Whamcloud - gitweb
LU-11115 lod: skip max_create_count=0 OST in QoS and RR algorithms
[fs/lustre-release.git] / lustre / lod / lod_qos.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,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License version 2 for more details.  A copy is
14  * included in the COPYING file that accompanied this code.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write to the Free Software
18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/lod/lod_qos.c
33  *
34  * Implementation of different allocation algorithm used
35  * to distribute objects and data among OSTs.
36  */
37
38 #define DEBUG_SUBSYSTEM S_LOV
39
40 #include <asm/div64.h>
41 #include <libcfs/libcfs.h>
42 #include <uapi/linux/lustre/lustre_idl.h>
43 #include <lustre_swab.h>
44 #include <obd_class.h>
45
46 #include "lod_internal.h"
47
48 /*
49  * force QoS policy (not RR) to be used for testing purposes
50  */
51 #define FORCE_QOS_
52
53 #define D_QOS   D_OTHER
54
55 #define QOS_DEBUG(fmt, ...)     CDEBUG(D_QOS, fmt, ## __VA_ARGS__)
56 #define QOS_CONSOLE(fmt, ...)   LCONSOLE(D_QOS, fmt, ## __VA_ARGS__)
57
58 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
59                        OST_TGT(lod,i)->ltd_statfs.os_bsize)
60
61 /**
62  * Add a new target to Quality of Service (QoS) target table.
63  *
64  * Add a new OST target to the structure representing an OSS. Resort the list
65  * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
66  * protected internally and no external locking is required.
67  *
68  * \param[in] lod               LOD device
69  * \param[in] ost_desc          OST description
70  *
71  * \retval 0                    on success
72  * \retval -ENOMEM              on error
73  */
74 int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
75 {
76         struct lod_qos_oss *oss = NULL, *temposs;
77         struct obd_export  *exp = ost_desc->ltd_exp;
78         int                 rc = 0, found = 0;
79         struct list_head   *list;
80         __u32 id = 0;
81         ENTRY;
82
83         down_write(&lod->lod_qos.lq_rw_sem);
84         /*
85          * a bit hacky approach to learn NID of corresponding connection
86          * but there is no official API to access information like this
87          * with OSD API.
88          */
89         list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
90                 if (obd_uuid_equals(&oss->lqo_uuid,
91                                     &exp->exp_connection->c_remote_uuid)) {
92                         found++;
93                         break;
94                 }
95                 if (oss->lqo_id > id)
96                         id = oss->lqo_id;
97         }
98
99         if (!found) {
100                 OBD_ALLOC_PTR(oss);
101                 if (!oss)
102                         GOTO(out, rc = -ENOMEM);
103                 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
104                        sizeof(oss->lqo_uuid));
105                 ++id;
106                 oss->lqo_id = id;
107         } else {
108                 /* Assume we have to move this one */
109                 list_del(&oss->lqo_oss_list);
110         }
111
112         oss->lqo_ost_count++;
113         ost_desc->ltd_qos.ltq_oss = oss;
114
115         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
116                obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
117                oss->lqo_ost_count);
118
119         /* Add sorted by # of OSTs.  Find the first entry that we're
120            bigger than... */
121         list = &lod->lod_qos.lq_oss_list;
122         list_for_each_entry(temposs, list, lqo_oss_list) {
123                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
124                         break;
125         }
126         /* ...and add before it.  If we're the first or smallest, temposs
127            points to the list head, and we add to the end. */
128         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
129
130         lod->lod_qos.lq_dirty = 1;
131         lod->lod_qos.lq_rr.lqr_dirty = 1;
132
133 out:
134         up_write(&lod->lod_qos.lq_rw_sem);
135         RETURN(rc);
136 }
137
138 /**
139  * Remove OST target from QoS table.
140  *
141  * Removes given OST target from QoS table and releases related OSS structure
142  * if no OSTs remain on the OSS.
143  *
144  * \param[in] lod               LOD device
145  * \param[in] ost_desc          OST description
146  *
147  * \retval 0                    on success
148  * \retval -ENOENT              if no OSS was found
149  */
150 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
151 {
152         struct lod_qos_oss *oss;
153         int                 rc = 0;
154         ENTRY;
155
156         down_write(&lod->lod_qos.lq_rw_sem);
157         oss = ost_desc->ltd_qos.ltq_oss;
158         if (!oss)
159                 GOTO(out, rc = -ENOENT);
160
161         oss->lqo_ost_count--;
162         if (oss->lqo_ost_count == 0) {
163                 CDEBUG(D_QOS, "removing OSS %s\n",
164                        obd_uuid2str(&oss->lqo_uuid));
165                 list_del(&oss->lqo_oss_list);
166                 ost_desc->ltd_qos.ltq_oss = NULL;
167                 OBD_FREE_PTR(oss);
168         }
169
170         lod->lod_qos.lq_dirty = 1;
171         lod->lod_qos.lq_rr.lqr_dirty = 1;
172 out:
173         up_write(&lod->lod_qos.lq_rw_sem);
174         RETURN(rc);
175 }
176
177 /**
178  * Check whether the target is available for new OST objects.
179  *
180  * Request statfs data from the given target and verify it's active and not
181  * read-only. If so, then it can be used to place new OST objects. This
182  * function also maintains the number of active/inactive targets and sets
183  * dirty flags if those numbers change so others can run re-balance procedures.
184  * No external locking is required.
185  *
186  * \param[in] env       execution environment for this thread
187  * \param[in] d         LOD device
188  * \param[in] index     index of OST target to check
189  * \param[out] sfs      buffer for statfs data
190  *
191  * \retval 0            if the target is good
192  * \retval negative     negated errno on error
193
194  */
195 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
196                                 int index, struct obd_statfs *sfs)
197 {
198         struct lod_tgt_desc *ost;
199         int                  rc;
200         ENTRY;
201
202         LASSERT(d);
203         ost = OST_TGT(d,index);
204         LASSERT(ost);
205
206         rc = dt_statfs(env, ost->ltd_ost, sfs);
207
208         if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
209             (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
210                 RETURN(-ENOSPC);
211
212         if (rc && rc != -ENOTCONN)
213                 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
214
215         /* If the OST is readonly then we can't allocate objects there */
216         if (sfs->os_state & OS_STATE_READONLY)
217                 rc = -EROFS;
218
219         /* object precreation is skipped on the OST with max_create_count=0 */
220         if (sfs->os_state & OS_STATE_NOPRECREATE)
221                 rc = -ENOBUFS;
222
223         /* check whether device has changed state (active, inactive) */
224         if (rc != 0 && ost->ltd_active) {
225                 /* turned inactive? */
226                 spin_lock(&d->lod_lock);
227                 if (ost->ltd_active) {
228                         ost->ltd_active = 0;
229                         if (rc == -ENOTCONN)
230                                 ost->ltd_connecting = 1;
231
232                         LASSERT(d->lod_desc.ld_active_tgt_count > 0);
233                         d->lod_desc.ld_active_tgt_count--;
234                         d->lod_qos.lq_dirty = 1;
235                         d->lod_qos.lq_rr.lqr_dirty = 1;
236                         CDEBUG(D_CONFIG, "%s: turns inactive\n",
237                                ost->ltd_exp->exp_obd->obd_name);
238                 }
239                 spin_unlock(&d->lod_lock);
240         } else if (rc == 0 && ost->ltd_active == 0) {
241                 /* turned active? */
242                 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
243                          "active tgt count %d, ost nr %d\n",
244                          d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
245                 spin_lock(&d->lod_lock);
246                 if (ost->ltd_active == 0) {
247                         ost->ltd_active = 1;
248                         ost->ltd_connecting = 0;
249                         d->lod_desc.ld_active_tgt_count++;
250                         d->lod_qos.lq_dirty = 1;
251                         d->lod_qos.lq_rr.lqr_dirty = 1;
252                         CDEBUG(D_CONFIG, "%s: turns active\n",
253                                ost->ltd_exp->exp_obd->obd_name);
254                 }
255                 spin_unlock(&d->lod_lock);
256         }
257
258         RETURN(rc);
259 }
260
261 /**
262  * Maintain per-target statfs data.
263  *
264  * The function refreshes statfs data for all the targets every N seconds.
265  * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
266  * initially.
267  *
268  * \param[in] env       execution environment for this thread
269  * \param[in] lod       LOD device
270  */
271 void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod)
272 {
273         struct obd_device *obd = lod2obd(lod);
274         struct ost_pool *osts = &(lod->lod_pool_info);
275         time64_t max_age;
276         unsigned int i;
277         u64 avail;
278         int idx;
279         ENTRY;
280
281         max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage;
282
283         if (obd->obd_osfs_age > max_age)
284                 /* statfs data are quite recent, don't need to refresh it */
285                 RETURN_EXIT;
286
287         down_write(&lod->lod_qos.lq_rw_sem);
288
289         if (obd->obd_osfs_age > max_age)
290                 goto out;
291
292         for (i = 0; i < osts->op_count; i++) {
293                 idx = osts->op_array[i];
294                 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
295                 if (lod_statfs_and_check(env, lod, idx,
296                                          &OST_TGT(lod, idx)->ltd_statfs))
297                         continue;
298                 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
299                         /* recalculate weigths */
300                         lod->lod_qos.lq_dirty = 1;
301         }
302         obd->obd_osfs_age = ktime_get_seconds();
303
304 out:
305         up_write(&lod->lod_qos.lq_rw_sem);
306         EXIT;
307 }
308
309 /**
310  * Calculate per-OST and per-OSS penalties
311  *
312  * Re-calculate penalties when the configuration changes, active targets
313  * change and after statfs refresh (all these are reflected by lq_dirty flag).
314  * On every OST and OSS: decay the penalty by half for every 8x the update
315  * interval that the device has been idle. That gives lots of time for the
316  * statfs information to be updated (which the penalty is only a proxy for),
317  * and avoids penalizing OSS/OSTs under light load.
318  * See lod_qos_calc_weight() for how penalties are factored into the weight.
