, "LPU64" "
966                                        "avail -> %d: rc = %d\n",
967                                        d->opd_obd->obd_name, msfs->os_blocks,
968                                        msfs->os_bfree, used, msfs->os_bavail,
969                                        d->opd_pre_status, rc);
970                         CDEBUG(D_INFO,
971                                "non-committed changes: %lu, in progress: %u\n",
972                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
973                 } else if (old == -ENOSPC) {
974                         d->opd_pre_status = 0;
975                         spin_lock(&d->opd_pre_lock);
976                         d->opd_pre_create_slow = 0;
977                         d->opd_pre_create_count = OST_MIN_PRECREATE;
978                         spin_unlock(&d->opd_pre_lock);
979                         wake_up(&d->opd_pre_waitq);
980                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
981                                " free, "LPU64" used, "LPU64" avail -> %d: "
982                                "rc = %d\n", d->opd_obd->obd_name,
983                                msfs->os_blocks, msfs->os_bfree, used,
984                                msfs->os_bavail, d->opd_pre_status, rc);
985                 }
986         }
987
988 out:
989         wake_up(&d->opd_pre_user_waitq);
990 }
991
992 /**
993  * Initialize FID for precreation
994  *
995  * For a just created new target, a new sequence should be taken.
996  * The function checks there is no IDIF in use (if the target was
997  * added with the older version of Lustre), then requests a new
998  * sequence from FLDB using the regular protocol. Then this new
999  * sequence is stored on a persisten storage synchronously to prevent
1000  * possible object leakage (for the detail see the description for
1001  * osp_precreate_rollover_new_seq()).
1002  *
1003  * \param[in] osp       OSP device
1004  *
1005  * \retval 0            on success
1006  * \retval negative     negated errno on error
1007  */
1008 int osp_init_pre_fid(struct osp_device *osp)
1009 {
1010         struct lu_env           env;
1011         struct osp_thread_info  *osi;
1012         struct lu_client_seq    *cli_seq;
1013         struct lu_fid           *last_fid;
1014         int                     rc;
1015         ENTRY;
1016
1017         LASSERT(osp->opd_pre != NULL);
1018
1019         /* Return if last_used fid has been initialized */
1020         if (!fid_is_zero(&osp->opd_last_used_fid))
1021                 RETURN(0);
1022
1023         rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1024         if (rc) {
1025                 CERROR("%s: init env error: rc = %d\n",
1026                        osp->opd_obd->obd_name, rc);
1027                 RETURN(rc);
1028         }
1029
1030         osi = osp_env_info(&env);
1031         last_fid = &osi->osi_fid;
1032         fid_zero(last_fid);
1033         /* For a freshed fs, it will allocate a new sequence first */
1034         if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1035                 cli_seq = osp->opd_obd->u.cli.cl_seq;
1036                 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1037                 if (rc != 0) {
1038                         CERROR("%s: alloc fid error: rc = %d\n",
1039                                osp->opd_obd->obd_name, rc);
1040                         GOTO(out, rc);
1041                 }
1042         } else {
1043                 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1044         }
1045         last_fid->f_oid = 1;
1046         last_fid->f_ver = 0;
1047
1048         spin_lock(&osp->opd_pre_lock);
1049         osp->opd_last_used_fid = *last_fid;
1050         osp->opd_pre_used_fid = *last_fid;
1051         osp->opd_pre_last_created_fid = *last_fid;
1052         spin_unlock(&osp->opd_pre_lock);
1053         rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1054         if (rc != 0) {
1055                 CERROR("%s: write fid error: rc = %d\n",
1056                        osp->opd_obd->obd_name, rc);
1057                 GOTO(out, rc);
1058         }
1059 out:
1060         lu_env_fini(&env);
1061         RETURN(rc);
1062 }
1063
1064 /**
1065  * The core of precreate functionality
1066  *
1067  * The function implements the main precreation loop. Basically it
1068  * involves connecting to the target, precerate FID initialization,
1069  * identifying and removing orphans, then serving precreation. As
1070  * part of the latter, the thread is responsible for statfs data
1071  * updates. The precreation is mostly driven by another threads
1072  * asking for new OST objects - those askers wake the thread when
1073  * the number of precreated objects reach low watermark.
