1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter_io.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * Author: Phil Schwan <phil@clusterfs.com>
11 * This file is part of Lustre, http://www.lustre.org.
13 * Lustre is free software; you can redistribute it and/or
14 * modify it under the terms of version 2 of the GNU General Public
15 * License as published by the Free Software Foundation.
17 * Lustre is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with Lustre; if not, write to the Free Software
24 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 #define DEBUG_SUBSYSTEM S_FILTER
29 #include <linux/config.h>
30 #include <linux/module.h>
31 #include <linux/pagemap.h> // XXX kill me soon
32 #include <linux/version.h>
34 #include <linux/obd_class.h>
35 #include <linux/lustre_fsfilt.h>
36 #include "filter_internal.h"
38 static int filter_start_page_read(struct obd_device *obd, struct inode *inode,
39 struct niobuf_local *lnb)
42 unsigned long index = lnb->offset >> PAGE_SHIFT;
44 page = fsfilt_getpage(obd, inode, index);
46 CERROR("page index %lu, rc = %ld\n", index, PTR_ERR(page));
49 lnb->rc = PTR_ERR(page);
58 static int filter_finish_page_read(struct niobuf_local *lnb)
60 if (lnb->page == NULL)
63 if (PageUptodate(lnb->page))
66 wait_on_page(lnb->page);
67 if (!PageUptodate(lnb->page)) {
68 CERROR("page index %lu/offset "LPX64" not uptodate\n",
69 lnb->page->index, lnb->offset);
70 GOTO(err_page, lnb->rc = -EIO);
72 if (PageError(lnb->page)) {
73 CERROR("page index %lu/offset "LPX64" has error\n",
74 lnb->page->index, lnb->offset);
75 GOTO(err_page, lnb->rc = -EIO);
81 page_cache_release(lnb->page);
86 /* Grab the dirty and seen grant announcements from the incoming obdo.
87 * We will later calculate the clients new grant and return it.
88 * Caller must hold osfs lock */
89 static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
91 struct filter_export_data *fed;
92 struct obd_device *obd = exp->exp_obd;
93 static unsigned long last_msg;
94 static int last_count;
98 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
100 if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
101 (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
102 oa->o_valid &= ~OBD_MD_FLGRANT;
107 fed = &exp->exp_filter_data;
109 /* Don't print this to the console the first time it happens, since
110 * it can happen legitimately on occasion, but only rarely. */
111 if (time_after(jiffies, last_msg + 60 * HZ)) {
115 if ((last_count & (-last_count)) == last_count)
119 /* Add some margin, since there is a small race if other RPCs arrive
120 * out-or-order and have already consumed some grant. We want to
121 * leave this here in case there is a large error in accounting. */
122 CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
123 "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
124 obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
125 oa->o_dropped, fed->fed_grant);
127 /* Update our accounting now so that statfs takes it into account.
128 * Note that fed_dirty is only approximate and can become incorrect
129 * if RPCs arrive out-of-order. No important calculations depend
130 * on fed_dirty however. */
131 obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
132 if (fed->fed_grant < oa->o_dropped) {
133 CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
134 obd->obd_name, exp->exp_client_uuid.uuid, exp,
135 oa->o_dropped, fed->fed_grant);
138 if (obd->u.filter.fo_tot_granted < oa->o_dropped) {
139 CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n",
140 obd->obd_name, exp->exp_client_uuid.uuid, exp,
141 oa->o_dropped, obd->u.filter.fo_tot_granted);
144 obd->u.filter.fo_tot_granted -= oa->o_dropped;
145 fed->fed_grant -= oa->o_dropped;
146 fed->fed_dirty = oa->o_dirty;
150 #define GRANT_FOR_LLOG(obd) 16
152 /* Figure out how much space is available between what we've granted
153 * and what remains in the filesystem. Compensate for ext3 indirect
154 * block overhead when computing how much free space is left ungranted.
