1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter_io_26.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
40 * Author: Phil Schwan <phil@clusterfs.com>
43 #ifndef AUTOCONF_INCLUDED
44 #include <linux/config.h>
46 #include <linux/module.h>
47 #include <linux/pagemap.h> // XXX kill me soon
48 #include <linux/version.h>
49 #include <linux/buffer_head.h>
51 #define DEBUG_SUBSYSTEM S_FILTER
53 #include <obd_class.h>
54 #include <lustre_fsfilt.h>
55 #include <lustre_quota.h>
56 #include "filter_internal.h"
58 /* 512byte block min */
59 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
61 cfs_atomic_t dr_numreqs; /* number of reqs being processed */
66 struct page **dr_pages;
67 unsigned long *dr_blocks;
68 unsigned int dr_ignore_quota:1;
69 struct filter_obd *dr_filter;
72 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
73 struct obd_export *exp)
75 struct filter_obd *filter = iobuf->dr_filter;
77 cfs_atomic_inc(&iobuf->dr_numreqs);
79 if (rw == OBD_BRW_READ) {
80 cfs_atomic_inc(&filter->fo_r_in_flight);
81 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
82 cfs_atomic_read(&filter->fo_r_in_flight));
83 lprocfs_oh_tally_log2(&filter->
84 fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
86 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
87 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
89 cfs_atomic_read(&filter-> \
91 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
92 nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
96 cfs_atomic_inc(&filter->fo_w_in_flight);
97 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
98 cfs_atomic_read(&filter->fo_w_in_flight));
99 lprocfs_oh_tally_log2(&filter->
100 fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
102 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
103 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
104 hist[BRW_W_RPC_HIST],
105 cfs_atomic_read(&filter-> \
107 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
108 nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
114 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
116 struct filter_obd *filter = iobuf->dr_filter;
118 /* CAVEAT EMPTOR: possibly in IRQ context
119 * DO NOT record procfs stats here!!! */
121 if (rw == OBD_BRW_READ)
122 cfs_atomic_dec(&filter->fo_r_in_flight);
124 cfs_atomic_dec(&filter->fo_w_in_flight);
126 if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
127 cfs_waitq_signal(&iobuf->dr_wait);
130 #ifdef HAVE_BIO_ENDIO_2ARG
131 #define DIO_RETURN(a)
132 static void dio_complete_routine(struct bio *bio, int error)
134 #define DIO_RETURN(a) return(a)
135 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
138 struct filter_iobuf *iobuf = bio->bi_private;
142 /* CAVEAT EMPTOR: possibly in IRQ context
143 * DO NOT record procfs stats here!!! */
145 #ifdef HAVE_BIO_ENDIO_2ARG
146 /* The "bi_size" check was needed for kernels < 2.6.24 in order to
147 * handle the case where a SCSI request error caused this callback
148 * to be called before all of the biovecs had been processed.
149 * Without this check the server thread will hang. In newer kernels
150 * the bio_end_io routine is never called for partial completions,
151 * so this check is no longer needed. */
152 if (bio->bi_size) /* Not complete */
156 if (unlikely(iobuf == NULL)) {
157 CERROR("***** bio->bi_private is NULL! This should never "
158 "happen. Normally, I would crash here, but instead I "
159 "will dump the bio contents to the console. Please "
160 "report this to <http://bugzilla.lustre.org/> , along "
161 "with any interesting messages leading up to this point "
162 "(like SCSI errors, perhaps). Because bi_private is "
163 "NULL, I can't wake up the thread that initiated this "
164 "IO - you will probably have to reboot this node.\n");
165 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
166 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
167 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
168 bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
169 bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
174 /* the check is outside of the cycle for performance reason -bzzz */
175 if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
176 bio_for_each_segment(bvl, bio, i) {
