4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * lustre/target/tgt_grant.c
31 * This file provides code related to grant space management on Lustre Targets
32 * (OSTs and MDTs). Grant is a mechanism used by client nodes to reserve disk
33 * space on a target for the data writeback cache. The Lustre client is thus
34 * assured that enough space will be available when flushing dirty pages
35 * asynchronously. Each client node is granted an initial amount of reserved
36 * space at connect time and gets additional space back from target in bulk
39 * We actually support three different cases:
40 * - The client supports the new grant parameters (i.e. OBD_CONNECT_GRANT_PARAM)
41 * which means that all grant overhead calculation happens on the client side.
42 * The server reports at connect time the backend filesystem block size, the
43 * maximum extent size as well as the extent insertion cost and it is then up
44 * to the osc layer to the track dirty extents and consume grant accordingly
45 * (see osc_cache.c). In each bulk write request, the client provides how much
46 * grant space was consumed for this RPC.
47 * - The client does not support OBD_CONNECT_GRANT_PARAM and always assumes a
48 * a backend file system block size of 4KB. We then have two cases:
49 * - If the block size is really 4KB, then the client can deal with grant
50 * allocation for partial block writes, but won't take extent insertion cost
51 * into account. For such clients, we inflate grant by 100% on the server
52 * side. It means that when 32MB of grant is hold by the client, 64MB of
53 * grant space is actually reserved on the server. All grant counters
54 * provided by such a client are inflated by 100%.
55 * - The backend filesystem block size is bigger than 4KB, which isn't
56 * supported by the client. In this case, we emulate a 4KB block size and
57 * consume one block size on the server for each 4KB of grant returned to
58 * client. With a 128KB blocksize, it means that 32MB dirty pages of 4KB
59 * on the client will actually consume 1GB of grant on the server.
60 * All grant counters provided by such a client are inflated by the block
63 * This file handles the core logic for:
64 * - grant allocation strategy
65 * - maintaining per-client as well as global grant space accounting
66 * - processing grant information packed in incoming requests
67 * - allocating server-side grant space for synchronous write RPCs which did not
68 * consume grant on the client side (OBD_BRW_FROM_GRANT flag not set). If not
69 * enough space is available, such RPCs fail with ENOSPC
71 * Author: Johann Lombardi <johann.lombardi@intel.com>
74 #define DEBUG_SUBSYSTEM S_FILTER
77 #include <obd_class.h>
79 #include "tgt_internal.h"
81 /* Clients typically hold 2x their max_rpcs_in_flight of grant space */
82 #define TGT_GRANT_SHRINK_LIMIT(exp) (2ULL * 8 * exp_max_brw_size(exp))
84 /* Helpers to inflate/deflate grants for clients that do not support the grant
86 static inline u64 tgt_grant_inflate(struct tg_grants_data *tgd, u64 val)
88 if (tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT)
89 /* Client does not support such large block size, grant
90 * is thus inflated. We already significantly overestimate
91 * overhead, no need to add the extent tax in this case */
92 return val << (tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT);
96 /* Companion of tgt_grant_inflate() */
97 static inline u64 tgt_grant_deflate(struct tg_grants_data *tgd, u64 val)
99 if (tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT)
100 return val >> (tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT);
104 /* Grant chunk is used as a unit for grant allocation. It should be inflated
105 * if the client does not support the grant paramaters.
106 * Check connection flag against \a data if not NULL. This is used during
107 * connection creation where exp->exp_connect_data isn't populated yet */
108 static inline u64 tgt_grant_chunk(struct obd_export *exp,
109 struct lu_target *lut,
110 struct obd_connect_data *data)
112 struct tg_grants_data *tgd = &lut->lut_tgd;
113 u64 chunk = exp_max_brw_size(exp);
116 if (exp->exp_obd->obd_self_export == exp)
117 /* Grant enough space to handle a big precreate request */
118 return OST_MAX_PRECREATE * lut->lut_dt_conf.ddp_inodespace / 2;
120 if ((data == NULL && !(exp_grant_param_supp(exp))) ||
121 (data != NULL && !OCD_HAS_FLAG(data, GRANT_PARAM)))
122 /* Try to grant enough space to send 2 full-size RPCs */
123 return tgt_grant_inflate(tgd, chunk) << 1;
125 /* Try to return enough to send two full-size RPCs
126 * = 2 * (BRW_size + #extents_in_BRW * grant_tax) */
127 tax = 1ULL << tgd->tgd_blockbits; /* block size */
128 tax *= lut->lut_dt_conf.ddp_max_extent_blks; /* max extent size */
129 tax = (chunk + tax - 1) / tax; /* #extents in a RPC */
130 tax *= lut->lut_dt_conf.ddp_extent_tax; /* extent tax for a RPC */
131 chunk = (chunk + tax) * 2; /* we said two full RPCs */
135 static int tgt_check_export_grants(struct obd_export *exp, u64 *dirty,
136 u64 *pending, u64 *granted, u64 maxsize)
138 struct tg_export_data *ted = &exp->exp_target_data;
141 if (ted->ted_grant < 0 || ted->ted_pending < 0 || ted->ted_dirty < 0)
143 CDEBUG_LIMIT(level, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
144 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
145 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
147 if (ted->ted_grant + ted->ted_pending > maxsize) {
148 CERROR("%s: cli %s/%p ted_grant(%ld) + ted_pending(%ld)"
149 " > maxsize(%llu)\n", exp->exp_obd->obd_name,
150 exp->exp_client_uuid.uuid, exp, ted->ted_grant,
151 ted->ted_pending, maxsize);
154 if (ted->ted_dirty > maxsize) {
155 CERROR("%s: cli %s/%p ted_dirty(%ld) > maxsize(%llu)\n",
156 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
157 exp, ted->ted_dirty, maxsize);
160 *granted += ted->ted_grant + ted->ted_pending;
161 *pending += ted->ted_pending;
162 *dirty += ted->ted_dirty;
167 * Perform extra sanity checks for grant accounting.
169 * This function scans the export list, sanity checks per-export grant counters
170 * and verifies accuracy of global grant accounting. If an inconsistency is
171 * found, a CERROR is printed with the function name \func that was passed as
172 * argument. LBUG is only called in case of serious counter corruption (i.e.
173 * value larger than the device size).
174 * Those sanity checks can be pretty expensive and are disabled if the OBD
175 * device has more than 100 connected exports.
