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, 2017, 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_CLASS
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, time64_t max_age, int *from_cache)
285 struct tg_grants_data *tgd = &lut->lut_tgd;
289 spin_lock(&tgd->tgd_osfs_lock);
290 if (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 osfs->os_namelen = min_t(__u32, osfs->os_namelen, NAME_MAX);
318 spin_lock(&tgd->tgd_grant_lock);
319 spin_lock(&tgd->tgd_osfs_lock);
320 /* calculate how much space was written while we released the
322 unstable = tgd->tgd_osfs_inflight - unstable;
323 tgd->tgd_osfs_unstable = 0;
325 /* some writes committed while we were running statfs
326 * w/o the tgd_osfs_lock. Those ones got added to
327 * the cached statfs data that we are about to crunch.
328 * Take them into account in the new statfs data */
329 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
330 unstable >> tgd->tgd_blockbits);
331 /* However, we don't really know if those writes got
332 * accounted in the statfs call, so tell
333 * tgt_grant_space_left() there is some uncertainty
334 * on the accounting of those writes.
335 * The purpose is to prevent spurious error messages in
336 * tgt_grant_space_left() since those writes might be
337 * accounted twice. */
338 tgd->tgd_osfs_unstable += unstable;
340 /* similarly, there is some uncertainty on write requests
341 * between prepare & commit */
342 tgd->tgd_osfs_unstable += tgd->tgd_tot_pending;
343 spin_unlock(&tgd->tgd_grant_lock);
345 /* finally udpate cached statfs data */
346 tgd->tgd_osfs = *osfs;
347 tgd->tgd_osfs_age = ktime_get_seconds();
349 tgd->tgd_statfs_inflight--; /* stop tracking */
350 if (tgd->tgd_statfs_inflight == 0)
351 tgd->tgd_osfs_inflight = 0;
352 spin_unlock(&tgd->tgd_osfs_lock);
357 /* use cached statfs data */
358 *osfs = tgd->tgd_osfs;
359 spin_unlock(&tgd->tgd_osfs_lock);
368 EXPORT_SYMBOL(tgt_statfs_internal);
371 * Update cached statfs information from the OSD layer
373 * Refresh statfs information cached in tgd::tgd_osfs if the cache is older
374 * than 1s or if force is set. The OSD layer is in charge of estimating data &
376 * This function can sleep so it should not be called with any spinlock held.
378 * \param[in] env LU environment passed by the caller
379 * \param[in] exp export used to print client info in debug
381 * \param[in] force force a refresh of statfs information
382 * \param[out] from_cache returns whether the statfs information are
385 static void tgt_grant_statfs(const struct lu_env *env, struct obd_export *exp,
386 int force, int *from_cache)
388 struct obd_device *obd = exp->exp_obd;
389 struct lu_target *lut = obd->u.obt.obt_lut;
390 struct tg_grants_data *tgd = &lut->lut_tgd;
391 struct tgt_thread_info *tti;
392 struct obd_statfs *osfs;
397 max_age = 0; /* get fresh statfs data */
399 max_age = ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS;
401 tti = tgt_th_info(env);
402 osfs = &tti->tti_u.osfs;
403 rc = tgt_statfs_internal(env, lut, osfs, max_age, from_cache);
410 CDEBUG(D_CACHE, "%s: cli %s/%p free: %llu avail: %llu\n",
411 obd->obd_name, exp->exp_client_uuid.uuid, exp,
412 osfs->os_bfree << tgd->tgd_blockbits,
413 osfs->os_bavail << tgd->tgd_blockbits);
417 * Figure out how much space is available on the backend filesystem after
418 * removing grant space already booked by clients.
420 * This is done by accessing cached statfs data previously populated by
421 * tgt_grant_statfs(), from which we withdraw the space already granted to
422 * clients and the reserved space.
423 * Caller must hold tgd_grant_lock spinlock.
425 * \param[in] exp export associated with the device for which the amount
426 * of available space is requested
427 * \retval amount of non-allocated space, in bytes
429 static u64 tgt_grant_space_left(struct obd_export *exp)
431 struct obd_device *obd = exp->exp_obd;
432 struct lu_target *lut = obd->u.obt.obt_lut;
433 struct tg_grants_data *tgd = &lut->lut_tgd;
441 assert_spin_locked(&tgd->tgd_grant_lock);
443 spin_lock(&tgd->tgd_osfs_lock);
444 /* get available space from cached statfs data */
445 left = tgd->tgd_osfs.os_bavail << tgd->tgd_blockbits;
446 unstable = tgd->tgd_osfs_unstable; /* those might be accounted twice */
447 spin_unlock(&tgd->tgd_osfs_lock);
449 reserved = left * tgd->tgd_reserved_pcnt / 100;
450 tot_granted = tgd->tgd_tot_granted + reserved;
452 if (left < tot_granted) {
453 int mask = (left + unstable <
454 tot_granted - tgd->tgd_tot_pending) ?
457 /* the below message is checked in sanityn.sh test_15 */
459 "%s: cli %s/%p left=%llu < tot_grant=%llu unstable=%llu pending=%llu dirty=%llu\n",
460 obd->obd_name, exp->exp_client_uuid.uuid, exp,
461 left, tot_granted, unstable,
462 tgd->tgd_tot_pending,
468 /* Withdraw space already granted to clients */
471 /* Align left on block size */
472 left &= ~((1ULL << tgd->tgd_blockbits) - 1);
475 "%s: cli %s/%p avail=%llu left=%llu unstable=%llu tot_grant=%llu pending=%llu\n",
476 obd->obd_name, exp->exp_client_uuid.uuid, exp, avail, left,
477 unstable, tot_granted, tgd->tgd_tot_pending);
483 * Process grant information from obdo structure packed in incoming BRW
484 * and inflate grant counters if required.
486 * Grab the dirty and seen grant announcements from the incoming obdo and
487 * inflate all grant counters passed in the request if the client does not
488 * support the grant parameters.