177 if (likely(error == 0))
178 SetPageUptodate(bvl->bv_page);
179 LASSERT(PageLocked(bvl->bv_page));
180 ClearPageConstant(bvl->bv_page);
182 record_finish_io(iobuf, OBD_BRW_READ, error);
184 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
185 bio_for_each_segment(bvl, bio, i) {
186 ClearPageConstant(bvl->bv_page);
189 record_finish_io(iobuf, OBD_BRW_WRITE, error);
192 /* any real error is good enough -bzzz */
193 if (error != 0 && iobuf->dr_error == 0)
194 iobuf->dr_error = error;
196 /* Completed bios used to be chained off iobuf->dr_bios and freed in
197 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
198 * mempool when serious on-disk fragmentation was encountered,
199 * deadlocking the OST. The bios are now released as soon as complete
200 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
205 static int can_be_merged(struct bio *bio, sector_t sector)
212 size = bio->bi_size >> 9;
213 return bio->bi_sector + size == sector ? 1 : 0;
216 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
217 int rw, int num_pages)
219 struct filter_iobuf *iobuf;
221 LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
223 OBD_ALLOC(iobuf, sizeof(*iobuf));
227 OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
228 if (iobuf->dr_pages == NULL)
231 OBD_ALLOC(iobuf->dr_blocks,
232 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
233 if (iobuf->dr_blocks == NULL)
236 iobuf->dr_filter = filter;
237 cfs_waitq_init(&iobuf->dr_wait);
238 cfs_atomic_set(&iobuf->dr_numreqs, 0);
239 iobuf->dr_max_pages = num_pages;
240 iobuf->dr_npages = 0;
246 OBD_FREE(iobuf->dr_pages,
247 num_pages * sizeof(*iobuf->dr_pages));
249 OBD_FREE(iobuf, sizeof(*iobuf));
251 RETURN(ERR_PTR(-ENOMEM));
254 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
256 iobuf->dr_npages = 0;
258 cfs_atomic_set(&iobuf->dr_numreqs, 0);
261 void filter_free_iobuf(struct filter_iobuf *iobuf)
263 int num_pages = iobuf->dr_max_pages;
265 filter_clear_iobuf(iobuf);
267 OBD_FREE(iobuf->dr_blocks,
268 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
269 OBD_FREE(iobuf->dr_pages,
270 num_pages * sizeof(*iobuf->dr_pages));
274 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
275 struct obd_trans_info *oti)
277 int thread_id = (oti && oti->oti_thread) ?
278 oti->oti_thread->t_id : -1;
280 if (unlikely(thread_id < 0)) {
281 filter_free_iobuf(iobuf);
285 LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
286 "iobuf mismatch for thread %d: pool %p iobuf %p\n",
287 thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
288 filter_clear_iobuf(iobuf);
291 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
292 struct inode *inode, struct page *page)
294 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
295 iobuf->dr_pages[iobuf->dr_npages++] = page;
300 int filter_do_bio(struct obd_export *exp, struct inode *inode,
301 struct filter_iobuf *iobuf, int rw)
303 struct obd_device *obd = exp->exp_obd;
304 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
305 struct page **pages = iobuf->dr_pages;
306 int npages = iobuf->dr_npages;
307 unsigned long *blocks = iobuf->dr_blocks;
308 int total_blocks = npages * blocks_per_page;
309 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
310 unsigned int blocksize = inode->i_sb->s_blocksize;
311 struct bio *bio = NULL;
313 unsigned long start_time = jiffies;
315 unsigned int page_offset;
324 LASSERT(iobuf->dr_npages == npages);
325 LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
327 for (page_idx = 0, block_idx = 0;
329 page_idx++, block_idx += blocks_per_page) {
331 page = pages[page_idx];
332 LASSERT (block_idx + blocks_per_page <= total_blocks);
334 for (i = 0, page_offset = 0;
336 i += nblocks, page_offset += blocksize * nblocks) {
340 if (blocks[block_idx + i] == 0) { /* hole */
341 LASSERT(rw == OBD_BRW_READ);
342 memset(kmap(page) + page_offset, 0, blocksize);
347 sector = (sector_t)blocks[block_idx + i] << sector_bits;
349 /* Additional contiguous file blocks? */
350 while (i + nblocks < blocks_per_page &&
351 (sector + (nblocks << sector_bits)) ==
352 ((sector_t)blocks[block_idx + i + nblocks] <<
356 /* I only set the page to be constant only if it
357 * is mapped to a contiguous underlying disk block(s).
