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 (exp->exp_obd->obd_self_export == exp)
142 CDEBUG(D_CACHE, "%s: processing self export: %ld %ld "
143 "%ld\n", exp->exp_obd->obd_name, ted->ted_grant,
144 ted->ted_pending, ted->ted_dirty);
146 if (ted->ted_grant < 0 || ted->ted_pending < 0 || ted->ted_dirty < 0)
148 CDEBUG_LIMIT(level, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
149 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
150 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
152 if (ted->ted_grant + ted->ted_pending > maxsize) {
153 CERROR("%s: cli %s/%p ted_grant(%ld) + ted_pending(%ld)"
154 " > maxsize(%llu)\n", exp->exp_obd->obd_name,
155 exp->exp_client_uuid.uuid, exp, ted->ted_grant,
156 ted->ted_pending, maxsize);
159 if (ted->ted_dirty > maxsize) {
160 CERROR("%s: cli %s/%p ted_dirty(%ld) > maxsize(%llu)\n",
161 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
162 exp, ted->ted_dirty, maxsize);
165 *granted += ted->ted_grant + ted->ted_pending;
166 *pending += ted->ted_pending;
167 *dirty += ted->ted_dirty;
172 * Perform extra sanity checks for grant accounting.
174 * This function scans the export list, sanity checks per-export grant counters
175 * and verifies accuracy of global grant accounting. If an inconsistency is
176 * found, a CERROR is printed with the function name \func that was passed as
177 * argument. LBUG is only called in case of serious counter corruption (i.e.
178 * value larger than the device size).
179 * Those sanity checks can be pretty expensive and are disabled if the OBD
180 * device has more than 100 connected exports.
182 * \param[in] obd OBD device for which grant accounting should be
184 * \param[in] func caller's function name
186 void tgt_grant_sanity_check(struct obd_device *obd, const char *func)
188 struct lu_target *lut = obd->u.obt.obt_lut;
189 struct tg_grants_data *tgd = &lut->lut_tgd;
190 struct obd_export *exp;
200 if (list_empty(&obd->obd_exports))
203 /* We don't want to do this for large machines that do lots of
204 * mounts or unmounts. It burns... */
205 if (obd->obd_num_exports > 100)
208 maxsize = tgd->tgd_osfs.os_blocks << tgd->tgd_blockbits;
210 spin_lock(&obd->obd_dev_lock);
211 spin_lock(&tgd->tgd_grant_lock);
212 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
213 error = tgt_check_export_grants(exp, &tot_dirty, &tot_pending,
214 &tot_granted, maxsize);
216 spin_unlock(&obd->obd_dev_lock);
217 spin_unlock(&tgd->tgd_grant_lock);
222 /* exports about to be unlinked should also be taken into account since
223 * they might still hold pending grant space to be released at
225 list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain) {
226 error = tgt_check_export_grants(exp, &tot_dirty, &tot_pending,
227 &tot_granted, maxsize);
229 spin_unlock(&obd->obd_dev_lock);
230 spin_unlock(&tgd->tgd_grant_lock);
235 fo_tot_granted = tgd->tgd_tot_granted;
236 fo_tot_pending = tgd->tgd_tot_pending;
237 fo_tot_dirty = tgd->tgd_tot_dirty;
238 spin_unlock(&obd->obd_dev_lock);
239 spin_unlock(&tgd->tgd_grant_lock);
241 if (tot_granted != fo_tot_granted)
242 CERROR("%s: tot_granted %llu != fo_tot_granted %llu\n",
243 func, tot_granted, fo_tot_granted);
244 if (tot_pending != fo_tot_pending)
245 CERROR("%s: tot_pending %llu != fo_tot_pending %llu\n",
246 func, tot_pending, fo_tot_pending);
247 if (tot_dirty != fo_tot_dirty)
248 CERROR("%s: tot_dirty %llu != fo_tot_dirty %llu\n",
249 func, tot_dirty, fo_tot_dirty);
250 if (tot_pending > tot_granted)
251 CERROR("%s: tot_pending %llu > tot_granted %llu\n",
252 func, tot_pending, tot_granted);
253 if (tot_granted > maxsize)
254 CERROR("%s: tot_granted %llu > maxsize %llu\n",
255 func, tot_granted, maxsize);
256 if (tot_dirty > maxsize)
257 CERROR("%s: tot_dirty %llu > maxsize %llu\n",
258 func, tot_dirty, maxsize);
260 EXPORT_SYMBOL(tgt_grant_sanity_check);
263 * Get file system statistics of target.
265 * Helper function for statfs(), also used by grant code.
266 * Implements caching for statistics to avoid calling OSD device each time.
268 * \param[in] env execution environment
269 * \param[in] lut LU target
270 * \param[out] osfs statistic data to return
271 * \param[in] max_age maximum age for cached data
272 * \param[in] from_cache show that data was get from cache or not
274 * \retval 0 if successful
275 * \retval negative value on error
277 int tgt_statfs_internal(const struct lu_env *env, struct lu_target *lut,
278 struct obd_statfs *osfs, __u64 max_age, int *from_cache)
280 struct tg_grants_data *tgd = &lut->lut_tgd;
284 spin_lock(&tgd->tgd_osfs_lock);
285 if (cfs_time_before_64(tgd->tgd_osfs_age, max_age) || max_age == 0) {
288 /* statfs data are too old, get up-to-date one.
289 * we must be cautious here since multiple threads might be
290 * willing to update statfs data concurrently and we must
291 * grant that cached statfs data are always consistent */
293 if (tgd->tgd_statfs_inflight == 0)
294 /* clear inflight counter if no users, although it would
295 * take a while to overflow this 64-bit counter ... */
296 tgd->tgd_osfs_inflight = 0;
297 /* notify tgt_grant_commit() that we want to track writes
298 * completed as of now */
299 tgd->tgd_statfs_inflight++;
300 /* record value of inflight counter before running statfs to
301 * compute the diff once statfs is completed */
302 unstable = tgd->tgd_osfs_inflight;
303 spin_unlock(&tgd->tgd_osfs_lock);
305 /* statfs can sleep ... hopefully not for too long since we can
306 * call it fairly often as space fills up */
307 rc = dt_statfs(env, lut->lut_bottom, osfs);
311 osfs->os_namelen = min_t(__u32, osfs->os_namelen, NAME_MAX);
313 spin_lock(&tgd->tgd_grant_lock);
314 spin_lock(&tgd->tgd_osfs_lock);
315 /* calculate how much space was written while we released the
317 unstable = tgd->tgd_osfs_inflight - unstable;
318 tgd->tgd_osfs_unstable = 0;
320 /* some writes committed while we were running statfs
321 * w/o the tgd_osfs_lock. Those ones got added to
322 * the cached statfs data that we are about to crunch.