489 * We will later calculate the client's new grant and return it.
490 * Caller must hold tgd_grant_lock spinlock.
492 * \param[in] env LU environment supplying osfs storage
493 * \param[in] exp export for which we received the request
494 * \param[in,out] oa incoming obdo sent by the client
496 static void tgt_grant_incoming(const struct lu_env *env, struct obd_export *exp,
497 struct obdo *oa, long chunk)
499 struct tg_export_data *ted = &exp->exp_target_data;
500 struct obd_device *obd = exp->exp_obd;
501 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
502 long long dirty, dropped;
505 assert_spin_locked(&tgd->tgd_grant_lock);
507 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
508 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
509 oa->o_valid &= ~OBD_MD_FLGRANT;
513 /* Add some margin, since there is a small race if other RPCs arrive
514 * out-or-order and have already consumed some grant. We want to
515 * leave this here in case there is a large error in accounting. */
517 "%s: cli %s/%p reports grant %llu dropped %u, local %lu\n",
518 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
519 oa->o_dropped, ted->ted_grant);
521 if ((long long)oa->o_dirty < 0)
524 /* inflate grant counters if required */
525 if (!exp_grant_param_supp(exp)) {
527 oa->o_grant = tgt_grant_inflate(tgd, oa->o_grant);
528 oa->o_dirty = tgt_grant_inflate(tgd, oa->o_dirty);
529 /* inflation can bump client's wish to >4GB which doesn't fit
530 * 32bit o_undirty, limit that .. */
531 tmp = tgt_grant_inflate(tgd, oa->o_undirty);
532 if (tmp >= OBD_MAX_GRANT)
533 tmp = OBD_MAX_GRANT & ~(1ULL << tgd->tgd_blockbits);
535 tmp = tgt_grant_inflate(tgd, oa->o_dropped);
536 if (tmp >= OBD_MAX_GRANT)
537 tmp = OBD_MAX_GRANT & ~(1ULL << tgd->tgd_blockbits);
542 dropped = oa->o_dropped;
544 /* Update our accounting now so that statfs takes it into account.
545 * Note that ted_dirty is only approximate and can become incorrect
546 * if RPCs arrive out-of-order. No important calculations depend
547 * on ted_dirty however, but we must check sanity to not assert. */
548 if (dirty > ted->ted_grant + 4 * chunk)
549 dirty = ted->ted_grant + 4 * chunk;
550 tgd->tgd_tot_dirty += dirty - ted->ted_dirty;
551 if (ted->ted_grant < dropped) {
553 "%s: cli %s/%p reports %llu dropped > grant %lu\n",
554 obd->obd_name, exp->exp_client_uuid.uuid, exp, dropped,
558 if (tgd->tgd_tot_granted < dropped) {
559 CERROR("%s: cli %s/%p reports %llu dropped > tot_grant %llu\n",
560 obd->obd_name, exp->exp_client_uuid.uuid, exp,
561 dropped, tgd->tgd_tot_granted);
564 tgd->tgd_tot_granted -= dropped;
565 ted->ted_grant -= dropped;
566 ted->ted_dirty = dirty;
568 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
569 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
570 obd->obd_name, exp->exp_client_uuid.uuid, exp,
571 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
572 spin_unlock(&tgd->tgd_grant_lock);
579 * Grant shrink request handler.
581 * Client nodes can explicitly release grant space (i.e. process called grant
582 * shrinking). This function proceeds with the shrink request when there is
583 * less ungranted space remaining than the amount all of the connected clients
584 * would consume if they used their full grant.
585 * Caller must hold tgd_grant_lock spinlock.
587 * \param[in] exp export releasing grant space
588 * \param[in,out] oa incoming obdo sent by the client
589 * \param[in] left_space remaining free space with space already granted
592 static void tgt_grant_shrink(struct obd_export *exp, struct obdo *oa,
595 struct tg_export_data *ted = &exp->exp_target_data;
596 struct obd_device *obd = exp->exp_obd;
597 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
600 assert_spin_locked(&tgd->tgd_grant_lock);
602 if (left_space >= tgd->tgd_tot_granted_clients *
603 TGT_GRANT_SHRINK_LIMIT(exp))
606 grant_shrink = oa->o_grant;
608 if (ted->ted_grant < grant_shrink) {
610 "%s: cli %s/%p wants %lu shrinked > grant %lu\n",
611 obd->obd_name, exp->exp_client_uuid.uuid, exp,
612 grant_shrink, ted->ted_grant);
613 grant_shrink = ted->ted_grant;
616 ted->ted_grant -= grant_shrink;
617 tgd->tgd_tot_granted -= grant_shrink;
619 CDEBUG(D_CACHE, "%s: cli %s/%p shrink %ld ted_grant %ld total %llu\n",
620 obd->obd_name, exp->exp_client_uuid.uuid, exp, grant_shrink,
621 ted->ted_grant, tgd->tgd_tot_granted);
623 /* client has just released some grant, don't grant any space back */
628 * Calculate how much space is required to write a given network buffer
630 * This function takes block alignment into account to estimate how much on-disk
631 * space will be required to successfully write the whole niobuf.
632 * Estimated space is inflated if the export does not support
633 * OBD_CONNECT_GRANT_PARAM and if the backend filesystem has a block size
634 * larger than the minimal supported page size (i.e. 4KB).