1074  * After a disconnect, the sequence above repeats. This is keep going
1075  * until the thread is requested to stop.
1076  *
1077  * \param[in] _arg      private data the thread (OSP device to handle)
1078  *
1079  * \retval 0            on success
1080  * \retval negative     negated errno on error
1081  */
1082 static int osp_precreate_thread(void *_arg)
1083 {
1084         struct osp_device       *d = _arg;
1085         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
1086         struct l_wait_info       lwi = { 0 };
1087         struct lu_env            env;
1088         int                      rc;
1089
1090         ENTRY;
1091
1092         rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1093         if (rc) {
1094                 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1095                        rc);
1096                 RETURN(rc);
1097         }
1098
1099         spin_lock(&d->opd_pre_lock);
1100         thread->t_flags = SVC_RUNNING;
1101         spin_unlock(&d->opd_pre_lock);
1102         wake_up(&thread->t_ctl_waitq);
1103
1104         while (osp_precreate_running(d)) {
1105                 /*
1106                  * need to be connected to OST
1107                  */
1108                 while (osp_precreate_running(d)) {
1109                         l_wait_event(d->opd_pre_waitq,
1110                                      !osp_precreate_running(d) ||
1111                                      d->opd_new_connection,
1112                                      &lwi);
1113
1114                         if (!d->opd_new_connection)
1115                                 continue;
1116
1117                         d->opd_new_connection = 0;
1118                         d->opd_got_disconnected = 0;
1119                         break;
1120                 }
1121
1122                 if (!osp_precreate_running(d))
1123                         break;
1124
1125                 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1126                 /* Sigh, fid client is not ready yet */
1127                 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1128                         continue;
1129
1130                 /* Init fid for osp_precreate if necessary */
1131                 rc = osp_init_pre_fid(d);
1132                 if (rc != 0) {
1133                         class_export_put(d->opd_exp);
1134                         d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1135                         CERROR("%s: init pre fid error: rc = %d\n",
1136                                d->opd_obd->obd_name, rc);
1137                         continue;
1138                 }
1139
1140                 osp_statfs_update(d);
1141
1142                 /*
1143                  * Clean up orphans or recreate missing objects.
1144                  */
1145                 rc = osp_precreate_cleanup_orphans(&env, d);
1146                 if (rc != 0)
1147                         continue;
1148                 /*
1149                  * connected, can handle precreates now
1150                  */
1151                 while (osp_precreate_running(d)) {
1152                         l_wait_event(d->opd_pre_waitq,
1153                                      !osp_precreate_running(d) ||
1154                                      osp_precreate_near_empty(&env, d) ||
1155                                      osp_statfs_need_update(d) ||
1156                                      d->opd_got_disconnected, &lwi);
1157
1158                         if (!osp_precreate_running(d))
1159                                 break;
1160
1161                         /* something happened to the connection
1162                          * have to start from the beginning */
1163                         if (d->opd_got_disconnected)
1164                                 break;
1165
1166                         if (osp_statfs_need_update(d))
1167                                 osp_statfs_update(d);
1168
1169                         /* To avoid handling different seq in precreate/orphan
1170                          * cleanup, it will hold precreate until current seq is
1171                          * used up. */
1172                         if (unlikely(osp_precreate_end_seq(&env, d) &&
1173                             !osp_create_end_seq(&env, d)))
1174                                 continue;