177 * \param[in] obd OBD device for which grant accounting should be
179 * \param[in] func caller's function name
181 void tgt_grant_sanity_check(struct obd_device *obd, const char *func)
183 struct lu_target *lut = obd->u.obt.obt_lut;
184 struct tg_grants_data *tgd = &lut->lut_tgd;
185 struct obd_export *exp;
186 struct tg_export_data *ted;
196 if (list_empty(&obd->obd_exports))
199 /* We don't want to do this for large machines that do lots of
200 * mounts or unmounts. It burns... */
201 if (obd->obd_num_exports > 100)
204 maxsize = tgd->tgd_osfs.os_blocks << tgd->tgd_blockbits;
206 spin_lock(&obd->obd_dev_lock);
207 spin_lock(&tgd->tgd_grant_lock);
208 exp = obd->obd_self_export;
209 ted = &exp->exp_target_data;
210 CDEBUG(D_CACHE, "%s: processing self export: %ld %ld "
211 "%ld\n", obd->obd_name, ted->ted_grant,
212 ted->ted_pending, ted->ted_dirty);
213 tot_granted += ted->ted_grant + ted->ted_pending;
214 tot_pending += ted->ted_pending;
215 tot_dirty += ted->ted_dirty;
217 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
218 error = tgt_check_export_grants(exp, &tot_dirty, &tot_pending,
219 &tot_granted, maxsize);
221 spin_unlock(&obd->obd_dev_lock);
222 spin_unlock(&tgd->tgd_grant_lock);
227 /* exports about to be unlinked should also be taken into account since
228 * they might still hold pending grant space to be released at
230 list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain) {
231 error = tgt_check_export_grants(exp, &tot_dirty, &tot_pending,
232 &tot_granted, maxsize);
234 spin_unlock(&obd->obd_dev_lock);
235 spin_unlock(&tgd->tgd_grant_lock);
240 fo_tot_granted = tgd->tgd_tot_granted;
241 fo_tot_pending = tgd->tgd_tot_pending;
242 fo_tot_dirty = tgd->tgd_tot_dirty;
243 spin_unlock(&obd->obd_dev_lock);
244 spin_unlock(&tgd->tgd_grant_lock);
246 if (tot_granted != fo_tot_granted)
247 CERROR("%s: tot_granted %llu != fo_tot_granted %llu\n",
248 func, tot_granted, fo_tot_granted);
249 if (tot_pending != fo_tot_pending)
250 CERROR("%s: tot_pending %llu != fo_tot_pending %llu\n",
251 func, tot_pending, fo_tot_pending);
252 if (tot_dirty != fo_tot_dirty)
253 CERROR("%s: tot_dirty %llu != fo_tot_dirty %llu\n",
254 func, tot_dirty, fo_tot_dirty);
255 if (tot_pending > tot_granted)
256 CERROR("%s: tot_pending %llu > tot_granted %llu\n",
257 func, tot_pending, tot_granted);
258 if (tot_granted > maxsize)
259 CERROR("%s: tot_granted %llu > maxsize %llu\n",
260 func, tot_granted, maxsize);
261 if (tot_dirty > maxsize)
262 CERROR("%s: tot_dirty %llu > maxsize %llu\n",
263 func, tot_dirty, maxsize);
265 EXPORT_SYMBOL(tgt_grant_sanity_check);
268 * Get file system statistics of target.
270 * Helper function for statfs(), also used by grant code.
271 * Implements caching for statistics to avoid calling OSD device each time.
273 * \param[in] env execution environment
274 * \param[in] lut LU target
275 * \param[out] osfs statistic data to return
276 * \param[in] max_age maximum age for cached data
277 * \param[in] from_cache show that data was get from cache or not
279 * \retval 0 if successful
280 * \retval negative value on error
282 int tgt_statfs_internal(const struct lu_env *env, struct lu_target *lut,
283 struct obd_statfs *osfs, __u64 max_age, int *from_cache)
285 struct tg_grants_data *tgd = &lut->lut_tgd;
289 spin_lock(&tgd->tgd_osfs_lock);
290 if (cfs_time_before_64(tgd->tgd_osfs_age, max_age) || max_age == 0) {
293 /* statfs data are too old, get up-to-date one.
294 * we must be cautious here since multiple threads might be
295 * willing to update statfs data concurrently and we must
296 * grant that cached statfs data are always consistent */
298 if (tgd->tgd_statfs_inflight == 0)
299 /* clear inflight counter if no users, although it would
300 * take a while to overflow this 64-bit counter ... */
301 tgd->tgd_osfs_inflight = 0;
302 /* notify tgt_grant_commit() that we want to track writes
303 * completed as of now */
304 tgd->tgd_statfs_inflight++;
305 /* record value of inflight counter before running statfs to
306 * compute the diff once statfs is completed */
307 unstable = tgd->tgd_osfs_inflight;
308 spin_unlock(&tgd->tgd_osfs_lock);
310 /* statfs can sleep ... hopefully not for too long since we can
311 * call it fairly often as space fills up */
312 rc = dt_statfs(env, lut->lut_bottom, osfs);
316 spin_lock(&tgd->tgd_grant_lock);
317 spin_lock(&tgd->tgd_osfs_lock);
318 /* calculate how much space was written while we released the
320 unstable = tgd->tgd_osfs_inflight - unstable;
321 tgd->tgd_osfs_unstable = 0;
323 /* some writes committed while we were running statfs
324 * w/o the tgd_osfs_lock. Those ones got added to
325 * the cached statfs data that we are about to crunch.
326 * Take them into account in the new statfs data */
327 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
328 unstable >> tgd->tgd_blockbits);
329 /* However, we don't really know if those writes got
330 * accounted in the statfs call, so tell
331 * tgt_grant_space_left() there is some uncertainty
332 * on the accounting of those writes.
333 * The purpose is to prevent spurious error messages in
334 * tgt_grant_space_left() since those writes might be
335 * accounted twice. */
336 tgd->tgd_osfs_unstable += unstable;
338 /* similarly, there is some uncertainty on write requests
339 * between prepare & commit */
340 tgd->tgd_osfs_unstable += tgd->tgd_tot_pending;
341 spin_unlock(&tgd->tgd_grant_lock);
343 /* finally udpate cached statfs data */
344 tgd->tgd_osfs = *osfs;
345 tgd->tgd_osfs_age = cfs_time_current_64();
347 tgd->tgd_statfs_inflight--; /* stop tracking */
348 if (tgd->tgd_statfs_inflight == 0)
349 tgd->tgd_osfs_inflight = 0;
350 spin_unlock(&tgd->tgd_osfs_lock);
355 /* use cached statfs data */
356 *osfs = tgd->tgd_osfs;
357 spin_unlock(&tgd->tgd_osfs_lock);
366 EXPORT_SYMBOL(tgt_statfs_internal);
369 * Update cached statfs information from the OSD layer
371 * Refresh statfs information cached in tgd::tgd_osfs if the cache is older
372 * than 1s or if force is set. The OSD layer is in charge of estimating data &
374 * This function can sleep so it should not be called with any spinlock held.