636 * \param[in] exp export associated which the write request
637 * if NULL, then size estimate is done for server-side
639 * \param[in] lut LU target handling the request
640 * \param[in] rnb network buffer to estimate size of
642 * \retval space (in bytes) that will be consumed to write the
645 static inline u64 tgt_grant_rnb_size(struct obd_export *exp,
646 struct lu_target *lut,
647 struct niobuf_remote *rnb)
649 struct tg_grants_data *tgd = &lut->lut_tgd;
654 if (exp && !exp_grant_param_supp(exp) &&
655 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT)
656 blksize = 1ULL << COMPAT_BSIZE_SHIFT;
658 blksize = 1ULL << tgd->tgd_blockbits;
660 /* The network buffer might span several blocks, align it on block
662 bytes = rnb->rnb_offset & (blksize - 1);
663 bytes += rnb->rnb_len;
664 end = bytes & (blksize - 1);
666 bytes += blksize - end;
668 if (exp == NULL || exp_grant_param_supp(exp)) {
669 /* add per-extent insertion cost */
673 max_ext = blksize * lut->lut_dt_conf.ddp_max_extent_blks;
674 nr_ext = (bytes + max_ext - 1) / max_ext;
675 bytes += nr_ext * lut->lut_dt_conf.ddp_extent_tax;
677 /* Inflate grant space if client does not support extent-based
678 * grant allocation */
679 bytes = tgt_grant_inflate(tgd, (u64)bytes);
686 * Validate grant accounting for each incoming remote network buffer.
688 * When clients have dirtied as much space as they've been granted they
689 * fall through to sync writes. These sync writes haven't been expressed
690 * in grants and need to error with ENOSPC when there isn't room in the
691 * filesystem for them after grants are taken into account. However,
692 * writeback of the dirty data that was already granted space can write
694 * The OBD_BRW_GRANTED flag will be set in the rnb_flags of each network
695 * buffer which has been granted enough space to proceed. Buffers without
696 * this flag will fail to be written with -ENOSPC (see tgt_preprw_write().
697 * Caller must hold tgd_grant_lock spinlock.
699 * \param[in] env LU environment passed by the caller
700 * \param[in] exp export identifying the client which sent the RPC
701 * \param[in] oa incoming obdo in which we should return the pack the
703 * \param[in,out] rnb the list of network buffers
704 * \param[in] niocount the number of network buffers in the list
705 * \param[in] left the remaining free space with space already granted
708 static void tgt_grant_check(const struct lu_env *env, struct obd_export *exp,
709 struct obdo *oa, struct niobuf_remote *rnb,
710 int niocount, u64 *left)
712 struct tg_export_data *ted = &exp->exp_target_data;
713 struct obd_device *obd = exp->exp_obd;
714 struct lu_target *lut = obd->u.obt.obt_lut;
715 struct tg_grants_data *tgd = &lut->lut_tgd;
716 unsigned long ungranted = 0;
717 unsigned long granted = 0;
723 assert_spin_locked(&tgd->tgd_grant_lock);
725 if (obd->obd_recovering) {
726 /* Replaying write. Grant info have been processed already so no
727 * need to do any enforcement here. It is worth noting that only
728 * bulk writes with all rnbs having OBD_BRW_FROM_GRANT can be
729 * replayed. If one page hasn't OBD_BRW_FROM_GRANT set, then
730 * the whole bulk is written synchronously */
732 CDEBUG(D_CACHE, "Replaying write, skipping accounting\n");
733 } else if ((oa->o_valid & OBD_MD_FLFLAGS) &&
734 (oa->o_flags & OBD_FL_RECOV_RESEND)) {
735 /* Recoverable resend, grant info have already been processed as
738 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
740 } else if (exp_grant_param_supp(exp) && oa->o_grant_used > 0) {
741 /* Client supports the new grant parameters and is telling us
742 * how much grant space it consumed for this bulk write.
743 * Although all rnbs are supposed to have the OBD_BRW_FROM_GRANT
744 * flag set, we will scan the rnb list and looks for non-cache
745 * I/O in case it changes in the future */
746 if (ted->ted_grant >= oa->o_grant_used) {
747 /* skip grant accounting for rnbs with
748 * OBD_BRW_FROM_GRANT and just used grant consumption
749 * claimed in the request */
750 granted = oa->o_grant_used;
753 /* client has used more grants for this request that
755 CERROR("%s: cli %s claims %lu GRANT, real grant %lu\n",
756 exp->exp_obd->obd_name,
757 exp->exp_client_uuid.uuid,
758 (unsigned long)oa->o_grant_used, ted->ted_grant);
760 /* check whether we can fill the gap with unallocated
762 if (*left > (oa->o_grant_used - ted->ted_grant)) {
763 /* ouf .. we are safe for now */
764 granted = ted->ted_grant;
765 ungranted = oa->o_grant_used - granted;
769 /* too bad, but we cannot afford to blow up our grant
770 * accounting. The loop below will handle each rnb in
775 for (i = 0; i < niocount; i++) {
778 if ((rnb[i].rnb_flags & OBD_BRW_FROM_GRANT)) {
780 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
784 /* compute how much grant space is actually needed for
785 * this rnb, inflate grant if required */
786 bytes = tgt_grant_rnb_size(exp, lut, &rnb[i]);
787 if (ted->ted_grant >= granted + bytes) {
789 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
793 CDEBUG(D_CACHE, "%s: cli %s/%p claims %ld+%d GRANT, "
794 "real grant %lu idx %d\n", obd->obd_name,
795 exp->exp_client_uuid.uuid, exp, granted, bytes,
799 if (obd->obd_recovering)
800 CERROR("%s: cli %s is replaying OST_WRITE while one rnb"
801 " hasn't OBD_BRW_FROM_GRANT set (0x%x)\n",
802 obd->obd_name, exp->exp_client_uuid.uuid,
805 /* Consume grant space on the server.
806 * Unlike above, tgt_grant_rnb_size() is called with exp = NULL
807 * so that the required grant space isn't inflated. This is
808 * done on purpose since the server can deal with large block
809 * size, unlike some clients */
810 bytes = tgt_grant_rnb_size(NULL, lut, &rnb[i]);
812 /* if enough space, pretend it was granted */
815 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
819 /* We can't check for already-mapped blocks here (make sense
820 * when backend filesystem does not use COW) as it requires
821 * dropping the grant lock.
822 * Instead, we clear OBD_BRW_GRANTED and in that case we need
823 * to go through and verify if all of the blocks not marked
824 * BRW_GRANTED are already mapped and we can ignore this error.