319  *
320  * \param[in] lod       LOD device
321  *
322  * \retval 0            on success
323  * \retval -EAGAIN      the number of OSTs isn't enough
324  */
325 static int lod_qos_calc_ppo(struct lod_device *lod)
326 {
327         struct lod_qos_oss *oss;
328         __u64               ba_max, ba_min, temp;
329         __u32               num_active;
330         unsigned int        i;
331         int                 rc, prio_wide;
332         time64_t            now, age;
333         ENTRY;
334
335         if (!lod->lod_qos.lq_dirty)
336                 GOTO(out, rc = 0);
337
338         num_active = lod->lod_desc.ld_active_tgt_count - 1;
339         if (num_active < 1)
340                 GOTO(out, rc = -EAGAIN);
341
342         /* find bavail on each OSS */
343         list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
344                             oss->lqo_bavail = 0;
345         lod->lod_qos.lq_active_oss_count = 0;
346
347         /*
348          * How badly user wants to select OSTs "widely" (not recently chosen
349          * and not on recent OSS's).  As opposed to "freely" (free space
350          * avail.) 0-256
351          */
352         prio_wide = 256 - lod->lod_qos.lq_prio_free;
353
354         ba_min = (__u64)(-1);
355         ba_max = 0;
356         now = ktime_get_real_seconds();
357         /* Calculate OST penalty per object
358          * (lod ref taken in lod_qos_prep_create())
359          */
360         cfs_foreach_bit(lod->lod_ost_bitmap, i) {
361                 LASSERT(OST_TGT(lod,i));
362                 temp = TGT_BAVAIL(i);
363                 if (!temp)
364                         continue;
365                 ba_min = min(temp, ba_min);
366                 ba_max = max(temp, ba_max);
367
368                 /* Count the number of usable OSS's */
369                 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
370                         lod->lod_qos.lq_active_oss_count++;
371                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
372
373                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
374                 temp >>= 1;
375                 do_div(temp, num_active);
376                 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
377                         (temp * prio_wide) >> 8;
378
379                 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
380                 if (lod->lod_qos.lq_reset ||
381                     age > 32 * lod->lod_desc.ld_qos_maxage)
382                         OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
383                 else if (age > lod->lod_desc.ld_qos_maxage)
384                         /* Decay OST penalty. */
385                         OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
386                                 (age / lod->lod_desc.ld_qos_maxage);
387         }
388
389         num_active = lod->lod_qos.lq_active_oss_count - 1;
390         if (num_active < 1) {
391                 /* If there's only 1 OSS, we can't penalize it, so instead
392                    we have to double the OST penalty */
393                 num_active = 1;
394                 cfs_foreach_bit(lod->lod_ost_bitmap, i)
395                         OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
396         }
397
398         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
399         list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
400                 temp = oss->lqo_bavail >> 1;
401                 do_div(temp, oss->lqo_ost_count * num_active);
402                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
403
404                 age = (now - oss->lqo_used) >> 3;
405                 if (lod->lod_qos.lq_reset ||
406                     age > 32 * lod->lod_desc.ld_qos_maxage)
407                         oss->lqo_penalty = 0;
408                 else if (age > lod->lod_desc.ld_qos_maxage)
409                         /* Decay OSS penalty. */
410                         oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
411         }
412
413         lod->lod_qos.lq_dirty = 0;
414         lod->lod_qos.lq_reset = 0;
415
416         /* If each ost has almost same free space,
417          * do rr allocation for better creation performance */
418         lod->lod_qos.lq_same_space = 0;
419         if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
420                 lod->lod_qos.lq_same_space = 1;
421                 /* Reset weights for the next time we enter qos mode */
422                 lod->lod_qos.lq_reset = 1;
423         }
424         rc = 0;
425
426 out:
427 #ifndef FORCE_QOS
428         if (!rc && lod->lod_qos.lq_same_space)
429                 RETURN(-EAGAIN);
430 #endif
431         RETURN(rc);
432 }
433
434 /**
435  * Calculate weight for a given OST target.
436  *
437  * The final OST weight is the number of bytes available minus the OST and
438  * OSS penalties.  See lod_qos_calc_ppo() for how penalties are calculated.
439  *
440  * \param[in] lod       LOD device, where OST targets are listed
441  * \param[in] i         OST target index
442  *
443  * \retval              0
444  */
445 static int lod_qos_calc_weight(struct lod_device *lod, int i)
446 {
447         __u64 temp, temp2;
448
449         temp = TGT_BAVAIL(i);
450         temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
451                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
452         if (temp < temp2)
453                 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
454         else
455                 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
456         return 0;
457 }
458
459 /**
460  * Re-calculate weights.
461  *
462  * The function is called when some OST target was used for a new object. In
463  * this case we should re-calculate all the weights to keep new allocations
464  * balanced well.
465  *
466  * \param[in] lod       LOD device
467  * \param[in] osts      OST pool where a new object was placed
468  * \param[in] index     OST target where a new object was placed
469  * \param[out] total_wt new total weight for the pool
470  *
471  * \retval              0
472  */
473 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
474                         __u32 index, __u64 *total_wt)
475 {
476         struct lod_tgt_desc *ost;
477         struct lod_qos_oss  *oss;
478         unsigned int j;
479         ENTRY;
480
481         ost = OST_TGT(lod,index);
482         LASSERT(ost);
483
484         /* Don't allocate on this devuce anymore, until the next alloc_qos */
485         ost->ltd_qos.ltq_usable = 0;
486
487         oss = ost->ltd_qos.ltq_oss;
488
489         /* Decay old penalty by half (we're adding max penalty, and don't
490            want it to run away.) */
491         ost->ltd_qos.ltq_penalty >>= 1;
492         oss->lqo_penalty >>= 1;
493
494         /* mark the OSS and OST as recently used */
495         ost->ltd_qos.ltq_used = oss->lqo_used = ktime_get_real_seconds();
496
497         /* Set max penalties for this OST and OSS */
498         ost->ltd_qos.ltq_penalty +=
499                 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
500         oss->lqo_penalty += oss->lqo_penalty_per_obj *
501                 lod->lod_qos.lq_active_oss_count;
502
503         /* Decrease all OSS penalties */
504         list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
505                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
506                         oss->lqo_penalty = 0;
507                 else
508                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
509         }
510
511         *total_wt = 0;
512         /* Decrease all OST penalties */
513         for (j = 0; j < osts->op_count; j++) {
514                 int i;
515
516                 i = osts->op_array[j];
517                 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
518                         continue;
519
520                 ost = OST_TGT(lod,i);
521                 LASSERT(ost);
522
523                 if (ost->ltd_qos.ltq_penalty <
524                                 ost->ltd_qos.ltq_penalty_per_obj)
525                         ost->ltd_qos.ltq_penalty = 0;
526                 else
527                         ost->ltd_qos.ltq_penalty -=
528                                 ost->ltd_qos.ltq_penalty_per_obj;
529
530                 lod_qos_calc_weight(lod, i);
531
532                 /* Recalc the total weight of usable osts */
533                 if (ost->ltd_qos.ltq_usable)
534                         *total_wt += ost->ltd_qos.ltq_weight;
535
536                 QOS_DEBUG("recalc tgt %d usable=%d avail=%llu"
537                           " ostppo=%llu ostp=%llu ossppo=%llu"
538                           " ossp=%llu wt=%llu\n",
539                           i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
540                           ost->ltd_qos.ltq_penalty_per_obj >> 10,
541                           ost->ltd_qos.ltq_penalty >> 10,
542                           ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
543                           ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
544                           ost->ltd_qos.ltq_weight >> 10);
545         }
546
547         RETURN(0);
548 }
549
550 void lod_qos_rr_init(struct lod_qos_rr *lqr)
551 {
552         spin_lock_init(&lqr->lqr_alloc);
553         lqr->lqr_dirty = 1;
554 }
555
556
557 #define LOV_QOS_EMPTY ((__u32)-1)
558
559 /**
560  * Calculate optimal round-robin order with regard to OSSes.
561  *
562  * Place all the OSTs from pool \a src_pool in a special array to be used for
563  * round-robin (RR) stripe allocation.  The placement algorithm interleaves
564  * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
565  * Resorts the targets when the number of active targets changes (because of
566  * a new target or activation/deactivation).
567  *
568  * \param[in] lod       LOD device
569  * \param[in] src_pool  OST pool
570  * \param[in] lqr       round-robin list
571  *
572  * \retval 0            on success
573  * \retval -ENOMEM      fails to allocate the array
574  */
575 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
576                            struct lod_qos_rr *lqr)
577 {
578         struct lod_qos_oss  *oss;
579         struct lod_tgt_desc *ost;
580         unsigned placed, real_count;
581         unsigned int i;
582         int rc;
583         ENTRY;
584
585         if (!lqr->lqr_dirty) {
586                 LASSERT(lqr->lqr_pool.op_size);
587                 RETURN(0);
588         }
589
590         /* Do actual allocation. */
591         down_write(&lod->lod_qos.lq_rw_sem);
592
593         /*
594          * Check again. While we were sleeping on @lq_rw_sem something could
595          * change.
596          */
597         if (!lqr->lqr_dirty) {
598                 LASSERT(lqr->lqr_pool.op_size);
599                 up_write(&lod->lod_qos.lq_rw_sem);
600                 RETURN(0);
601         }
602
603         real_count = src_pool->op_count;
604
605         /* Zero the pool array */
606         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
607            deleting from the pool. The lq_rw_sem insures that nobody else
608            is reading. */
609         lqr->lqr_pool.op_count = real_count;
610         rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
611         if (rc) {
612                 up_write(&lod->lod_qos.lq_rw_sem);
613                 RETURN(rc);
614         }
615         for (i = 0; i < lqr->lqr_pool.op_count; i++)
616                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
617
618         /* Place all the OSTs from 1 OSS at the same time. */
619         placed = 0;
620         list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
621                 int j = 0;
622
623                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
624                         int next;
625
626                         if (!cfs_bitmap_check(lod->lod_ost_bitmap,
627                                                 src_pool->op_array[i]))
628                                 continue;
629
630                         ost = OST_TGT(lod,src_pool->op_array[i]);
631                         LASSERT(ost && ost->ltd_ost);
632                         if (ost->ltd_qos.ltq_oss != oss)
633                                 continue;
634
635                         /* Evenly space these OSTs across arrayspace */
636                         next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
637                         while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
638                                 next = (next + 1) % lqr->lqr_pool.op_count;
639
640                         lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
641                         j++;
642                         placed++;
643                 }
644         }
645
646         lqr->lqr_dirty = 0;
647         up_write(&lod->lod_qos.lq_rw_sem);
648
649         if (placed != real_count) {
650                 /* This should never happen */
651                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
652                                    "round-robin list (%d of %d).\n",
653                                    placed, real_count);
654                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
655                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
656                                  lqr->lqr_pool.op_array[i]);
657                 }
658                 lqr->lqr_dirty = 1;
659                 RETURN(-EAGAIN);
660         }
661
662 #if 0
663         for (i = 0; i < lqr->lqr_pool.op_count; i++)
664                 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
665 #endif
666
667         RETURN(0);
668 }
669
670 /**
671  * Instantiate and declare creation of a new object.
672  *
673  * The function instantiates LU representation for a new object on the
674  * specified device. Also it declares an intention to create that
675  * object on the storage target.
676  *
677  * Note lu_object_anon() is used which is a trick with regard to LU/OSD
678  * infrastructure - in the existing precreation framework we can't assign FID
679  * at this moment, we do this later once a transaction is started. So the
680  * special method instantiates FID-less object in the cache and later it
681  * will get a FID and proper placement in LU cache.