323 * Take them into account in the new statfs data */
324 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
325 unstable >> tgd->tgd_blockbits);
326 /* However, we don't really know if those writes got
327 * accounted in the statfs call, so tell
328 * tgt_grant_space_left() there is some uncertainty
329 * on the accounting of those writes.
330 * The purpose is to prevent spurious error messages in
331 * tgt_grant_space_left() since those writes might be
332 * accounted twice. */
333 tgd->tgd_osfs_unstable += unstable;
335 /* similarly, there is some uncertainty on write requests
336 * between prepare & commit */
337 tgd->tgd_osfs_unstable += tgd->tgd_tot_pending;
338 spin_unlock(&tgd->tgd_grant_lock);
340 /* finally udpate cached statfs data */
341 tgd->tgd_osfs = *osfs;
342 tgd->tgd_osfs_age = cfs_time_current_64();
344 tgd->tgd_statfs_inflight--; /* stop tracking */
345 if (tgd->tgd_statfs_inflight == 0)
346 tgd->tgd_osfs_inflight = 0;
347 spin_unlock(&tgd->tgd_osfs_lock);
352 /* use cached statfs data */
353 *osfs = tgd->tgd_osfs;
354 spin_unlock(&tgd->tgd_osfs_lock);
363 EXPORT_SYMBOL(tgt_statfs_internal);
366 * Update cached statfs information from the OSD layer
368 * Refresh statfs information cached in tgd::tgd_osfs if the cache is older
369 * than 1s or if force is set. The OSD layer is in charge of estimating data &
371 * This function can sleep so it should not be called with any spinlock held.
373 * \param[in] env LU environment passed by the caller
374 * \param[in] exp export used to print client info in debug
376 * \param[in] force force a refresh of statfs information
377 * \param[out] from_cache returns whether the statfs information are
380 static void tgt_grant_statfs(const struct lu_env *env, struct obd_export *exp,
381 int force, int *from_cache)
383 struct obd_device *obd = exp->exp_obd;
384 struct lu_target *lut = obd->u.obt.obt_lut;
385 struct tg_grants_data *tgd = &lut->lut_tgd;
386 struct tgt_thread_info *tti;
387 struct obd_statfs *osfs;
392 max_age = 0; /* get fresh statfs data */
394 max_age = cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS);
396 tti = tgt_th_info(env);
397 osfs = &tti->tti_u.osfs;
398 rc = tgt_statfs_internal(env, lut, osfs, max_age, from_cache);
405 CDEBUG(D_CACHE, "%s: cli %s/%p free: %llu avail: %llu\n",
406 obd->obd_name, exp->exp_client_uuid.uuid, exp,
407 osfs->os_bfree << tgd->tgd_blockbits,
408 osfs->os_bavail << tgd->tgd_blockbits);
412 * Figure out how much space is available on the backend filesystem after
413 * removing grant space already booked by clients.
415 * This is done by accessing cached statfs data previously populated by
416 * tgt_grant_statfs(), from which we withdraw the space already granted to
417 * clients and the reserved space.
418 * Caller must hold tgd_grant_lock spinlock.
420 * \param[in] exp export associated with the device for which the amount
421 * of available space is requested
422 * \retval amount of non-allocated space, in bytes
424 static u64 tgt_grant_space_left(struct obd_export *exp)
426 struct obd_device *obd = exp->exp_obd;
427 struct lu_target *lut = obd->u.obt.obt_lut;
428 struct tg_grants_data *tgd = &lut->lut_tgd;
436 assert_spin_locked(&tgd->tgd_grant_lock);
438 spin_lock(&tgd->tgd_osfs_lock);
439 /* get available space from cached statfs data */
440 left = tgd->tgd_osfs.os_bavail << tgd->tgd_blockbits;
441 unstable = tgd->tgd_osfs_unstable; /* those might be accounted twice */
442 spin_unlock(&tgd->tgd_osfs_lock);
444 reserved = left * tgd->tgd_reserved_pcnt / 100;
445 tot_granted = tgd->tgd_tot_granted + reserved;
447 if (left < tot_granted) {
448 int mask = (left + unstable <
449 tot_granted - tgd->tgd_tot_pending) ?
452 CDEBUG_LIMIT(mask, "%s: cli %s/%p left %llu < tot_grant "
453 "%llu unstable %llu pending %llu "
455 obd->obd_name, exp->exp_client_uuid.uuid, exp,
456 left, tot_granted, unstable,
457 tgd->tgd_tot_pending,
463 /* Withdraw space already granted to clients */
466 /* Align left on block size */
467 left &= ~((1ULL << tgd->tgd_blockbits) - 1);
469 CDEBUG(D_CACHE, "%s: cli %s/%p avail %llu left %llu unstable "
470 "%llu tot_grant %llu pending %llu\n", obd->obd_name,
471 exp->exp_client_uuid.uuid, exp, avail, left, unstable,
472 tot_granted, tgd->tgd_tot_pending);
478 * Process grant information from obdo structure packed in incoming BRW
479 * and inflate grant counters if required.
481 * Grab the dirty and seen grant announcements from the incoming obdo and
482 * inflate all grant counters passed in the request if the client does not
483 * support the grant parameters.
484 * We will later calculate the client's new grant and return it.
485 * Caller must hold tgd_grant_lock spinlock.
487 * \param[in] env LU environment supplying osfs storage
488 * \param[in] exp export for which we received the request
489 * \param[in,out] oa incoming obdo sent by the client
491 static void tgt_grant_incoming(const struct lu_env *env, struct obd_export *exp,
492 struct obdo *oa, long chunk)
494 struct tg_export_data *ted = &exp->exp_target_data;
495 struct obd_device *obd = exp->exp_obd;
496 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
501 assert_spin_locked(&tgd->tgd_grant_lock);
503 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
504 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
505 oa->o_valid &= ~OBD_MD_FLGRANT;
509 /* Add some margin, since there is a small race if other RPCs arrive
510 * out-or-order and have already consumed some grant. We want to
511 * leave this here in case there is a large error in accounting. */
513 "%s: cli %s/%p reports grant %llu dropped %u, local %lu\n",
514 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
515 oa->o_dropped, ted->ted_grant);
517 if ((long long)oa->o_dirty < 0)
520 /* inflate grant counters if required */
521 if (!exp_grant_param_supp(exp)) {
522 oa->o_grant = tgt_grant_inflate(tgd, oa->o_grant);
523 oa->o_dirty = tgt_grant_inflate(tgd, oa->o_dirty);
524 oa->o_dropped = tgt_grant_inflate(tgd, (u64)oa->o_dropped);
525 oa->o_undirty = tgt_grant_inflate(tgd, oa->o_undirty);
529 dropped = oa->o_dropped;
531 /* Update our accounting now so that statfs takes it into account.
