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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ofd/ofd_grant.c
38 * Author: Johann Lombardi <johann@whamcloud..com>
41 #define DEBUG_SUBSYSTEM S_FILTER
43 #include "ofd_internal.h"
45 /* At least enough to send a couple of 1MB RPCs, even if not max sized */
46 #define OFD_GRANT_CHUNK (2ULL * DT_MAX_BRW_SIZE)
48 /* Clients typically hold 2x their max_rpcs_in_flight of grant space */
49 #define OFD_GRANT_SHRINK_LIMIT(exp) (2ULL * 8 * exp_max_brw_size(exp))
51 static inline obd_size ofd_grant_from_cli(struct obd_export *exp,
52 struct ofd_device *ofd, obd_size val)
54 if (ofd_grant_compat(exp, ofd))
55 /* clients not supporting OBD_CONNECT_GRANT_PARAM actually
56 * consume 4KB of grant per block, we should thus inflate
57 * the grant counters to reflect what was actually consumed */
58 return val << (ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT);
62 static inline obd_size ofd_grant_to_cli(struct obd_export *exp,
63 struct ofd_device *ofd, obd_size val)
65 if (ofd_grant_compat(exp, ofd))
66 return val >> (ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT);
70 static inline obd_size ofd_grant_chunk(struct obd_export *exp,
71 struct ofd_device *ofd)
73 if (ofd_obd(ofd)->obd_self_export == exp)
74 /* Grant enough space to handle a big precreate request */
75 return OST_MAX_PRECREATE * ofd->ofd_dt_conf.ddp_inodespace;
77 if (ofd_grant_compat(exp, ofd))
78 /* Try to grant enough space to send a full-size RPC */
79 return exp_max_brw_size(exp) <<
80 (ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT);
82 /* Try to return enough to send two full RPCs, if needed */
83 return exp_max_brw_size(exp) * 2;
87 * Perform extra sanity checks for grant accounting. This is done at connect,
88 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
89 * always get rid of it or turn it off when we know accounting is good.
91 * \param obd - is the device to check
92 * \param func - is the function to call if an inconsistency is found
94 void ofd_grant_sanity_check(struct obd_device *obd, const char *func)
96 struct filter_export_data *fed;
97 struct ofd_device *ofd = ofd_dev(obd->obd_lu_dev);
98 struct obd_export *exp;
100 obd_size tot_dirty = 0;
101 obd_size tot_pending = 0;
102 obd_size tot_granted = 0;
103 obd_size fo_tot_dirty, fo_tot_pending;
104 obd_size fo_tot_granted;
106 if (cfs_list_empty(&obd->obd_exports))
109 /* We don't want to do this for large machines that do lots of
110 * mounts or unmounts. It burns... */
111 if (obd->obd_num_exports > 100)
114 maxsize = ofd->ofd_osfs.os_blocks << ofd->ofd_blockbits;
116 spin_lock(&obd->obd_dev_lock);
117 spin_lock(&ofd->ofd_grant_lock);
118 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
121 fed = &exp->exp_filter_data;
123 if (obd->obd_self_export == exp)
124 CDEBUG(D_CACHE, "%s: processing self export: %ld %ld "
125 "%ld\n", obd->obd_name, fed->fed_grant,
126 fed->fed_pending, fed->fed_dirty);
128 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
131 if (fed->fed_grant + fed->fed_pending > maxsize) {
132 CERROR("%s: cli %s/%p fed_grant(%ld) + fed_pending(%ld)"
133 " > maxsize("LPU64")\n", obd->obd_name,
134 exp->exp_client_uuid.uuid, exp, fed->fed_grant,
135 fed->fed_pending, maxsize);
136 spin_unlock(&obd->obd_dev_lock);
137 spin_unlock(&ofd->ofd_grant_lock);
140 if (fed->fed_dirty > maxsize) {
141 CERROR("%s: cli %s/%p fed_dirty(%ld) > maxsize("LPU64
142 ")\n", obd->obd_name, exp->exp_client_uuid.uuid,
143 exp, fed->fed_dirty, maxsize);
144 spin_unlock(&obd->obd_dev_lock);
145 spin_unlock(&ofd->ofd_grant_lock);
148 CDEBUG_LIMIT(error ? D_ERROR : D_CACHE, "%s: cli %s/%p dirty "
149 "%ld pend %ld grant %ld\n", obd->obd_name,
150 exp->exp_client_uuid.uuid, exp, fed->fed_dirty,
151 fed->fed_pending, fed->fed_grant);
152 tot_granted += fed->fed_grant + fed->fed_pending;
153 tot_pending += fed->fed_pending;
154 tot_dirty += fed->fed_dirty;
156 spin_unlock(&obd->obd_dev_lock);
157 fo_tot_granted = ofd->ofd_tot_granted;
158 fo_tot_pending = ofd->ofd_tot_pending;