682  *
683  * \param[in] env       execution environment for this thread
684  * \param[in] d         LOD device
685  * \param[in] ost_idx   OST target index where the object is being created
686  * \param[in] th        transaction handle
687  *
688  * \retval              object ptr on success, ERR_PTR() otherwise
689  */
690 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
691                                                    struct lod_device *d,
692                                                    __u32 ost_idx,
693                                                    struct thandle *th)
694 {
695         struct lod_tgt_desc *ost;
696         struct lu_object *o, *n;
697         struct lu_device *nd;
698         struct dt_object *dt;
699         int               rc;
700         ENTRY;
701
702         LASSERT(d);
703         LASSERT(ost_idx < d->lod_osts_size);
704         ost = OST_TGT(d,ost_idx);
705         LASSERT(ost);
706         LASSERT(ost->ltd_ost);
707
708         nd = &ost->ltd_ost->dd_lu_dev;
709
710         /*
711          * allocate anonymous object with zero fid, real fid
712          * will be assigned by OSP within transaction
713          * XXX: to be fixed with fully-functional OST fids
714          */
715         o = lu_object_anon(env, nd, NULL);
716         if (IS_ERR(o))
717                 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
718
719         n = lu_object_locate(o->lo_header, nd->ld_type);
720         if (unlikely(n == NULL)) {
721                 CERROR("can't find slice\n");
722                 lu_object_put(env, o);
723                 GOTO(out, dt = ERR_PTR(-EINVAL));
724         }
725
726         dt = container_of(n, struct dt_object, do_lu);
727
728         rc = lod_sub_declare_create(env, dt, NULL, NULL, NULL, th);
729         if (rc < 0) {
730                 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
731                        ost_idx, rc);
732                 lu_object_put(env, o);
733                 dt = ERR_PTR(rc);
734         }
735
736 out:
737         RETURN(dt);
738 }
739
740 /**
741  * Calculate a minimum acceptable stripe count.
742  *
743  * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
744  * all stripes or 3/4 of stripes.
745  *
746  * \param[in] stripe_count      number of stripes requested
747  * \param[in] flags             0 or LOV_USES_DEFAULT_STRIPE
748  *
749  * \retval                      acceptable stripecount
750  */
751 static int min_stripe_count(__u32 stripe_count, int flags)
752 {
753         return (flags & LOV_USES_DEFAULT_STRIPE ?
754                 stripe_count - (stripe_count / 4) : stripe_count);
755 }
756
757 #define LOV_CREATE_RESEED_MULT 30
758 #define LOV_CREATE_RESEED_MIN  2000
759
760 /**
761  * Initialize temporary OST-in-use array.
762  *
763  * Allocate or extend the array used to mark targets already assigned to a new
764  * striping so they are not used more than once.
765  *
766  * \param[in] env       execution environment for this thread
767  * \param[in] stripes   number of items needed in the array
768  *
769  * \retval 0            on success
770  * \retval -ENOMEM      on error
771  */
772 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
773                                            __u32 stripes)
774 {
775         struct lod_thread_info *info = lod_env_info(env);
776
777         if (info->lti_ea_store_size < sizeof(int) * stripes)
778                 lod_ea_store_resize(info, stripes * sizeof(int));
779         if (info->lti_ea_store_size < sizeof(int) * stripes) {
780                 CERROR("can't allocate memory for ost-in-use array\n");
781                 return -ENOMEM;
782         }
783         memset(info->lti_ea_store, -1, sizeof(int) * stripes);
784         return 0;
785 }
786
787 /**
788  * Remember a target in the array of used targets.
789  *
790  * Mark the given target as used for a new striping being created. The status
791  * of an OST in a striping can be checked with lod_qos_is_ost_used().
792  *
793  * \param[in] env       execution environment for this thread
794  * \param[in] idx       index in the array
795  * \param[in] ost       OST target index to mark as used
796  */
797 static inline void lod_qos_ost_in_use(const struct lu_env *env,
798                                       int idx, int ost)
799 {
800         struct lod_thread_info *info = lod_env_info(env);
801         int *osts = info->lti_ea_store;
802
803         LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
804         osts[idx] = ost;
805 }
806
807 /**
808  * Check is OST used in a striping.
809  *
810  * Checks whether OST with the given index is marked as used in the temporary
811  * array (see lod_qos_ost_in_use()).
812  *
813  * \param[in] env       execution environment for this thread
814  * \param[in] ost       OST target index to check
815  * \param[in] stripes   the number of items used in the array already
816  *
817  * \retval 0            not used
818  * \retval 1            used
819  */
820 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
821 {
822         struct lod_thread_info *info = lod_env_info(env);
823         int *osts = info->lti_ea_store;
824         __u32 j;
825
826         for (j = 0; j < stripes; j++) {
827                 if (osts[j] == ost)
828                         return 1;
829         }
830         return 0;
831 }
832
833 /**
834  * Check is OST used in a composite layout
835  *
836  * \param[in] inuse     all inuse ost indexs
837  * \param[in] ost       OST target index to check
838  *
839  * \retval 0            not used
840  * \retval 1            used
841  */
842 static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost)
843 {
844         __u32 j;
845         LASSERT(inuse != NULL);
846
847         if (inuse->op_size == 0)
848                 return 0;
849
850         LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <= inuse->op_size);
851         for (j = 0; j < inuse->op_count; j++) {
852                 if (inuse->op_array[j] == ost)
853                         return 1;
854         }
855         return 0;
856 }
857
858 /**
859  * Mark the given target as used for a composite layout
860  *
861  * \param[in] inuse     inuse ost index array
862  * \param[in] idx       index in the array
863  */
864 static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost)
865 {
866         LASSERT(inuse != NULL);
867         if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) {
868                 LASSERTF(inuse->op_count * sizeof(inuse->op_array[0]) <
869                          inuse->op_size,
870                          "count %d size %u", inuse->op_count, inuse->op_size);
871                 inuse->op_array[inuse->op_count] = ost;
872                 inuse->op_count++;
873         }
874 }
875
876 static inline void lod_avoid_update(struct lod_object *lo,
877                                     struct lod_avoid_guide *lag)
878 {
879         if (!lod_is_flr(lo))
880                 return;
881
882         lag->lag_ost_avail--;
883 }
884
885 static inline bool lod_should_avoid_ost(struct lod_object *lo,
886                                         struct lod_avoid_guide *lag,
887                                         __u32 index)
888 {
889         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
890         struct lod_tgt_desc *ost = OST_TGT(lod, index);
891         struct lod_qos_oss *lqo = ost->ltd_qos.ltq_oss;
892         bool used = false;
893         int i;
894
895         if (!cfs_bitmap_check(lod->lod_ost_bitmap, index))
896                 return true;
897
898         /**
899          * we've tried our best, all available OSTs have been used in
900          * overlapped components in the other mirror
901          */
902         if (lag->lag_ost_avail == 0)
903                 return false;
904
905         /* check OSS use */
906         for (i = 0; i < lag->lag_oaa_count; i++) {
907                 if (lag->lag_oss_avoid_array[i] == lqo->lqo_id) {
908                         used = true;
909                         break;
910                 }
911         }
912         /**
913          * if the OSS which OST[index] resides has not been used, we'd like to
914          * use it
915          */
916         if (!used)
917                 return false;
918
919         /* if the OSS has been used, check whether the OST has been used */
920         if (!cfs_bitmap_check(lag->lag_ost_avoid_bitmap, index))
921                 used = false;
922         else
923                 QOS_DEBUG("OST%d: has been used in overlapped component "
924                           "in other mirror\n", index);
925         return used;
926 }
927
928 static int lod_check_and_reserve_ost(const struct lu_env *env,
929                                      struct lod_object *lo,
930                                      struct obd_statfs *sfs, __u32 ost_idx,
931                                      __u32 speed, __u32 *s_idx,
932                                      struct dt_object **stripe,
933                                      __u32 *ost_indices,
934                                      struct thandle *th,
935                                      struct ost_pool *inuse)
936 {
937         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
938         struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
939         struct dt_object   *o;
940         __u32 stripe_idx = *s_idx;
941         int rc;
942
943         rc = lod_statfs_and_check(env, lod, ost_idx, sfs);
944         if (rc) {
945                 /* this OSP doesn't feel well */
946                 goto out_return;
947         }
948
949         /*
950          * We expect number of precreated objects in f_ffree at
951          * the first iteration, skip OSPs with no objects ready
952          */
953         if (sfs->os_fprecreated == 0 && speed == 0) {
954                 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
955                 goto out_return;
956         }
957
958         /*
959          * try to use another OSP if this one is degraded
960          */
961         if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
962                 QOS_DEBUG("#%d: degraded\n", ost_idx);
963                 goto out_return;
964         }
965
966         /*
967          * try not allocate on OST which has been used by other
968          * component
969          */
970         if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
971                 QOS_DEBUG("#%d: used by other component\n", ost_idx);
972                 goto out_return;
973         }
974
975         /**
976          * try not allocate OSTs used by conflicting component of other mirrors
977          * for the first and second time.
978          */
979         if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) {
980                 QOS_DEBUG("#%d: used by overlapped component of other mirror\n",
981                           ost_idx);
982                 goto out_return;
983         }
984         /*
985          * do not put >1 objects on a single OST
986          */
987         if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
988                 goto out_return;
989
990         o = lod_qos_declare_object_on(env, lod, ost_idx, th);
991         if (IS_ERR(o)) {
992                 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
993                        ost_idx, (int) PTR_ERR(o));
994                 rc = PTR_ERR(o);
995                 goto out_return;
996         }
997
998         /*
999          * We've successfully declared (reserved) an object
1000          */
1001         lod_avoid_update(lo, lag);
1002         lod_qos_ost_in_use(env, stripe_idx, ost_idx);
1003         lod_comp_ost_in_use(inuse, ost_idx);
1004         stripe[stripe_idx] = o;
1005         ost_indices[stripe_idx] = ost_idx;
1006         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
1007         stripe_idx++;
1008         *s_idx = stripe_idx;
1009
1010 out_return:
1011         return rc;
1012 }
1013
1014 /**
1015  * Allocate a striping using round-robin algorithm.
1016  *
1017  * Allocates a new striping using round-robin algorithm. The function refreshes
1018  * all the internal structures (statfs cache, array of available OSTs sorted
1019  * with regard to OSS, etc). The number of stripes required is taken from the
1020  * object (must be prepared by the caller), but can change if the flag
1021  * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
1022  * is trying to create a striping on the object in parallel. All the internal
1023  * structures (like pools, etc) are protected and no additional locking is
1024  * required. The function succeeds even if a single stripe is allocated. To save
1025  * time we give priority to targets which already have objects precreated.