826 rnb[i].rnb_flags &= ~OBD_BRW_GRANTED;
827 CDEBUG(D_CACHE, "%s: cli %s/%p idx %d no space for %d\n",
828 obd->obd_name, exp->exp_client_uuid.uuid, exp, i, bytes);
831 /* record in o_grant_used the actual space reserved for the I/O, will be
832 * used later in tgt_grant_commmit() */
833 oa->o_grant_used = granted + ungranted;
835 /* record space used for the I/O, will be used in tgt_grant_commmit() */
836 /* Now substract what the clients has used already. We don't subtract
837 * this from the tot_granted yet, so that other client's can't grab
838 * that space before we have actually allocated our blocks. That
839 * happens in tgt_grant_commit() after the writes are done. */
840 ted->ted_grant -= granted;
841 ted->ted_pending += oa->o_grant_used;
842 tgd->tgd_tot_granted += ungranted;
843 tgd->tgd_tot_pending += oa->o_grant_used;
846 "%s: cli %s/%p granted: %lu ungranted: %lu grant: %lu dirty: %lu"
847 "\n", obd->obd_name, exp->exp_client_uuid.uuid, exp,
848 granted, ungranted, ted->ted_grant, ted->ted_dirty);
850 if (obd->obd_recovering || (oa->o_valid & OBD_MD_FLGRANT) == 0)
851 /* don't update dirty accounting during recovery or
852 * if grant information got discarded (e.g. during resend) */
855 if (ted->ted_dirty < granted) {
856 CWARN("%s: cli %s/%p claims granted %lu > ted_dirty %lu\n",
857 obd->obd_name, exp->exp_client_uuid.uuid, exp,
858 granted, ted->ted_dirty);
859 granted = ted->ted_dirty;
861 tgd->tgd_tot_dirty -= granted;
862 ted->ted_dirty -= granted;
864 if (ted->ted_dirty < 0 || ted->ted_grant < 0 || ted->ted_pending < 0) {
865 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
866 obd->obd_name, exp->exp_client_uuid.uuid, exp,
867 ted->ted_dirty, ted->ted_pending, ted->ted_grant);
868 spin_unlock(&tgd->tgd_grant_lock);
875 * Allocate additional grant space to a client
877 * Calculate how much grant space to return to client, based on how much space
878 * is currently free and how much of that is already granted.
879 * Caller must hold tgd_grant_lock spinlock.
881 * \param[in] exp export of the client which sent the request
882 * \param[in] curgrant current grant claimed by the client
883 * \param[in] want how much grant space the client would like to
885 * \param[in] left remaining free space with granted space taken
887 * \param[in] conservative if set to true, the server should be cautious
888 * and limit how much space is granted back to the
889 * client. Otherwise, the server should try hard to
890 * satisfy the client request.
892 * \retval amount of grant space allocated
894 static long tgt_grant_alloc(struct obd_export *exp, u64 curgrant,
895 u64 want, u64 left, long chunk,
898 struct obd_device *obd = exp->exp_obd;
899 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
900 struct tg_export_data *ted = &exp->exp_target_data;
905 /* When tgd_grant_compat_disable is set, we don't grant any space to
906 * clients not supporting OBD_CONNECT_GRANT_PARAM.
907 * Otherwise, space granted to such a client is inflated since it
908 * consumes PAGE_SIZE of grant space per block */
909 if ((obd->obd_self_export != exp && !exp_grant_param_supp(exp) &&
910 tgd->tgd_grant_compat_disable) || left == 0 || exp->exp_failed)
913 if (want > OBD_MAX_GRANT) {
914 CERROR("%s: client %s/%p requesting > max (%lu), %llu\n",
915 obd->obd_name, exp->exp_client_uuid.uuid, exp,
916 OBD_MAX_GRANT, want);
920 /* Grant some fraction of the client's requested grant space so that
921 * they are not always waiting for write credits (not all of it to
922 * avoid overgranting in face of multiple RPCs in flight). This
923 * essentially will be able to control the OSC_MAX_RIF for a client.
925 * If we do have a large disparity between what the client thinks it
926 * has and what we think it has, don't grant very much and let the
927 * client consume its grant first. Either it just has lots of RPCs
928 * in flight, or it was evicted and its grants will soon be used up. */
929 if (curgrant >= want || curgrant >= ted->ted_grant + chunk)
932 if (obd->obd_recovering)
933 conservative = false;
936 /* don't grant more than 1/8th of the remaining free space in
939 grant = min(want - curgrant, left);
940 /* round grant up to the next block size */
941 grant = (grant + (1 << tgd->tgd_blockbits) - 1) &
942 ~((1ULL << tgd->tgd_blockbits) - 1);
947 /* Limit to grant_chunk if not reconnect/recovery */
948 if ((grant > chunk) && conservative)
952 * Limit grant so that export' grant does not exceed what the
953 * client would like to have by more than grants for 2 full
956 if (want + chunk <= ted->ted_grant)
958 if (ted->ted_grant + grant > want + chunk)
959 grant = want + chunk - ted->ted_grant;
961 tgd->tgd_tot_granted += grant;
962 ted->ted_grant += grant;
964 if (ted->ted_grant < 0) {
965 CERROR("%s: cli %s/%p grant %ld want %llu current %llu\n",
966 obd->obd_name, exp->exp_client_uuid.uuid, exp,
967 ted->ted_grant, want, curgrant);
968 spin_unlock(&tgd->tgd_grant_lock);
973 "%s: cli %s/%p wants: %llu current grant %llu"
974 " granting: %llu\n", obd->obd_name, exp->exp_client_uuid.uuid,
975 exp, want, curgrant, grant);
977 "%s: cli %s/%p tot cached:%llu granted:%llu"
978 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
979 exp, tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
980 obd->obd_num_exports);
986 * Handle grant space allocation on client connection & reconnection.
988 * A new non-readonly connection gets an initial grant allocation equals to
989 * tgt_grant_chunk() (i.e. twice the max BRW size in most of the cases).
990 * On reconnection, grant counters between client & target are resynchronized
991 * and additional space might be granted back if possible.
993 * \param[in] env LU environment provided by the caller
994 * \param[in] exp client's export which is (re)connecting
995 * \param[in,out] data obd_connect_data structure sent by the client in the
997 * \param[in] new_conn must set to true if this is a new connection and false
1000 void tgt_grant_connect(const struct lu_env *env, struct obd_export *exp,
1001 struct obd_connect_data *data, bool new_conn)
1003 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1004 struct tg_grants_data *tgd = &lut->lut_tgd;
1005 struct tg_export_data *ted = &exp->exp_target_data;
1010 int force = 0; /* can use cached data */
1012 /* don't grant space to client with read-only access */
1013 if (OCD_HAS_FLAG(data, RDONLY) ||
1014 (!OCD_HAS_FLAG(data, GRANT_PARAM) &&
1015 tgd->tgd_grant_compat_disable)) {
1016 data->ocd_grant = 0;
1017 data->ocd_connect_flags &= ~(OBD_CONNECT_GRANT |
1018 OBD_CONNECT_GRANT_PARAM);
1022 if (OCD_HAS_FLAG(data, GRANT_PARAM))
1023 want = data->ocd_grant;
1025 want = tgt_grant_inflate(tgd, data->ocd_grant);
1026 chunk = tgt_grant_chunk(exp, lut, data);
1028 tgt_grant_statfs(env, exp, force, &from_cache);
1030 spin_lock(&tgd->tgd_grant_lock);
1032 /* Grab free space from cached info and take out space already granted
1033 * to clients as well as reserved space */
1034 left = tgt_grant_space_left(exp);
1036 /* get fresh statfs data if we are short in ungranted space */
1037 if (from_cache && left < 32 * chunk) {
1038 spin_unlock(&tgd->tgd_grant_lock);
1039 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
1044 tgt_grant_alloc(exp, (u64)ted->ted_grant, want, left, chunk, new_conn);
1046 /* return to client its current grant */
1047 if (OCD_HAS_FLAG(data, GRANT_PARAM))
1048 data->ocd_grant = ted->ted_grant;
1051 data->ocd_grant = tgt_grant_deflate(tgd, (u64)ted->ted_grant);
1053 /* reset dirty accounting */
1054 tgd->tgd_tot_dirty -= ted->ted_dirty;
1057 if (new_conn && OCD_HAS_FLAG(data, GRANT))
1058 tgd->tgd_tot_granted_clients++;
1060 spin_unlock(&tgd->tgd_grant_lock);
1062 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: %llu left: %llu\n",
1063 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1064 exp, data->ocd_grant, want, left);
1068 EXPORT_SYMBOL(tgt_grant_connect);
1071 * Release all grant space attached to a given export.
1073 * Remove a client from the grant accounting totals. We also remove
1074 * the export from the obd device under the osfs and dev locks to ensure
1075 * that the tgt_grant_sanity_check() calculations are always valid.
1076 * The client should do something similar when it invalidates its import.
1078 * \param[in] exp client's export to remove from grant accounting
1080 void tgt_grant_discard(struct obd_export *exp)
1082 struct obd_device *obd = exp->exp_obd;
1083 struct lu_target *lut = class_exp2tgt(exp);
1084 struct tg_export_data *ted = &exp->exp_target_data;
1085 struct tg_grants_data *tgd;
1090 tgd = &lut->lut_tgd;
1091 spin_lock(&tgd->tgd_grant_lock);
1092 LASSERTF(tgd->tgd_tot_granted >= ted->ted_grant,
1093 "%s: tot_granted %llu cli %s/%p ted_grant %ld\n",
1094 obd->obd_name, tgd->tgd_tot_granted,
1095 exp->exp_client_uuid.uuid, exp, ted->ted_grant);
1096 tgd->tgd_tot_granted -= ted->ted_grant;
1098 LASSERTF(tgd->tgd_tot_pending >= ted->ted_pending,
1099 "%s: tot_pending %llu cli %s/%p ted_pending %ld\n",
1100 obd->obd_name, tgd->tgd_tot_pending,
1101 exp->exp_client_uuid.uuid, exp, ted->ted_pending);
1102 /* tgd_tot_pending is handled in tgt_grant_commit as bulk
1104 LASSERTF(tgd->tgd_tot_dirty >= ted->ted_dirty,
1105 "%s: tot_dirty %llu cli %s/%p ted_dirty %ld\n",
1106 obd->obd_name, tgd->tgd_tot_dirty,
1107 exp->exp_client_uuid.uuid, exp, ted->ted_dirty);
1108 tgd->tgd_tot_dirty -= ted->ted_dirty;
1110 spin_unlock(&tgd->tgd_grant_lock);
1112 EXPORT_SYMBOL(tgt_grant_discard);
1115 * Process grant information from incoming bulk read request.
1117 * Extract grant information packed in obdo structure (OBD_MD_FLGRANT set in
1118 * o_valid). Bulk reads usually comes with grant announcements (number of dirty
1119 * blocks, remaining amount of grant space, ...) and could also include a grant
1120 * shrink request. Unlike bulk write, no additional grant space is returned on
1121 * bulk read request.
1123 * \param[in] env is the lu environment provided by the caller
1124 * \param[in] exp is the export of the client which sent the request
1125 * \param[in,out] oa is the incoming obdo sent by the client
1127 void tgt_grant_prepare_read(const struct lu_env *env,
1128 struct obd_export *exp, struct obdo *oa)
1130 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1131 struct tg_grants_data *tgd = &lut->lut_tgd;
1140 if ((oa->o_valid & OBD_MD_FLGRANT) == 0)
1141 /* The read request does not contain any grant
1145 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1146 (oa->o_flags & OBD_FL_SHRINK_GRANT)) {
1147 /* To process grant shrink request, we need to know how much
1148 * available space remains on the backend filesystem.