159 fo_tot_dirty = ofd->ofd_tot_dirty;
161 if (tot_granted != fo_tot_granted)
162 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
163 func, tot_granted, fo_tot_granted);
164 if (tot_pending != fo_tot_pending)
165 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
166 func, tot_pending, fo_tot_pending);
167 if (tot_dirty != fo_tot_dirty)
168 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
169 func, tot_dirty, fo_tot_dirty);
170 if (tot_pending > tot_granted)
171 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
172 func, tot_pending, tot_granted);
173 if (tot_granted > maxsize)
174 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
175 func, tot_granted, maxsize);
176 if (tot_dirty > maxsize)
177 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
178 func, tot_dirty, maxsize);
179 spin_unlock(&ofd->ofd_grant_lock);
183 * Get fresh statfs information from the OSD layer if the cache is older than 1s
184 * or if force is set. The OSD layer is in charge of estimating data & metadata
187 * \param env - is the lu environment passed by the caller
188 * \param exp - export used to print client info in debug messages
189 * \param force - is used to force a refresh of statfs information
190 * \param from_cache - returns whether the statfs information are
193 static void ofd_grant_statfs(const struct lu_env *env, struct obd_export *exp,
194 int force, int *from_cache)
196 struct obd_device *obd = exp->exp_obd;
197 struct ofd_device *ofd = ofd_exp(exp);
198 struct obd_statfs *osfs = &ofd_info(env)->fti_u.osfs;
203 max_age = 0; /* get fresh statfs data */
205 max_age = cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS);
207 rc = ofd_statfs_internal(env, ofd, osfs, max_age, from_cache);
213 CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64"\n",
214 obd->obd_name, exp->exp_client_uuid.uuid, exp,
215 osfs->os_bfree << ofd->ofd_blockbits,
216 osfs->os_bavail << ofd->ofd_blockbits);
220 * Figure out how much space is available on the backend filesystem.
221 * This is done by accessing cached statfs data previously populated by
222 * ofd_grant_statfs(), from which we withdraw the space already granted to
223 * clients and the reserved space.
225 * \param exp - export which received the write request
227 static obd_size ofd_grant_space_left(struct obd_export *exp)
229 struct obd_device *obd = exp->exp_obd;
230 struct ofd_device *ofd = ofd_exp(exp);
231 obd_size tot_granted;
232 obd_size left, avail;
236 LASSERT_SPIN_LOCKED(&ofd->ofd_grant_lock);
238 spin_lock(&ofd->ofd_osfs_lock);
239 /* get available space from cached statfs data */
240 left = ofd->ofd_osfs.os_bavail << ofd->ofd_blockbits;
241 unstable = ofd->ofd_osfs_unstable; /* those might be accounted twice */
242 spin_unlock(&ofd->ofd_osfs_lock);
244 tot_granted = ofd->ofd_tot_granted;
246 if (left < tot_granted) {
247 int mask = (left + unstable <
248 tot_granted - ofd->ofd_tot_pending) ?
251 CDEBUG_LIMIT(mask, "%s: cli %s/%p left "LPU64" < tot_grant "
252 LPU64" unstable "LPU64" pending "LPU64"\n",
253 obd->obd_name, exp->exp_client_uuid.uuid, exp,
254 left, tot_granted, unstable,
255 ofd->ofd_tot_pending);
260 /* Withdraw space already granted to clients */
263 /* If the left space is below the grant threshold x available space,
264 * stop granting space to clients.
265 * The purpose of this threshold is to keep some error margin on the
266 * overhead estimate made by the OSD layer. If we grant all the free
267 * space, we have no way (grant space cannot be revoked yet) to
268 * adjust if the write overhead has been underestimated. */
269 left -= min_t(obd_size, left, ofd_grant_reserved(ofd, avail));
271 /* Align left on block size */
272 left &= ~((1ULL << ofd->ofd_blockbits) - 1);
274 CDEBUG(D_CACHE, "%s: cli %s/%p avail "LPU64" left "LPU64" unstable "
275 LPU64" tot_grant "LPU64" pending "LPU64"\n", obd->obd_name,
276 exp->exp_client_uuid.uuid, exp, avail, left, unstable,
277 tot_granted, ofd->ofd_tot_pending);
283 * Grab the dirty and seen grant announcements from the incoming obdo.