1026  * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
1027  *
1028  * \param[in] env               execution environment for this thread
1029  * \param[in] lo                LOD object
1030  * \param[out] stripe           striping created
1031  * \param[out] ost_indices      ost indices of striping created
1032  * \param[in] flags             allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
1033  * \param[in] th                transaction handle
1034  * \param[in] comp_idx          index of ldo_comp_entries
1035  * \param[in|out] inuse         array of inuse ost index
1036  *
1037  * \retval 0            on success
1038  * \retval -ENOSPC      if not enough OSTs are found
1039  * \retval negative     negated errno for other failures
1040  */
1041 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
1042                         struct dt_object **stripe, __u32 *ost_indices,
1043                         int flags, struct thandle *th, int comp_idx,
1044                         struct ost_pool *inuse)
1045 {
1046         struct lod_layout_component *lod_comp;
1047         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1048         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1049         struct pool_desc  *pool = NULL;
1050         struct ost_pool   *osts;
1051         struct lod_qos_rr *lqr;
1052         unsigned int    i, array_idx;
1053         __u32 ost_start_idx_temp;
1054         __u32 stripe_idx = 0;
1055         __u32 stripe_count, stripe_count_min, ost_idx;
1056         int rc, speed = 0, ost_connecting = 0;
1057         ENTRY;
1058
1059         LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1060         lod_comp = &lo->ldo_comp_entries[comp_idx];
1061         stripe_count = lod_comp->llc_stripe_count;
1062         stripe_count_min = min_stripe_count(stripe_count, flags);
1063
1064         if (lod_comp->llc_pool != NULL)
1065                 pool = lod_find_pool(m, lod_comp->llc_pool);
1066
1067         if (pool != NULL) {
1068                 down_read(&pool_tgt_rw_sem(pool));
1069                 osts = &(pool->pool_obds);
1070                 lqr = &(pool->pool_rr);
1071         } else {
1072                 osts = &(m->lod_pool_info);
1073                 lqr = &(m->lod_qos.lq_rr);
1074         }
1075
1076         rc = lod_qos_calc_rr(m, osts, lqr);
1077         if (rc)
1078                 GOTO(out, rc);
1079
1080         rc = lod_qos_ost_in_use_clear(env, stripe_count);
1081         if (rc)
1082                 GOTO(out, rc);
1083
1084         down_read(&m->lod_qos.lq_rw_sem);
1085         spin_lock(&lqr->lqr_alloc);
1086         if (--lqr->lqr_start_count <= 0) {
1087                 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
1088                 lqr->lqr_start_count =
1089                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
1090                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
1091         } else if (stripe_count_min >= osts->op_count ||
1092                         lqr->lqr_start_idx > osts->op_count) {
1093                 /* If we have allocated from all of the OSTs, slowly
1094                  * precess the next start if the OST/stripe count isn't
1095                  * already doing this for us. */
1096                 lqr->lqr_start_idx %= osts->op_count;
1097                 if (stripe_count > 1 && (osts->op_count % stripe_count) != 1)
1098                         ++lqr->lqr_offset_idx;
1099         }
1100         ost_start_idx_temp = lqr->lqr_start_idx;
1101
1102 repeat_find:
1103
1104         QOS_DEBUG("pool '%s' want %d start_idx %d start_count %d offset %d "
1105                   "active %d count %d\n",
1106                   lod_comp->llc_pool ? lod_comp->llc_pool : "",
1107                   stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
1108                   lqr->lqr_offset_idx, osts->op_count, osts->op_count);
1109
1110         for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
1111                 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
1112                                 osts->op_count;
1113                 ++lqr->lqr_start_idx;
1114                 ost_idx = lqr->lqr_pool.op_array[array_idx];
1115
1116                 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1117                           i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1118                           stripe_idx, array_idx, ost_idx);
1119
1120                 if ((ost_idx == LOV_QOS_EMPTY) ||
1121                     !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1122                         continue;
1123
1124                 /* Fail Check before osc_precreate() is called
1125                    so we can only 'fail' single OSC. */
1126                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1127                         continue;
1128
1129                 spin_unlock(&lqr->lqr_alloc);
1130                 rc = lod_check_and_reserve_ost(env, lo, sfs, ost_idx, speed,
1131                                                &stripe_idx, stripe, ost_indices,
1132                                                th, inuse);
1133                 spin_lock(&lqr->lqr_alloc);
1134
1135                 if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
1136                         ost_connecting = 1;
1137         }
1138         if ((speed < 2) && (stripe_idx < stripe_count_min)) {
1139                 /* Try again, allowing slower OSCs */
1140                 speed++;
1141                 lqr->lqr_start_idx = ost_start_idx_temp;
1142
1143                 ost_connecting = 0;
1144                 goto repeat_find;
1145         }
1146
1147         spin_unlock(&lqr->lqr_alloc);
1148         up_read(&m->lod_qos.lq_rw_sem);
1149
1150         if (stripe_idx) {
1151                 lod_comp->llc_stripe_count = stripe_idx;
1152                 /* at least one stripe is allocated */
1153                 rc = 0;
1154         } else {
1155                 /* nobody provided us with a single object */
1156                 if (ost_connecting)
1157                         rc = -EINPROGRESS;
1158                 else
1159                         rc = -ENOSPC;
1160         }
1161
1162 out:
1163         if (pool != NULL) {
1164                 up_read(&pool_tgt_rw_sem(pool));
1165                 /* put back ref got by lod_find_pool() */
1166                 lod_pool_putref(pool);
1167         }
1168
1169         RETURN(rc);
1170 }
1171
1172 /**
1173  * Allocate a specific striping layout on a user defined set of OSTs.
1174  *
1175  * Allocates new striping using the OST index range provided by the data from
1176  * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1177  * OSTs are not considered. The exact order of OSTs requested by the user
1178  * is respected as much as possible depending on OST status. The number of
1179  * stripes needed and stripe offset are taken from the object. If that number
1180  * can not be met, then the function returns a failure and then it's the
1181  * caller's responsibility to release the stripes allocated. All the internal
1182  * structures are protected, but no concurrent allocation is allowed on the
1183  * same objects.
1184  *
1185  * \param[in] env               execution environment for this thread
1186  * \param[in] lo                LOD object
1187  * \param[out] stripe           striping created
1188  * \param[out] ost_indices      ost indices of striping created
1189  * \param[in] th                transaction handle
1190  * \param[in] comp_idx          index of ldo_comp_entries
1191  * \param[in|out] inuse         array of inuse ost index
1192  *
1193  * \retval 0            on success
1194  * \retval -ENODEV      OST index does not exist on file system
1195  * \retval -EINVAL      requested OST index is invalid
1196  * \retval negative     negated errno on error
1197  */
1198 static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
1199                               struct dt_object **stripe, __u32 *ost_indices,
1200                               struct thandle *th, int comp_idx,
1201                               struct ost_pool *inuse)
1202 {
1203         struct lod_layout_component *lod_comp;
1204         struct lod_device       *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1205         struct obd_statfs       *sfs = &lod_env_info(env)->lti_osfs;
1206         struct dt_object        *o;
1207         unsigned int            array_idx = 0;
1208         int                     stripe_count = 0;
1209         int                     i;
1210         int                     rc = -EINVAL;
1211         ENTRY;
1212
1213         /* for specific OSTs layout */
1214         LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1215         lod_comp = &lo->ldo_comp_entries[comp_idx];
1216         LASSERT(lod_comp->llc_ostlist.op_array);
1217
1218         rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1219         if (rc < 0)
1220                 RETURN(rc);
1221
1222         for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1223                 if (lod_comp->llc_ostlist.op_array[i] ==
1224                     lod_comp->llc_stripe_offset) {
1225                         array_idx = i;
1226                         break;
1227                 }
1228         }
1229         if (i == lod_comp->llc_stripe_count) {
1230                 CDEBUG(D_OTHER,
1231                        "%s: start index %d not in the specified list of OSTs\n",
1232                        lod2obd(m)->obd_name, lod_comp->llc_stripe_offset);
1233                 RETURN(-EINVAL);
1234         }
1235
1236         for (i = 0; i < lod_comp->llc_stripe_count;
1237              i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
1238                 __u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
1239
1240                 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1241                         rc = -ENODEV;
1242                         break;
1243                 }
1244
1245                 /*
1246                  * do not put >1 objects on a single OST
1247                  */
1248                 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1249                         rc = -EINVAL;
1250                         break;
1251                 }
1252
1253                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1254                 if (rc < 0) /* this OSP doesn't feel well */
1255                         break;
1256
1257                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1258                 if (IS_ERR(o)) {
1259                         rc = PTR_ERR(o);
1260                         CDEBUG(D_OTHER,
1261                                "%s: can't declare new object on #%u: %d\n",
1262                                lod2obd(m)->obd_name, ost_idx, rc);
1263                         break;
1264                 }
1265
1266                 /*
1267                  * We've successfully declared (reserved) an object
1268                  */
1269                 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1270                 lod_comp_ost_in_use(inuse, ost_idx);
1271                 stripe[stripe_count] = o;
1272                 ost_indices[stripe_count] = ost_idx;
1273                 stripe_count++;
1274         }
1275
1276         RETURN(rc);
1277 }
1278
1279 /**
1280  * Allocate a striping on a predefined set of OSTs.
1281  *
1282  * Allocates new layout starting from OST index in lo->ldo_stripe_offset.
1283  * Full OSTs are not considered. The exact order of OSTs is not important and
1284  * varies depending on OST status. The allocation procedure prefers the targets
1285  * with precreated objects ready. The number of stripes needed and stripe
1286  * offset are taken from the object. If that number cannot be met, then the
1287  * function returns an error and then it's the caller's responsibility to
1288  * release the stripes allocated. All the internal structures are protected,
1289  * but no concurrent allocation is allowed on the same objects.
1290  *
1291  * \param[in] env               execution environment for this thread
1292  * \param[in] lo                LOD object
1293  * \param[out] stripe           striping created
1294  * \param[out] ost_indices      ost indices of striping created
1295  * \param[in] flags             not used
1296  * \param[in] th                transaction handle
1297  * \param[in] comp_idx          index of ldo_comp_entries
1298  * \param[in|out]inuse          array of inuse ost index
1299  *
1300  * \retval 0            on success
1301  * \retval -ENOSPC      if no OST objects are available at all
1302  * \retval -EFBIG       if not enough OST objects are found
1303  * \retval -EINVAL      requested offset is invalid
1304  * \retval negative     errno on failure
1305  */
1306 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1307                               struct dt_object **stripe, __u32 *ost_indices,
1308                               int flags, struct thandle *th, int comp_idx,
1309                               struct ost_pool *inuse)
1310 {
1311         struct lod_layout_component *lod_comp;
1312         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1313         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1314         struct dt_object  *o;
1315         __u32              ost_idx;
1316         unsigned int       i, array_idx, ost_count;
1317         int                rc, stripe_num = 0;
1318         int                speed = 0;
1319         struct pool_desc  *pool = NULL;
1320         struct ost_pool   *osts;
1321         ENTRY;
1322
1323         LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1324         lod_comp = &lo->ldo_comp_entries[comp_idx];
1325
1326         rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1327         if (rc)
1328                 GOTO(out, rc);
1329
1330         if (lod_comp->llc_pool != NULL)
1331                 pool = lod_find_pool(m, lod_comp->llc_pool);
1332
1333         if (pool != NULL) {
1334                 down_read(&pool_tgt_rw_sem(pool));
1335                 osts = &(pool->pool_obds);
1336         } else {
1337                 osts = &(m->lod_pool_info);
1338         }
1339
1340         ost_count = osts->op_count;
1341
1342 repeat_find:
1343         /* search loi_ost_idx in ost array */
1344         array_idx = 0;
1345         for (i = 0; i < ost_count; i++) {
1346                 if (osts->op_array[i] == lod_comp->llc_stripe_offset) {
1347                         array_idx = i;
1348                         break;
1349                 }
1350         }
1351         if (i == ost_count) {
1352                 CERROR("Start index %d not found in pool '%s'\n",
1353                        lod_comp->llc_stripe_offset,
1354                        lod_comp->llc_pool ? lod_comp->llc_pool : "");
1355                 GOTO(out, rc = -EINVAL);
1356         }
1357
1358         for (i = 0; i < ost_count;
1359                         i++, array_idx = (array_idx + 1) % ost_count) {
1360                 ost_idx = osts->op_array[array_idx];
1361
1362                 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1363                         continue;
1364
1365                 /* Fail Check before osc_precreate() is called
1366                    so we can only 'fail' single OSC. */
1367                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1368                         continue;
1369
1370                 /*
1371                  * do not put >1 objects on a single OST
1372                  */
1373                 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1374                         continue;
1375
1376                 /*
1377                  * try not allocate on the OST used by other component
1378                  */
1379                 if (speed == 0 && i != 0 &&
1380                     lod_comp_is_ost_used(inuse, ost_idx))
1381                         continue;
1382
1383                 /* Drop slow OSCs if we can, but not for requested start idx.