358 * It will then make sure the corresponding device
359 * cache of raid5 will be overwritten by this page.
361 if ((rw == OBD_BRW_WRITE) &&
362 (nblocks == blocks_per_page) &&
363 mapping_cap_page_constant_write(inode->i_mapping))
364 SetPageConstant(page);
367 can_be_merged(bio, sector) &&
368 bio_add_page(bio, page,
369 blocksize * nblocks, page_offset) != 0)
370 continue; /* added this frag OK */
373 struct request_queue *q =
374 bdev_get_queue(bio->bi_bdev);
376 /* Dang! I have to fragment this I/O */
377 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
378 "sectors %d(%d) psg %d(%d) hsg %d(%d) "
379 "sector %llu next %llu\n",
381 bio->bi_vcnt, bio->bi_max_vecs,
382 bio->bi_size >> 9, queue_max_sectors(q),
383 bio_phys_segments(q, bio),
384 queue_max_phys_segments(q),
385 bio_hw_segments(q, bio),
386 queue_max_hw_segments(q),
387 (unsigned long long)bio->bi_sector,
388 (unsigned long long)sector);
390 record_start_io(iobuf, rw, bio->bi_size, exp);
391 rc = fsfilt_send_bio(rw, obd, inode, bio);
393 CERROR("Can't send bio: %d\n", rc);
394 record_finish_io(iobuf, rw, rc);
400 /* allocate new bio, limited by max BIO size, b=9945 */
401 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
402 (npages - page_idx) *
405 CERROR("Can't allocate bio %u*%u = %u pages\n",
406 (npages - page_idx), blocks_per_page,
407 (npages - page_idx) * blocks_per_page);
412 bio->bi_bdev = inode->i_sb->s_bdev;
413 bio->bi_sector = sector;
414 bio->bi_end_io = dio_complete_routine;
415 bio->bi_private = iobuf;
417 rc = bio_add_page(bio, page,
418 blocksize * nblocks, page_offset);
424 record_start_io(iobuf, rw, bio->bi_size, exp);
425 rc = fsfilt_send_bio(rw, obd, inode, bio);
430 CERROR("Can't send bio: %d\n", rc);
431 record_finish_io(iobuf, rw, rc);
436 cfs_wait_event(iobuf->dr_wait,
437 cfs_atomic_read(&iobuf->dr_numreqs) == 0);
439 if (rw == OBD_BRW_READ) {
440 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
441 hist[BRW_R_DIO_FRAGS],
443 lprocfs_oh_tally_log2(&obd->u.filter.
444 fo_filter_stats.hist[BRW_R_IO_TIME],
445 jiffies - start_time);
446 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
447 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
448 hist[BRW_R_DIO_FRAGS],
450 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
451 nid_brw_stats->hist[BRW_R_IO_TIME],
452 jiffies - start_time);
455 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
456 hist[BRW_W_DIO_FRAGS], frags);
457 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
459 jiffies - start_time);
460 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
461 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
462 hist[BRW_W_DIO_FRAGS],
464 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
465 nid_brw_stats->hist[BRW_W_IO_TIME],
466 jiffies - start_time);
471 rc = iobuf->dr_error;
475 /* Must be called with i_mutex taken for writes; this will drop it */
476 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
477 struct obd_export *exp, struct iattr *attr,
478 struct obd_trans_info *oti, void **wait_handle)
480 struct obd_device *obd = exp->exp_obd;
481 struct inode *inode = dchild->d_inode;
482 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
484 cfs_semaphore_t *sem;
487 LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
488 iobuf->dr_npages, iobuf->dr_max_pages);
489 LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
491 if (rw == OBD_BRW_READ) {
492 if (iobuf->dr_npages == 0)
497 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
498 LASSERT(iobuf->dr_npages > 0);
500 sem = &obd->u.filter.fo_alloc_lock;
502 lquota_enforce(filter_quota_interface_ref, obd,
503 iobuf->dr_ignore_quota);
506 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
507 iobuf->dr_npages, iobuf->dr_blocks,
508 obdfilter_created_scratchpad, create, sem);
510 if (rw == OBD_BRW_WRITE) {
512 filter_tally(exp, iobuf->dr_pages,
513 iobuf->dr_npages, iobuf->dr_blocks,
515 if (attr->ia_size > i_size_read(inode))
516 attr->ia_valid |= ATTR_SIZE;
517 rc = fsfilt_setattr(obd, dchild,
518 oti->oti_handle, attr, 0);
521 UNLOCK_INODE_MUTEX(inode);
523 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
525 CERROR("can't close transaction: %d\n", rc2);
531 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
534 rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
539 } else if (rc == 0) {
540 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
541 iobuf->dr_blocks, blocks_per_page, 0);
544 RETURN(filter_do_bio(exp, inode, iobuf, rw));
547 /* See if there are unallocated parts in given file region */
548 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
550 sector_t (*fs_bmap)(struct address_space *, sector_t) =
551 inode->i_mapping->a_ops->bmap;
554 /* We can't know if we are overwriting or not */
558 offset >>= inode->i_blkbits;
559 len >>= inode->i_blkbits;
561 for (j = 0; j <= len; j++)
562 if (fs_bmap(inode->i_mapping, offset + j) == 0)
569 * interesting use cases on how it interacts with VM:
571 * - vm writeout -- shouldn't see our pages as we don't mark them dirty
572 * though vm can find partial page left dirty by truncate. in this
573 * usual writeout is used unless our write rewrite that page - then we
574 * drop PG_dirty with PG_lock held.
579 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
580 int objcount, struct obd_ioobj *obj,
581 struct niobuf_remote *nb, int niocount,
582 struct niobuf_local *res, struct obd_trans_info *oti,
585 struct niobuf_local *lnb;
586 struct filter_iobuf *iobuf = NULL;
587 struct lvfs_run_ctxt saved;
588 struct fsfilt_objinfo fso;
589 struct iattr iattr = { 0 };
590 struct inode *inode = res->dentry->d_inode;
591 unsigned long now = jiffies;
592 int i, err, cleanup_phase = 0;
593 struct obd_device *obd = exp->exp_obd;
594 struct filter_obd *fo = &obd->u.filter;
595 void *wait_handle = NULL;
597 unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
598 int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
599 int sync_journal_commit = obd->u.filter.fo_syncjournal;
602 LASSERT(oti != NULL);
603 LASSERT(objcount == 1);
604 LASSERT(current->journal_info == NULL);
609 iobuf = filter_iobuf_get(&obd->u.filter, oti);
611 GOTO(cleanup, rc = PTR_ERR(iobuf));
614 fso.fso_dentry = res->dentry;
615 fso.fso_bufcnt = obj->ioo_bufcnt;
617 iobuf->dr_ignore_quota = 0;
618 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
620 __u32 flags = lnb->flags;
622 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
623 /* If overwriting an existing block,
624 * we don't need a grant */
625 if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
631 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
632 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
636 LASSERT(PageLocked(lnb->page));
637 LASSERT(!PageWriteback(lnb->page));
639 /* since write & truncate are serialized by the i_alloc_sem,
640 * even partial truncate should not leave dirty pages in
642 LASSERT(!PageDirty(lnb->page));
644 SetPageUptodate(lnb->page);
646 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
649 total_size += lnb->len;
651 /* we expect these pages to be in offset order, but we'll
653 this_size = lnb->offset + lnb->len;
654 if (this_size > iattr.ia_size)
655 iattr.ia_size = this_size;
657 /* if one page is a write-back page from client cache and
658 * not from direct_io, or it's written by root, then mark
659 * the whole io request as ignore quota request, remote
660 * client can not break through quota. */
661 if (exp_connect_rmtclient(exp))
662 flags &= ~OBD_BRW_NOQUOTA;
663 if ((flags & OBD_BRW_NOQUOTA) ||
664 (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
666 iobuf->dr_ignore_quota = 1;
668 if (!(lnb->flags & OBD_BRW_ASYNC)) {
669 sync_journal_commit = 1;
673 /* we try to get enough quota to write here, and let ldiskfs
674 * decide if it is out of quota or not b=14783 */
675 rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
676 rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
677 (void *)inode, obj->ioo_bufcnt);
681 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
684 ll_vfs_dq_init(inode);
685 fsfilt_check_slow(obd, now, "quota init");