532 * Note that ted_dirty is only approximate and can become incorrect
533 * if RPCs arrive out-of-order. No important calculations depend
534 * on ted_dirty however, but we must check sanity to not assert. */
535 if (dirty > ted->ted_grant + 4 * chunk)
536 dirty = ted->ted_grant + 4 * chunk;
537 tgd->tgd_tot_dirty += dirty - ted->ted_dirty;
538 if (ted->ted_grant < dropped) {
540 "%s: cli %s/%p reports %lu dropped > grant %lu\n",
541 obd->obd_name, exp->exp_client_uuid.uuid, exp, dropped,
545 if (tgd->tgd_tot_granted < dropped) {
546 CERROR("%s: cli %s/%p reports %lu dropped > tot_grant %llu\n",
547 obd->obd_name, exp->exp_client_uuid.uuid, exp,
548 dropped, tgd->tgd_tot_granted);
551 tgd->tgd_tot_granted -= dropped;
552 ted->ted_grant -= dropped;
553 ted->ted_dirty = dirty;
555 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
556 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
557 obd->obd_name, exp->exp_client_uuid.uuid, exp,
558 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
559 spin_unlock(&tgd->tgd_grant_lock);
566 * Grant shrink request handler.
568 * Client nodes can explicitly release grant space (i.e. process called grant
569 * shrinking). This function proceeds with the shrink request when there is
570 * less ungranted space remaining than the amount all of the connected clients
571 * would consume if they used their full grant.
572 * Caller must hold tgd_grant_lock spinlock.
574 * \param[in] exp export releasing grant space
575 * \param[in,out] oa incoming obdo sent by the client
576 * \param[in] left_space remaining free space with space already granted
579 static void tgt_grant_shrink(struct obd_export *exp, struct obdo *oa,
582 struct tg_export_data *ted = &exp->exp_target_data;
583 struct obd_device *obd = exp->exp_obd;
584 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
587 assert_spin_locked(&tgd->tgd_grant_lock);
589 if (left_space >= tgd->tgd_tot_granted_clients *
590 TGT_GRANT_SHRINK_LIMIT(exp))
593 grant_shrink = oa->o_grant;
595 ted->ted_grant -= grant_shrink;
596 tgd->tgd_tot_granted -= grant_shrink;
598 CDEBUG(D_CACHE, "%s: cli %s/%p shrink %ld ted_grant %ld total %llu\n",
599 obd->obd_name, exp->exp_client_uuid.uuid, exp, grant_shrink,
600 ted->ted_grant, tgd->tgd_tot_granted);
602 /* client has just released some grant, don't grant any space back */
607 * Calculate how much space is required to write a given network buffer
609 * This function takes block alignment into account to estimate how much on-disk
610 * space will be required to successfully write the whole niobuf.
611 * Estimated space is inflated if the export does not support
612 * OBD_CONNECT_GRANT_PARAM and if the backend filesystem has a block size
613 * larger than the minimal supported page size (i.e. 4KB).
615 * \param[in] exp export associated which the write request
616 * if NULL, then size estimate is done for server-side
618 * \param[in] lut LU target handling the request
619 * \param[in] rnb network buffer to estimate size of
621 * \retval space (in bytes) that will be consumed to write the
624 static inline u64 tgt_grant_rnb_size(struct obd_export *exp,
625 struct lu_target *lut,
626 struct niobuf_remote *rnb)
628 struct tg_grants_data *tgd = &lut->lut_tgd;
633 if (exp && !exp_grant_param_supp(exp) &&
634 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT)
635 blksize = 1ULL << COMPAT_BSIZE_SHIFT;
637 blksize = 1ULL << tgd->tgd_blockbits;
639 /* The network buffer might span several blocks, align it on block
641 bytes = rnb->rnb_offset & (blksize - 1);
642 bytes += rnb->rnb_len;
643 end = bytes & (blksize - 1);
645 bytes += blksize - end;
647 if (exp == NULL || exp_grant_param_supp(exp)) {
648 /* add per-extent insertion cost */
652 max_ext = blksize * lut->lut_dt_conf.ddp_max_extent_blks;
653 nr_ext = (bytes + max_ext - 1) / max_ext;
654 bytes += nr_ext * lut->lut_dt_conf.ddp_extent_tax;
656 /* Inflate grant space if client does not support extent-based
657 * grant allocation */
658 bytes = tgt_grant_inflate(tgd, (u64)bytes);
665 * Validate grant accounting for each incoming remote network buffer.
667 * When clients have dirtied as much space as they've been granted they
668 * fall through to sync writes. These sync writes haven't been expressed
669 * in grants and need to error with ENOSPC when there isn't room in the
670 * filesystem for them after grants are taken into account. However,
671 * writeback of the dirty data that was already granted space can write
673 * The OBD_BRW_GRANTED flag will be set in the rnb_flags of each network
674 * buffer which has been granted enough space to proceed. Buffers without
675 * this flag will fail to be written with -ENOSPC (see tgt_preprw_write().
676 * Caller must hold tgd_grant_lock spinlock.