376 * \param[in] env LU environment passed by the caller
377 * \param[in] exp export used to print client info in debug
379 * \param[in] force force a refresh of statfs information
380 * \param[out] from_cache returns whether the statfs information are
383 static void tgt_grant_statfs(const struct lu_env *env, struct obd_export *exp,
384 int force, int *from_cache)
386 struct obd_device *obd = exp->exp_obd;
387 struct lu_target *lut = obd->u.obt.obt_lut;
388 struct tg_grants_data *tgd = &lut->lut_tgd;
389 struct tgt_thread_info *tti;
390 struct obd_statfs *osfs;
395 max_age = 0; /* get fresh statfs data */
397 max_age = cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS);
399 tti = tgt_th_info(env);
400 osfs = &tti->tti_u.osfs;
401 rc = tgt_statfs_internal(env, lut, osfs, max_age, from_cache);
408 CDEBUG(D_CACHE, "%s: cli %s/%p free: %llu avail: %llu\n",
409 obd->obd_name, exp->exp_client_uuid.uuid, exp,
410 osfs->os_bfree << tgd->tgd_blockbits,
411 osfs->os_bavail << tgd->tgd_blockbits);
415 * Figure out how much space is available on the backend filesystem after
416 * removing grant space already booked by clients.
418 * This is done by accessing cached statfs data previously populated by
419 * tgt_grant_statfs(), from which we withdraw the space already granted to
420 * clients and the reserved space.
421 * Caller must hold tgd_grant_lock spinlock.
423 * \param[in] exp export associated with the device for which the amount
424 * of available space is requested
425 * \retval amount of non-allocated space, in bytes
427 static u64 tgt_grant_space_left(struct obd_export *exp)
429 struct obd_device *obd = exp->exp_obd;
430 struct lu_target *lut = obd->u.obt.obt_lut;
431 struct tg_grants_data *tgd = &lut->lut_tgd;
438 assert_spin_locked(&tgd->tgd_grant_lock);
440 spin_lock(&tgd->tgd_osfs_lock);
441 /* get available space from cached statfs data */
442 left = tgd->tgd_osfs.os_bavail << tgd->tgd_blockbits;
443 unstable = tgd->tgd_osfs_unstable; /* those might be accounted twice */
444 spin_unlock(&tgd->tgd_osfs_lock);
446 tot_granted = tgd->tgd_tot_granted;
448 if (left < tot_granted) {
449 int mask = (left + unstable <
450 tot_granted - tgd->tgd_tot_pending) ?
453 CDEBUG_LIMIT(mask, "%s: cli %s/%p left %llu < tot_grant "
454 "%llu unstable %llu pending %llu "
456 obd->obd_name, exp->exp_client_uuid.uuid, exp,
457 left, tot_granted, unstable,
458 tgd->tgd_tot_pending,
464 /* Withdraw space already granted to clients */
467 /* Align left on block size */
468 left &= ~((1ULL << tgd->tgd_blockbits) - 1);
470 CDEBUG(D_CACHE, "%s: cli %s/%p avail %llu left %llu unstable "
471 "%llu tot_grant %llu pending %llu\n", obd->obd_name,
472 exp->exp_client_uuid.uuid, exp, avail, left, unstable,
473 tot_granted, tgd->tgd_tot_pending);
479 * Process grant information from obdo structure packed in incoming BRW
480 * and inflate grant counters if required.
482 * Grab the dirty and seen grant announcements from the incoming obdo and
483 * inflate all grant counters passed in the request if the client does not
484 * support the grant parameters.
485 * We will later calculate the client's new grant and return it.
486 * Caller must hold tgd_grant_lock spinlock.
488 * \param[in] env LU environment supplying osfs storage
489 * \param[in] exp export for which we received the request
490 * \param[in,out] oa incoming obdo sent by the client
492 static void tgt_grant_incoming(const struct lu_env *env, struct obd_export *exp,
493 struct obdo *oa, long chunk)
495 struct tg_export_data *ted = &exp->exp_target_data;
496 struct obd_device *obd = exp->exp_obd;
497 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
502 assert_spin_locked(&tgd->tgd_grant_lock);
504 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
505 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
506 oa->o_valid &= ~OBD_MD_FLGRANT;
510 /* Add some margin, since there is a small race if other RPCs arrive
511 * out-or-order and have already consumed some grant. We want to
512 * leave this here in case there is a large error in accounting. */
514 "%s: cli %s/%p reports grant %llu dropped %u, local %lu\n",
515 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
516 oa->o_dropped, ted->ted_grant);
518 if ((long long)oa->o_dirty < 0)
521 /* inflate grant counters if required */
522 if (!exp_grant_param_supp(exp)) {
523 oa->o_grant = tgt_grant_inflate(tgd, oa->o_grant);
524 oa->o_dirty = tgt_grant_inflate(tgd, oa->o_dirty);
525 oa->o_dropped = tgt_grant_inflate(tgd, (u64)oa->o_dropped);
526 oa->o_undirty = tgt_grant_inflate(tgd, oa->o_undirty);
530 dropped = oa->o_dropped;
532 /* Update our accounting now so that statfs takes it into account.
533 * Note that ted_dirty is only approximate and can become incorrect
534 * if RPCs arrive out-of-order. No important calculations depend
535 * on ted_dirty however, but we must check sanity to not assert. */
536 if (dirty > ted->ted_grant + 4 * chunk)
537 dirty = ted->ted_grant + 4 * chunk;
538 tgd->tgd_tot_dirty += dirty - ted->ted_dirty;
539 if (ted->ted_grant < dropped) {
541 "%s: cli %s/%p reports %lu dropped > grant %lu\n",
542 obd->obd_name, exp->exp_client_uuid.uuid, exp, dropped,
546 if (tgd->tgd_tot_granted < dropped) {
547 CERROR("%s: cli %s/%p reports %lu dropped > tot_grant %llu\n",
548 obd->obd_name, exp->exp_client_uuid.uuid, exp,
549 dropped, tgd->tgd_tot_granted);
552 tgd->tgd_tot_granted -= dropped;
553 ted->ted_grant -= dropped;
554 ted->ted_dirty = dirty;
556 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
557 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
558 obd->obd_name, exp->exp_client_uuid.uuid, exp,
559 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
560 spin_unlock(&tgd->tgd_grant_lock);
567 * Grant shrink request handler.