1384                  *
1385                  * This means "if OSC is slow and it is not the requested
1386                  * start OST, then it can be skipped, otherwise skip it only
1387                  * if it is inactive/recovering/out-of-space." */
1388
1389                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1390                 if (rc) {
1391                         /* this OSP doesn't feel well */
1392                         continue;
1393                 }
1394
1395                 /*
1396                  * We expect number of precreated objects at the first
1397                  * iteration.  Skip OSPs with no objects ready.  Don't apply
1398                  * this logic to OST specified with stripe_offset.
1399                  */
1400                 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1401                         continue;
1402
1403                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1404                 if (IS_ERR(o)) {
1405                         CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1406                                ost_idx, (int) PTR_ERR(o));
1407                         continue;
1408                 }
1409
1410                 /*
1411                  * We've successfully declared (reserved) an object
1412                  */
1413                 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1414                 lod_comp_ost_in_use(inuse, ost_idx);
1415                 stripe[stripe_num] = o;
1416                 ost_indices[stripe_num] = ost_idx;
1417                 stripe_num++;
1418
1419                 /* We have enough stripes */
1420                 if (stripe_num == lod_comp->llc_stripe_count)
1421                         GOTO(out, rc = 0);
1422         }
1423         if (speed < 2) {
1424                 /* Try again, allowing slower OSCs */
1425                 speed++;
1426                 goto repeat_find;
1427         }
1428
1429         /* If we were passed specific striping params, then a failure to
1430          * meet those requirements is an error, since we can't reallocate
1431          * that memory (it might be part of a larger array or something).
1432          */
1433         CERROR("can't lstripe objid "DFID": have %d want %u\n",
1434                PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1435                lod_comp->llc_stripe_count);
1436         rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1437 out:
1438         if (pool != NULL) {
1439                 up_read(&pool_tgt_rw_sem(pool));
1440                 /* put back ref got by lod_find_pool() */
1441                 lod_pool_putref(pool);
1442         }
1443
1444         RETURN(rc);
1445 }
1446
1447 /**
1448  * Check whether QoS allocation should be used.
1449  *
1450  * A simple helper to decide when QoS allocation should be used:
1451  * if it's just a single available target or the used space is
1452  * evenly distributed among the targets at the moment, then QoS
1453  * allocation algorithm should not be used.
1454  *
1455  * \param[in] lod       LOD device
1456  *
1457  * \retval 0            should not be used
1458  * \retval 1            should be used
1459  */
1460 static inline int lod_qos_is_usable(struct lod_device *lod)
1461 {
1462 #ifdef FORCE_QOS
1463         /* to be able to debug QoS code */
1464         return 1;
1465 #endif
1466
1467         /* Detect -EAGAIN early, before expensive lock is taken. */
1468         if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1469                 return 0;
1470
1471         if (lod->lod_desc.ld_active_tgt_count < 2)
1472                 return 0;
1473
1474         return 1;
1475 }
1476
1477 /**
1478  * Allocate a striping using an algorithm with weights.
1479  *
1480  * The function allocates OST objects to create a striping. The algorithm
1481  * used is based on weights (currently only using the free space), and it's
1482  * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1483  * configuration (# of stripes, offset, pool) is taken from the object and
1484  * is prepared by the caller.
1485  *
1486  * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1487  * be met due to too few OSTs, then allocation fails. If the flag is passed
1488  * fewer than 3/4 of the requested number of stripes can be allocated, then
1489  * allocation fails.
1490  *
1491  * No concurrent allocation is allowed on the object and this must be ensured
1492  * by the caller. All the internal structures are protected by the function.
1493  *
1494  * The algorithm has two steps: find available OSTs and calculate their
1495  * weights, then select the OSTs with their weights used as the probability.
1496  * An OST with a higher weight is proportionately more likely to be selected
1497  * than one with a lower weight.
1498  *
1499  * \param[in] env               execution environment for this thread
1500  * \param[in] lo                LOD object
1501  * \param[out] stripe           striping created
1502  * \param[out] ost_indices      ost indices of striping created
1503  * \param[in] flags             0 or LOV_USES_DEFAULT_STRIPE
1504  * \param[in] th                transaction handle
1505  * \param[in] comp_idx          index of ldo_comp_entries
1506  * \param[in|out]inuse          array of inuse ost index
1507  *
1508  * \retval 0            on success
1509  * \retval -EAGAIN      not enough OSTs are found for specified stripe count
1510  * \retval -EINVAL      requested OST index is invalid
1511  * \retval negative     errno on failure
1512  */
1513 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1514                          struct dt_object **stripe, __u32 *ost_indices,
1515                          int flags, struct thandle *th, int comp_idx,
1516                          struct ost_pool *inuse)
1517 {
1518         struct lod_layout_component *lod_comp;
1519         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1520         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1521         struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
1522         struct lod_tgt_desc *ost;
1523         struct dt_object *o;
1524         __u64 total_weight = 0;
1525         struct pool_desc *pool = NULL;
1526         struct ost_pool *osts;
1527         unsigned int i;
1528         __u32 nfound, good_osts, stripe_count, stripe_count_min;
1529         __u32 inuse_old_count = inuse->op_count;
1530         int rc = 0;
1531         ENTRY;
1532
1533         LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1534         lod_comp = &lo->ldo_comp_entries[comp_idx];
1535         stripe_count = lod_comp->llc_stripe_count;
1536         stripe_count_min = min_stripe_count(stripe_count, flags);
1537         if (stripe_count_min < 1)
1538                 RETURN(-EINVAL);
1539
1540         if (lod_comp->llc_pool != NULL)
1541                 pool = lod_find_pool(lod, lod_comp->llc_pool);
1542
1543         if (pool != NULL) {
1544                 down_read(&pool_tgt_rw_sem(pool));
1545                 osts = &(pool->pool_obds);
1546         } else {
1547                 osts = &(lod->lod_pool_info);
1548         }
1549
1550         /* Detect -EAGAIN early, before expensive lock is taken. */
1551         if (!lod_qos_is_usable(lod))
1552                 GOTO(out_nolock, rc = -EAGAIN);
1553
1554         /* Do actual allocation, use write lock here. */
1555         down_write(&lod->lod_qos.lq_rw_sem);
1556
1557         /*
1558          * Check again, while we were sleeping on @lq_rw_sem things could
1559          * change.
1560          */
1561         if (!lod_qos_is_usable(lod))
1562                 GOTO(out, rc = -EAGAIN);
1563
1564         rc = lod_qos_calc_ppo(lod);
1565         if (rc)
1566                 GOTO(out, rc);
1567
1568         rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1569         if (rc)
1570                 GOTO(out, rc);
1571
1572         good_osts = 0;
1573         /* Find all the OSTs that are valid stripe candidates */
1574         for (i = 0; i < osts->op_count; i++) {
1575                 if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i]))
1576                         continue;
1577
1578                 ost = OST_TGT(lod, osts->op_array[i]);
1579                 ost->ltd_qos.ltq_usable = 0;
1580
1581                 rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
1582                 if (rc) {
1583                         /* this OSP doesn't feel well */
1584                         continue;
1585                 }
1586
1587                 if (sfs->os_state & OS_STATE_DEGRADED)
1588                         continue;
1589
1590                 /* Fail Check before osc_precreate() is called
1591                    so we can only 'fail' single OSC. */
1592                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1593                                    osts->op_array[i] == 0)
1594                         continue;
1595
1596                 ost->ltd_qos.ltq_usable = 1;
1597                 lod_qos_calc_weight(lod, osts->op_array[i]);
1598                 total_weight += ost->ltd_qos.ltq_weight;
1599
1600                 good_osts++;
1601         }
1602
1603         QOS_DEBUG("found %d good osts\n", good_osts);
1604
1605         if (good_osts < stripe_count_min)
1606                 GOTO(out, rc = -EAGAIN);
1607
1608         /* We have enough osts */
1609         if (good_osts < stripe_count)
1610                 stripe_count = good_osts;
1611
1612         /* Find enough OSTs with weighted random allocation. */
1613         nfound = 0;
1614         while (nfound < stripe_count) {
1615                 __u64 rand, cur_weight;
1616
1617                 cur_weight = 0;
1618                 rc = -ENOSPC;
1619
1620                 if (total_weight) {
1621 #if BITS_PER_LONG == 32
1622                         rand = cfs_rand() % (unsigned)total_weight;
1623                         /* If total_weight > 32-bit, first generate the high
1624                          * 32 bits of the random number, then add in the low
1625                          * 32 bits (truncated to the upper limit, if needed) */
1626                         if (total_weight > 0xffffffffULL)
1627                                 rand = (__u64)(cfs_rand() %
1628                                         (unsigned)(total_weight >> 32)) << 32;
1629                         else
1630                                 rand = 0;
1631
1632                         if (rand == (total_weight & 0xffffffff00000000ULL))
1633                                 rand |= cfs_rand() % (unsigned)total_weight;
1634                         else
1635                                 rand |= cfs_rand();
1636
1637 #else
1638                         rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1639                                 total_weight;
1640 #endif
1641                 } else {
1642                         rand = 0;
1643                 }
1644
1645                 /* On average, this will hit larger-weighted OSTs more often.