678 * \param[in] env LU environment passed by the caller
679 * \param[in] exp export identifying the client which sent the RPC
680 * \param[in] oa incoming obdo in which we should return the pack the
682 * \param[in,out] rnb the list of network buffers
683 * \param[in] niocount the number of network buffers in the list
684 * \param[in] left the remaining free space with space already granted
687 static void tgt_grant_check(const struct lu_env *env, struct obd_export *exp,
688 struct obdo *oa, struct niobuf_remote *rnb,
689 int niocount, u64 *left)
691 struct tg_export_data *ted = &exp->exp_target_data;
692 struct obd_device *obd = exp->exp_obd;
693 struct lu_target *lut = obd->u.obt.obt_lut;
694 struct tg_grants_data *tgd = &lut->lut_tgd;
695 unsigned long ungranted = 0;
696 unsigned long granted = 0;
702 assert_spin_locked(&tgd->tgd_grant_lock);
704 if (obd->obd_recovering) {
705 /* Replaying write. Grant info have been processed already so no
706 * need to do any enforcement here. It is worth noting that only
707 * bulk writes with all rnbs having OBD_BRW_FROM_GRANT can be
708 * replayed. If one page hasn't OBD_BRW_FROM_GRANT set, then
709 * the whole bulk is written synchronously */
711 CDEBUG(D_CACHE, "Replaying write, skipping accounting\n");
712 } else if ((oa->o_valid & OBD_MD_FLFLAGS) &&
713 (oa->o_flags & OBD_FL_RECOV_RESEND)) {
714 /* Recoverable resend, grant info have already been processed as
717 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
719 } else if (exp_grant_param_supp(exp) && oa->o_grant_used > 0) {
720 /* Client supports the new grant parameters and is telling us
721 * how much grant space it consumed for this bulk write.
722 * Although all rnbs are supposed to have the OBD_BRW_FROM_GRANT
723 * flag set, we will scan the rnb list and looks for non-cache
724 * I/O in case it changes in the future */
725 if (ted->ted_grant >= oa->o_grant_used) {
726 /* skip grant accounting for rnbs with
727 * OBD_BRW_FROM_GRANT and just used grant consumption
728 * claimed in the request */
729 granted = oa->o_grant_used;
732 /* client has used more grants for this request that
734 CERROR("%s: cli %s claims %lu GRANT, real grant %lu\n",
735 exp->exp_obd->obd_name,
736 exp->exp_client_uuid.uuid,
737 (unsigned long)oa->o_grant_used, ted->ted_grant);
739 /* check whether we can fill the gap with unallocated
741 if (*left > (oa->o_grant_used - ted->ted_grant)) {
742 /* ouf .. we are safe for now */
743 granted = ted->ted_grant;
744 ungranted = oa->o_grant_used - granted;
748 /* too bad, but we cannot afford to blow up our grant
749 * accounting. The loop below will handle each rnb in
754 for (i = 0; i < niocount; i++) {
757 if ((rnb[i].rnb_flags & OBD_BRW_FROM_GRANT)) {
759 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
763 /* compute how much grant space is actually needed for
764 * this rnb, inflate grant if required */
765 bytes = tgt_grant_rnb_size(exp, lut, &rnb[i]);
766 if (ted->ted_grant >= granted + bytes) {
768 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
772 CDEBUG(D_CACHE, "%s: cli %s/%p claims %ld+%d GRANT, "
773 "real grant %lu idx %d\n", obd->obd_name,
774 exp->exp_client_uuid.uuid, exp, granted, bytes,
778 if (obd->obd_recovering)
779 CERROR("%s: cli %s is replaying OST_WRITE while one rnb"
780 " hasn't OBD_BRW_FROM_GRANT set (0x%x)\n",
781 obd->obd_name, exp->exp_client_uuid.uuid,
784 /* Consume grant space on the server.
785 * Unlike above, tgt_grant_rnb_size() is called with exp = NULL
786 * so that the required grant space isn't inflated. This is
787 * done on purpose since the server can deal with large block
788 * size, unlike some clients */
789 bytes = tgt_grant_rnb_size(NULL, lut, &rnb[i]);
791 /* if enough space, pretend it was granted */
794 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
798 /* We can't check for already-mapped blocks here (make sense
799 * when backend filesystem does not use COW) as it requires
800 * dropping the grant lock.
801 * Instead, we clear OBD_BRW_GRANTED and in that case we need
802 * to go through and verify if all of the blocks not marked
803 * BRW_GRANTED are already mapped and we can ignore this error.
805 rnb[i].rnb_flags &= ~OBD_BRW_GRANTED;
806 CDEBUG(D_CACHE, "%s: cli %s/%p idx %d no space for %d\n",
807 obd->obd_name, exp->exp_client_uuid.uuid, exp, i, bytes);
810 /* record in o_grant_used the actual space reserved for the I/O, will be
811 * used later in tgt_grant_commmit() */
812 oa->o_grant_used = granted + ungranted;
814 /* record space used for the I/O, will be used in tgt_grant_commmit() */
815 /* Now substract what the clients has used already. We don't subtract
816 * this from the tot_granted yet, so that other client's can't grab
817 * that space before we have actually allocated our blocks. That
818 * happens in tgt_grant_commit() after the writes are done. */
819 ted->ted_grant -= granted;
820 ted->ted_pending += oa->o_grant_used;
821 tgd->tgd_tot_granted += ungranted;
822 tgd->tgd_tot_pending += oa->o_grant_used;
825 "%s: cli %s/%p granted: %lu ungranted: %lu grant: %lu dirty: %lu"
826 "\n", obd->obd_name, exp->exp_client_uuid.uuid, exp,
827 granted, ungranted, ted->ted_grant, ted->ted_dirty);
829 if (obd->obd_recovering || (oa->o_valid & OBD_MD_FLGRANT) == 0)
830 /* don't update dirty accounting during recovery or
831 * if grant information got discarded (e.g. during resend) */
834 if (ted->ted_dirty < granted) {
835 CWARN("%s: cli %s/%p claims granted %lu > ted_dirty %lu\n",
836 obd->obd_name, exp->exp_client_uuid.uuid, exp,
837 granted, ted->ted_dirty);
838 granted = ted->ted_dirty;
840 tgd->tgd_tot_dirty -= granted;
841 ted->ted_dirty -= granted;
843 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
844 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
845 obd->obd_name, exp->exp_client_uuid.uuid, exp,
846 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
847 spin_unlock(&tgd->tgd_grant_lock);
854 * Allocate additional grant space to a client
856 * Calculate how much grant space to return to client, based on how much space
857 * is currently free and how much of that is already granted.
858 * Caller must hold tgd_grant_lock spinlock.
860 * \param[in] exp export of the client which sent the request
861 * \param[in] curgrant current grant claimed by the client
862 * \param[in] want how much grant space the client would like to
864 * \param[in] left remaining free space with granted space taken
866 * \param[in] conservative if set to true, the server should be cautious
867 * and limit how much space is granted back to the
868 * client. Otherwise, the server should try hard to
869 * satisfy the client request.