1175
1176                         if (unlikely(osp_precreate_end_seq(&env, d) &&
1177                                      osp_create_end_seq(&env, d))) {
1178                                 LCONSOLE_INFO("%s:"LPX64" is used up."
1179                                               " Update to new seq\n",
1180                                               d->opd_obd->obd_name,
1181                                          fid_seq(&d->opd_pre_last_created_fid));
1182                                 rc = osp_precreate_rollover_new_seq(&env, d);
1183                                 if (rc)
1184                                         continue;
1185                         }
1186
1187                         if (osp_precreate_near_empty(&env, d)) {
1188                                 rc = osp_precreate_send(&env, d);
1189                                 /* osp_precreate_send() sets opd_pre_status
1190                                  * in case of error, that prevent the using of
1191                                  * failed device. */
1192                                 if (rc < 0 && rc != -ENOSPC &&
1193                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
1194                                         CERROR("%s: cannot precreate objects:"
1195                                                " rc = %d\n",
1196                                                d->opd_obd->obd_name, rc);
1197                         }
1198                 }
1199         }
1200
1201         thread->t_flags = SVC_STOPPED;
1202         lu_env_fini(&env);
1203         wake_up(&thread->t_ctl_waitq);
1204
1205         RETURN(0);
1206 }
1207
1208 /**
1209  * Check when to stop to wait for precreate objects.
1210  *
1211  * The caller wanting a new OST object can't wait undefinitely. The
1212  * function checks for few conditions including available new OST
1213  * objects, disconnected OST, lack of space with no pending destroys,
1214  * etc. IOW, it checks whether the current OSP state is good to keep
1215  * waiting or it's better to give up.
1216  *
1217  * \param[in] env       LU environment provided by the caller
1218  * \param[in] d         OSP device
1219  *
1220  * \retval              0 - keep waiting, 1 - no luck
1221  */
1222 static int osp_precreate_ready_condition(const struct lu_env *env,
1223                                          struct osp_device *d)
1224 {
1225         if (d->opd_pre_recovering)
1226                 return 0;
1227
1228         /* ready if got enough precreated objects */
1229         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1230         if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1231                 return 1;
1232
1233         /* ready if OST reported no space and no destroys in progress */
1234         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1235             d->opd_pre_status == -ENOSPC)
1236                 return 1;
1237
1238         /* Bail out I/O fails to OST */
1239         if (d->opd_pre_status != 0 &&
1240             d->opd_pre_status != -EAGAIN &&
1241             d->opd_pre_status != -ENODEV &&
1242             d->opd_pre_status != -ENOSPC) {
1243                 /* DEBUG LU-3230 */
1244                 if (d->opd_pre_status != -EIO)
1245                         CERROR("%s: precreate failed opd_pre_status %d\n",
1246                                d->opd_obd->obd_name, d->opd_pre_status);
1247                 return 1;
1248         }
1249
1250         return 0;
1251 }
1252
1253 static int osp_precreate_timeout_condition(void *data)
1254 {
1255         struct osp_device *d = data;
1256
1257         CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1258               "reserved="LPU64", syn_changes=%lu, "
1259               "syn_rpc_in_progress=%d, status=%d\n",
1260               d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1261               PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1262               d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1263               d->opd_pre_status);
1264
1265         return 1;
1266 }
1267
1268 /**
1269  * Reserve object in precreate pool
1270  *
1271  * When the caller wants to create a new object on this target (target
1272  * represented by the given OSP), it should declare this intention using
1273  * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1274  * to reserve an object in the existing precreated pool or wait up to
1275  * obd_timeout for the available object to appear in the pool (a dedicated
1276  * thread will be doing real precreation in background). The object can be
1277  * consumed later with osp_precreate_get_fid() or be released with call to
1278  * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1279  * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1280  * If the space on the target is short and there is a pending object destroy,
1281  * then the function forces local commit to speedup space release (see
1282  * osp_sync.c for the details).