1646                  * 0-weight OSTs will always get used last (only when rand=0) */
1647                 for (i = 0; i < osts->op_count; i++) {
1648                         __u32 idx = osts->op_array[i];
1649
1650                         if (lod_should_avoid_ost(lo, lag, idx))
1651                                 continue;
1652
1653                         ost = OST_TGT(lod, idx);
1654
1655                         if (!ost->ltd_qos.ltq_usable)
1656                                 continue;
1657
1658                         cur_weight += ost->ltd_qos.ltq_weight;
1659                         QOS_DEBUG("stripe_count=%d nfound=%d cur_weight=%llu "
1660                                   "rand=%llu total_weight=%llu\n",
1661                                   stripe_count, nfound, cur_weight, rand,
1662                                   total_weight);
1663
1664                         if (cur_weight < rand)
1665                                 continue;
1666
1667                         QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1668                         /*
1669                          * do not put >1 objects on a single OST
1670                          */
1671                         if (lod_qos_is_ost_used(env, idx, nfound) ||
1672                             lod_comp_is_ost_used(inuse, idx))
1673                                 continue;
1674
1675                         o = lod_qos_declare_object_on(env, lod, idx, th);
1676                         if (IS_ERR(o)) {
1677                                 QOS_DEBUG("can't declare object on #%u: %d\n",
1678                                           idx, (int) PTR_ERR(o));
1679                                 continue;
1680                         }
1681
1682                         lod_avoid_update(lo, lag);
1683                         lod_qos_ost_in_use(env, nfound, idx);
1684                         lod_comp_ost_in_use(inuse, idx);
1685                         stripe[nfound] = o;
1686                         ost_indices[nfound] = idx;
1687                         lod_qos_used(lod, osts, idx, &total_weight);
1688                         nfound++;
1689                         rc = 0;
1690                         break;
1691                 }
1692
1693                 if (rc) {
1694                         /* no OST found on this iteration, give up */
1695                         break;
1696                 }
1697         }
1698
1699         if (unlikely(nfound != stripe_count)) {
1700                 /*
1701                  * when the decision to use weighted algorithm was made
1702                  * we had enough appropriate OSPs, but this state can
1703                  * change anytime (no space on OST, broken connection, etc)
1704                  * so it's possible OSP won't be able to provide us with
1705                  * an object due to just changed state
1706                  */
1707                 QOS_DEBUG("%s: wanted %d objects, found only %d\n",
1708                           lod2obd(lod)->obd_name, stripe_count, nfound);
1709                 for (i = 0; i < nfound; i++) {
1710                         LASSERT(stripe[i] != NULL);
1711                         dt_object_put(env, stripe[i]);
1712                         stripe[i] = NULL;
1713                 }
1714                 LASSERTF(nfound <= inuse->op_count,
1715                          "nfound:%d, op_count:%u\n", nfound, inuse->op_count);
1716                 inuse->op_count = inuse_old_count;
1717
1718                 /* makes sense to rebalance next time */
1719                 lod->lod_qos.lq_dirty = 1;
1720                 lod->lod_qos.lq_same_space = 0;
1721
1722                 rc = -EAGAIN;
1723         }
1724
1725 out:
1726         up_write(&lod->lod_qos.lq_rw_sem);
1727
1728 out_nolock:
1729         if (pool != NULL) {
1730                 up_read(&pool_tgt_rw_sem(pool));
1731                 /* put back ref got by lod_find_pool() */
1732                 lod_pool_putref(pool);
1733         }
1734
1735         RETURN(rc);
1736 }
1737
1738 /**
1739  * Find largest stripe count the caller can use.
1740  *
1741  * Find the maximal possible stripe count not greater than \a stripe_count.
1742  * Sometimes suggested stripecount can't be reached for a number of reasons:
1743  * lack of enough active OSTs or the backend does not support EAs that large.
1744  * If the passed one is 0, then the filesystem's default one is used.
1745  *
1746  * \param[in] lod       LOD device
1747  * \param[in] lo        The lod_object
1748  * \param[in] stripe_count      count the caller would like to use
1749  *
1750  * \retval              the maximum usable stripe count
1751  */
1752 __u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
1753                            __u16 stripe_count)
1754 {
1755         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1756
1757         if (!stripe_count)
1758                 stripe_count = lod->lod_desc.ld_default_stripe_count;
1759         if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1760                 stripe_count = lod->lod_desc.ld_active_tgt_count;
1761         if (!stripe_count)
1762                 stripe_count = 1;
1763
1764         /* stripe count is based on whether OSD can handle larger EA sizes */
1765         if (lod->lod_osd_max_easize > 0) {
1766                 unsigned int easize = lod->lod_osd_max_easize;
1767                 int i;
1768
1769                 if (lo->ldo_is_composite) {
1770                         struct lod_layout_component *lod_comp;
1771                         unsigned int header_sz = sizeof(struct lov_comp_md_v1);
1772
1773                         header_sz += sizeof(struct lov_comp_md_entry_v1) *
1774                                         lo->ldo_comp_cnt;
1775                         for (i = 0; i < lo->ldo_comp_cnt; i++) {
1776                                 lod_comp = &lo->ldo_comp_entries[i];
1777                                 if (lod_comp->llc_flags & LCME_FL_INIT)
1778                                         header_sz += lov_mds_md_size(
1779                                                 lod_comp->llc_stripe_count,
1780                                                 LOV_MAGIC_V3);
1781                         }
1782                         if (easize > header_sz)
1783                                 easize -= header_sz;
1784                         else
1785                                 easize = 0;
1786                 }
1787
1788                 max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
1789         }
1790
1791         return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1792 }
1793
1794 /**
1795  * Create in-core respresentation for a fully-defined striping
1796  *
1797  * When the caller passes a fully-defined striping (i.e. everything including
1798  * OST object FIDs are defined), then we still need to instantiate LU-cache
1799  * with the objects representing the stripes defined. This function completes
1800  * that task.
1801  *
1802  * \param[in] env       execution environment for this thread
1803  * \param[in] mo        LOD object
1804  * \param[in] buf       buffer containing the striping
1805  *
1806  * \retval 0            on success
1807  * \retval negative     negated errno on error
1808  */
1809 int lod_use_defined_striping(const struct lu_env *env,
1810                              struct lod_object *mo,
1811                              const struct lu_buf *buf)
1812 {
1813         struct lod_layout_component *lod_comp;
1814         struct lov_mds_md_v1   *v1 = buf->lb_buf;
1815         struct lov_mds_md_v3   *v3 = buf->lb_buf;
1816         struct lov_comp_md_v1  *comp_v1 = NULL;
1817         struct lov_ost_data_v1 *objs;
1818         __u32   magic;
1819         __u16   comp_cnt;
1820         __u16   mirror_cnt;
1821         int     rc = 0, i;
1822         ENTRY;
1823
1824         magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED;
1825
1826         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1827             magic != LOV_MAGIC_COMP_V1)
1828                 RETURN(-EINVAL);
1829
1830         if (magic == LOV_MAGIC_COMP_V1) {
1831                 comp_v1 = buf->lb_buf;
1832                 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1833                 if (comp_cnt == 0)
1834                         RETURN(-EINVAL);
1835                 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1836                 mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1837                                         LCM_FL_FLR_MASK;
1838                 mo->ldo_is_composite = 1;
1839         } else {
1840                 mo->ldo_is_composite = 0;
1841                 comp_cnt = 1;
1842                 mirror_cnt = 0;
1843         }
1844         mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1845
1846         rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt);
1847         if (rc)
1848                 RETURN(rc);
1849
1850         for (i = 0; i < comp_cnt; i++) {
1851                 struct lu_extent *ext;
1852                 char    *pool_name;
1853                 __u32   offs;
1854
1855                 lod_comp = &mo->ldo_comp_entries[i];
1856
1857                 if (mo->ldo_is_composite) {
1858                         offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1859                         v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1860                         magic = le32_to_cpu(v1->lmm_magic);
1861
1862                         ext = &comp_v1->lcm_entries[i].lcme_extent;
1863                         lod_comp->llc_extent.e_start =
1864                                 le64_to_cpu(ext->e_start);
1865                         lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1866                         lod_comp->llc_flags =
1867                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1868                         lod_comp->llc_id =
1869                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1870                         if (lod_comp->llc_id == LCME_ID_INVAL)
1871                                 GOTO(out, rc = -EINVAL);
1872                 }
1873
1874                 pool_name = NULL;
1875                 if (magic == LOV_MAGIC_V1) {
1876                         objs = &v1->lmm_objects[0];
1877                 } else if (magic == LOV_MAGIC_V3) {
1878                         objs = &v3->lmm_objects[0];
1879                         if (v3->lmm_pool_name[0] != '\0')
1880                                 pool_name = v3->lmm_pool_name;
1881                 } else {
1882                         CDEBUG(D_LAYOUT, "Invalid magic %x\n", magic);
1883                         GOTO(out, rc = -EINVAL);
1884                 }
1885
1886                 lod_comp->llc_pattern = le32_to_cpu(v1->lmm_pattern);
1887                 lod_comp->llc_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1888                 lod_comp->llc_stripe_count = le16_to_cpu(v1->lmm_stripe_count);
1889                 lod_comp->llc_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1890                 /**
1891                  * The stripe_offset of an uninit-ed component is stored in
1892                  * the lmm_layout_gen
1893                  */
1894                 if (mo->ldo_is_composite && !lod_comp_inited(lod_comp))
1895                         lod_comp->llc_stripe_offset = lod_comp->llc_layout_gen;
1896                 lod_obj_set_pool(mo, i, pool_name);
1897
1898                 if ((!mo->ldo_is_composite || lod_comp_inited(lod_comp)) &&
1899                     !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1900                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1901                         rc = lod_initialize_objects(env, mo, objs, i);
1902                         if (rc)
1903                                 GOTO(out, rc);
1904                 }
1905         }
1906
1907         rc = lod_fill_mirrors(mo);
1908         if (rc)
1909                 GOTO(out, rc);
1910 out:
1911         if (rc)
1912                 lod_object_free_striping(env, mo);
1913
1914         RETURN(rc);
1915 }
1916
1917 /**
1918  * Parse suggested striping configuration.
1919  *
1920  * The caller gets a suggested striping configuration from a number of sources
1921  * including per-directory default and applications. Then it needs to verify
1922  * the suggested striping is valid, apply missing bits and store the resulting
1923  * configuration in the object to be used by the allocator later. Must not be
1924  * called concurrently against the same object. It's OK to provide a
1925  * fully-defined striping.