871 * \retval amount of grant space allocated
873 static long tgt_grant_alloc(struct obd_export *exp, u64 curgrant,
874 u64 want, u64 left, long chunk,
877 struct obd_device *obd = exp->exp_obd;
878 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
879 struct tg_export_data *ted = &exp->exp_target_data;
884 /* When tgd_grant_compat_disable is set, we don't grant any space to
885 * clients not supporting OBD_CONNECT_GRANT_PARAM.
886 * Otherwise, space granted to such a client is inflated since it
887 * consumes PAGE_SIZE of grant space per block */
888 if ((obd->obd_self_export != exp && !exp_grant_param_supp(exp) &&
889 tgd->tgd_grant_compat_disable) || left == 0 || exp->exp_failed)
892 if (want > 0x7fffffff) {
893 CERROR("%s: client %s/%p requesting > 2GB grant %llu\n",
894 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
898 /* Grant some fraction of the client's requested grant space so that
899 * they are not always waiting for write credits (not all of it to
900 * avoid overgranting in face of multiple RPCs in flight). This
901 * essentially will be able to control the OSC_MAX_RIF for a client.
903 * If we do have a large disparity between what the client thinks it
904 * has and what we think it has, don't grant very much and let the
905 * client consume its grant first. Either it just has lots of RPCs
906 * in flight, or it was evicted and its grants will soon be used up. */
907 if (curgrant >= want || curgrant >= ted->ted_grant + chunk)
910 if (obd->obd_recovering)
911 conservative = false;
914 /* don't grant more than 1/8th of the remaining free space in
917 grant = min(want - curgrant, left);
918 /* round grant up to the next block size */
919 grant = (grant + (1 << tgd->tgd_blockbits) - 1) &
920 ~((1ULL << tgd->tgd_blockbits) - 1);
925 /* Limit to grant_chunk if not reconnect/recovery */
926 if ((grant > chunk) && conservative)
929 tgd->tgd_tot_granted += grant;
930 ted->ted_grant += grant;
932 if (ted->ted_grant < 0) {
933 CERROR("%s: cli %s/%p grant %ld want %llu current %llu\n",
934 obd->obd_name, exp->exp_client_uuid.uuid, exp,
935 ted->ted_grant, want, curgrant);
936 spin_unlock(&tgd->tgd_grant_lock);
941 "%s: cli %s/%p wants: %llu current grant %llu"
942 " granting: %llu\n", obd->obd_name, exp->exp_client_uuid.uuid,
943 exp, want, curgrant, grant);
945 "%s: cli %s/%p tot cached:%llu granted:%llu"
946 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
947 exp, tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
948 obd->obd_num_exports);
954 * Handle grant space allocation on client connection & reconnection.
956 * A new non-readonly connection gets an initial grant allocation equals to
957 * tgt_grant_chunk() (i.e. twice the max BRW size in most of the cases).
958 * On reconnection, grant counters between client & target are resynchronized
959 * and additional space might be granted back if possible.
961 * \param[in] env LU environment provided by the caller
962 * \param[in] exp client's export which is (re)connecting
963 * \param[in,out] data obd_connect_data structure sent by the client in the
965 * \param[in] new_conn must set to true if this is a new connection and false
968 void tgt_grant_connect(const struct lu_env *env, struct obd_export *exp,
969 struct obd_connect_data *data, bool new_conn)
971 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
972 struct tg_grants_data *tgd = &lut->lut_tgd;
973 struct tg_export_data *ted = &exp->exp_target_data;
978 int force = 0; /* can use cached data */
980 /* don't grant space to client with read-only access */
981 if (OCD_HAS_FLAG(data, RDONLY) ||
982 (!OCD_HAS_FLAG(data, GRANT_PARAM) &&
983 tgd->tgd_grant_compat_disable)) {
985 data->ocd_connect_flags &= ~(OBD_CONNECT_GRANT |
986 OBD_CONNECT_GRANT_PARAM);
990 if (OCD_HAS_FLAG(data, GRANT_PARAM))
991 want = data->ocd_grant;
993 want = tgt_grant_inflate(tgd, data->ocd_grant);
994 chunk = tgt_grant_chunk(exp, lut, data);
996 tgt_grant_statfs(env, exp, force, &from_cache);
998 spin_lock(&tgd->tgd_grant_lock);
1000 /* Grab free space from cached info and take out space already granted
1001 * to clients as well as reserved space */
1002 left = tgt_grant_space_left(exp);
1004 /* get fresh statfs data if we are short in ungranted space */
1005 if (from_cache && left < 32 * chunk) {
1006 spin_unlock(&tgd->tgd_grant_lock);
1007 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
1012 tgt_grant_alloc(exp, (u64)ted->ted_grant, want, left, chunk, new_conn);
1014 /* return to client its current grant */
1015 if (OCD_HAS_FLAG(data, GRANT_PARAM))
1016 data->ocd_grant = ted->ted_grant;
1019 data->ocd_grant = tgt_grant_deflate(tgd, (u64)ted->ted_grant);
1021 /* reset dirty accounting */
1022 tgd->tgd_tot_dirty -= ted->ted_dirty;
1025 if (new_conn && OCD_HAS_FLAG(data, GRANT))
1026 tgd->tgd_tot_granted_clients++;
1028 spin_unlock(&tgd->tgd_grant_lock);
1030 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: %llu left: %llu\n",
1031 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1032 exp, data->ocd_grant, want, left);
1036 EXPORT_SYMBOL(tgt_grant_connect);
1039 * Release all grant space attached to a given export.
1041 * Remove a client from the grant accounting totals. We also remove
1042 * the export from the obd device under the osfs and dev locks to ensure
1043 * that the tgt_grant_sanity_check() calculations are always valid.
1044 * The client should do something similar when it invalidates its import.
1046 * \param[in] exp client's export to remove from grant accounting
1048 void tgt_grant_discard(struct obd_export *exp)
1050 struct obd_device *obd = exp->exp_obd;
1051 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1052 struct tg_export_data *ted = &exp->exp_target_data;
1054 spin_lock(&tgd->tgd_grant_lock);
1055 LASSERTF(tgd->tgd_tot_granted >= ted->ted_grant,
1056 "%s: tot_granted %llu cli %s/%p ted_grant %ld\n",
1057 obd->obd_name, tgd->tgd_tot_granted,
1058 exp->exp_client_uuid.uuid, exp, ted->ted_grant);
1059 tgd->tgd_tot_granted -= ted->ted_grant;
1061 LASSERTF(tgd->tgd_tot_pending >= ted->ted_pending,
1062 "%s: tot_pending %llu cli %s/%p ted_pending %ld\n",
1063 obd->obd_name, tgd->tgd_tot_pending,
1064 exp->exp_client_uuid.uuid, exp, ted->ted_pending);
1065 /* tgd_tot_pending is handled in tgt_grant_commit as bulk
1067 LASSERTF(tgd->tgd_tot_dirty >= ted->ted_dirty,
1068 "%s: tot_dirty %llu cli %s/%p ted_dirty %ld\n",
1069 obd->obd_name, tgd->tgd_tot_dirty,
1070 exp->exp_client_uuid.uuid, exp, ted->ted_dirty);
1071 tgd->tgd_tot_dirty -= ted->ted_dirty;
1073 spin_unlock(&tgd->tgd_grant_lock);
1075 EXPORT_SYMBOL(tgt_grant_discard);
1078 * Process grant information from incoming bulk read request.