156 * Caller must hold obd_osfs_lock. */
157 obd_size filter_grant_space_left(struct obd_export *exp)
159 struct obd_device *obd = exp->exp_obd;
160 int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
161 obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
162 int rc, statfs_done = 0;
164 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
166 if (time_before(obd->obd_osfs_age, jiffies - HZ)) {
168 rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
169 if (rc) /* N.B. statfs can't really fail */
174 avail = obd->obd_osfs.os_bavail;
175 left = avail - (avail >> (blockbits - 3)); /* (d)indirect */
176 if (left > GRANT_FOR_LLOG(obd)) {
177 left = (left - GRANT_FOR_LLOG(obd)) << blockbits;
179 left = 0 /* << blockbits */;
182 if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) {
183 CDEBUG(D_CACHE, "fs has no space left and statfs too old\n");
187 if (left >= tot_granted) {
190 static unsigned long next;
191 if (left < tot_granted - obd->u.filter.fo_tot_pending &&
192 time_after(jiffies, next)) {
193 spin_unlock(&obd->obd_osfs_lock);
194 CERROR("%s: cli %s/%p grant "LPU64" > available "
195 LPU64" and pending "LPU64"\n", obd->obd_name,
196 exp->exp_client_uuid.uuid, exp, tot_granted,
197 left, obd->u.filter.fo_tot_pending);
199 portals_debug_dumplog();
200 next = jiffies + 20 * HZ;
201 spin_lock(&obd->obd_osfs_lock);
206 CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64
207 " left: "LPU64" pending: "LPU64"\n", obd->obd_name,
208 exp->exp_client_uuid.uuid, exp,
209 obd->obd_osfs.os_bfree << blockbits, avail << blockbits,
210 tot_granted, left, obd->u.filter.fo_tot_pending);
215 /* Calculate how much grant space to allocate to this client, based on how
216 * much space is currently free and how much of that is already granted.
218 * Caller must hold obd_osfs_lock. */
219 long filter_grant(struct obd_export *exp, obd_size current_grant,
220 obd_size want, obd_size fs_space_left)
222 struct obd_device *obd = exp->exp_obd;
223 struct filter_export_data *fed = &exp->exp_filter_data;
224 int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
227 LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
229 /* Grant some fraction of the client's requested grant space so that
230 * they are not always waiting for write credits (not all of it to
231 * avoid overgranting in face of multiple RPCs in flight). This
232 * essentially will be able to control the OSC_MAX_RIF for a client.
234 * If we do have a large disparity between what the client thinks it
235 * has and what we think it has, don't grant very much and let the
236 * client consume its grant first. Either it just has lots of RPCs
237 * in flight, or it was evicted and its grants will soon be used up. */
238 if (current_grant < want &&
239 current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
240 grant = min((want >> blockbits) / 2,
241 (fs_space_left >> blockbits) / 8);
245 if (grant > FILTER_GRANT_CHUNK)
246 grant = FILTER_GRANT_CHUNK;
248 obd->u.filter.fo_tot_granted += grant;
249 fed->fed_grant += grant;
253 CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
254 obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
256 "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
257 " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
258 exp, obd->u.filter.fo_tot_dirty,
259 obd->u.filter.fo_tot_granted, obd->obd_num_exports);
264 static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
265 int objcount, struct obd_ioobj *obj,
266 int niocount, struct niobuf_remote *nb,
267 struct niobuf_local *res,
268 struct obd_trans_info *oti)
270 struct obd_device *obd = exp->exp_obd;
271 struct lvfs_run_ctxt saved;
273 struct niobuf_remote *rnb;
274 struct niobuf_local *lnb = NULL;
275 struct fsfilt_objinfo *fso;
276 struct dentry *dentry;
278 int rc = 0, i, j, tot_bytes = 0, cleanup_phase = 0;
279 unsigned long now = jiffies;
282 /* We are currently not supporting multi-obj BRW_READ RPCS at all.