1149 * Shrink requests are not so common, we always get fresh
1150 * statfs information. */
1151 tgt_grant_statfs(env, exp, 1, NULL);
1153 /* protect all grant counters */
1154 spin_lock(&tgd->tgd_grant_lock);
1156 /* Grab free space from cached statfs data and take out space
1157 * already granted to clients as well as reserved space */
1158 left = tgt_grant_space_left(exp);
1160 /* all set now to proceed with shrinking */
1163 /* no grant shrinking request packed in the obdo and
1164 * since we don't grant space back on reads, no point
1165 * in running statfs, so just skip it and process
1166 * incoming grant data directly. */
1167 spin_lock(&tgd->tgd_grant_lock);
1171 /* extract incoming grant information provided by the client and
1172 * inflate grant counters if required */
1173 tgt_grant_incoming(env, exp, oa, tgt_grant_chunk(exp, lut, NULL));
1175 /* unlike writes, we don't return grants back on reads unless a grant
1176 * shrink request was packed and we decided to turn it down. */
1178 tgt_grant_shrink(exp, oa, left);
1182 if (!exp_grant_param_supp(exp))
1183 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1184 spin_unlock(&tgd->tgd_grant_lock);
1187 EXPORT_SYMBOL(tgt_grant_prepare_read);
1190 * Process grant information from incoming bulk write request.
1192 * This function extracts client's grant announcements from incoming bulk write
1193 * request and attempts to allocate grant space for network buffers that need it
1194 * (i.e. OBD_BRW_FROM_GRANT not set in rnb_fags).
1195 * Network buffers which aren't granted the OBD_BRW_GRANTED flag should not
1196 * proceed further and should fail with -ENOSPC.
1197 * Whenever possible, additional grant space will be returned to the client
1198 * in the bulk write reply.
1199 * tgt_grant_prepare_write() must be called before writting any buffers to
1200 * the backend storage. This function works in pair with tgt_grant_commit()
1201 * which must be invoked once all buffers have been written to disk in order
1202 * to release space from the pending grant counter.
1204 * \param[in] env LU environment provided by the caller
1205 * \param[in] exp export of the client which sent the request
1206 * \param[in] oa incoming obdo sent by the client
1207 * \param[in] rnb list of network buffers
1208 * \param[in] niocount number of network buffers in the list
1210 void tgt_grant_prepare_write(const struct lu_env *env,
1211 struct obd_export *exp, struct obdo *oa,
1212 struct niobuf_remote *rnb, int niocount)
1214 struct obd_device *obd = exp->exp_obd;
1215 struct lu_target *lut = obd->u.obt.obt_lut;
1216 struct tg_grants_data *tgd = &lut->lut_tgd;
1219 int force = 0; /* can use cached data intially */
1220 long chunk = tgt_grant_chunk(exp, lut, NULL);
1225 /* get statfs information from OSD layer */
1226 tgt_grant_statfs(env, exp, force, &from_cache);
1228 spin_lock(&tgd->tgd_grant_lock); /* protect all grant counters */
1230 /* Grab free space from cached statfs data and take out space already
1231 * granted to clients as well as reserved space */
1232 left = tgt_grant_space_left(exp);
1234 /* Get fresh statfs data if we are short in ungranted space */
1235 if (from_cache && left < 32 * chunk) {
1236 spin_unlock(&tgd->tgd_grant_lock);
1237 CDEBUG(D_CACHE, "%s: fs has no space left and statfs too old\n",
1243 /* When close to free space exhaustion, trigger a sync to force
1244 * writeback cache to consume required space immediately and release as
1245 * much space as possible. */
1246 if (!obd->obd_recovering && force != 2 && left < chunk) {
1247 bool from_grant = true;
1250 /* That said, it is worth running a sync only if some pages did
1251 * not consume grant space on the client and could thus fail
1252 * with ENOSPC later in tgt_grant_check() */
1253 for (i = 0; i < niocount; i++)
1254 if (!(rnb[i].rnb_flags & OBD_BRW_FROM_GRANT))
1258 /* at least one network buffer requires acquiring grant
1259 * space on the server */
1260 spin_unlock(&tgd->tgd_grant_lock);
1261 /* discard errors, at least we tried ... */
1262 dt_sync(env, lut->lut_bottom);
1268 /* extract incoming grant information provided by the client,
1269 * and inflate grant counters if required */
1270 tgt_grant_incoming(env, exp, oa, chunk);
1273 tgt_grant_check(env, exp, oa, rnb, niocount, &left);
1275 if (!(oa->o_valid & OBD_MD_FLGRANT)) {
1276 spin_unlock(&tgd->tgd_grant_lock);
1280 /* if OBD_FL_SHRINK_GRANT is set, the client is willing to release some
1282 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1283 (oa->o_flags & OBD_FL_SHRINK_GRANT))
1284 tgt_grant_shrink(exp, oa, left);
1286 /* grant more space back to the client if possible */
1287 oa->o_grant = tgt_grant_alloc(exp, oa->o_grant, oa->o_undirty,
1290 if (!exp_grant_param_supp(exp))
1291 oa->o_grant = tgt_grant_deflate(tgd, oa->o_grant);
1292 spin_unlock(&tgd->tgd_grant_lock);
1295 EXPORT_SYMBOL(tgt_grant_prepare_write);
1298 * Consume grant space reserved for object creation.
1300 * Grant space is allocated to the local self export for object precreation.
1301 * This is required to prevent object precreation from consuming grant space
1302 * allocated to client nodes for the data writeback cache.