1283  *
1284  * \param[in] env       LU environment provided by the caller
1285  * \param[in] d         OSP device
1286  *
1287  * \retval              0 on success
1288  * \retval              -ENOSPC when no space on OST
1289  * \retval              -EAGAIN try later, slow precreation in progress
1290  * \retval              -EIO when no access to OST
1291  */
1292 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1293 {
1294         struct l_wait_info       lwi;
1295         cfs_time_t               expire = cfs_time_shift(obd_timeout);
1296         int                      precreated, rc;
1297
1298         ENTRY;
1299
1300         LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1301                  "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1302                  PFID(&d->opd_pre_used_fid));
1303
1304         /*
1305          * wait till:
1306          *  - preallocation is done
1307          *  - no free space expected soon
1308          *  - can't connect to OST for too long (obd_timeout)
1309          *  - OST can allocate fid sequence.
1310          */
1311         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1312                 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1313
1314                 /*
1315                  * increase number of precreations
1316                  */
1317                 precreated = osp_objs_precreated(env, d);
1318                 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1319                     d->opd_pre_create_slow == 0 &&
1320                     precreated <= (d->opd_pre_create_count / 4 + 1)) {
1321                         spin_lock(&d->opd_pre_lock);
1322                         d->opd_pre_create_slow = 1;
1323                         d->opd_pre_create_count *= 2;
1324                         spin_unlock(&d->opd_pre_lock);
1325                 }
1326
1327                 spin_lock(&d->opd_pre_lock);
1328                 precreated = osp_objs_precreated(env, d);
1329                 if (precreated > d->opd_pre_reserved &&
1330                     !d->opd_pre_recovering) {
1331                         d->opd_pre_reserved++;
1332                         spin_unlock(&d->opd_pre_lock);
1333                         rc = 0;
1334
1335                         /* XXX: don't wake up if precreation is in progress */
1336                         if (osp_precreate_near_empty_nolock(env, d) &&
1337                            !osp_precreate_end_seq_nolock(env, d))
1338                                 wake_up(&d->opd_pre_waitq);
1339
1340                         break;
1341                 }
1342                 spin_unlock(&d->opd_pre_lock);
1343
1344                 /*
1345                  * all precreated objects have been used and no-space
1346                  * status leave us no chance to succeed very soon
1347                  * but if there is destroy in progress, then we should
1348                  * wait till that is done - some space might be released
1349                  */
1350                 if (unlikely(rc == -ENOSPC)) {
1351                         if (d->opd_syn_changes) {
1352                                 /* force local commit to release space */
1353                                 dt_commit_async(env, d->opd_storage);
1354                         }
1355                         if (d->opd_syn_rpc_in_progress) {
1356                                 /* just wait till destroys are done */
1357                                 /* see l_wait_even() few lines below */
1358                         }
1359                         if (d->opd_syn_changes +
1360                             d->opd_syn_rpc_in_progress == 0) {
1361                                 /* no hope for free space */
1362                                 break;
1363                         }
1364                 }
1365
1366                 /* XXX: don't wake up if precreation is in progress */
1367                 wake_up(&d->opd_pre_waitq);
1368
1369                 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1370                                 osp_precreate_timeout_condition, d);
1371                 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1372                         rc = -ETIMEDOUT;
1373                         break;
1374                 }
1375
1376                 l_wait_event(d->opd_pre_user_waitq,
1377                              osp_precreate_ready_condition(env, d), &lwi);
1378         }
1379
1380         RETURN(rc);
1381 }
1382
1383 /**
1384  * Get a FID from precreation pool
1385  *
1386  * The function is a companion for osp_precreate_reserve() - it assigns
1387  * a specific FID from the precreate. The function should be called only
1388  * if the call to osp_precreate_reserve() was successful. The function
1389  * updates a local storage to remember the highest object ID referenced
1390  * by the node in the given sequence.
1391  *
1392  * A very importan details: this is supposed to be called once the
1393  * transaction is started, so on-disk update will be atomic with the
1394  * data (like LOVEA) refering this object. Then the object won't be leaked:
1395  * either it's referenced by the committed transaction or it's a subject
1396  * to the orphan cleanup procedure.