283 * When we do this function's dentry cleanup will need to be fixed */
284 LASSERT(objcount == 1);
285 LASSERT(obj->ioo_bufcnt > 0);
287 if (oa && oa->o_valid & OBD_MD_FLGRANT) {
288 spin_lock(&obd->obd_osfs_lock);
289 filter_grant_incoming(exp, oa);
292 /* Reads do not increase grants */
293 oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
294 filter_grant_space_left(exp));
298 spin_unlock(&obd->obd_osfs_lock);
301 OBD_ALLOC(fso, objcount * sizeof(*fso));
305 memset(res, 0, niocount * sizeof(*res));
307 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
308 for (i = 0, o = obj; i < objcount; i++, o++) {
309 LASSERT(o->ioo_bufcnt);
311 dentry = filter_oa2dentry(obd, oa);
313 GOTO(cleanup, rc = PTR_ERR(dentry));
315 if (dentry->d_inode == NULL) {
316 CERROR("trying to BRW to non-existent file "LPU64"\n",
319 GOTO(cleanup, rc = -ENOENT);
322 fso[i].fso_dentry = dentry;
323 fso[i].fso_bufcnt = o->ioo_bufcnt;
326 if (time_after(jiffies, now + 15 * HZ))
327 CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ);
329 CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
332 for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
333 dentry = fso[i].fso_dentry;
334 inode = dentry->d_inode;
336 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
337 lnb->dentry = dentry;
338 lnb->offset = rnb->offset;
340 lnb->flags = rnb->flags;
342 if (inode->i_size <= rnb->offset) {
343 /* If there's no more data, abort early.
344 * lnb->page == NULL and lnb->rc == 0, so it's
345 * easy to detect later. */
348 rc = filter_start_page_read(obd, inode, lnb);
352 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
353 "page err %u@"LPU64" %u/%u %p: rc %d\n",
354 lnb->len, lnb->offset, j, o->ioo_bufcnt,
360 if (inode->i_size < lnb->offset + lnb->len - 1)
361 lnb->rc = inode->i_size - lnb->offset;
365 tot_bytes += lnb->rc;
366 if (lnb->rc < lnb->len) {
367 /* short read, be sure to wait on it */
374 if (time_after(jiffies, now + 15 * HZ))
375 CERROR("slow start_page_read %lus\n", (jiffies - now) / HZ);
377 CDEBUG(D_INFO, "start_page_read: %lu jiffies\n",
380 lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
381 while (lnb-- > res) {
382 rc = filter_finish_page_read(lnb);
384 CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len,
385 lnb->offset, (int)(lnb - res), lnb->dentry, rc);
391 if (time_after(jiffies, now + 15 * HZ))
392 CERROR("slow finish_page_read %lus\n", (jiffies - now) / HZ);
394 CDEBUG(D_INFO, "finish_page_read: %lu jiffies\n",
397 filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
402 switch (cleanup_phase) {
404 for (lnb = res; lnb < (res + niocount); lnb++) {
406 page_cache_release(lnb->page);
408 if (res->dentry != NULL)
411 CERROR("NULL dentry in cleanup -- tell CFS\n");
413 OBD_FREE(fso, objcount * sizeof(*fso));
414 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
419 /* When clients have dirtied as much space as they've been granted they
420 * fall through to sync writes. These sync writes haven't been expressed
421 * in grants and need to error with ENOSPC when there isn't room in the
422 * filesystem for them after grants are taken into account. However,
423 * writeback of the dirty data that was already granted space can write
426 * Caller must hold obd_osfs_lock. */
427 static int filter_grant_check(struct obd_export *exp, int objcount,
428 struct fsfilt_objinfo *fso, int niocount,
429 struct niobuf_remote *rnb,
430 struct niobuf_local *lnb, obd_size *left,
433 struct filter_export_data *fed = &exp->exp_filter_data;
434 int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
435 unsigned long used = 0, ungranted = 0, using;
436 int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
438 LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
440 for (obj = 0; obj < objcount; obj++) {
441 for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
444 /* FIXME: this is calculated with PAGE_SIZE on client */
446 bytes += rnb[n].offset & (blocksize - 1);
447 tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
449 bytes += blocksize - tmp;
451 if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
452 if (fed->fed_grant < used + bytes) {
454 "%s: cli %s/%p claims %ld+%d "
455 "GRANT, real grant %lu idx %d\n",
456 exp->exp_obd->obd_name,
457 exp->exp_client_uuid.uuid, exp,
458 used, bytes, fed->fed_grant, n);
462 rnb[n].flags |= OBD_BRW_GRANTED;
463 lnb[n].lnb_grant_used = bytes;
464 CDEBUG(0, "idx %d used=%lu\n", n, used);
469 if (*left > ungranted) {
470 /* if enough space, pretend it was granted */
472 rnb[n].flags |= OBD_BRW_GRANTED;
473 CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
478 /* We can't check for already-mapped blocks here, as
479 * it requires dropping the osfs lock to do the bmap.