284 * We will later calculate the client's new grant and return it.
285 * Caller must hold ofd_grant_lock spinlock.
287 * \param env - is the lu environment supplying osfs storage
288 * \param exp - is the export for which we received the request
289 * \paral oa - is the incoming obdo sent by the client
291 static void ofd_grant_incoming(const struct lu_env *env, struct obd_export *exp,
294 struct filter_export_data *fed;
295 struct ofd_device *ofd = ofd_exp(exp);
296 struct obd_device *obd = exp->exp_obd;
297 long dirty, dropped, grant_chunk;
300 LASSERT_SPIN_LOCKED(&ofd->ofd_grant_lock);
302 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
303 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
304 oa->o_valid &= ~OBD_MD_FLGRANT;
308 fed = &exp->exp_filter_data;
310 /* Add some margin, since there is a small race if other RPCs arrive
311 * out-or-order and have already consumed some grant. We want to
312 * leave this here in case there is a large error in accounting. */
314 "%s: cli %s/%p reports grant "LPU64" dropped %u, local %lu\n",
315 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
316 oa->o_dropped, fed->fed_grant);
318 if ((long long)oa->o_dirty < 0)
321 dirty = ofd_grant_from_cli(exp, ofd, oa->o_dirty);
322 dropped = ofd_grant_from_cli(exp, ofd, (obd_size)oa->o_dropped);
323 grant_chunk = ofd_grant_chunk(exp, ofd);
325 /* Update our accounting now so that statfs takes it into account.
326 * Note that fed_dirty is only approximate and can become incorrect
327 * if RPCs arrive out-of-order. No important calculations depend
328 * on fed_dirty however, but we must check sanity to not assert. */
329 if (dirty > fed->fed_grant + 4 * grant_chunk)
330 dirty = fed->fed_grant + 4 * grant_chunk;
331 ofd->ofd_tot_dirty += dirty - fed->fed_dirty;
332 if (fed->fed_grant < dropped) {
334 "%s: cli %s/%p reports %lu dropped > grant %lu\n",
335 obd->obd_name, exp->exp_client_uuid.uuid, exp, dropped,
339 if (ofd->ofd_tot_granted < dropped) {
340 CERROR("%s: cli %s/%p reports %lu dropped > tot_grant "LPU64
341 "\n", obd->obd_name, exp->exp_client_uuid.uuid, exp,
342 dropped, ofd->ofd_tot_granted);
345 ofd->ofd_tot_granted -= dropped;
346 fed->fed_grant -= dropped;
347 fed->fed_dirty = dirty;
349 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
350 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
351 obd->obd_name, exp->exp_client_uuid.uuid, exp,
352 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
353 spin_unlock(&ofd->ofd_grant_lock);
360 * Called when the client is able to release some grants. Proceed with the
361 * shrink request when there is less ungranted space remaining
362 * than the amount all of the connected clients would consume if they
363 * used their full grant.
365 * \param exp - is the export for which we received the request
366 * \paral oa - is the incoming obdo sent by the client
367 * \param left_space - is the remaining free space with space already granted
370 static void ofd_grant_shrink(struct obd_export *exp,
371 struct obdo *oa, obd_size left_space)
373 struct filter_export_data *fed;
374 struct ofd_device *ofd = ofd_exp(exp);
375 struct obd_device *obd = exp->exp_obd;
378 LASSERT_SPIN_LOCKED(&ofd->ofd_grant_lock);
380 if (left_space >= ofd->ofd_tot_granted_clients *
381 OFD_GRANT_SHRINK_LIMIT(exp))
384 grant_shrink = ofd_grant_from_cli(exp, ofd, oa->o_grant);
386 fed = &exp->exp_filter_data;
387 fed->fed_grant -= grant_shrink;
388 ofd->ofd_tot_granted -= grant_shrink;
390 CDEBUG(D_CACHE, "%s: cli %s/%p shrink %ld fed_grant %ld total "
391 LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
392 exp, grant_shrink, fed->fed_grant, ofd->ofd_tot_granted);
394 /* client has just released some grant, don't grant any space back */
399 * Calculate how much space is required to write a given network buffer
401 static inline int ofd_grant_rnb_size(struct obd_export *exp,
402 struct ofd_device *ofd,
403 struct niobuf_remote *rnb)
405 obd_size blocksize, bytes, end;
407 if (exp && ofd_grant_compat(exp, ofd))
408 blocksize = 1ULL << COMPAT_BSIZE_SHIFT;
410 blocksize = 1ULL << ofd->ofd_blockbits;
412 /* The network buffer might span several blocks, align it on block
414 bytes = rnb->rnb_offset & (blocksize - 1);
415 bytes += rnb->rnb_len;
416 end = bytes & (blocksize - 1);
418 bytes += blocksize - end;
420 /* Apply per-export pecularities if one is given */
421 bytes = ofd_grant_from_cli(exp, ofd, (obd_size)bytes);
427 * When clients have dirtied as much space as they've been granted they
428 * fall through to sync writes. These sync writes haven't been expressed
429 * in grants and need to error with ENOSPC when there isn't room in the
430 * filesystem for them after grants are taken into account. However,
431 * writeback of the dirty data that was already granted space can write
433 * Caller must hold ofd_grant_lock spinlock.