1926  *
1927  * \param[in] env       execution environment for this thread
1928  * \param[in] lo        LOD object
1929  * \param[in] buf       buffer containing the striping
1930  *
1931  * \retval 0            on success
1932  * \retval negative     negated errno on error
1933  */
1934 int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
1935                          const struct lu_buf *buf)
1936 {
1937         struct lod_layout_component *lod_comp;
1938         struct lod_device       *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1939         struct lov_desc         *desc = &d->lod_desc;
1940         struct lov_user_md_v1   *v1 = NULL;
1941         struct lov_user_md_v3   *v3 = NULL;
1942         struct lov_comp_md_v1   *comp_v1 = NULL;
1943         char    def_pool[LOV_MAXPOOLNAME + 1];
1944         __u32   magic;
1945         __u16   comp_cnt;
1946         __u16   mirror_cnt;
1947         int     i, rc;
1948         ENTRY;
1949
1950         if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1951                 RETURN(0);
1952
1953         memset(def_pool, 0, sizeof(def_pool));
1954         if (lo->ldo_comp_entries != NULL)
1955                 lod_layout_get_pool(lo->ldo_comp_entries, lo->ldo_comp_cnt,
1956                                     def_pool, sizeof(def_pool));
1957
1958         /* free default striping info */
1959         lod_free_comp_entries(lo);
1960
1961         rc = lod_verify_striping(d, lo, buf, false);
1962         if (rc)
1963                 RETURN(-EINVAL);
1964
1965         v3 = buf->lb_buf;
1966         v1 = buf->lb_buf;
1967         comp_v1 = buf->lb_buf;
1968         magic = v1->lmm_magic;
1969
1970         if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
1971                 /* try to use as fully defined striping */
1972                 rc = lod_use_defined_striping(env, lo, buf);
1973                 RETURN(rc);
1974         }
1975
1976         switch (magic) {
1977         case __swab32(LOV_USER_MAGIC_V1):
1978                 lustre_swab_lov_user_md_v1(v1);
1979                 magic = v1->lmm_magic;
1980                 /* fall through */
1981         case LOV_USER_MAGIC_V1:
1982                 break;
1983         case __swab32(LOV_USER_MAGIC_V3):
1984                 lustre_swab_lov_user_md_v3(v3);
1985                 magic = v3->lmm_magic;
1986                 /* fall through */
1987         case LOV_USER_MAGIC_V3:
1988                 break;
1989         case __swab32(LOV_USER_MAGIC_SPECIFIC):
1990                 lustre_swab_lov_user_md_v3(v3);
1991                 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1992                                                 v3->lmm_stripe_count);
1993                 magic = v3->lmm_magic;
1994                 /* fall through */
1995         case LOV_USER_MAGIC_SPECIFIC:
1996                 break;
1997         case __swab32(LOV_USER_MAGIC_COMP_V1):
1998                 lustre_swab_lov_comp_md_v1(comp_v1);
1999                 magic = comp_v1->lcm_magic;
2000                 /* fall trhough */
2001         case LOV_USER_MAGIC_COMP_V1:
2002                 break;
2003         default:
2004                 CERROR("%s: unrecognized magic %X\n",
2005                        lod2obd(d)->obd_name, magic);
2006                 RETURN(-EINVAL);
2007         }
2008
2009         lustre_print_user_md(D_OTHER, v1, "parse config");
2010
2011         if (magic == LOV_USER_MAGIC_COMP_V1) {
2012                 comp_cnt = comp_v1->lcm_entry_count;
2013                 if (comp_cnt == 0)
2014                         RETURN(-EINVAL);
2015                 mirror_cnt =  comp_v1->lcm_mirror_count + 1;
2016                 if (mirror_cnt > 1)
2017                         lo->ldo_flr_state = LCM_FL_RDONLY;
2018                 lo->ldo_is_composite = 1;
2019         } else {
2020                 comp_cnt = 1;
2021                 mirror_cnt = 0;
2022                 lo->ldo_is_composite = 0;
2023         }
2024
2025         rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
2026         if (rc)
2027                 RETURN(rc);
2028
2029         LASSERT(lo->ldo_comp_entries);
2030
2031         for (i = 0; i < comp_cnt; i++) {
2032                 struct pool_desc        *pool;
2033                 struct lu_extent        *ext;
2034                 char    *pool_name;
2035
2036                 lod_comp = &lo->ldo_comp_entries[i];
2037
2038                 if (lo->ldo_is_composite) {
2039                         v1 = (struct lov_user_md *)((char *)comp_v1 +
2040                                         comp_v1->lcm_entries[i].lcme_offset);
2041                         ext = &comp_v1->lcm_entries[i].lcme_extent;
2042                         lod_comp->llc_extent = *ext;
2043                         lod_comp->llc_flags =
2044                                 comp_v1->lcm_entries[i].lcme_flags &
2045                                         LCME_USER_FLAGS;
2046                 }
2047
2048                 pool_name = NULL;
2049                 if (v1->lmm_magic == LOV_USER_MAGIC_V3 ||
2050                     v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
2051                         int j;
2052
2053                         v3 = (struct lov_user_md_v3 *)v1;
2054                         if (v3->lmm_pool_name[0] != '\0')
2055                                 pool_name = v3->lmm_pool_name;
2056
2057                         if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
2058                                 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
2059                                         v3->lmm_stripe_offset =
2060                                                 v3->lmm_objects[0].l_ost_idx;
2061
2062                                 /* copy ost list from lmm */
2063                                 lod_comp->llc_ostlist.op_count =
2064                                         v3->lmm_stripe_count;
2065                                 lod_comp->llc_ostlist.op_size =
2066                                         v3->lmm_stripe_count * sizeof(__u32);
2067                                 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
2068                                           lod_comp->llc_ostlist.op_size);
2069                                 if (!lod_comp->llc_ostlist.op_array)
2070                                         GOTO(free_comp, rc = -ENOMEM);
2071
2072                                 for (j = 0; j < v3->lmm_stripe_count; j++)
2073                                         lod_comp->llc_ostlist.op_array[j] =
2074                                                 v3->lmm_objects[j].l_ost_idx;
2075                         }
2076                 }
2077
2078                 if (pool_name == NULL && def_pool[0] != '\0')
2079                         pool_name = def_pool;
2080
2081                 if (v1->lmm_pattern == 0)
2082                         v1->lmm_pattern = LOV_PATTERN_RAID0;
2083                 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 &&
2084                     lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) {
2085                         CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
2086                                lod2obd(d)->obd_name, v1->lmm_pattern);
2087                         GOTO(free_comp, rc = -EINVAL);
2088                 }
2089
2090                 lod_comp->llc_pattern = v1->lmm_pattern;
2091                 lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
2092                 if (v1->lmm_stripe_size)
2093                         lod_comp->llc_stripe_size = v1->lmm_stripe_size;
2094
2095                 lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
2096                 if (v1->lmm_stripe_count ||
2097                     lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT)
2098                         lod_comp->llc_stripe_count = v1->lmm_stripe_count;
2099
2100                 lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
2101                 lod_obj_set_pool(lo, i, pool_name);
2102
2103                 LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) ==
2104                              LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0));
2105
2106                 if (pool_name == NULL)
2107                         continue;
2108
2109                 /* In the function below, .hs_keycmp resolves to
2110                  * pool_hashkey_keycmp() */
2111                 /* coverity[overrun-buffer-val] */
2112                 pool = lod_find_pool(d, pool_name);
2113                 if (pool == NULL)
2114                         continue;
2115
2116                 if (lod_comp->llc_stripe_offset != LOV_OFFSET_DEFAULT) {
2117                         rc = lod_check_index_in_pool(
2118                                         lod_comp->llc_stripe_offset, pool);
2119                         if (rc < 0) {
2120                                 lod_pool_putref(pool);
2121                                 CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n",
2122                                        lod2obd(d)->obd_name,
2123                                        lod_comp->llc_stripe_offset);
2124                                 GOTO(free_comp, rc = -EINVAL);
2125                         }
2126                 }
2127
2128                 if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
2129                         lod_comp->llc_stripe_count = pool_tgt_count(pool);
2130
2131                 lod_pool_putref(pool);
2132         }
2133
2134         RETURN(0);
2135
2136 free_comp:
2137         lod_free_comp_entries(lo);
2138         RETURN(rc);
2139 }
2140
2141 /**
2142  * prepare enough OST avoidance bitmap space
2143  */
2144 int lod_prepare_avoidance(const struct lu_env *env, struct lod_object *lo)
2145 {
2146         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
2147         struct lod_tgt_descs *ltds = &lod->lod_ost_descs;
2148         struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
2149         struct cfs_bitmap *bitmap = NULL;
2150         __u32 *new_oss = NULL;
2151
2152         lag->lag_ost_avail = ltds->ltd_tgtnr;
2153
2154         /* reset OSS avoid guide array */
2155         lag->lag_oaa_count = 0;
2156         if (lag->lag_oss_avoid_array && lag->lag_oaa_size < ltds->ltd_tgtnr) {
2157                 OBD_FREE(lag->lag_oss_avoid_array,
2158                          sizeof(__u32) * lag->lag_oaa_size);
2159                 lag->lag_oss_avoid_array = NULL;
2160                 lag->lag_oaa_size = 0;
2161         }
2162
2163         /* init OST avoid guide bitmap */
2164         if (lag->lag_ost_avoid_bitmap) {
2165                 if (ltds->ltd_tgtnr <= lag->lag_ost_avoid_bitmap->size) {
2166                         CFS_RESET_BITMAP(lag->lag_ost_avoid_bitmap);
2167                 } else {
2168                         CFS_FREE_BITMAP(lag->lag_ost_avoid_bitmap);
2169                         lag->lag_ost_avoid_bitmap = NULL;
2170                 }
2171         }
2172
2173         if (!lag->lag_ost_avoid_bitmap) {
2174                 bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr);
2175                 if (!bitmap)
2176                         return -ENOMEM;
2177         }
2178
2179         if (!lag->lag_oss_avoid_array) {
2180                 /**
2181                  * usually there are multiple OSTs in one OSS, but we don't
2182                  * know the exact OSS number, so we choose a safe option,
2183                  * using OST count to allocate the array to store the OSS
2184                  * id.
2185                  */
2186                 OBD_ALLOC(new_oss, sizeof(*new_oss) * ltds->ltd_tgtnr);
2187                 if (!new_oss) {
2188                         CFS_FREE_BITMAP(bitmap);
2189                         return -ENOMEM;
2190                 }
2191         }
2192
2193         if (new_oss) {
2194                 lag->lag_oss_avoid_array = new_oss;
2195                 lag->lag_oaa_size = ltds->ltd_tgtnr;
2196         }
2197         if (bitmap)
2198                 lag->lag_ost_avoid_bitmap = bitmap;
2199
2200         return 0;
2201 }
2202
2203 /**
2204  * Collect information of used OSTs and OSSs in the overlapped components
2205  * of other mirrors
2206  */
2207 void lod_collect_avoidance(struct lod_object *lo, struct lod_avoid_guide *lag,
2208                            int comp_idx)
2209 {
2210         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
2211         struct lod_layout_component *lod_comp = &lo->ldo_comp_entries[comp_idx];
2212         struct cfs_bitmap *bitmap = lag->lag_ost_avoid_bitmap;
2213         int i, j;
2214
2215         /* iterate mirrors */
2216         for (i = 0; i < lo->ldo_mirror_count; i++) {
2217                 struct lod_layout_component *comp;
2218
2219                 /**
2220                  * skip mirror containing component[comp_idx], we only
2221                  * collect OSTs info of conflicting component in other mirrors,
2222                  * so that during read, if OSTs of a mirror's component are
2223                  * not available, we still have other mirror with different
2224                  * OSTs to read the data.
2225                  */
2226                 comp = &lo->ldo_comp_entries[lo->ldo_mirrors[i].lme_start];
2227                 if (comp->llc_id == LCME_ID_INVAL ||
2228                     mirror_id_of(comp->llc_id) ==
2229                                                 mirror_id_of(lod_comp->llc_id))
2230                         continue;
2231
2232                 /* iterate components of a mirror */
2233                 lod_foreach_mirror_comp(comp, lo, i) {
2234                         /* skip non-overlapped or un-instantiated components */
2235                         if (!lu_extent_is_overlapped(&comp->llc_extent,
2236                                                      &lod_comp->llc_extent) ||
2237                             !lod_comp_inited(comp) || !comp->llc_stripe)
2238                                 continue;
2239
2240                         /**
2241                          * collect used OSTs index and OSS info from a
2242                          * component
2243                          */
2244                         for (j = 0; j < comp->llc_stripe_count; j++) {
2245                                 struct lod_tgt_desc *ost;
2246                                 struct lod_qos_oss *lqo;
2247                                 int k;
2248
2249                                 ost = OST_TGT(lod, comp->llc_ost_indices[j]);
2250                                 lqo = ost->ltd_qos.ltq_oss;
2251
2252                                 if (cfs_bitmap_check(bitmap, ost->ltd_index))
2253                                         continue;
2254
2255                                 cfs_bitmap_set(bitmap, ost->ltd_index);
2256                                 lag->lag_ost_avail--;
2257
2258                                 for (k = 0; k < lag->lag_oaa_count; k++) {
2259                                         if (lag->lag_oss_avoid_array[k] ==
2260                                             lqo->lqo_id)
2261                                                 break;
2262                                 }
2263                                 if (k == lag->lag_oaa_count) {
2264                                         lag->lag_oss_avoid_array[k] =
2265                                                                 lqo->lqo_id;
2266                                         lag->lag_oaa_count++;
2267                                 }
2268                         }
2269                 }
2270         }
2271 }
2272
2273 /**
2274  * Create a striping for an obejct.
2275  *
2276  * The function creates a new striping for the object. The function tries QoS
2277  * algorithm first unless free space is distributed evenly among OSTs, but
2278  * by default RR algorithm is preferred due to internal concurrency (QoS is
2279  * serialized). The caller must ensure no concurrent calls to the function
2280  * are made against the same object.
2281  *
2282  * \param[in] env       execution environment for this thread
2283  * \param[in] lo        LOD object
2284  * \param[in] attr      attributes OST objects will be declared with
2285  * \param[in] th        transaction handle
2286  * \param[in] comp_idx  index of ldo_comp_entries
2287  * \param[in|out] inuse array of inuse ost index
2288  *
2289  * \retval 0            on success
2290  * \retval negative     negated errno on error
2291  */
2292 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
2293                         struct lu_attr *attr, struct thandle *th,
2294                         int comp_idx, struct ost_pool *inuse)
2295 {
2296         struct lod_layout_component *lod_comp;
2297         struct lod_device      *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2298         int                     stripe_len;
2299         int                     flag = LOV_USES_ASSIGNED_STRIPE;
2300         int                     i, rc = 0;
2301         struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
2302         struct dt_object **stripe = NULL;
2303         __u32 *ost_indices = NULL;
2304         ENTRY;
2305
2306         LASSERT(lo);
2307         LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
2308         lod_comp = &lo->ldo_comp_entries[comp_idx];
2309
2310         /* A released component is being created */
2311         if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
2312                 RETURN(0);
2313
2314         /* A Data-on-MDT component is being created */
2315         if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT)
2316                 RETURN(0);
2317
2318         if (likely(lod_comp->llc_stripe == NULL)) {
2319                 /*
2320                  * no striping has been created so far
2321                  */
2322                 LASSERT(lod_comp->llc_stripe_count);
2323                 /*
2324                  * statfs and check OST targets now, since ld_active_tgt_count
2325                  * could be changed if some OSTs are [de]activated manually.
2326                  */
2327                 lod_qos_statfs_update(env, d);
2328                 stripe_len = lod_get_stripe_count(d, lo,
2329                                                   lod_comp->llc_stripe_count);
2330                 if (stripe_len == 0)
2331                         GOTO(out, rc = -ERANGE);
2332                 lod_comp->llc_stripe_count = stripe_len;
2333                 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
2334                 if (stripe == NULL)
2335                         GOTO(out, rc = -ENOMEM);
2336                 OBD_ALLOC(ost_indices, sizeof(*ost_indices) * stripe_len);
2337                 if (!ost_indices)
2338                         GOTO(out, rc = -ENOMEM);
2339
2340                 lod_getref(&d->lod_ost_descs);
2341                 /* XXX: support for non-0 files w/o objects */
2342                 CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
2343                                 d->lod_desc.ld_tgt_count, stripe_len);
2344
2345                 if (lod_comp->llc_ostlist.op_array) {
2346                         rc = lod_alloc_ost_list(env, lo, stripe, ost_indices,
2347                                                 th, comp_idx, inuse);
2348                 } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
2349                         /**
2350                          * collect OSTs and OSSs used in other mirrors whose
2351                          * components cross the ldo_comp_entries[comp_idx]
2352                          */
2353                         rc = lod_prepare_avoidance(env, lo);
2354                         if (rc)
2355                                 GOTO(put_ldts, rc);
2356
2357                         lod_collect_avoidance(lo, lag, comp_idx);
2358
2359                         rc = lod_alloc_qos(env, lo, stripe, ost_indices, flag,
2360                                            th, comp_idx, inuse);
2361                         if (rc == -EAGAIN)
2362                                 rc = lod_alloc_rr(env, lo, stripe, ost_indices,
2363                                                   flag, th, comp_idx, inuse);
2364                 } else {
2365                         rc = lod_alloc_specific(env, lo, stripe, ost_indices,
2366                                                 flag, th, comp_idx, inuse);
2367                 }
2368 put_ldts:
2369                 lod_putref(d, &d->lod_ost_descs);
2370                 if (rc < 0) {
2371                         for (i = 0; i < stripe_len; i++)
2372                                 if (stripe[i] != NULL)
2373                                         dt_object_put(env, stripe[i]);
2374                         lod_comp->llc_stripe_count = 0;
2375                 } else {
2376                         lod_comp->llc_stripe = stripe;
2377                         lod_comp->llc_ost_indices = ost_indices;
2378                         lod_comp->llc_stripes_allocated = stripe_len;
2379                 }
2380         } else {
2381                 /*
2382                  * lod_qos_parse_config() found supplied buf as a predefined
2383                  * striping (not a hint), so it allocated all the object
2384                  * now we need to create them
2385                  */
2386                 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
2387                         struct dt_object  *o;
2388
2389                         o = lod_comp->llc_stripe[i];
2390                         LASSERT(o);
2391
2392                         rc = lod_sub_declare_create(env, o, attr, NULL,
2393                                                     NULL, th);
2394                         if (rc < 0) {
2395                                 CERROR("can't declare create: %d\n", rc);
2396                                 break;
2397                         }
2398                 }
2399                 /**
2400                  * Clear LCME_FL_INIT for the component so that
2401                  * lod_striping_create() can create the striping objects
2402                  * in replay.
2403                  */
2404                 lod_comp_unset_init(lod_comp);
2405         }
2406
2407 out:
2408         if (rc < 0) {
2409                 if (stripe)
2410                         OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
2411                 if (ost_indices)
2412                         OBD_FREE(ost_indices,
2413                                  sizeof(*ost_indices) * stripe_len);
2414         }
2415         RETURN(rc);
2416 }
2417
2418 int lod_obj_stripe_set_inuse_cb(const struct lu_env *env,
2419                                 struct lod_object *lo,
2420                                 struct dt_object *dt, struct thandle *th,
2421                                 int comp_idx, int stripe_idx,
2422                                 struct lod_obj_stripe_cb_data *data)
2423 {
2424         struct lod_thread_info  *info = lod_env_info(env);
2425         struct lod_device       *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2426         struct lu_fid   *fid = &info->lti_fid;
2427         __u32   index;
2428         int     rc, type = LU_SEQ_RANGE_OST;
2429
2430         *fid = *lu_object_fid(&dt->do_lu);
2431         rc = lod_fld_lookup(env, d, fid, &index, &type);
2432         if (rc < 0) {
2433                 CERROR("%s: fail to locate "DFID": rc = %d\n",
2434                        lod2obd(d)->obd_name, PFID(fid), rc);
2435                 return rc;
2436         }
2437         lod_comp_ost_in_use(data->locd_inuse, index);
2438         return 0;
2439 }
2440
2441 /**
2442  * Resize per-thread ost list to hold OST target index list already used.
2443  *
2444  * \param[in,out] inuse         structure contains ost list array
2445  * \param[in] cnt               total stripe count of all components
2446  * \param[in] max               array's max size if @max > 0
2447  *
2448  * \retval 0            on success
2449  * \retval -ENOMEM      reallocation failed
2450  */
2451 static int lod_inuse_resize(struct ost_pool *inuse, __u16 cnt, __u16 max)
2452 {
2453         __u32 *array;
2454         __u32 new = cnt * sizeof(inuse->op_array[0]);
2455
2456         inuse->op_count = 0;
2457
2458         if (new <= inuse->op_size)
2459                 return 0;
2460
2461         if (max)
2462                 new = min_t(__u32, new, max);
2463
2464         OBD_ALLOC(array, new);
2465         if (!array)
2466                 return -ENOMEM;
2467
2468         if (inuse->op_array)
2469                 OBD_FREE(inuse->op_array, inuse->op_size);
2470
2471         inuse->op_array = array;
2472         inuse->op_size = new;
2473
2474         return 0;
2475 }
2476
2477 int lod_prepare_inuse(const struct lu_env *env, struct lod_object *lo)
2478 {
2479         struct lod_thread_info *info = lod_env_info(env);
2480         struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2481         struct ost_pool *inuse = &info->lti_inuse_osts;
2482         struct lod_obj_stripe_cb_data data = { { 0 } };
2483         __u32 stripe_count = 0;
2484         int i;
2485         int rc;
2486
2487         for (i = 0; i < lo->ldo_comp_cnt; i++)
2488                 stripe_count += lod_comp_entry_stripe_count(lo,
2489                                         &lo->ldo_comp_entries[i], false);
2490         rc = lod_inuse_resize(inuse, stripe_count, d->lod_osd_max_easize);
2491         if (rc)
2492                 return rc;
2493
2494         data.locd_inuse = inuse;
2495         data.locd_stripe_cb = lod_obj_stripe_set_inuse_cb;
2496         return lod_obj_for_each_stripe(env, lo, NULL, &data);
2497 }
2498
2499 int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
2500                        struct lu_attr *attr, const struct lu_buf *buf,
2501                        struct thandle *th)
2502
2503 {
2504         struct lod_thread_info *info = lod_env_info(env);
2505         struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2506         struct ost_pool *inuse = &info->lti_inuse_osts;
2507         uint64_t size = 0;
2508         int i;
2509         int rc;
2510         ENTRY;
2511
2512         LASSERT(lo);
2513
2514         /* no OST available */
2515         /* XXX: should we be waiting a bit to prevent failures during
2516          * cluster initialization? */
2517         if (d->lod_ostnr == 0)
2518                 RETURN(-EIO);
2519
2520         /*
2521          * by this time, the object's ldo_stripe_count and ldo_stripe_size
2522          * contain default value for striping: taken from the parent
2523          * or from filesystem defaults
2524          *
2525          * in case the caller is passing lovea with new striping config,
2526          * we may need to parse lovea and apply new configuration
2527          */
2528         rc = lod_qos_parse_config(env, lo, buf);
2529         if (rc)
2530                 RETURN(rc);
2531
2532         if (attr->la_valid & LA_SIZE)
2533                 size = attr->la_size;
2534
2535         /* prepare inuse */
2536         rc = lod_prepare_inuse(env, lo);
2537         if (rc)
2538                 RETURN(rc);
2539
2540         /**
2541          * prepare OST object creation for the component covering file's
2542          * size, the 1st component (including plain layout file) is always
2543          * instantiated.
2544          */
2545         for (i = 0; i < lo->ldo_comp_cnt; i++) {
2546                 struct lod_layout_component *lod_comp;
2547                 struct lu_extent *extent;
2548
2549                 lod_comp = &lo->ldo_comp_entries[i];
2550                 extent = &lod_comp->llc_extent;
2551                 CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
2552                        size, extent->e_start, extent->e_end);
2553                 if (!lo->ldo_is_composite || size >= extent->e_start) {
2554                         rc = lod_qos_prep_create(env, lo, attr, th, i, inuse);
2555                         if (rc)
2556                                 break;
2557                 }
2558         }
2559
2560         RETURN(rc);
2561 }