1080 * Extract grant information packed in obdo structure (OBD_MD_FLGRANT set in
1081 * o_valid). Bulk reads usually comes with grant announcements (number of dirty
1082 * blocks, remaining amount of grant space, ...) and could also include a grant
1083 * shrink request. Unlike bulk write, no additional grant space is returned on
1084 * bulk read request.
1086 * \param[in] env is the lu environment provided by the caller
1087 * \param[in] exp is the export of the client which sent the request
1088 * \param[in,out] oa is the incoming obdo sent by the client
1090 void tgt_grant_prepare_read(const struct lu_env *env,
1091 struct obd_export *exp, struct obdo *oa)
1093 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1094 struct tg_grants_data *tgd = &lut->lut_tgd;
1103 if ((oa->o_valid & OBD_MD_FLGRANT) == 0)
1104 /* The read request does not contain any grant
1108 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1109 (oa->o_flags & OBD_FL_SHRINK_GRANT)) {
1110 /* To process grant shrink request, we need to know how much
1111 * available space remains on the backend filesystem.
1112 * Shrink requests are not so common, we always get fresh
1113 * statfs information. */
1114 tgt_grant_statfs(env, exp, 1, NULL);
1116 /* protect all grant counters */
1117 spin_lock(&tgd->tgd_grant_lock);
1119 /* Grab free space from cached statfs data and take out space
1120 * already granted to clients as well as reserved space */
1121 left = tgt_grant_space_left(exp);
1123 /* all set now to proceed with shrinking */
1126 /* no grant shrinking request packed in the obdo and
1127 * since we don't grant space back on reads, no point
1128 * in running statfs, so just skip it and process
1129 * incoming grant data directly. */
1130 spin_lock(&tgd->tgd_grant_lock);
1134 /* extract incoming grant information provided by the client and
1135 * inflate grant counters if required */
1136 tgt_grant_incoming(env, exp, oa, tgt_grant_chunk(exp, lut, NULL));
1138 /* unlike writes, we don't return grants back on reads unless a grant
1139 * shrink request was packed and we decided to turn it down. */
1141 tgt_grant_shrink(exp, oa, left);
1145 if (!exp_grant_param_supp(exp))
1146 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1147 spin_unlock(&tgd->tgd_grant_lock);
1150 EXPORT_SYMBOL(tgt_grant_prepare_read);
1153 * Process grant information from incoming bulk write request.
1155 * This function extracts client's grant announcements from incoming bulk write
1156 * request and attempts to allocate grant space for network buffers that need it
1157 * (i.e. OBD_BRW_FROM_GRANT not set in rnb_fags).
1158 * Network buffers which aren't granted the OBD_BRW_GRANTED flag should not
1159 * proceed further and should fail with -ENOSPC.
1160 * Whenever possible, additional grant space will be returned to the client
1161 * in the bulk write reply.
1162 * tgt_grant_prepare_write() must be called before writting any buffers to
1163 * the backend storage. This function works in pair with tgt_grant_commit()
1164 * which must be invoked once all buffers have been written to disk in order
1165 * to release space from the pending grant counter.
1167 * \param[in] env LU environment provided by the caller
1168 * \param[in] exp export of the client which sent the request
1169 * \param[in] oa incoming obdo sent by the client
1170 * \param[in] rnb list of network buffers
1171 * \param[in] niocount number of network buffers in the list
1173 void tgt_grant_prepare_write(const struct lu_env *env,
1174 struct obd_export *exp, struct obdo *oa,
1175 struct niobuf_remote *rnb, int niocount)
1177 struct obd_device *obd = exp->exp_obd;
1178 struct lu_target *lut = obd->u.obt.obt_lut;
1179 struct tg_grants_data *tgd = &lut->lut_tgd;
1182 int force = 0; /* can use cached data intially */
1183 long chunk = tgt_grant_chunk(exp, lut, NULL);
1188 /* get statfs information from OSD layer */
1189 tgt_grant_statfs(env, exp, force, &from_cache);
1191 spin_lock(&tgd->tgd_grant_lock); /* protect all grant counters */
1193 /* Grab free space from cached statfs data and take out space already
1194 * granted to clients as well as reserved space */
1195 left = tgt_grant_space_left(exp);
1197 /* Get fresh statfs data if we are short in ungranted space */
1198 if (from_cache && left < 32 * chunk) {
1199 spin_unlock(&tgd->tgd_grant_lock);
1200 CDEBUG(D_CACHE, "%s: fs has no space left and statfs too old\n",
1206 /* When close to free space exhaustion, trigger a sync to force
1207 * writeback cache to consume required space immediately and release as
1208 * much space as possible. */
1209 if (!obd->obd_recovering && force != 2 && left < chunk) {
1210 bool from_grant = true;
1213 /* That said, it is worth running a sync only if some pages did
1214 * not consume grant space on the client and could thus fail
1215 * with ENOSPC later in tgt_grant_check() */
1216 for (i = 0; i < niocount; i++)
1217 if (!(rnb[i].rnb_flags & OBD_BRW_FROM_GRANT))
1221 /* at least one network buffer requires acquiring grant
1222 * space on the server */
1223 spin_unlock(&tgd->tgd_grant_lock);
1224 /* discard errors, at least we tried ... */
1225 dt_sync(env, lut->lut_bottom);
1231 /* extract incoming grant information provided by the client,
1232 * and inflate grant counters if required */
1233 tgt_grant_incoming(env, exp, oa, chunk);
1236 tgt_grant_check(env, exp, oa, rnb, niocount, &left);
1238 if (!(oa->o_valid & OBD_MD_FLGRANT)) {
1239 spin_unlock(&tgd->tgd_grant_lock);
1243 /* if OBD_FL_SHRINK_GRANT is set, the client is willing to release some
1245 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1246 (oa->o_flags & OBD_FL_SHRINK_GRANT))
1247 tgt_grant_shrink(exp, oa, left);
1249 /* grant more space back to the client if possible */
1250 oa->o_grant = tgt_grant_alloc(exp, oa->o_grant, oa->o_undirty,
1253 if (!exp_grant_param_supp(exp))
1254 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1255 spin_unlock(&tgd->tgd_grant_lock);
1258 EXPORT_SYMBOL(tgt_grant_prepare_write);
1261 * Consume grant space reserved for object creation.