435 * \param env - is the lu environment passed by the caller
436 * \param exp - is the export identifying the client which sent the RPC
437 * \param oa - is the incoming obdo in which we should return the pack the
439 * \param rnb - is the list of network buffers
440 * \param niocont - is the number of network buffers in the list
441 * \param left - is the remaining free space with space already granted
444 static void ofd_grant_check(const struct lu_env *env, struct obd_export *exp,
445 struct obdo *oa, struct niobuf_remote *rnb,
446 int niocount, obd_size *left)
448 struct filter_export_data *fed = &exp->exp_filter_data;
449 struct obd_device *obd = exp->exp_obd;
450 struct ofd_device *ofd = ofd_exp(exp);
451 unsigned long ungranted = 0;
452 unsigned long granted = 0;
455 struct ofd_thread_info *info = ofd_info(env);
459 LASSERT_SPIN_LOCKED(&ofd->ofd_grant_lock);
461 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462 (oa->o_flags & OBD_FL_RECOV_RESEND)) {
464 CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
468 for (i = 0; i < niocount; i++) {
471 if (obd->obd_recovering) {
472 /* Replaying write. Grant info have been processed
473 * already so no need to do any enforcement here.
474 * It is worth noting that only bulk writes with all
475 * rnbs having OBD_BRW_FROM_GRANT can be replayed.
476 * If one page hasn't OBD_BRW_FROM_GRANT set, then
477 * the whole bulk is written synchronously */
478 if (rnb[i].rnb_flags & OBD_BRW_FROM_GRANT) {
479 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
482 CERROR("%s: cli %s is replaying OST_WRITE "
483 "while one rnb hasn't OBD_BRW_FROM_GRANT"
484 " set (0x%x)\n", exp->exp_obd->obd_name,
485 exp->exp_client_uuid.uuid,
489 } else if ((oa->o_valid & OBD_MD_FLGRANT) &&
490 (rnb[i].rnb_flags & OBD_BRW_FROM_GRANT)) {
492 /* This is a recoverable resend so grant
493 * information have already been processed */
494 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
498 /* inflate consumed space if needed */
499 bytes = ofd_grant_rnb_size(exp, ofd, &rnb[i]);
500 if (fed->fed_grant < granted + bytes) {
501 CDEBUG(D_CACHE, "%s: cli %s/%p claims %ld+%d "
502 "GRANT, real grant %lu idx %d\n",
503 exp->exp_obd->obd_name,
504 exp->exp_client_uuid.uuid, exp,
505 granted, bytes, fed->fed_grant, i);
508 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
513 /* Consume grant space on the server.
514 * Unlike above, ofd_grant_rnb_size() is called with exp = NULL
515 * so that the required grant space isn't inflated. This is
516 * done on purpose since the server can deal with large block
517 * size, unlike some clients */
518 bytes = ofd_grant_rnb_size(NULL, ofd, &rnb[i]);
519 if (*left > ungranted + bytes) {
520 /* if enough space, pretend it was granted */
522 rnb[i].rnb_flags |= OBD_BRW_GRANTED;
526 /* We can't check for already-mapped blocks here (make sense
527 * when backend filesystem does not use COW) as it requires
528 * dropping the grant lock.
529 * Instead, we clear ~OBD_BRW_GRANTED and in that case we need
530 * to go through and verify if all of the blocks not marked
531 * BRW_GRANTED are already mapped and we can ignore this error.