1397  *
1398  * \param[in] env       LU environment provided by the caller
1399  * \param[in] d         OSP device
1400  * \param[out] fid      generated FID
1401  *
1402  * \retval 0            on success
1403  * \retval negative     negated errno on error
1404  */
1405 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1406                           struct lu_fid *fid)
1407 {
1408         /* grab next id from the pool */
1409         spin_lock(&d->opd_pre_lock);
1410
1411         LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1412                              &d->opd_pre_last_created_fid) < 0,
1413                  "next fid "DFID" last created fid "DFID"\n",
1414                  PFID(&d->opd_pre_used_fid),
1415                  PFID(&d->opd_pre_last_created_fid));
1416
1417         d->opd_pre_used_fid.f_oid++;
1418         memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1419         d->opd_pre_reserved--;
1420         /*
1421          * last_used_id must be changed along with getting new id otherwise
1422          * we might miscalculate gap causing object loss or leak
1423          */
1424         osp_update_last_fid(d, fid);
1425         spin_unlock(&d->opd_pre_lock);
1426
1427         /*
1428          * probably main thread suspended orphan cleanup till
1429          * all reservations are released, see comment in
1430          * osp_precreate_thread() just before orphan cleanup
1431          */
1432         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1433                 wake_up(&d->opd_pre_waitq);
1434
1435         return 0;
1436 }
1437
1438 /*
1439  * Set size regular attribute on an object
1440  *
1441  * When a striping is created late, it's possible that size is already
1442  * initialized on the file. Then the new striping should inherit size
1443  * from the file. The function sets size on the object using the regular
1444  * protocol (OST_PUNCH).
1445  * XXX: should be re-implemented using OUT ?
1446  *
1447  * \param[in] env       LU environment provided by the caller
1448  * \param[in] dt        object
1449  * \param[in] size      size to set.
1450  *
1451  * \retval 0            on success
1452  * \retval negative     negated errno on error
1453  */
1454 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1455                         __u64 size)
1456 {
1457         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
1458         struct ptlrpc_request   *req = NULL;
1459         struct obd_import       *imp;
1460         struct ost_body         *body;
1461         struct obdo             *oa = NULL;
1462         int                      rc;
1463
1464         ENTRY;
1465
1466         imp = d->opd_obd->u.cli.cl_import;
1467         LASSERT(imp);
1468
1469         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1470         if (req == NULL)
1471                 RETURN(-ENOMEM);
1472
1473         /* XXX: capa support? */
1474         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1475         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1476         if (rc) {
1477                 ptlrpc_request_free(req);
1478                 RETURN(rc);
1479         }
1480
1481         /*
1482          * XXX: decide how do we do here with resend
1483          * if we don't resend, then client may see wrong file size
1484          * if we do resend, then MDS thread can get stuck for quite long
1485          * and if we don't resend, then client will also get -EWOULDBLOCK !!
1486          * (see LU-7975 and sanity/test_27F use cases)
1487          * but let's decide not to resend/delay this truncate request to OST
1488          * and allow Client to decide to resend, in a less agressive way from
1489          * after_reply(), by returning -EINPROGRESS instead of
1490          * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1491          * end of this routine
1492          */
1493         req->rq_no_resend = req->rq_no_delay = 1;
1494
1495         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1496         ptlrpc_at_set_req_timeout(req);
1497
1498         OBD_ALLOC_PTR(oa);
1499         if (oa == NULL)
1500                 GOTO(out, rc = -ENOMEM);
1501
1502         rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1503         LASSERT(rc == 0);
1504         oa->o_size = size;
1505         oa->o_blocks = OBD_OBJECT_EOF;
1506         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1507                       OBD_MD_FLID | OBD_MD_FLGROUP;
1508
1509         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1510         LASSERT(body);
1511         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1512
1513         /* XXX: capa support? */
1514         /* osc_pack_capa(req, body, capa); */
1515
1516         ptlrpc_request_set_replen(req);
1517
1518         rc = ptlrpc_queue_wait(req);
1519         if (rc) {
1520                 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1521                  * since we have decided not to resend/delay, but this could
1522                  * lead to wrong size to be seen at Client side and even process
1523                  * trying to open to exit/fail if not itself handling -EAGAIN.