569 * Client nodes can explicitly release grant space (i.e. process called grant
570 * shrinking). This function proceeds with the shrink request when there is
571 * less ungranted space remaining than the amount all of the connected clients
572 * would consume if they used their full grant.
573 * Caller must hold tgd_grant_lock spinlock.
575 * \param[in] exp export releasing grant space
576 * \param[in,out] oa incoming obdo sent by the client
577 * \param[in] left_space remaining free space with space already granted
580 static void tgt_grant_shrink(struct obd_export *exp, struct obdo *oa,
583 struct tg_export_data *ted = &exp->exp_target_data;
584 struct obd_device *obd = exp->exp_obd;
585 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
588 assert_spin_locked(&tgd->tgd_grant_lock);
590 if (left_space >= tgd->tgd_tot_granted_clients *
591 TGT_GRANT_SHRINK_LIMIT(exp))
594 grant_shrink = oa->o_grant;
596 ted->ted_grant -= grant_shrink;
597 tgd->tgd_tot_granted -= grant_shrink;
599 CDEBUG(D_CACHE, "%s: cli %s/%p shrink %ld ted_grant %ld total %llu\n",
600 obd->obd_name, exp->exp_client_uuid.uuid, exp, grant_shrink,
601 ted->ted_grant, tgd->tgd_tot_granted);
603 /* client has just released some grant, don't grant any space back */
608 * Calculate how much space is required to write a given network buffer
610 * This function takes block alignment into account to estimate how much on-disk
611 * space will be required to successfully write the whole niobuf.
612 * Estimated space is inflated if the export does not support
613 * OBD_CONNECT_GRANT_PARAM and if the backend filesystem has a block size
614 * larger than the minimal supported page size (i.e. 4KB).
616 * \param[in] exp export associated which the write request
617 * if NULL, then size estimate is done for server-side
619 * \param[in] lut LU target handling the request
620 * \param[in] rnb network buffer to estimate size of
622 * \retval space (in bytes) that will be consumed to write the
625 static inline u64 tgt_grant_rnb_size(struct obd_export *exp,
626 struct lu_target *lut,
627 struct niobuf_remote *rnb)
629 struct tg_grants_data *tgd = &lut->lut_tgd;
634 if (exp && !exp_grant_param_supp(exp) &&
635 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT)
636 blksize = 1ULL << COMPAT_BSIZE_SHIFT;
638 blksize = 1ULL << tgd->tgd_blockbits;
640 /* The network buffer might span several blocks, align it on block
642 bytes = rnb->rnb_offset & (blksize - 1);
643 bytes += rnb->rnb_len;
644 end = bytes & (blksize - 1);
646 bytes += blksize - end;
648 if (exp == NULL || exp_grant_param_supp(exp)) {
649 /* add per-extent insertion cost */
653 max_ext = blksize * lut->lut_dt_conf.ddp_max_extent_blks;
654 nr_ext = (bytes + max_ext - 1) / max_ext;
655 bytes += nr_ext * lut->lut_dt_conf.ddp_extent_tax;
657 /* Inflate grant space if client does not support extent-based
658 * grant allocation */
659 bytes = tgt_grant_inflate(tgd, (u64)bytes);
666 * Validate grant accounting for each incoming remote network buffer.
668 * When clients have dirtied as much space as they've been granted they
669 * fall through to sync writes. These sync writes haven't been expressed
670 * in grants and need to error with ENOSPC when there isn't room in the
671 * filesystem for them after grants are taken into account. However,
672 * writeback of the dirty data that was already granted space can write
674 * The OBD_BRW_GRANTED flag will be set in the rnb_flags of each network
675 * buffer which has been granted enough space to proceed. Buffers without
676 * this flag will fail to be written with -ENOSPC (see tgt_preprw_write().
677 * Caller must hold tgd_grant_lock spinlock.
679 * \param[in] env LU environment passed by the caller
680 * \param[in] exp export identifying the client which sent the RPC
681 * \param[in] oa incoming obdo in which we should return the pack the
683 * \param[in,out] rnb the list of network buffers
684 * \param[in] niocount the number of network buffers in the list
685 * \param[in] left the remaining free space with space already granted
688 static void tgt_grant_check(const struct lu_env *env, struct obd_export *exp,
689 struct obdo *oa, struct niobuf_remote *rnb,
690 int niocount, u64 *left)
692 struct tg_export_data *ted = &exp->exp_target_data;
693 struct obd_device *obd = exp->exp_obd;
694 struct lu_target *lut = obd->u.obt.obt_lut;
695 struct tg_grants_data *tgd = &lut->lut_tgd;
696 unsigned long ungranted = 0;
697 unsigned long granted = 0;
703 assert_spin_locked(&tgd->tgd_grant_lock);
705 if (obd->obd_recovering) {
706 /* Replaying write. Grant info have been processed already so no
707 * need to do any enforcement here. It is worth noting that only
708 * bulk writes with all rnbs having OBD_BRW_FROM_GRANT can be
709 * replayed. If one page hasn't OBD_BRW_FROM_GRANT set, then
710 * the whole bulk is written synchronously */
712 CDEBUG(D_CACHE, "Replaying write, skipping accounting\n");
713 } else if ((oa->o_valid & OBD_MD_FLFLAGS) &&
714 (oa->o_flags & OBD_FL_RECOV_RESEND)) {
715 /* Recoverable resend, grant info have already been processed as
718 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
720 } else if (exp_grant_param_supp(exp) && oa->o_grant_used > 0) {
721 /* Client supports the new grant parameters and is telling us
722 * how much grant space it consumed for this bulk write.
723 * Although all rnbs are supposed to have the OBD_BRW_FROM_GRANT
724 * flag set, we will scan the rnb list and looks for non-cache
725 * I/O in case it changes in the future */
726 if (ted->ted_grant >= oa->o_grant_used) {
727 /* skip grant accounting for rnbs with
728 * OBD_BRW_FROM_GRANT and just used grant consumption
729 * claimed in the request */
730 granted = oa->o_grant_used;
733 /* client has used more grants for this request that
735 CERROR("%s: cli %s claims %lu GRANT, real grant %lu\n",
736 exp->exp_obd->obd_name,
737 exp->exp_client_uuid.uuid,
738 (unsigned long)oa->o_grant_used, ted->ted_grant);
740 /* check whether we can fill the gap with unallocated
742 if (*left > (oa->o_grant_used - ted->ted_grant)) {
743 /* ouf .. we are safe for now */
744 granted = ted->ted_grant;
745 ungranted = oa->o_grant_used - granted;
749 /* too bad, but we cannot afford to blow up our grant
750 * accounting. The loop below will handle each rnb in
755 for (i = 0; i < niocount; i++) {
758 if ((rnb[i].rnb_flags & OBD_BRW_FROM_GRANT)) {
760 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
764 /* compute how much grant space is actually needed for
765 * this rnb, inflate grant if required */
766 bytes = tgt_grant_rnb_size(exp, lut, &rnb[i]);
767 if (ted->ted_grant >= granted + bytes) {
769 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
773 CDEBUG(D_CACHE, "%s: cli %s/%p claims %ld+%d GRANT, "
774 "real grant %lu idx %d\n", obd->obd_name,
775 exp->exp_client_uuid.uuid, exp, granted, bytes,
779 if (obd->obd_recovering)
780 CERROR("%s: cli %s is replaying OST_WRITE while one rnb"
781 " hasn't OBD_BRW_FROM_GRANT set (0x%x)\n",
782 obd->obd_name, exp->exp_client_uuid.uuid,
785 /* Consume grant space on the server.