533 rnb[i].rnb_flags &= ~OBD_BRW_GRANTED;
534 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
535 exp->exp_obd->obd_name,
536 exp->exp_client_uuid.uuid, exp, i, bytes);
539 /* record space used for the I/O, will be used in ofd_grant_commmit() */
540 /* Now substract what the clients has used already. We don't subtract
541 * this from the tot_granted yet, so that other client's can't grab
542 * that space before we have actually allocated our blocks. That
543 * happens in ofd_grant_commit() after the writes are done. */
544 info->fti_used = granted + ungranted;
546 fed->fed_grant -= granted;
547 fed->fed_pending += info->fti_used;
548 ofd->ofd_tot_granted += ungranted;
549 ofd->ofd_tot_pending += info->fti_used;
552 "%s: cli %s/%p granted: %lu ungranted: %lu grant: %lu dirty: %lu"
553 "\n", obd->obd_name, exp->exp_client_uuid.uuid, exp,
554 granted, ungranted, fed->fed_grant, fed->fed_dirty);
556 if (obd->obd_recovering)
557 /* don't update dirty accounting during recovery */
560 if (fed->fed_dirty < granted) {
561 CWARN("%s: cli %s/%p claims granted %lu > fed_dirty %lu\n",
562 obd->obd_name, exp->exp_client_uuid.uuid, exp,
563 granted, fed->fed_dirty);
564 granted = fed->fed_dirty;
566 ofd->ofd_tot_dirty -= granted;
567 fed->fed_dirty -= granted;
569 if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
570 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
571 obd->obd_name, exp->exp_client_uuid.uuid, exp,
572 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
573 spin_unlock(&ofd->ofd_grant_lock);
580 * Calculate how much grant space to return to client, based on how much space
581 * is currently free and how much of that is already granted.
582 * Caller must hold ofd_grant_lock spinlock.
584 * \param exp - is the export of the client which sent the request
585 * \param curgrant - is the current grant claimed by the client
586 * \param want - is how much grant space the client would like to have
587 * \param left - is the remaining free space with granted space taken out
589 static long ofd_grant(struct obd_export *exp, obd_size curgrant,
590 obd_size want, obd_size left)
592 struct obd_device *obd = exp->exp_obd;
593 struct ofd_device *ofd = ofd_exp(exp);
594 struct filter_export_data *fed = &exp->exp_filter_data;
600 if (ofd_grant_prohibit(exp, ofd) || left == 0 || exp->exp_failed)
603 if (want > 0x7fffffff) {
604 CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
605 obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
609 /* client not supporting OBD_CONNECT_GRANT_PARAM works with a 4KB block
610 * size while the reality is different */
611 curgrant = ofd_grant_from_cli(exp, ofd, curgrant);
612 want = ofd_grant_from_cli(exp, ofd, want);
613 grant_chunk = ofd_grant_chunk(exp, ofd);
615 /* Grant some fraction of the client's requested grant space so that
616 * they are not always waiting for write credits (not all of it to
617 * avoid overgranting in face of multiple RPCs in flight). This
618 * essentially will be able to control the OSC_MAX_RIF for a client.
620 * If we do have a large disparity between what the client thinks it
621 * has and what we think it has, don't grant very much and let the
622 * client consume its grant first. Either it just has lots of RPCs
623 * in flight, or it was evicted and its grants will soon be used up. */
624 if (curgrant >= want || curgrant >= fed->fed_grant + grant_chunk)
627 if (!obd->obd_recovering)
628 /* don't grant more than 1/8th of the remaining free space in
631 grant = min(want, left);
632 /* align grant on block size */
633 grant &= ~((1ULL << ofd->ofd_blockbits) - 1);
638 /* Limit to ofd_grant_chunk() if client is not reconnecting */
639 if ((grant > grant_chunk) && (!obd->obd_recovering))
642 ofd->ofd_tot_granted += grant;
643 fed->fed_grant += grant;
645 if (fed->fed_grant < 0) {
646 CERROR("%s: cli %s/%p grant %ld want "LPU64" current "LPU64"\n",
647 obd->obd_name, exp->exp_client_uuid.uuid, exp,
648 fed->fed_grant, want, curgrant);
649 spin_unlock(&ofd->ofd_grant_lock);
654 "%s: cli %s/%p wants: "LPU64" current grant "LPU64
655 " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
656 exp, want, curgrant, grant);
658 "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
659 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
660 exp, ofd->ofd_tot_dirty, ofd->ofd_tot_granted,
661 obd->obd_num_exports);
663 RETURN(ofd_grant_to_cli(exp, ofd, grant));
667 * Client connection or reconnection.
669 * \param env - is the lu environment provided by the caller
670 * \param exp - is the client's export which is reconnecting
671 * \param want - is how much the client would like to get
673 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
676 struct ofd_device *ofd = ofd_exp(exp);
677 struct filter_export_data *fed = &exp->exp_filter_data;
681 int force = 0; /* can use cached data */
683 /* don't grant space to client with read-only access */
684 if ((exp_connect_flags(exp) & OBD_CONNECT_RDONLY) ||
685 ofd_grant_prohibit(exp, ofd))
689 ofd_grant_statfs(env, exp, force, &from_cache);
691 spin_lock(&ofd->ofd_grant_lock);
693 /* Grab free space from cached info and take out space already granted
694 * to clients as well as reserved space */
695 left = ofd_grant_space_left(exp);
697 /* get fresh statfs data if we are short in ungranted space */
698 if (from_cache && left < 32 * ofd_grant_chunk(exp, ofd)) {
699 spin_unlock(&ofd->ofd_grant_lock);
700 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
705 ofd_grant(exp, ofd_grant_to_cli(exp, ofd, (obd_size)fed->fed_grant),
708 /* return to client its current grant */
709 grant = ofd_grant_to_cli(exp, ofd, (obd_size)fed->fed_grant);
710 ofd->ofd_tot_granted_clients++;
712 spin_unlock(&ofd->ofd_grant_lock);
714 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %ld want: "LPU64" left: "
715 LPU64"\n", exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
716 exp, grant, want, left);
722 * Remove a client from the grant accounting totals. We also remove
723 * the export from the obd device under the osfs and dev locks to ensure
724 * that the ofd_grant_sanity_check() calculations are always valid.
725 * The client should do something similar when it invalidates its import.
727 * \param exp - is the client's export to remove from grant accounting
729 void ofd_grant_discard(struct obd_export *exp)
731 struct obd_device *obd = exp->exp_obd;
732 struct ofd_device *ofd = ofd_exp(exp);
733 struct filter_export_data *fed = &exp->exp_filter_data;
735 spin_lock(&ofd->ofd_grant_lock);
736 LASSERTF(ofd->ofd_tot_granted >= fed->fed_grant,
737 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
738 obd->obd_name, ofd->ofd_tot_granted,
739 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
740 ofd->ofd_tot_granted -= fed->fed_grant;
742 LASSERTF(ofd->ofd_tot_pending >= fed->fed_pending,
743 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
744 obd->obd_name, ofd->ofd_tot_pending,
745 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
746 /* ofd_tot_pending is handled in ofd_grant_commit as bulk
748 LASSERTF(ofd->ofd_tot_dirty >= fed->fed_dirty,
749 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
750 obd->obd_name, ofd->ofd_tot_dirty,
751 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
752 ofd->ofd_tot_dirty -= fed->fed_dirty;
754 spin_unlock(&ofd->ofd_grant_lock);
758 * Called at prepare time when handling read request. This function extracts
759 * incoming grant information from the obdo and processes the grant shrink
762 * \param env - is the lu environment provided by the caller
763 * \param exp - is the export of the client which sent the request
764 * \paral oa - is the incoming obdo sent by the client
766 void ofd_grant_prepare_read(const struct lu_env *env,
767 struct obd_export *exp, struct obdo *oa)
769 struct ofd_device *ofd = ofd_exp(exp);
776 if ((oa->o_valid & OBD_MD_FLGRANT) == 0)
777 /* The read request does not contain any grant
781 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
782 (oa->o_flags & OBD_FL_SHRINK_GRANT)) {
783 /* To process grant shrink request, we need to know how much
784 * available space remains on the backend filesystem.
785 * Shrink requests are not so common, we always get fresh
786 * statfs information. */
787 ofd_grant_statfs(env, exp, 1, NULL);
789 /* protect all grant counters */
790 spin_lock(&ofd->ofd_grant_lock);
792 /* Grab free space from cached statfs data and take out space
793 * already granted to clients as well as reserved space */
794 left = ofd_grant_space_left(exp);
796 /* all set now to proceed with shrinking */
799 /* no grant shrinking request packed in the obdo and
800 * since we don't grant space back on reads, no point
801 * in running statfs, so just skip it and process
802 * incoming grant data directly. */
803 spin_lock(&ofd->ofd_grant_lock);
807 /* extract incoming grant infomation provided by the client */
808 ofd_grant_incoming(env, exp, oa);
810 /* unlike writes, we don't return grants back on reads unless a grant
811 * shrink request was packed and we decided to turn it down. */
813 ofd_grant_shrink(exp, oa, left);
817 spin_unlock(&ofd->ofd_grant_lock);
821 * Called at write prepare time to handle incoming grant, check that we have
822 * enough space and grant some space back to the client if possible.
824 * \param env - is the lu environment provided by the caller
825 * \param exp - is the export of the client which sent the request
826 * \paral oa - is the incoming obdo sent by the client
827 * \param rnb - is the list of network buffers
828 * \param niocont - is the number of network buffers in the list
830 void ofd_grant_prepare_write(const struct lu_env *env,
831 struct obd_export *exp, struct obdo *oa,
832 struct niobuf_remote *rnb, int niocount)
834 struct obd_device *obd = exp->exp_obd;
835 struct ofd_device *ofd = ofd_exp(exp);
838 int force = 0; /* can use cached data intially */
844 /* get statfs information from OSD layer */
845 ofd_grant_statfs(env, exp, force, &from_cache);
847 spin_lock(&ofd->ofd_grant_lock); /* protect all grant counters */
849 /* Grab free space from cached statfs data and take out space already
850 * granted to clients as well as reserved space */
851 left = ofd_grant_space_left(exp);
853 /* Get fresh statfs data if we are short in ungranted space */
854 if (from_cache && left < 32 * ofd_grant_chunk(exp, ofd)) {
855 spin_unlock(&ofd->ofd_grant_lock);
856 CDEBUG(D_CACHE, "%s: fs has no space left and statfs too old\n",
862 /* When close to free space exhaustion, trigger a sync to force
863 * writeback cache to consume required space immediately and release as
864 * much space as possible. */
865 if (!obd->obd_recovering && force != 2 && left < OFD_GRANT_CHUNK) {
866 bool from_grant = true;
869 /* That said, it is worth running a sync only if some pages did
870 * not consume grant space on the client and could thus fail
871 * with ENOSPC later in ofd_grant_check() */
872 for (i = 0; i < niocount; i++)
873 if (!(rnb[i].rnb_flags & OBD_BRW_FROM_GRANT))
877 /* at least one network buffer requires acquiring grant
878 * space on the server */
879 spin_unlock(&ofd->ofd_grant_lock);
880 /* discard errors, at least we tried ... */
881 rc = dt_sync(env, ofd->ofd_osd);
887 /* extract incoming grant information provided by the client */
888 ofd_grant_incoming(env, exp, oa);
891 ofd_grant_check(env, exp, oa, rnb, niocount, &left);
893 if (!(oa->o_valid & OBD_MD_FLGRANT)) {
894 spin_unlock(&ofd->ofd_grant_lock);
898 /* if OBD_FL_SHRINK_GRANT is set, the client is willing to release some
900 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
901 (oa->o_flags & OBD_FL_SHRINK_GRANT))
902 ofd_grant_shrink(exp, oa, left);
904 /* grant more space back to the client if possible */
905 oa->o_grant = ofd_grant(exp, oa->o_grant, oa->o_undirty, left);
906 spin_unlock(&ofd->ofd_grant_lock);
910 * Called during object precreation to consume grant space.
911 * More space is granted for precreation if possible.
913 * \param env - is the lu environment provided by the caller
914 * \param exp - is the export holding the grant space for precreation (= self
916 * \paral nr - is the number of objects the caller wants to create objects
918 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr)
920 struct ofd_thread_info *info = ofd_info(env);
921 struct ofd_device *ofd = ofd_exp(exp);
922 struct filter_export_data *fed = &exp->exp_filter_data;
924 unsigned long wanted;
930 if (exp->exp_obd->obd_recovering ||
931 ofd->ofd_dt_conf.ddp_inodespace == 0)
932 /* don't enforce grant during recovery */
935 /* Update statfs data if required */
936 ofd_grant_statfs(env, exp, 1, NULL);
938 /* protect all grant counters */
939 spin_lock(&ofd->ofd_grant_lock);
941 /* fail precreate request if there is not enough blocks available for
943 if (ofd->ofd_osfs.os_bavail - (fed->fed_grant >> ofd->ofd_blockbits) <
944 (ofd->ofd_osfs.os_blocks >> 10)) {
945 spin_unlock(&ofd->ofd_grant_lock);
946 CDEBUG(D_RPCTRACE, "%s: not enough space for create "LPU64"\n",
947 ofd_obd(ofd)->obd_name,
948 ofd->ofd_osfs.os_bavail * ofd->ofd_osfs.os_blocks);
952 /* Grab free space from cached statfs data and take out space
953 * already granted to clients as well as reserved space */
954 left = ofd_grant_space_left(exp);
956 /* compute how much space is required to handle the precreation
958 wanted = *nr * ofd->ofd_dt_conf.ddp_inodespace;
959 if (wanted > fed->fed_grant + left) {
960 /* that's beyond what remains, adjust the number of objects that
961 * can be safely precreated */
962 wanted = fed->fed_grant + left;
963 *nr = wanted / ofd->ofd_dt_conf.ddp_inodespace;
965 /* we really have no space any more for precreation,
966 * fail the precreate request with ENOSPC */
967 spin_unlock(&ofd->ofd_grant_lock);
970 /* compute space needed for the new number of creations */
971 wanted = *nr * ofd->ofd_dt_conf.ddp_inodespace;
973 LASSERT(wanted <= fed->fed_grant + left);
975 if (wanted <= fed->fed_grant) {
976 /* we've enough grant space to handle this precreate request */
977 fed->fed_grant -= wanted;
979 /* we need to take some space from the ungranted pool */
980 ofd->ofd_tot_granted += wanted - fed->fed_grant;
981 left -= wanted - fed->fed_grant;
984 info->fti_used = wanted;
985 fed->fed_pending += info->fti_used;
986 ofd->ofd_tot_pending += info->fti_used;
988 /* grant more space (twice as much as needed for this request) for
989 * precreate purpose if possible */
990 ofd_grant(exp, fed->fed_grant, wanted * 2, left);
991 spin_unlock(&ofd->ofd_grant_lock);
996 * Called at commit time to update pending grant counter for writes in flight
998 * \param env - is the lu environment provided by the caller
999 * \param exp - is the export of the client which sent the request
1001 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp,
1004 struct ofd_device *ofd = ofd_exp(exp);
1005 struct ofd_thread_info *info = ofd_info(env);
1006 unsigned long pending;
1010 /* get space accounted in tot_pending for the I/O, set in
1011 * ofd_grant_check() */
1012 pending = info->fti_used;
1016 spin_lock(&ofd->ofd_grant_lock);
1017 /* Don't update statfs data for errors raised before commit (e.g.
1018 * bulk transfer failed, ...) since we know those writes have not been
1019 * processed. For other errors hit during commit, we cannot really tell
1020 * whether or not something was written, so we update statfs data.
1021 * In any case, this should not be fatal since we always get fresh
1022 * statfs data before failing a request with ENOSPC */
1024 spin_lock(&ofd->ofd_osfs_lock);
1025 /* Take pending out of cached statfs data */
1026 ofd->ofd_osfs.os_bavail -= min_t(obd_size,
1027 ofd->ofd_osfs.os_bavail,
1028 pending >> ofd->ofd_blockbits);
1029 if (ofd->ofd_statfs_inflight)
1030 /* someone is running statfs and want to be notified of
1031 * writes happening meanwhile */
1032 ofd->ofd_osfs_inflight += pending;
1033 spin_unlock(&ofd->ofd_osfs_lock);
1036 if (exp->exp_filter_data.fed_pending < pending) {
1037 CERROR("%s: cli %s/%p fed_pending(%lu) < grant_used(%lu)\n",
1038 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1039 exp->exp_filter_data.fed_pending, pending);
1040 spin_unlock(&ofd->ofd_grant_lock);
1043 exp->exp_filter_data.fed_pending -= pending;
1045 if (ofd->ofd_tot_granted < pending) {
1046 CERROR("%s: cli %s/%p tot_granted("LPU64") < grant_used(%lu)"
1047 "\n", exp->exp_obd->obd_name,
1048 exp->exp_client_uuid.uuid, exp, ofd->ofd_tot_granted,
1050 spin_unlock(&ofd->ofd_grant_lock);
1053 ofd->ofd_tot_granted -= pending;
1055 if (ofd->ofd_tot_pending < pending) {
1056 CERROR("%s: cli %s/%p tot_pending("LPU64") < grant_used(%lu)"
1057 "\n", exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1058 exp, ofd->ofd_tot_pending, pending);
1059 spin_unlock(&ofd->ofd_grant_lock);
1062 ofd->ofd_tot_pending -= pending;
1063 spin_unlock(&ofd->ofd_grant_lock);