687 LOCK_INODE_MUTEX(inode);
688 fsfilt_check_slow(obd, now, "i_mutex");
689 oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
691 if (IS_ERR(oti->oti_handle)) {
692 UNLOCK_INODE_MUTEX(inode);
693 rc = PTR_ERR(oti->oti_handle);
694 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
695 "error starting transaction: rc = %d\n", rc);
696 oti->oti_handle = NULL;
699 /* have to call fsfilt_commit() from this point on */
701 fsfilt_check_slow(obd, now, "brw_start");
703 i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
705 /* If the inode still has SUID+SGID bits set (see filter_precreate())
706 * then we will accept the UID+GID if sent by the client for
707 * initializing the ownership of this inode. We only allow this to
708 * happen once (so clear these bits) and later only allow setattr. */
709 if (inode->i_mode & S_ISUID)
711 if (inode->i_mode & S_ISGID)
714 iattr_from_obdo(&iattr, oa, i);
715 if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
718 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
719 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
721 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
723 iattr.ia_valid |= ATTR_MODE;
724 iattr.ia_mode = inode->i_mode;
725 if (iattr.ia_valid & ATTR_UID)
726 iattr.ia_mode &= ~S_ISUID;
727 if (iattr.ia_valid & ATTR_GID)
728 iattr.ia_mode &= ~S_ISGID;
730 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
732 /* To avoid problems with quotas, UID and GID must be set
733 * in the inode before filter_direct_io() - see bug 10357. */
734 save = iattr.ia_valid;
735 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
736 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
737 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
738 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
739 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
742 /* filter_direct_io drops i_mutex */
743 rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
744 oti, sync_journal_commit ? &wait_handle : NULL);
746 obdo_from_inode(oa, inode, NULL, rc == 0 ? FILTER_VALID_FLAGS : 0 |
747 OBD_MD_FLUID |OBD_MD_FLGID);
749 lquota_getflag(filter_quota_interface_ref, obd, oa);
751 fsfilt_check_slow(obd, now, "direct_io");
754 err = fsfilt_commit_wait(obd, inode, wait_handle);
759 CERROR("Failure to commit OST transaction (%d)?\n", err);
764 if (obd->obd_replayable && !rc && wait_handle)
765 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
766 "oti_transno "LPU64" last_committed "LPU64"\n",
767 oti->oti_transno, obd->obd_last_committed);
769 fsfilt_check_slow(obd, now, "commitrw commit");
772 lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
775 filter_grant_commit(exp, niocount, res);
777 switch (cleanup_phase) {
779 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
780 LASSERT(current->journal_info == NULL);
782 filter_iobuf_put(&obd->u.filter, iobuf, oti);
785 * lnb->page automatically returns back into per-thread page
791 /* trigger quota pre-acquire */
792 err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
794 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
795 "(rc:%d, uid:%u, gid:%u)\n",
796 err, qcids[USRQUOTA], qcids[GRPQUOTA]);
797 if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
798 qcids[USRQUOTA] = oa->o_uid;
799 qcids[GRPQUOTA] = oa->o_gid;
800 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
801 NULL, rc, FSFILT_OP_CREATE);
802 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
803 "(rc:%d, uid:%u, gid:%u)\n",
804 err, qcids[USRQUOTA], qcids[GRPQUOTA]);
807 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
808 if (lnb->page == NULL)
812 /* If the write has failed, the page cache may
813 * not be consitent with what is on disk, so
814 * force pages to be reread next time it is
816 ClearPageUptodate(lnb->page);
818 LASSERT(PageLocked(lnb->page));
819 unlock_page(lnb->page);
821 page_cache_release(lnb->page);
827 if (fo->fo_writethrough_cache == 0 ||
828 i_size_read(inode) > fo->fo_readcache_max_filesize)
829 filter_release_cache(obd, obj, nb, inode);
830 up_read(&inode->i_alloc_sem);