1524                  * So it should be better to return -EINPROGRESS instead and
1525                  * leave the decision to resend at Client side in after_reply()
1526                  */
1527                 if (rc == -EWOULDBLOCK) {
1528                         rc = -EINPROGRESS;
1529                         CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1530                                "-EWOULDBLOCK/-EAGAIN to allow Client to "
1531                                "resend\n");
1532                 } else {
1533                         CERROR("can't punch object: %d\n", rc);
1534                 }
1535         }
1536 out:
1537         ptlrpc_req_finished(req);
1538         if (oa)
1539                 OBD_FREE_PTR(oa);
1540         RETURN(rc);
1541 }
1542
1543 /**
1544  * Initialize precreation functionality of OSP
1545  *
1546  * Prepares all the internal structures and starts the precreate thread
1547  *
1548  * \param[in] d         OSP device
1549  *
1550  * \retval 0            on success
1551  * \retval negative     negated errno on error
1552  */
1553 int osp_init_precreate(struct osp_device *d)
1554 {
1555         struct l_wait_info       lwi = { 0 };
1556         struct task_struct              *task;
1557
1558         ENTRY;
1559
1560         OBD_ALLOC_PTR(d->opd_pre);
1561         if (d->opd_pre == NULL)
1562                 RETURN(-ENOMEM);
1563
1564         /* initially precreation isn't ready */
1565         d->opd_pre_status = -EAGAIN;
1566         fid_zero(&d->opd_pre_used_fid);
1567         d->opd_pre_used_fid.f_oid = 1;
1568         fid_zero(&d->opd_pre_last_created_fid);
1569         d->opd_pre_last_created_fid.f_oid = 1;
1570         d->opd_pre_reserved = 0;
1571         d->opd_got_disconnected = 1;
1572         d->opd_pre_create_slow = 0;
1573         d->opd_pre_create_count = OST_MIN_PRECREATE;
1574         d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1575         d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1576
1577         spin_lock_init(&d->opd_pre_lock);
1578         init_waitqueue_head(&d->opd_pre_waitq);
1579         init_waitqueue_head(&d->opd_pre_user_waitq);
1580         init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1581
1582         /*
1583          * Initialize statfs-related things
1584          */
1585         d->opd_statfs_maxage = 5; /* default update interval */
1586         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1587         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1588                (unsigned long long)cfs_time_current(),
1589                (unsigned long long)d->opd_statfs_fresh_till);
1590         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1591
1592         /*
1593          * start thread handling precreation and statfs updates
1594          */
1595         task = kthread_run(osp_precreate_thread, d,
1596                            "osp-pre-%u-%u", d->opd_index, d->opd_group);
1597         if (IS_ERR(task)) {
1598                 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1599                 RETURN(PTR_ERR(task));
1600         }
1601
1602         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1603                      osp_precreate_running(d) || osp_precreate_stopped(d),
1604                      &lwi);
1605
1606         RETURN(0);
1607 }
1608
1609 /**
1610  * Finish precreate functionality of OSP
1611  *
1612  *
1613  * Asks all the activity (the thread, update timer) to stop, then
1614  * wait till that is done.
1615  *
1616  * \param[in] d         OSP device
1617  */
1618 void osp_precreate_fini(struct osp_device *d)
1619 {
1620         struct ptlrpc_thread *thread;
1621
1622         ENTRY;
1623
1624         cfs_timer_disarm(&d->opd_statfs_timer);
1625
1626         if (d->opd_pre == NULL)
1627                 RETURN_EXIT;
1628
1629         thread = &d->opd_pre_thread;
1630
1631         thread->t_flags = SVC_STOPPING;
1632         wake_up(&d->opd_pre_waitq);
1633
1634         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1635
1636         OBD_FREE_PTR(d->opd_pre);
1637         d->opd_pre = NULL;
1638
1639         EXIT;
1640 }
1641