1263 * Grant space is allocated to the local self export for object precreation.
1264 * This is required to prevent object precreation from consuming grant space
1265 * allocated to client nodes for the data writeback cache.
1266 * This function consumes enough space to create \a nr objects and allocates
1267 * more grant space to the self export for future precreation requests, if
1270 * \param[in] env LU environment provided by the caller
1271 * \param[in] exp export holding the grant space for precreation (= self
1273 * \param[in] nr number of objects to be created
1275 * \retval >= 0 amount of grant space allocated to the precreate request
1276 * \retval -ENOSPC on failure
1278 long tgt_grant_create(const struct lu_env *env, struct obd_export *exp, s64 *nr)
1280 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1281 struct tg_grants_data *tgd = &lut->lut_tgd;
1282 struct tg_export_data *ted = &exp->exp_target_data;
1284 unsigned long wanted;
1285 unsigned long granted;
1288 if (exp->exp_obd->obd_recovering ||
1289 lut->lut_dt_conf.ddp_inodespace == 0)
1290 /* don't enforce grant during recovery */
1293 /* Update statfs data if required */
1294 tgt_grant_statfs(env, exp, 1, NULL);
1296 /* protect all grant counters */
1297 spin_lock(&tgd->tgd_grant_lock);
1299 /* fail precreate request if there is not enough blocks available for
1301 if (tgd->tgd_osfs.os_bavail - (ted->ted_grant >> tgd->tgd_blockbits) <
1302 (tgd->tgd_osfs.os_blocks >> 10)) {
1303 spin_unlock(&tgd->tgd_grant_lock);
1304 CDEBUG(D_RPCTRACE, "%s: not enough space for create %llu\n",
1305 exp->exp_obd->obd_name,
1306 tgd->tgd_osfs.os_bavail * tgd->tgd_osfs.os_blocks);
1310 /* Grab free space from cached statfs data and take out space
1311 * already granted to clients as well as reserved space */
1312 left = tgt_grant_space_left(exp);
1314 /* compute how much space is required to handle the precreation
1316 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1317 if (wanted > ted->ted_grant + left) {
1318 /* that's beyond what remains, adjust the number of objects that
1319 * can be safely precreated */
1320 wanted = ted->ted_grant + left;
1321 *nr = wanted / lut->lut_dt_conf.ddp_inodespace;
1323 /* we really have no space any more for precreation,
1324 * fail the precreate request with ENOSPC */
1325 spin_unlock(&tgd->tgd_grant_lock);
1328 /* compute space needed for the new number of creations */
1329 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1331 LASSERT(wanted <= ted->ted_grant + left);
1333 if (wanted <= ted->ted_grant) {
1334 /* we've enough grant space to handle this precreate request */
1335 ted->ted_grant -= wanted;
1337 /* we need to take some space from the ungranted pool */
1338 tgd->tgd_tot_granted += wanted - ted->ted_grant;
1339 left -= wanted - ted->ted_grant;
1343 ted->ted_pending += granted;
1344 tgd->tgd_tot_pending += granted;
1346 /* grant more space for precreate purpose if possible. */
1347 wanted = OST_MAX_PRECREATE * lut->lut_dt_conf.ddp_inodespace / 2;
1348 if (wanted > ted->ted_grant) {
1351 /* always try to book enough space to handle a large precreate
1353 chunk = tgt_grant_chunk(exp, lut, NULL);
1354 wanted -= ted->ted_grant;
1355 tgt_grant_alloc(exp, ted->ted_grant, wanted, left, chunk,
1358 spin_unlock(&tgd->tgd_grant_lock);
1361 EXPORT_SYMBOL(tgt_grant_create);
1364 * Release grant space added to the pending counter by tgt_grant_prepare_write()
1366 * Update pending grant counter once buffers have been written to the disk.
1368 * \param[in] exp export of the client which sent the request
1369 * \param[in] pending amount of reserved space to be released
1370 * \param[in] rc return code of pre-commit operations
1372 void tgt_grant_commit(struct obd_export *exp, unsigned long pending,
1375 struct tg_grants_data *tgd = &exp->exp_obd->u.obt.obt_lut->lut_tgd;
1379 /* get space accounted in tot_pending for the I/O, set in
1380 * tgt_grant_check() */
1384 spin_lock(&tgd->tgd_grant_lock);
1385 /* Don't update statfs data for errors raised before commit (e.g.
1386 * bulk transfer failed, ...) since we know those writes have not been
1387 * processed. For other errors hit during commit, we cannot really tell
1388 * whether or not something was written, so we update statfs data.
1389 * In any case, this should not be fatal since we always get fresh
1390 * statfs data before failing a request with ENOSPC */
1392 spin_lock(&tgd->tgd_osfs_lock);
1393 /* Take pending out of cached statfs data */
1394 tgd->tgd_osfs.os_bavail -= min_t(u64,
1395 tgd->tgd_osfs.os_bavail,
1396 pending >> tgd->tgd_blockbits);
1397 if (tgd->tgd_statfs_inflight)
1398 /* someone is running statfs and want to be notified of
1399 * writes happening meanwhile */
1400 tgd->tgd_osfs_inflight += pending;
1401 spin_unlock(&tgd->tgd_osfs_lock);
1404 if (exp->exp_target_data.ted_pending < pending) {
1405 CERROR("%s: cli %s/%p ted_pending(%lu) < grant_used(%lu)\n",
1406 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1407 exp->exp_target_data.ted_pending, pending);
1408 spin_unlock(&tgd->tgd_grant_lock);
1411 exp->exp_target_data.ted_pending -= pending;
1413 if (tgd->tgd_tot_granted < pending) {
1414 CERROR("%s: cli %s/%p tot_granted(%llu) < grant_used(%lu)\n",
1415 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1416 tgd->tgd_tot_granted, pending);
1417 spin_unlock(&tgd->tgd_grant_lock);
1420 tgd->tgd_tot_granted -= pending;
1422 if (tgd->tgd_tot_pending < pending) {
1423 CERROR("%s: cli %s/%p tot_pending(%llu) < grant_used(%lu)\n",
1424 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1425 tgd->tgd_tot_pending, pending);
1426 spin_unlock(&tgd->tgd_grant_lock);
1429 tgd->tgd_tot_pending -= pending;
1430 spin_unlock(&tgd->tgd_grant_lock);
1433 EXPORT_SYMBOL(tgt_grant_commit);
1435 struct tgt_grant_cb {
1436 /* commit callback structure */
1437 struct dt_txn_commit_cb tgc_cb;
1438 /* export associated with the bulk write */
1439 struct obd_export *tgc_exp;
1440 /* pending grant to be released */
1441 unsigned long tgc_granted;
1445 * Callback function for grant releasing
1447 * Release grant space reserved by the client node.