480 * Instead, we return ENOSPC and in that case we need
481 * to go through and verify if all of the blocks not
482 * marked BRW_GRANTED are already mapped and we can
483 * ignore this error. */
485 rnb[n].flags &= OBD_BRW_GRANTED;
486 CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
487 exp->exp_obd->obd_name,
488 exp->exp_client_uuid.uuid, exp, n, bytes);
492 /* Now substract what client have used already. We don't subtract
493 * this from the tot_granted yet, so that other client's can't grab
494 * that space before we have actually allocated our blocks. That
495 * happens in filter_grant_commit() after the writes are done. */
497 fed->fed_grant -= used;
498 fed->fed_pending += used;
499 exp->exp_obd->u.filter.fo_tot_pending += used;
502 "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
503 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
504 ungranted, fed->fed_grant, fed->fed_dirty);
506 /* Rough calc in case we don't refresh cached statfs data */
507 using = (used + ungranted + 1 ) >>
508 exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
509 if (exp->exp_obd->obd_osfs.os_bavail > using)
510 exp->exp_obd->obd_osfs.os_bavail -= using;
512 exp->exp_obd->obd_osfs.os_bavail = 0;
514 if (fed->fed_dirty < used) {
515 CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n",
516 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
517 used, fed->fed_dirty);
518 used = fed->fed_dirty;
520 exp->exp_obd->u.filter.fo_tot_dirty -= used;
521 fed->fed_dirty -= used;
526 static int filter_start_page_write(struct obd_device *obd, struct inode *inode,
527 struct niobuf_local *lnb)
531 if (lnb->len != PAGE_SIZE)
532 return filter_start_page_read(obd, inode, lnb);
534 page = alloc_pages(GFP_HIGHUSER, 0);
536 CERROR("no memory for a temp page\n");
537 RETURN(lnb->rc = -ENOMEM);
540 POISON_PAGE(page, 0xf1);
541 if (lnb->len != PAGE_SIZE) {
542 memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
546 page->index = lnb->offset >> PAGE_SHIFT;
552 static void filter_abort_page_write(struct niobuf_local *lnb)
554 LASSERT(lnb->page != NULL);
556 if (lnb->len != PAGE_SIZE)
557 page_cache_release(lnb->page);
559 __free_pages(lnb->page, 0);
562 /* a helper for both the 2.4 and 2.6 commitrw paths which are both built
563 * up by our shared filter_preprw_write() */
564 void filter_release_write_page(struct filter_obd *filter, struct inode *inode,
565 struct niobuf_local *lnb, int rc)
567 if (lnb->len != PAGE_SIZE)
568 return filter_release_read_page(filter, inode, lnb->page);
571 flip_into_page_cache(inode, lnb->page);
572 __free_page(lnb->page);
575 /* If we ever start to support multi-object BRW RPCs, we will need to get locks
576 * on mulitple inodes. That isn't all, because there still exists the
577 * possibility of a truncate starting a new transaction while holding the ext3
578 * rwsem = write while some writes (which have started their transactions here)
579 * blocking on the ext3 rwsem = read => lock inversion.