786 * Unlike above, tgt_grant_rnb_size() is called with exp = NULL
787 * so that the required grant space isn't inflated. This is
788 * done on purpose since the server can deal with large block
789 * size, unlike some clients */
790 bytes = tgt_grant_rnb_size(NULL, lut, &rnb[i]);
792 /* if enough space, pretend it was granted */
795 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
799 /* We can't check for already-mapped blocks here (make sense
800 * when backend filesystem does not use COW) as it requires
801 * dropping the grant lock.
802 * Instead, we clear OBD_BRW_GRANTED and in that case we need
803 * to go through and verify if all of the blocks not marked
804 * BRW_GRANTED are already mapped and we can ignore this error.
806 rnb[i].rnb_flags &= ~OBD_BRW_GRANTED;
807 CDEBUG(D_CACHE, "%s: cli %s/%p idx %d no space for %d\n",
808 obd->obd_name, exp->exp_client_uuid.uuid, exp, i, bytes);
811 /* record in o_grant_used the actual space reserved for the I/O, will be
812 * used later in tgt_grant_commmit() */
813 oa->o_grant_used = granted + ungranted;
815 /* record space used for the I/O, will be used in tgt_grant_commmit() */
816 /* Now substract what the clients has used already. We don't subtract
817 * this from the tot_granted yet, so that other client's can't grab
818 * that space before we have actually allocated our blocks. That
819 * happens in tgt_grant_commit() after the writes are done. */
820 ted->ted_grant -= granted;
821 ted->ted_pending += oa->o_grant_used;
822 tgd->tgd_tot_granted += ungranted;
823 tgd->tgd_tot_pending += oa->o_grant_used;
826 "%s: cli %s/%p granted: %lu ungranted: %lu grant: %lu dirty: %lu"
827 "\n", obd->obd_name, exp->exp_client_uuid.uuid, exp,
828 granted, ungranted, ted->ted_grant, ted->ted_dirty);
830 if (obd->obd_recovering || (oa->o_valid & OBD_MD_FLGRANT) == 0)
831 /* don't update dirty accounting during recovery or
832 * if grant information got discarded (e.g. during resend) */
835 if (ted->ted_dirty < granted) {
836 CWARN("%s: cli %s/%p claims granted %lu > ted_dirty %lu\n",
837 obd->obd_name, exp->exp_client_uuid.uuid, exp,
838 granted, ted->ted_dirty);
839 granted = ted->ted_dirty;
841 tgd->tgd_tot_dirty -= granted;
842 ted->ted_dirty -= granted;
844 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
845 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
846 obd->obd_name, exp->exp_client_uuid.uuid, exp,
847 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
848 spin_unlock(&tgd->tgd_grant_lock);
855 * Allocate additional grant space to a client
857 * Calculate how much grant space to return to client, based on how much space
858 * is currently free and how much of that is already granted.
859 * Caller must hold tgd_grant_lock spinlock.
861 * \param[in] exp export of the client which sent the request
862 * \param[in] curgrant current grant claimed by the client
863 * \param[in] want how much grant space the client would like to
865 * \param[in] left remaining free space with granted space taken
867 * \param[in] conservative if set to true, the server should be cautious
868 * and limit how much space is granted back to the
869 * client. Otherwise, the server should try hard to
870 * satisfy the client request.
872 * \retval amount of grant space allocated
874 static long tgt_grant_alloc(struct obd_export *exp, u64 curgrant,
875 u64 want, u64 left, long chunk,
878 struct obd_device *obd = exp->exp_obd;
879 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
880 struct tg_export_data *ted = &exp->exp_target_data;
885 /* When tgd_grant_compat_disable is set, we don't grant any space to
886 * clients not supporting OBD_CONNECT_GRANT_PARAM.
887 * Otherwise, space granted to such a client is inflated since it
888 * consumes PAGE_SIZE of grant space per block */
889 if ((obd->obd_self_export != exp && !exp_grant_param_supp(exp) &&
890 tgd->tgd_grant_compat_disable) || left == 0 || exp->exp_failed)
893 if (want > 0x7fffffff) {
894 CERROR("%s: client %s/%p requesting > 2GB grant %llu\n",
895 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
899 /* Grant some fraction of the client's requested grant space so that
900 * they are not always waiting for write credits (not all of it to
901 * avoid overgranting in face of multiple RPCs in flight). This
902 * essentially will be able to control the OSC_MAX_RIF for a client.
904 * If we do have a large disparity between what the client thinks it
905 * has and what we think it has, don't grant very much and let the
906 * client consume its grant first. Either it just has lots of RPCs
907 * in flight, or it was evicted and its grants will soon be used up. */
908 if (curgrant >= want || curgrant >= ted->ted_grant + chunk)
911 if (obd->obd_recovering)
912 conservative = false;
915 /* don't grant more than 1/8th of the remaining free space in
918 grant = min(want - curgrant, left);
919 /* round grant up to the next block size */
920 grant = (grant + (1 << tgd->tgd_blockbits) - 1) &
921 ~((1ULL << tgd->tgd_blockbits) - 1);
926 /* Limit to grant_chunk if not reconnect/recovery */
927 if ((grant > chunk) && conservative)
930 tgd->tgd_tot_granted += grant;
931 ted->ted_grant += grant;
933 if (ted->ted_grant < 0) {
934 CERROR("%s: cli %s/%p grant %ld want %llu current %llu\n",
935 obd->obd_name, exp->exp_client_uuid.uuid, exp,
936 ted->ted_grant, want, curgrant);
937 spin_unlock(&tgd->tgd_grant_lock);
942 "%s: cli %s/%p wants: %llu current grant %llu"
943 " granting: %llu\n", obd->obd_name, exp->exp_client_uuid.uuid,
944 exp, want, curgrant, grant);
946 "%s: cli %s/%p tot cached:%llu granted:%llu"
947 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
948 exp, tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
949 obd->obd_num_exports);
955 * Handle grant space allocation on client connection & reconnection.
957 * A new non-readonly connection gets an initial grant allocation equals to
958 * tgt_grant_chunk() (i.e. twice the max BRW size in most of the cases).
959 * On reconnection, grant counters between client & target are resynchronized
960 * and additional space might be granted back if possible.
962 * \param[in] env LU environment provided by the caller
963 * \param[in] exp client's export which is (re)connecting
964 * \param[in,out] data obd_connect_data structure sent by the client in the
966 * \param[in] new_conn must set to true if this is a new connection and false
969 void tgt_grant_connect(const struct lu_env *env, struct obd_export *exp,
970 struct obd_connect_data *data, bool new_conn)
972 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
973 struct tg_grants_data *tgd = &lut->lut_tgd;
974 struct tg_export_data *ted = &exp->exp_target_data;
979 int force = 0; /* can use cached data */
981 /* don't grant space to client with read-only access */
982 if (OCD_HAS_FLAG(data, RDONLY) ||
983 (!OCD_HAS_FLAG(data, GRANT_PARAM) &&
984 tgd->tgd_grant_compat_disable)) {
986 data->ocd_connect_flags &= ~(OBD_CONNECT_GRANT |
987 OBD_CONNECT_GRANT_PARAM);
991 if (OCD_HAS_FLAG(data, GRANT_PARAM))
992 want = data->ocd_grant;
994 want = tgt_grant_inflate(tgd, data->ocd_grant);
995 chunk = tgt_grant_chunk(exp, lut, data);
997 tgt_grant_statfs(env, exp, force, &from_cache);
999 spin_lock(&tgd->tgd_grant_lock);
1001 /* Grab free space from cached info and take out space already granted
1002 * to clients as well as reserved space */
1003 left = tgt_grant_space_left(exp);
1005 /* get fresh statfs data if we are short in ungranted space */
1006 if (from_cache && left < 32 * chunk) {
1007 spin_unlock(&tgd->tgd_grant_lock);
1008 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
1013 tgt_grant_alloc(exp, (u64)ted->ted_grant, want, left, chunk, new_conn);
1015 /* return to client its current grant */
1016 if (OCD_HAS_FLAG(data, GRANT_PARAM))
1017 data->ocd_grant = ted->ted_grant;
1020 data->ocd_grant = tgt_grant_deflate(tgd, (u64)ted->ted_grant);
1022 /* reset dirty accounting */
1023 tgd->tgd_tot_dirty -= ted->ted_dirty;
1026 if (new_conn && OCD_HAS_FLAG(data, GRANT))
1027 tgd->tgd_tot_granted_clients++;
1029 spin_unlock(&tgd->tgd_grant_lock);
1031 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: %llu left: %llu\n",
1032 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1033 exp, data->ocd_grant, want, left);
1037 EXPORT_SYMBOL(tgt_grant_connect);
1040 * Release all grant space attached to a given export.
1042 * Remove a client from the grant accounting totals. We also remove
1043 * the export from the obd device under the osfs and dev locks to ensure
1044 * that the tgt_grant_sanity_check() calculations are always valid.
1045 * The client should do something similar when it invalidates its import.
1047 * \param[in] exp client's export to remove from grant accounting
1049 void tgt_grant_discard(struct obd_export *exp)
1051 struct obd_device *obd = exp->exp_obd;
1052 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1053 struct tg_export_data *ted = &exp->exp_target_data;
1055 spin_lock(&tgd->tgd_grant_lock);
1056 LASSERTF(tgd->tgd_tot_granted >= ted->ted_grant,
1057 "%s: tot_granted %llu cli %s/%p ted_grant %ld\n",
1058 obd->obd_name, tgd->tgd_tot_granted,
1059 exp->exp_client_uuid.uuid, exp, ted->ted_grant);
1060 tgd->tgd_tot_granted -= ted->ted_grant;
1062 LASSERTF(tgd->tgd_tot_pending >= ted->ted_pending,
1063 "%s: tot_pending %llu cli %s/%p ted_pending %ld\n",
1064 obd->obd_name, tgd->tgd_tot_pending,
1065 exp->exp_client_uuid.uuid, exp, ted->ted_pending);
1066 /* tgd_tot_pending is handled in tgt_grant_commit as bulk
1068 LASSERTF(tgd->tgd_tot_dirty >= ted->ted_dirty,
1069 "%s: tot_dirty %llu cli %s/%p ted_dirty %ld\n",
1070 obd->obd_name, tgd->tgd_tot_dirty,
1071 exp->exp_client_uuid.uuid, exp, ted->ted_dirty);
1072 tgd->tgd_tot_dirty -= ted->ted_dirty;
1074 spin_unlock(&tgd->tgd_grant_lock);
1076 EXPORT_SYMBOL(tgt_grant_discard);
1079 * Process grant information from incoming bulk read request.
1081 * Extract grant information packed in obdo structure (OBD_MD_FLGRANT set in
1082 * o_valid). Bulk reads usually comes with grant announcements (number of dirty
1083 * blocks, remaining amount of grant space, ...) and could also include a grant
1084 * shrink request. Unlike bulk write, no additional grant space is returned on
1085 * bulk read request.
1087 * \param[in] env is the lu environment provided by the caller
1088 * \param[in] exp is the export of the client which sent the request
1089 * \param[in,out] oa is the incoming obdo sent by the client
1091 void tgt_grant_prepare_read(const struct lu_env *env,
1092 struct obd_export *exp, struct obdo *oa)
1094 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1095 struct tg_grants_data *tgd = &lut->lut_tgd;
1104 if ((oa->o_valid & OBD_MD_FLGRANT) == 0)
1105 /* The read request does not contain any grant
1109 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1110 (oa->o_flags & OBD_FL_SHRINK_GRANT)) {
1111 /* To process grant shrink request, we need to know how much
1112 * available space remains on the backend filesystem.
1113 * Shrink requests are not so common, we always get fresh
1114 * statfs information. */
1115 tgt_grant_statfs(env, exp, 1, NULL);
1117 /* protect all grant counters */
1118 spin_lock(&tgd->tgd_grant_lock);
1120 /* Grab free space from cached statfs data and take out space
1121 * already granted to clients as well as reserved space */
1122 left = tgt_grant_space_left(exp);
1124 /* all set now to proceed with shrinking */
1127 /* no grant shrinking request packed in the obdo and
1128 * since we don't grant space back on reads, no point
1129 * in running statfs, so just skip it and process
1130 * incoming grant data directly. */
1131 spin_lock(&tgd->tgd_grant_lock);
1135 /* extract incoming grant information provided by the client and
1136 * inflate grant counters if required */
1137 tgt_grant_incoming(env, exp, oa, tgt_grant_chunk(exp, lut, NULL));
1139 /* unlike writes, we don't return grants back on reads unless a grant
1140 * shrink request was packed and we decided to turn it down. */
1142 tgt_grant_shrink(exp, oa, left);
1146 if (!exp_grant_param_supp(exp))
1147 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1148 spin_unlock(&tgd->tgd_grant_lock);
1151 EXPORT_SYMBOL(tgt_grant_prepare_read);
1154 * Process grant information from incoming bulk write request.
1156 * This function extracts client's grant announcements from incoming bulk write
1157 * request and attempts to allocate grant space for network buffers that need it
1158 * (i.e. OBD_BRW_FROM_GRANT not set in rnb_fags).
1159 * Network buffers which aren't granted the OBD_BRW_GRANTED flag should not
1160 * proceed further and should fail with -ENOSPC.
1161 * Whenever possible, additional grant space will be returned to the client
1162 * in the bulk write reply.
1163 * tgt_grant_prepare_write() must be called before writting any buffers to
1164 * the backend storage. This function works in pair with tgt_grant_commit()
1165 * which must be invoked once all buffers have been written to disk in order
1166 * to release space from the pending grant counter.
1168 * \param[in] env LU environment provided by the caller
1169 * \param[in] exp export of the client which sent the request
1170 * \param[in] oa incoming obdo sent by the client
1171 * \param[in] rnb list of network buffers
1172 * \param[in] niocount number of network buffers in the list
1174 void tgt_grant_prepare_write(const struct lu_env *env,
1175 struct obd_export *exp, struct obdo *oa,
1176 struct niobuf_remote *rnb, int niocount)
1178 struct obd_device *obd = exp->exp_obd;
1179 struct lu_target *lut = obd->u.obt.obt_lut;
1180 struct tg_grants_data *tgd = &lut->lut_tgd;
1183 int force = 0; /* can use cached data intially */
1184 long chunk = tgt_grant_chunk(exp, lut, NULL);
1189 /* get statfs information from OSD layer */
1190 tgt_grant_statfs(env, exp, force, &from_cache);
1192 spin_lock(&tgd->tgd_grant_lock); /* protect all grant counters */
1194 /* Grab free space from cached statfs data and take out space already
1195 * granted to clients as well as reserved space */
1196 left = tgt_grant_space_left(exp);
1198 /* Get fresh statfs data if we are short in ungranted space */
1199 if (from_cache && left < 32 * chunk) {
1200 spin_unlock(&tgd->tgd_grant_lock);
1201 CDEBUG(D_CACHE, "%s: fs has no space left and statfs too old\n",
1207 /* When close to free space exhaustion, trigger a sync to force
1208 * writeback cache to consume required space immediately and release as
1209 * much space as possible. */
1210 if (!obd->obd_recovering && force != 2 && left < chunk) {
1211 bool from_grant = true;
1214 /* That said, it is worth running a sync only if some pages did
1215 * not consume grant space on the client and could thus fail
1216 * with ENOSPC later in tgt_grant_check() */
1217 for (i = 0; i < niocount; i++)
1218 if (!(rnb[i].rnb_flags & OBD_BRW_FROM_GRANT))
1222 /* at least one network buffer requires acquiring grant
1223 * space on the server */
1224 spin_unlock(&tgd->tgd_grant_lock);
1225 /* discard errors, at least we tried ... */
1226 dt_sync(env, lut->lut_bottom);
1232 /* extract incoming grant information provided by the client,
1233 * and inflate grant counters if required */
1234 tgt_grant_incoming(env, exp, oa, chunk);
1237 tgt_grant_check(env, exp, oa, rnb, niocount, &left);
1239 if (!(oa->o_valid & OBD_MD_FLGRANT)) {
1240 spin_unlock(&tgd->tgd_grant_lock);
1244 /* if OBD_FL_SHRINK_GRANT is set, the client is willing to release some
1246 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1247 (oa->o_flags & OBD_FL_SHRINK_GRANT))
1248 tgt_grant_shrink(exp, oa, left);
1250 /* grant more space back to the client if possible */
1251 oa->o_grant = tgt_grant_alloc(exp, oa->o_grant, oa->o_undirty,
1254 if (!exp_grant_param_supp(exp))
1255 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1256 spin_unlock(&tgd->tgd_grant_lock);
1259 EXPORT_SYMBOL(tgt_grant_prepare_write);
1262 * Consume grant space reserved for object creation.
1264 * Grant space is allocated to the local self export for object precreation.
1265 * This is required to prevent object precreation from consuming grant space
1266 * allocated to client nodes for the data writeback cache.
1267 * This function consumes enough space to create \a nr objects and allocates
1268 * more grant space to the self export for future precreation requests, if
1271 * \param[in] env LU environment provided by the caller
1272 * \param[in] exp export holding the grant space for precreation (= self
1274 * \param[in] nr number of objects to be created
1276 * \retval >= 0 amount of grant space allocated to the precreate request
1277 * \retval -ENOSPC on failure
1279 long tgt_grant_create(const struct lu_env *env, struct obd_export *exp, s64 *nr)
1281 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1282 struct tg_grants_data *tgd = &lut->lut_tgd;
1283 struct tg_export_data *ted = &exp->exp_target_data;
1285 unsigned long wanted;
1286 unsigned long granted;
1289 if (exp->exp_obd->obd_recovering ||
1290 lut->lut_dt_conf.ddp_inodespace == 0)
1291 /* don't enforce grant during recovery */
1294 /* Update statfs data if required */
1295 tgt_grant_statfs(env, exp, 1, NULL);
1297 /* protect all grant counters */
1298 spin_lock(&tgd->tgd_grant_lock);
1300 /* fail precreate request if there is not enough blocks available for
1302 if (tgd->tgd_osfs.os_bavail - (ted->ted_grant >> tgd->tgd_blockbits) <
1303 (tgd->tgd_osfs.os_blocks >> 10)) {
1304 spin_unlock(&tgd->tgd_grant_lock);
1305 CDEBUG(D_RPCTRACE, "%s: not enough space for create %llu\n",
1306 exp->exp_obd->obd_name,
1307 tgd->tgd_osfs.os_bavail * tgd->tgd_osfs.os_blocks);
1311 /* Grab free space from cached statfs data and take out space
1312 * already granted to clients as well as reserved space */
1313 left = tgt_grant_space_left(exp);
1315 /* compute how much space is required to handle the precreation
1317 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1318 if (wanted > ted->ted_grant + left) {
1319 /* that's beyond what remains, adjust the number of objects that
1320 * can be safely precreated */
1321 wanted = ted->ted_grant + left;
1322 *nr = wanted / lut->lut_dt_conf.ddp_inodespace;
1324 /* we really have no space any more for precreation,
1325 * fail the precreate request with ENOSPC */
1326 spin_unlock(&tgd->tgd_grant_lock);
1329 /* compute space needed for the new number of creations */
1330 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1332 LASSERT(wanted <= ted->ted_grant + left);
1334 if (wanted <= ted->ted_grant) {
1335 /* we've enough grant space to handle this precreate request */
1336 ted->ted_grant -= wanted;
1338 /* we need to take some space from the ungranted pool */
1339 tgd->tgd_tot_granted += wanted - ted->ted_grant;
1340 left -= wanted - ted->ted_grant;
1344 ted->ted_pending += granted;
1345 tgd->tgd_tot_pending += granted;
1347 /* grant more space for precreate purpose if possible. */
1348 wanted = OST_MAX_PRECREATE * lut->lut_dt_conf.ddp_inodespace / 2;
1349 if (wanted > ted->ted_grant) {
1352 /* always try to book enough space to handle a large precreate
1354 chunk = tgt_grant_chunk(exp, lut, NULL);
1355 wanted -= ted->ted_grant;
1356 tgt_grant_alloc(exp, ted->ted_grant, wanted, left, chunk,
1359 spin_unlock(&tgd->tgd_grant_lock);
1362 EXPORT_SYMBOL(tgt_grant_create);
1365 * Release grant space added to the pending counter by tgt_grant_prepare_write()
1367 * Update pending grant counter once buffers have been written to the disk.
1369 * \param[in] exp export of the client which sent the request
1370 * \param[in] pending amount of reserved space to be released
1371 * \param[in] rc return code of pre-commit operations
1373 void tgt_grant_commit(struct obd_export *exp, unsigned long pending,
1376 struct tg_grants_data *tgd = &exp->exp_obd->u.obt.obt_lut->lut_tgd;
1380 /* get space accounted in tot_pending for the I/O, set in
1381 * tgt_grant_check() */
1385 spin_lock(&tgd->tgd_grant_lock);
1386 /* Don't update statfs data for errors raised before commit (e.g.
1387 * bulk transfer failed, ...) since we know those writes have not been
1388 * processed. For other errors hit during commit, we cannot really tell
1389 * whether or not something was written, so we update statfs data.
1390 * In any case, this should not be fatal since we always get fresh
1391 * statfs data before failing a request with ENOSPC */
1393 spin_lock(&tgd->tgd_osfs_lock);
1394 /* Take pending out of cached statfs data */
1395 tgd->tgd_osfs.os_bavail -= min_t(u64,
1396 tgd->tgd_osfs.os_bavail,
1397 pending >> tgd->tgd_blockbits);
1398 if (tgd->tgd_statfs_inflight)
1399 /* someone is running statfs and want to be notified of
1400 * writes happening meanwhile */
1401 tgd->tgd_osfs_inflight += pending;
1402 spin_unlock(&tgd->tgd_osfs_lock);
1405 if (exp->exp_target_data.ted_pending < pending) {
1406 CERROR("%s: cli %s/%p ted_pending(%lu) < grant_used(%lu)\n",
1407 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1408 exp->exp_target_data.ted_pending, pending);
1409 spin_unlock(&tgd->tgd_grant_lock);
1412 exp->exp_target_data.ted_pending -= pending;
1414 if (tgd->tgd_tot_granted < pending) {
1415 CERROR("%s: cli %s/%p tot_granted(%llu) < grant_used(%lu)\n",
1416 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1417 tgd->tgd_tot_granted, pending);
1418 spin_unlock(&tgd->tgd_grant_lock);
1421 tgd->tgd_tot_granted -= pending;
1423 if (tgd->tgd_tot_pending < pending) {
1424 CERROR("%s: cli %s/%p tot_pending(%llu) < grant_used(%lu)\n",
1425 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1426 tgd->tgd_tot_pending, pending);
1427 spin_unlock(&tgd->tgd_grant_lock);
1430 tgd->tgd_tot_pending -= pending;
1431 spin_unlock(&tgd->tgd_grant_lock);
1434 EXPORT_SYMBOL(tgt_grant_commit);
1436 struct tgt_grant_cb {
1437 /* commit callback structure */
1438 struct dt_txn_commit_cb tgc_cb;
1439 /* export associated with the bulk write */
1440 struct obd_export *tgc_exp;
1441 /* pending grant to be released */
1442 unsigned long tgc_granted;
1446 * Callback function for grant releasing
1448 * Release grant space reserved by the client node.
1450 * \param[in] env execution environment
1451 * \param[in] th transaction handle
1452 * \param[in] cb callback data
1453 * \param[in] err error code
1455 static void tgt_grant_commit_cb(struct lu_env *env, struct thandle *th,
1456 struct dt_txn_commit_cb *cb, int err)
1458 struct tgt_grant_cb *tgc;
1460 tgc = container_of(cb, struct tgt_grant_cb, tgc_cb);
1462 tgt_grant_commit(tgc->tgc_exp, tgc->tgc_granted, err);
1463 class_export_cb_put(tgc->tgc_exp);
1468 * Add callback for grant releasing
1470 * Register a commit callback to release grant space.
1472 * \param[in] th transaction handle
1473 * \param[in] exp OBD export of client
1474 * \param[in] granted amount of grant space to be released upon commit
1476 * \retval 0 on successful callback adding
1477 * \retval negative value on error
1479 int tgt_grant_commit_cb_add(struct thandle *th, struct obd_export *exp,
1480 unsigned long granted)
1482 struct tgt_grant_cb *tgc;
1483 struct dt_txn_commit_cb *dcb;
1491 tgc->tgc_exp = class_export_cb_get(exp);
1492 tgc->tgc_granted = granted;
1495 dcb->dcb_func = tgt_grant_commit_cb;
1496 INIT_LIST_HEAD(&dcb->dcb_linkage);
1497 strlcpy(dcb->dcb_name, "tgt_grant_commit_cb", sizeof(dcb->dcb_name));
1499 rc = dt_trans_cb_add(th, dcb);
1501 class_export_cb_put(tgc->tgc_exp);
1507 EXPORT_SYMBOL(tgt_grant_commit_cb_add);