1449 * \param[in] env execution environment
1450 * \param[in] th transaction handle
1451 * \param[in] cb callback data
1452 * \param[in] err error code
1454 static void tgt_grant_commit_cb(struct lu_env *env, struct thandle *th,
1455 struct dt_txn_commit_cb *cb, int err)
1457 struct tgt_grant_cb *tgc;
1459 tgc = container_of(cb, struct tgt_grant_cb, tgc_cb);
1461 tgt_grant_commit(tgc->tgc_exp, tgc->tgc_granted, err);
1462 class_export_cb_put(tgc->tgc_exp);
1467 * Add callback for grant releasing
1469 * Register a commit callback to release grant space.
1471 * \param[in] th transaction handle
1472 * \param[in] exp OBD export of client
1473 * \param[in] granted amount of grant space to be released upon commit
1475 * \retval 0 on successful callback adding
1476 * \retval negative value on error
1478 int tgt_grant_commit_cb_add(struct thandle *th, struct obd_export *exp,
1479 unsigned long granted)
1481 struct tgt_grant_cb *tgc;
1482 struct dt_txn_commit_cb *dcb;
1490 tgc->tgc_exp = class_export_cb_get(exp);
1491 tgc->tgc_granted = granted;
1494 dcb->dcb_func = tgt_grant_commit_cb;
1495 INIT_LIST_HEAD(&dcb->dcb_linkage);
1496 strlcpy(dcb->dcb_name, "tgt_grant_commit_cb", sizeof(dcb->dcb_name));
1498 rc = dt_trans_cb_add(th, dcb);
1500 class_export_cb_put(tgc->tgc_exp);
1506 EXPORT_SYMBOL(tgt_grant_commit_cb_add);
1509 * Show estimate of total amount of dirty data on clients.
1511 * \param[in] m seq_file handle
1512 * \param[in] data unused for single entry
1514 * \retval 0 on success
1515 * \retval negative value on error
1517 int tgt_tot_dirty_seq_show(struct seq_file *m, void *data)
1519 struct obd_device *obd = m->private;
1520 struct tg_grants_data *tgd;
1522 LASSERT(obd != NULL);
1523 tgd = &obd->u.obt.obt_lut->lut_tgd;
1524 seq_printf(m, "%llu\n", tgd->tgd_tot_dirty);
1527 EXPORT_SYMBOL(tgt_tot_dirty_seq_show);
1530 * Show total amount of space granted to clients.
1532 * \param[in] m seq_file handle
1533 * \param[in] data unused for single entry
1535 * \retval 0 on success
1536 * \retval negative value on error
1538 int tgt_tot_granted_seq_show(struct seq_file *m, void *data)
1540 struct obd_device *obd = m->private;
1541 struct tg_grants_data *tgd;
1543 LASSERT(obd != NULL);
1544 tgd = &obd->u.obt.obt_lut->lut_tgd;
1545 seq_printf(m, "%llu\n", tgd->tgd_tot_granted);
1548 EXPORT_SYMBOL(tgt_tot_granted_seq_show);
1551 * Show total amount of space used by IO in progress.
1553 * \param[in] m seq_file handle
1554 * \param[in] data unused for single entry
1556 * \retval 0 on success
1557 * \retval negative value on error
1559 int tgt_tot_pending_seq_show(struct seq_file *m, void *data)
1561 struct obd_device *obd = m->private;
1562 struct tg_grants_data *tgd;
1564 LASSERT(obd != NULL);
1565 tgd = &obd->u.obt.obt_lut->lut_tgd;
1566 seq_printf(m, "%llu\n", tgd->tgd_tot_pending);
1569 EXPORT_SYMBOL(tgt_tot_pending_seq_show);
1572 * Show if grants compatibility mode is disabled.
1574 * When tgd_grant_compat_disable is set, we don't grant any space to clients
1575 * not supporting OBD_CONNECT_GRANT_PARAM. Otherwise, space granted to such
1576 * a client is inflated since it consumes PAGE_SIZE of grant space per
1577 * block, (i.e. typically 4kB units), but underlaying file system might have
1578 * block size bigger than page size, e.g. ZFS. See LU-2049 for details.
1580 * \param[in] m seq_file handle
1581 * \param[in] data unused for single entry
1583 * \retval 0 on success
1584 * \retval negative value on error
1586 int tgt_grant_compat_disable_seq_show(struct seq_file *m, void *data)
1588 struct obd_device *obd = m->private;
1589 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1591 seq_printf(m, "%u\n", tgd->tgd_grant_compat_disable);
1594 EXPORT_SYMBOL(tgt_grant_compat_disable_seq_show);
1597 * Change grant compatibility mode.
1599 * Setting tgd_grant_compat_disable prohibit any space granting to clients
1600 * not supporting OBD_CONNECT_GRANT_PARAM. See details above.
1602 * \param[in] file proc file
1603 * \param[in] buffer string which represents mode
1604 * 1: disable compatibility mode
1605 * 0: enable compatibility mode
1606 * \param[in] count \a buffer length
1607 * \param[in] off unused for single entry
1609 * \retval \a count on success
1610 * \retval negative number on error
1612 ssize_t tgt_grant_compat_disable_seq_write(struct file *file,
1613 const char __user *buffer,
1614 size_t count, loff_t *off)
1616 struct seq_file *m = file->private_data;
1617 struct obd_device *obd = m->private;
1618 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1622 rc = lprocfs_str_to_s64(buffer, count, &val);
1629 tgd->tgd_grant_compat_disable = !!val;
1633 EXPORT_SYMBOL(tgt_grant_compat_disable_seq_write);