581 * The handling gets very ugly when dealing with locked pages. It may be easier
582 * to just get rid of the locked page code (which has problems of its own) and
583 * either discover we do not need it anymore (i.e. it was a symptom of another
584 * bug) or ensure we get the page locks in an appropriate order. */
585 static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa,
586 int objcount, struct obd_ioobj *obj,
587 int niocount, struct niobuf_remote *nb,
588 struct niobuf_local *res,
589 struct obd_trans_info *oti)
591 struct lvfs_run_ctxt saved;
592 struct niobuf_remote *rnb;
593 struct niobuf_local *lnb;
594 struct fsfilt_objinfo fso;
595 struct dentry *dentry;
597 unsigned long now = jiffies;
598 int rc = 0, i, tot_bytes = 0, cleanup_phase = 1;
600 LASSERT(objcount == 1);
601 LASSERT(obj->ioo_bufcnt > 0);
603 memset(res, 0, niocount * sizeof(*res));
605 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
606 dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
609 GOTO(cleanup, rc = PTR_ERR(dentry));
611 if (dentry->d_inode == NULL) {
612 CERROR("trying to BRW to non-existent file "LPU64"\n",
615 GOTO(cleanup, rc = -ENOENT);
618 fso.fso_dentry = dentry;
619 fso.fso_bufcnt = obj->ioo_bufcnt;
621 if (time_after(jiffies, now + 15 * HZ))
622 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
624 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
627 spin_lock(&exp->exp_obd->obd_osfs_lock);
629 filter_grant_incoming(exp, oa);
632 left = filter_grant_space_left(exp);
634 rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
635 &left, dentry->d_inode);
636 if (oa && oa->o_valid & OBD_MD_FLGRANT)
637 oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
639 spin_unlock(&exp->exp_obd->obd_osfs_lock);
646 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
648 /* We still set up for ungranted pages so that granted pages
649 * can be written to disk as they were promised, and portals
650 * needs to keep the pages all aligned properly. */
651 lnb->dentry = dentry;
652 lnb->offset = rnb->offset;
654 lnb->flags = rnb->flags;
656 rc = filter_start_page_write(exp->exp_obd, dentry->d_inode,lnb);
658 CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
659 lnb->len, lnb->offset,
660 i, obj->ioo_bufcnt, dentry, rc);
662 filter_abort_page_write(lnb);
667 tot_bytes += lnb->len;
670 while (lnb-- > res) {
671 if (lnb->len == PAGE_SIZE)
673 rc = filter_finish_page_read(lnb);
675 CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len,
676 lnb->offset, (int)(lnb - res), lnb->dentry, rc);
681 if (time_after(jiffies, now + 15 * HZ))
682 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
684 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
687 lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
691 switch(cleanup_phase) {
693 spin_lock(&exp->exp_obd->obd_osfs_lock);
695 filter_grant_incoming(exp, oa);
696 spin_unlock(&exp->exp_obd->obd_osfs_lock);
699 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
703 int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
704 int objcount, struct obd_ioobj *obj, int niocount,
705 struct niobuf_remote *nb, struct niobuf_local *res,
706 struct obd_trans_info *oti)
708 if (cmd == OBD_BRW_WRITE)
709 return filter_preprw_write(cmd, exp, oa, objcount, obj,
710 niocount, nb, res, oti);
712 if (cmd == OBD_BRW_READ)
713 return filter_preprw_read(cmd, exp, oa, objcount, obj,
714 niocount, nb, res, oti);
720 void filter_release_read_page(struct filter_obd *filter, struct inode *inode,
726 (inode->i_size > filter->fo_readcache_max_filesize))
729 /* drop from cache like truncate_list_pages() */
730 if (drop && !TryLockPage(page)) {
732 ll_truncate_complete_page(page);
735 page_cache_release(page);
738 static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa,
739 int objcount, struct obd_ioobj *obj,
740 int niocount, struct niobuf_local *res,
741 struct obd_trans_info *oti, int rc)
744 struct niobuf_local *lnb;
746 struct inode *inode = NULL;
749 if (res->dentry != NULL)
750 inode = res->dentry->d_inode;
752 for (i = 0, o = obj, lnb = res; i < objcount; i++, o++) {
753 for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) {
754 if (lnb->page == NULL)
756 filter_release_read_page(&exp->exp_obd->u.filter,
761 if (res->dentry != NULL)
766 void flip_into_page_cache(struct inode *inode, struct page *new_page)
768 struct page *old_page;
772 /* the dlm is protecting us from read/write concurrency, so we
773 * expect this find_lock_page to return quickly. even if we