1303 * This function consumes enough space to create \a nr objects and allocates
1304 * more grant space to the self export for future precreation requests, if
1307 * \param[in] env LU environment provided by the caller
1308 * \param[in] exp export holding the grant space for precreation (= self
1310 * \param[in] nr number of objects to be created
1312 * \retval >= 0 amount of grant space allocated to the precreate request
1313 * \retval -ENOSPC on failure
1315 long tgt_grant_create(const struct lu_env *env, struct obd_export *exp, s64 *nr)
1317 struct lu_target *lut = exp->exp_obd->u.obt.obt_lut;
1318 struct tg_grants_data *tgd = &lut->lut_tgd;
1319 struct tg_export_data *ted = &exp->exp_target_data;
1321 unsigned long wanted;
1322 unsigned long granted;
1325 if (exp->exp_obd->obd_recovering ||
1326 lut->lut_dt_conf.ddp_inodespace == 0)
1327 /* don't enforce grant during recovery */
1330 /* Update statfs data if required */
1331 tgt_grant_statfs(env, exp, 1, NULL);
1333 /* protect all grant counters */
1334 spin_lock(&tgd->tgd_grant_lock);
1336 /* fail precreate request if there is not enough blocks available for
1338 if (tgd->tgd_osfs.os_bavail - (ted->ted_grant >> tgd->tgd_blockbits) <
1339 (tgd->tgd_osfs.os_blocks >> 10)) {
1340 spin_unlock(&tgd->tgd_grant_lock);
1341 CDEBUG(D_RPCTRACE, "%s: not enough space for create %llu\n",
1342 exp->exp_obd->obd_name,
1343 tgd->tgd_osfs.os_bavail * tgd->tgd_osfs.os_blocks);
1347 /* Grab free space from cached statfs data and take out space
1348 * already granted to clients as well as reserved space */
1349 left = tgt_grant_space_left(exp);
1351 /* compute how much space is required to handle the precreation
1353 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1354 if (wanted > ted->ted_grant + left) {
1355 /* that's beyond what remains, adjust the number of objects that
1356 * can be safely precreated */
1357 wanted = ted->ted_grant + left;
1358 *nr = wanted / lut->lut_dt_conf.ddp_inodespace;
1360 /* we really have no space any more for precreation,
1361 * fail the precreate request with ENOSPC */
1362 spin_unlock(&tgd->tgd_grant_lock);
1365 /* compute space needed for the new number of creations */
1366 wanted = *nr * lut->lut_dt_conf.ddp_inodespace;
1368 LASSERT(wanted <= ted->ted_grant + left);
1370 if (wanted <= ted->ted_grant) {
1371 /* we've enough grant space to handle this precreate request */
1372 ted->ted_grant -= wanted;
1374 /* we need to take some space from the ungranted pool */
1375 tgd->tgd_tot_granted += wanted - ted->ted_grant;
1376 left -= wanted - ted->ted_grant;
1380 ted->ted_pending += granted;
1381 tgd->tgd_tot_pending += granted;
1383 /* grant more space for precreate purpose if possible. */
1384 wanted = OST_MAX_PRECREATE * lut->lut_dt_conf.ddp_inodespace / 2;
1385 if (wanted > ted->ted_grant) {
1388 /* always try to book enough space to handle a large precreate
1390 chunk = tgt_grant_chunk(exp, lut, NULL);
1391 wanted -= ted->ted_grant;
1392 tgt_grant_alloc(exp, ted->ted_grant, wanted, left, chunk,
1395 spin_unlock(&tgd->tgd_grant_lock);
1398 EXPORT_SYMBOL(tgt_grant_create);
1401 * Release grant space added to the pending counter by tgt_grant_prepare_write()
1403 * Update pending grant counter once buffers have been written to the disk.
1405 * \param[in] exp export of the client which sent the request
1406 * \param[in] pending amount of reserved space to be released
1407 * \param[in] rc return code of pre-commit operations
1409 void tgt_grant_commit(struct obd_export *exp, unsigned long pending,
1412 struct tg_grants_data *tgd = &exp->exp_obd->u.obt.obt_lut->lut_tgd;
1416 /* get space accounted in tot_pending for the I/O, set in
1417 * tgt_grant_check() */
1421 spin_lock(&tgd->tgd_grant_lock);
1422 /* Don't update statfs data for errors raised before commit (e.g.
1423 * bulk transfer failed, ...) since we know those writes have not been
1424 * processed. For other errors hit during commit, we cannot really tell
1425 * whether or not something was written, so we update statfs data.
1426 * In any case, this should not be fatal since we always get fresh
1427 * statfs data before failing a request with ENOSPC */
1429 spin_lock(&tgd->tgd_osfs_lock);
1430 /* Take pending out of cached statfs data */
1431 tgd->tgd_osfs.os_bavail -= min_t(u64,
1432 tgd->tgd_osfs.os_bavail,
1433 pending >> tgd->tgd_blockbits);
1434 if (tgd->tgd_statfs_inflight)
1435 /* someone is running statfs and want to be notified of
1436 * writes happening meanwhile */
1437 tgd->tgd_osfs_inflight += pending;
1438 spin_unlock(&tgd->tgd_osfs_lock);
1441 if (exp->exp_target_data.ted_pending < pending) {
1442 CERROR("%s: cli %s/%p ted_pending(%lu) < grant_used(%lu)\n",
1443 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1444 exp->exp_target_data.ted_pending, pending);
1445 spin_unlock(&tgd->tgd_grant_lock);
1448 exp->exp_target_data.ted_pending -= pending;
1450 if (tgd->tgd_tot_granted < pending) {
1451 CERROR("%s: cli %s/%p tot_granted(%llu) < grant_used(%lu)\n",
1452 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1453 tgd->tgd_tot_granted, pending);
1454 spin_unlock(&tgd->tgd_grant_lock);
1457 tgd->tgd_tot_granted -= pending;
1459 if (tgd->tgd_tot_pending < pending) {
1460 CERROR("%s: cli %s/%p tot_pending(%llu) < grant_used(%lu)\n",
1461 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1462 tgd->tgd_tot_pending, pending);
1463 spin_unlock(&tgd->tgd_grant_lock);
1466 tgd->tgd_tot_pending -= pending;
1467 spin_unlock(&tgd->tgd_grant_lock);
1470 EXPORT_SYMBOL(tgt_grant_commit);
1472 struct tgt_grant_cb {
1473 /* commit callback structure */
1474 struct dt_txn_commit_cb tgc_cb;
1475 /* export associated with the bulk write */
1476 struct obd_export *tgc_exp;
1477 /* pending grant to be released */
1478 unsigned long tgc_granted;
1482 * Callback function for grant releasing
1484 * Release grant space reserved by the client node.
1486 * \param[in] env execution environment
1487 * \param[in] th transaction handle
1488 * \param[in] cb callback data
1489 * \param[in] err error code
1491 static void tgt_grant_commit_cb(struct lu_env *env, struct thandle *th,
1492 struct dt_txn_commit_cb *cb, int err)
1494 struct tgt_grant_cb *tgc;
1496 tgc = container_of(cb, struct tgt_grant_cb, tgc_cb);
1498 tgt_grant_commit(tgc->tgc_exp, tgc->tgc_granted, err);
1499 class_export_cb_put(tgc->tgc_exp);
1504 * Add callback for grant releasing
1506 * Register a commit callback to release grant space.
1508 * \param[in] th transaction handle
1509 * \param[in] exp OBD export of client
1510 * \param[in] granted amount of grant space to be released upon commit
1512 * \retval 0 on successful callback adding
1513 * \retval negative value on error
1515 int tgt_grant_commit_cb_add(struct thandle *th, struct obd_export *exp,
1516 unsigned long granted)
1518 struct tgt_grant_cb *tgc;
1519 struct dt_txn_commit_cb *dcb;
1527 tgc->tgc_exp = class_export_cb_get(exp);
1528 tgc->tgc_granted = granted;
1531 dcb->dcb_func = tgt_grant_commit_cb;
1532 INIT_LIST_HEAD(&dcb->dcb_linkage);
1533 strlcpy(dcb->dcb_name, "tgt_grant_commit_cb", sizeof(dcb->dcb_name));
1535 rc = dt_trans_cb_add(th, dcb);
1537 class_export_cb_put(tgc->tgc_exp);
1543 EXPORT_SYMBOL(tgt_grant_commit_cb_add);
1546 * Show estimate of total amount of dirty data on clients.
1548 * @kobj kobject embedded in obd_device
1550 * @buf buf used by sysfs to print out data
1552 * Return: 0 on success
1553 * negative value on error
1555 ssize_t tot_dirty_show(struct kobject *kobj, struct attribute *attr,
1558 struct obd_device *obd = container_of(kobj, struct obd_device,
1560 struct tg_grants_data *tgd;
1562 tgd = &obd->u.obt.obt_lut->lut_tgd;
1563 return scnprintf(buf, PAGE_SIZE, "%llu\n", tgd->tgd_tot_dirty);
1565 EXPORT_SYMBOL(tot_dirty_show);
1568 * Show total amount of space granted to clients.
1570 * @kobj kobject embedded in obd_device
1572 * @buf buf used by sysfs to print out data
1574 * Return: 0 on success
1575 * negative value on error
1577 ssize_t tot_granted_show(struct kobject *kobj, struct attribute *attr,
1580 struct obd_device *obd = container_of(kobj, struct obd_device,
1582 struct tg_grants_data *tgd;
1584 tgd = &obd->u.obt.obt_lut->lut_tgd;
1585 return scnprintf(buf, PAGE_SIZE, "%llu\n", tgd->tgd_tot_granted);
1587 EXPORT_SYMBOL(tot_granted_show);
1590 * Show total amount of space used by IO in progress.
1592 * @kobj kobject embedded in obd_device
1594 * @buf buf used by sysfs to print out data
1596 * Return: 0 on success
1597 * negative value on error
1599 ssize_t tot_pending_show(struct kobject *kobj, struct attribute *attr,
1602 struct obd_device *obd = container_of(kobj, struct obd_device,
1604 struct tg_grants_data *tgd;
1606 tgd = &obd->u.obt.obt_lut->lut_tgd;
1607 return scnprintf(buf, PAGE_SIZE, "%llu\n", tgd->tgd_tot_pending);
1609 EXPORT_SYMBOL(tot_pending_show);
1612 * Show if grants compatibility mode is disabled.
1614 * When tgd_grant_compat_disable is set, we don't grant any space to clients
1615 * not supporting OBD_CONNECT_GRANT_PARAM. Otherwise, space granted to such
1616 * a client is inflated since it consumes PAGE_SIZE of grant space per
1617 * block, (i.e. typically 4kB units), but underlaying file system might have
1618 * block size bigger than page size, e.g. ZFS. See LU-2049 for details.
1620 * @kobj kobject embedded in obd_device
1622 * @buf buf used by sysfs to print out data
1624 * Return: string length of @buf output on success
1626 ssize_t grant_compat_disable_show(struct kobject *kobj, struct attribute *attr,
1629 struct obd_device *obd = container_of(kobj, struct obd_device,
1631 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1633 return scnprintf(buf, PAGE_SIZE, "%u\n", tgd->tgd_grant_compat_disable);
1635 EXPORT_SYMBOL(grant_compat_disable_show);
1638 * Change grant compatibility mode.
1640 * Setting tgd_grant_compat_disable prohibit any space granting to clients
1641 * not supporting OBD_CONNECT_GRANT_PARAM. See details above.
1643 * @kobj kobject embedded in obd_device
1645 * @buffer string which represents mode
1646 * 1: disable compatibility mode
1647 * 0: enable compatibility mode
1648 * @count @buffer length
1650 * Return: @count on success
1651 * negative number on error
1653 ssize_t grant_compat_disable_store(struct kobject *kobj,
1654 struct attribute *attr,
1655 const char *buffer, size_t count)
1657 struct obd_device *obd = container_of(kobj, struct obd_device,
1659 struct tg_grants_data *tgd = &obd->u.obt.obt_lut->lut_tgd;
1663 rc = kstrtobool(buffer, &val);
1667 tgd->tgd_grant_compat_disable = val;
1671 EXPORT_SYMBOL(grant_compat_disable_store);