774 * race with another writer it won't be doing much work with
775 * the page locked. we do this 'cause t_c_p expects a
776 * locked page, and it wants to grab the pagecache lock
778 old_page = find_lock_page(inode->i_mapping, new_page->index);
780 ll_truncate_complete_page(old_page);
781 unlock_page(old_page);
782 page_cache_release(old_page);
785 #if 0 /* this should be a /proc tunable someday */
786 /* racing o_directs (no locking ioctl) could race adding
787 * their pages, so we repeat the page invalidation unless
788 * we successfully added our new page */
789 rc = add_to_page_cache_unique(new_page, inode->i_mapping,
791 page_hash(inode->i_mapping,
794 /* add_to_page_cache clears uptodate|dirty and locks
796 SetPageUptodate(new_page);
797 unlock_page(new_page);
805 void filter_grant_commit(struct obd_export *exp, int niocount,
806 struct niobuf_local *res)
808 struct filter_obd *filter = &exp->exp_obd->u.filter;
809 struct niobuf_local *lnb = res;
810 unsigned long pending = 0;
813 spin_lock(&exp->exp_obd->obd_osfs_lock);
814 for (i = 0, lnb = res; i < niocount; i++, lnb++)
815 pending += lnb->lnb_grant_used;
817 LASSERTF(exp->exp_filter_data.fed_pending >= pending,
818 "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n",
819 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
820 exp->exp_filter_data.fed_pending, pending);
821 exp->exp_filter_data.fed_pending -= pending;
822 LASSERTF(filter->fo_tot_granted >= pending,
823 "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n",
824 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
825 exp->exp_obd->u.filter.fo_tot_granted, pending);
826 filter->fo_tot_granted -= pending;
827 LASSERTF(filter->fo_tot_pending >= pending,
828 "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n",
829 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
830 filter->fo_tot_pending, pending);
831 filter->fo_tot_pending -= pending;
833 spin_unlock(&exp->exp_obd->obd_osfs_lock);
836 int filter_write_extents(struct obd_export *exp, struct obd_ioobj *obj,
837 int niocount, struct niobuf_local *local,
840 struct lvfs_run_ctxt saved;
841 struct dentry *dentry;
842 struct niobuf_local *lnb;
849 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
851 dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
853 if (IS_ERR(dentry)) {
854 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
855 RETURN (PTR_ERR(dentry));
858 if (dentry->d_inode == NULL) {
859 CERROR("trying to write extents to non-existent file "LPU64"\n",
861 GOTO(cleanup, rc = -ENOENT);
864 for (i = 0, lnb = local; i < obj->ioo_bufcnt; i++, lnb++) {
866 offset = lnb->offset;
868 } else if (lnb->offset == (offset + len)) {
871 rc = fsfilt_write_extents(exp->exp_obd, dentry,
874 CERROR("write exts off "LPU64" num %u rc:%d\n",
878 offset = lnb->offset;
883 rc = fsfilt_write_extents(exp->exp_obd, dentry,
886 CERROR("write exts off "LPU64" num %u rc:%d\n",
893 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
897 int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
898 int objcount, struct obd_ioobj *obj, int niocount,
899 struct niobuf_local *res, struct obd_trans_info *oti,int rc)
901 if (cmd == OBD_BRW_WRITE)
902 return filter_commitrw_write(exp, oa, objcount, obj, niocount,
904 if (cmd == OBD_BRW_READ)
905 return filter_commitrw_read(exp, oa, objcount, obj, niocount,
911 int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
912 struct lov_stripe_md *lsm, obd_count oa_bufs,
913 struct brw_page *pga, struct obd_trans_info *oti)
915 struct obd_ioobj ioo;
916 struct niobuf_local *lnb;
917 struct niobuf_remote *rnb;
922 OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local));
923 OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote));
925 if (lnb == NULL || rnb == NULL)
926 GOTO(out, ret = -ENOMEM);
928 for (i = 0; i < oa_bufs; i++) {
929 rnb[i].offset = pga[i].off;
930 rnb[i].len = pga[i].count;
933 obdo_to_ioobj(oa, &ioo);
934 ioo.ioo_bufcnt = oa_bufs;
936 ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
940 for (i = 0; i < oa_bufs; i++) {
941 void *virt = kmap(pga[i].pg);
942 obd_off off = pga[i].off & ~PAGE_MASK;
943 void *addr = kmap(lnb[i].page);
945 /* 2 kmaps == vanishingly small deadlock opportunity */
947 if (cmd & OBD_BRW_WRITE)
948 memcpy(addr + off, virt + off, pga[i].count);
950 memcpy(virt + off, addr + off, pga[i].count);
956 ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
960 OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local));
962 OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote));