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) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 #include <linux/module.h>
44 #include <linux/pagemap.h> // XXX kill me soon
45 #include <linux/version.h>
46 #include <linux/buffer_head.h>
48 #define DEBUG_SUBSYSTEM S_FILTER
50 #include <obd_class.h>
51 #include <lustre_fsfilt.h>
52 #include <lustre_quota.h>
53 #include "filter_internal.h"
55 /* 512byte block min */
56 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
58 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
59 struct obd_export *exp)
61 struct filter_obd *filter = iobuf->dr_filter;
63 cfs_atomic_inc(&iobuf->dr_numreqs);
65 if (rw == OBD_BRW_READ) {
66 cfs_atomic_inc(&filter->fo_r_in_flight);
67 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
68 cfs_atomic_read(&filter->fo_r_in_flight));
69 lprocfs_oh_tally_log2(&filter->
70 fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
72 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
73 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
75 cfs_atomic_read(&filter-> \
77 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
78 nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
82 cfs_atomic_inc(&filter->fo_w_in_flight);
83 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
84 cfs_atomic_read(&filter->fo_w_in_flight));
85 lprocfs_oh_tally_log2(&filter->
86 fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
88 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
89 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
91 cfs_atomic_read(&filter-> \
93 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
94 nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
100 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
102 struct filter_obd *filter = iobuf->dr_filter;
104 /* CAVEAT EMPTOR: possibly in IRQ context
105 * DO NOT record procfs stats here!!! */
107 if (rw == OBD_BRW_READ)
108 cfs_atomic_dec(&filter->fo_r_in_flight);
110 cfs_atomic_dec(&filter->fo_w_in_flight);
112 if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
113 cfs_waitq_signal(&iobuf->dr_wait);
116 #ifdef HAVE_BIO_ENDIO_2ARG
117 #define DIO_RETURN(a)
118 static void dio_complete_routine(struct bio *bio, int error)
120 #define DIO_RETURN(a) return(a)
121 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
124 struct filter_iobuf *iobuf = bio->bi_private;
128 /* CAVEAT EMPTOR: possibly in IRQ context
129 * DO NOT record procfs stats here!!! */
131 #ifdef HAVE_BIO_ENDIO_2ARG
132 /* The "bi_size" check was needed for kernels < 2.6.24 in order to
133 * handle the case where a SCSI request error caused this callback
134 * to be called before all of the biovecs had been processed.
135 * Without this check the server thread will hang. In newer kernels
136 * the bio_end_io routine is never called for partial completions,
137 * so this check is no longer needed. */
138 if (bio->bi_size) /* Not complete */
142 if (unlikely(iobuf == NULL)) {
143 CERROR("***** bio->bi_private is NULL! This should never "
144 "happen. Normally, I would crash here, but instead I "
145 "will dump the bio contents to the console. Please "
146 "report this to <http://bugs.whamcloud.com/>, along "
147 "with any interesting messages leading up to this point "
148 "(like SCSI errors, perhaps). Because bi_private is "
149 "NULL, I can't wake up the thread that initiated this "
150 "IO - you will probably have to reboot this node.\n");
151 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
152 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
153 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
154 bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
155 bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
160 /* the check is outside of the cycle for performance reason -bzzz */
161 if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
162 bio_for_each_segment(bvl, bio, i) {
163 if (likely(error == 0))
164 SetPageUptodate(bvl->bv_page);
165 LASSERT(PageLocked(bvl->bv_page));
166 ClearPageConstant(bvl->bv_page);
168 record_finish_io(iobuf, OBD_BRW_READ, error);
170 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
171 bio_for_each_segment(bvl, bio, i) {
172 ClearPageConstant(bvl->bv_page);
175 record_finish_io(iobuf, OBD_BRW_WRITE, error);
178 /* any real error is good enough -bzzz */
179 if (error != 0 && iobuf->dr_error == 0)
180 iobuf->dr_error = error;
182 /* Completed bios used to be chained off iobuf->dr_bios and freed in
183 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
184 * mempool when serious on-disk fragmentation was encountered,
185 * deadlocking the OST. The bios are now released as soon as complete
186 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
191 static int can_be_merged(struct bio *bio, sector_t sector)
198 size = bio->bi_size >> 9;
199 return bio->bi_sector + size == sector ? 1 : 0;
202 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
203 int rw, int num_pages)
205 struct filter_iobuf *iobuf;
207 LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
209 OBD_ALLOC(iobuf, sizeof(*iobuf));
213 OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
214 if (iobuf->dr_pages == NULL)
217 OBD_ALLOC(iobuf->dr_blocks,
218 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
219 if (iobuf->dr_blocks == NULL)
222 CFS_INIT_HLIST_NODE(&iobuf->dr_hlist);
223 iobuf->dr_filter = filter;
224 cfs_waitq_init(&iobuf->dr_wait);
225 cfs_atomic_set(&iobuf->dr_numreqs, 0);
226 iobuf->dr_max_pages = num_pages;
227 iobuf->dr_npages = 0;
233 OBD_FREE(iobuf->dr_pages,
234 num_pages * sizeof(*iobuf->dr_pages));
236 OBD_FREE(iobuf, sizeof(*iobuf));
238 RETURN(ERR_PTR(-ENOMEM));
241 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
243 iobuf->dr_npages = 0;
245 cfs_atomic_set(&iobuf->dr_numreqs, 0);
248 void filter_free_iobuf(struct filter_iobuf *iobuf)
250 int num_pages = iobuf->dr_max_pages;
252 filter_clear_iobuf(iobuf);
254 LASSERT(cfs_hlist_unhashed(&iobuf->dr_hlist));
255 OBD_FREE(iobuf->dr_blocks,
256 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
257 OBD_FREE(iobuf->dr_pages,
258 num_pages * sizeof(*iobuf->dr_pages));
262 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
263 struct obd_trans_info *oti)
265 int thread_id = (oti && oti->oti_thread) ?
266 oti->oti_thread->t_id : -1;
268 if (unlikely(thread_id < 0)) {
269 filter_free_iobuf(iobuf);
273 filter_clear_iobuf(iobuf);
276 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
277 struct inode *inode, struct page *page)
279 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
280 iobuf->dr_pages[iobuf->dr_npages++] = page;
285 int filter_do_bio(struct obd_export *exp, struct inode *inode,
286 struct filter_iobuf *iobuf, int rw)
288 struct obd_device *obd = exp->exp_obd;
289 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
290 struct page **pages = iobuf->dr_pages;
291 int npages = iobuf->dr_npages;
292 unsigned long *blocks = iobuf->dr_blocks;
293 int total_blocks = npages * blocks_per_page;
294 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
295 unsigned int blocksize = inode->i_sb->s_blocksize;
296 struct bio *bio = NULL;
298 unsigned long start_time = jiffies;
300 unsigned int page_offset;
309 LASSERT(iobuf->dr_npages == npages);
310 LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
312 for (page_idx = 0, block_idx = 0;
314 page_idx++, block_idx += blocks_per_page) {
316 page = pages[page_idx];
317 LASSERT (block_idx + blocks_per_page <= total_blocks);
319 for (i = 0, page_offset = 0;
321 i += nblocks, page_offset += blocksize * nblocks) {
325 if (blocks[block_idx + i] == 0) { /* hole */
326 LASSERT(rw == OBD_BRW_READ);
327 memset(kmap(page) + page_offset, 0, blocksize);
332 sector = (sector_t)blocks[block_idx + i] << sector_bits;
334 /* Additional contiguous file blocks? */
335 while (i + nblocks < blocks_per_page &&
336 (sector + (nblocks << sector_bits)) ==
337 ((sector_t)blocks[block_idx + i + nblocks] <<
341 /* I only set the page to be constant only if it
342 * is mapped to a contiguous underlying disk block(s).
343 * It will then make sure the corresponding device
344 * cache of raid5 will be overwritten by this page.
346 if ((rw == OBD_BRW_WRITE) &&
347 (nblocks == blocks_per_page) &&
348 mapping_cap_page_constant_write(inode->i_mapping))
349 SetPageConstant(page);
352 can_be_merged(bio, sector) &&
353 bio_add_page(bio, page,
354 blocksize * nblocks, page_offset) != 0)
355 continue; /* added this frag OK */
358 struct request_queue *q =
359 bdev_get_queue(bio->bi_bdev);
361 /* Dang! I have to fragment this I/O */
362 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
363 "sectors %d(%d) psg %d(%d) hsg %d(%d) "
364 "sector %llu next %llu\n",
366 bio->bi_vcnt, bio->bi_max_vecs,
367 bio->bi_size >> 9, queue_max_sectors(q),
368 bio_phys_segments(q, bio),
369 queue_max_phys_segments(q),
370 bio_hw_segments(q, bio),
371 queue_max_hw_segments(q),
372 (unsigned long long)bio->bi_sector,
373 (unsigned long long)sector);
375 record_start_io(iobuf, rw, bio->bi_size, exp);
376 rc = fsfilt_send_bio(rw, obd, inode, bio);
378 CERROR("Can't send bio: %d\n", rc);
379 record_finish_io(iobuf, rw, rc);
385 /* allocate new bio, limited by max BIO size, b=9945 */
386 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
387 (npages - page_idx) *
390 CERROR("Can't allocate bio %u*%u = %u pages\n",
391 (npages - page_idx), blocks_per_page,
392 (npages - page_idx) * blocks_per_page);
397 bio->bi_bdev = inode->i_sb->s_bdev;
398 bio->bi_sector = sector;
399 bio->bi_rw = (rw == OBD_BRW_READ ? READ : WRITE);
400 bio->bi_end_io = dio_complete_routine;
401 bio->bi_private = iobuf;
403 rc = bio_add_page(bio, page,
404 blocksize * nblocks, page_offset);
410 record_start_io(iobuf, rw, bio->bi_size, exp);
411 rc = fsfilt_send_bio(rw, obd, inode, bio);
416 CERROR("Can't send bio: %d\n", rc);
417 record_finish_io(iobuf, rw, rc);
422 cfs_wait_event(iobuf->dr_wait,
423 cfs_atomic_read(&iobuf->dr_numreqs) == 0);
425 if (rw == OBD_BRW_READ) {
426 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
427 hist[BRW_R_DIO_FRAGS],
429 lprocfs_oh_tally_log2(&obd->u.filter.
430 fo_filter_stats.hist[BRW_R_IO_TIME],
431 jiffies - start_time);
432 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
433 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
434 hist[BRW_R_DIO_FRAGS],
436 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
437 nid_brw_stats->hist[BRW_R_IO_TIME],
438 jiffies - start_time);
441 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
442 hist[BRW_W_DIO_FRAGS], frags);
443 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
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_W_DIO_FRAGS],
450 lprocfs_oh_tally_log2(&exp->exp_nid_stats->
451 nid_brw_stats->hist[BRW_W_IO_TIME],
452 jiffies - start_time);
457 rc = iobuf->dr_error;
461 /* Must be called with i_mutex taken for writes; this will drop it */
462 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
463 struct obd_export *exp, struct iattr *attr,
464 struct obd_trans_info *oti, void **wait_handle)
466 struct obd_device *obd = exp->exp_obd;
467 struct inode *inode = dchild->d_inode;
468 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
473 LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
474 iobuf->dr_npages, iobuf->dr_max_pages);
475 LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
477 if (rw == OBD_BRW_READ) {
478 if (iobuf->dr_npages == 0)
483 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
484 LASSERT(iobuf->dr_npages > 0);
486 mutex = &obd->u.filter.fo_alloc_lock;
488 lquota_enforce(filter_quota_interface_ref, obd,
489 iobuf->dr_ignore_quota);
492 if (rw == OBD_BRW_WRITE &&
493 OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
496 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
497 iobuf->dr_npages, iobuf->dr_blocks,
498 obdfilter_created_scratchpad, create, mutex);
501 if (rw == OBD_BRW_WRITE) {
503 filter_tally(exp, iobuf->dr_pages,
504 iobuf->dr_npages, iobuf->dr_blocks,
506 if (attr->ia_size > i_size_read(inode))
507 attr->ia_valid |= ATTR_SIZE;
508 rc = fsfilt_setattr(obd, dchild,
509 oti->oti_handle, attr, 0);
512 mutex_unlock(&inode->i_mutex);
514 /* Force commit to make the just-deleted blocks
515 * reusable. LU-456 */
517 fsfilt_commit(obd, inode, oti->oti_handle, 1);
521 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
523 CERROR("can't close transaction: %d\n", rc2);
529 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
532 rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
537 } else if (rc == 0) {
538 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
539 iobuf->dr_blocks, blocks_per_page, 0);
542 RETURN(filter_do_bio(exp, inode, iobuf, rw));
545 /* See if there are unallocated parts in given file region */
546 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
548 sector_t (*fs_bmap)(struct address_space *, sector_t) =
549 inode->i_mapping->a_ops->bmap;
552 /* We can't know if we are overwriting or not */
556 offset >>= inode->i_blkbits;
557 len >>= inode->i_blkbits;
559 for (j = 0; j <= len; j++)
560 if (fs_bmap(inode->i_mapping, offset + j) == 0)
567 * interesting use cases on how it interacts with VM:
569 * - vm writeout -- shouldn't see our pages as we don't mark them dirty
570 * though vm can find partial page left dirty by truncate. in this
571 * usual writeout is used unless our write rewrite that page - then we
572 * drop PG_dirty with PG_lock held.
577 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
578 int objcount, struct obd_ioobj *obj,
579 struct niobuf_remote *nb, int niocount,
580 struct niobuf_local *res, struct obd_trans_info *oti,
583 struct niobuf_local *lnb;
584 struct filter_iobuf *iobuf = NULL;
585 struct lvfs_run_ctxt saved;
586 struct fsfilt_objinfo fso;
587 struct iattr iattr = { 0 };
588 struct inode *inode = res->dentry->d_inode;
589 unsigned long now = jiffies;
590 int i, err, cleanup_phase = 0;
591 struct obd_device *obd = exp->exp_obd;
592 struct filter_obd *fo = &obd->u.filter;
593 void *wait_handle = NULL;
595 unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
596 int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
597 int sync_journal_commit = obd->u.filter.fo_syncjournal;
601 LASSERT(oti != NULL);
602 LASSERT(objcount == 1);
603 LASSERT(current->journal_info == NULL);
608 iobuf = filter_iobuf_get(&obd->u.filter, oti);
610 GOTO(cleanup, rc = PTR_ERR(iobuf));
613 fso.fso_dentry = res->dentry;
614 fso.fso_bufcnt = obj->ioo_bufcnt;
616 iobuf->dr_ignore_quota = 0;
617 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
619 __u32 flags = lnb->flags;
621 if (filter_range_is_mapped(inode, lnb->lnb_file_offset,
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->lnb_file_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 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
682 GOTO(cleanup, rc = -EINPROGRESS);
684 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
687 fsfilt_check_slow(obd, now, "quota init");
690 mutex_lock(&inode->i_mutex);
691 fsfilt_check_slow(obd, now, "i_mutex");
692 oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
694 if (IS_ERR(oti->oti_handle)) {
695 mutex_unlock(&inode->i_mutex);
696 rc = PTR_ERR(oti->oti_handle);
697 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
698 "error starting transaction: rc = %d\n", rc);
699 oti->oti_handle = NULL;
702 /* have to call fsfilt_commit() from this point on */
704 fsfilt_check_slow(obd, now, "brw_start");
706 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
707 ll_vfs_dq_init(inode);
709 i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
711 /* If the inode still has SUID+SGID bits set (see filter_precreate())
712 * then we will accept the UID+GID if sent by the client for
713 * initializing the ownership of this inode. We only allow this to
714 * happen once (so clear these bits) and later only allow setattr. */
715 if (inode->i_mode & S_ISUID)
717 if (inode->i_mode & S_ISGID)
720 iattr_from_obdo(&iattr, oa, i);
721 if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
724 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
725 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
727 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
729 iattr.ia_valid |= ATTR_MODE;
730 iattr.ia_mode = inode->i_mode;
731 if (iattr.ia_valid & ATTR_UID)
732 iattr.ia_mode &= ~S_ISUID;
733 if (iattr.ia_valid & ATTR_GID)
734 iattr.ia_mode &= ~S_ISGID;
736 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
738 /* To avoid problems with quotas, UID and GID must be set
739 * in the inode before filter_direct_io() - see bug 10357. */
740 save = iattr.ia_valid;
741 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
742 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
743 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
744 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
745 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
748 CDEBUG(D_INODE, "FID "DFID" to write: s="LPU64" m="LPU64" a="LPU64
749 " c="LPU64" b="LPU64"\n",
750 oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
751 oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
753 /* filter_direct_io drops i_mutex */
754 rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
755 oti, sync_journal_commit ? &wait_handle : NULL);
756 if (rc == -ENOSPC && retries++ < 3) {
757 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
759 oti->oti_handle = NULL;
760 fsfilt_check_slow(obd, now, "direct_io");
764 obdo_from_inode(oa, inode, (rc == 0 ? FILTER_VALID_FLAGS : 0) |
765 OBD_MD_FLUID | OBD_MD_FLGID);
767 CDEBUG(D_INODE, "FID "DFID" after write: s="LPU64" m="LPU64" a="LPU64
768 " c="LPU64" b="LPU64"\n",
769 oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
770 oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
772 lquota_getflag(filter_quota_interface_ref, obd, oa);
774 fsfilt_check_slow(obd, now, "direct_io");
777 err = fsfilt_commit_wait(obd, inode, wait_handle);
782 CERROR("Failure to commit OST transaction (%d)?\n", err);
787 /* In rare cases fsfilt_commit_wait() will wake up and return after
788 * the transaction has finished its work and updated j_commit_sequence
789 * but the commit callbacks have not been run yet. Wait here until
790 * that is finished so that clients requesting sync IO don't see the
791 * reply transno < last_committed. LU-753 */
792 if (unlikely(obd->obd_replayable && !rc && wait_handle &&
793 oti->oti_transno > obd->obd_last_committed)) {
795 struct l_wait_info lwi =
796 LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
797 (cfs_duration_t)((HZ + 4)/5),
801 oti->oti_transno <= obd->obd_last_committed,
804 /* commit callback isn't done after waiting for 5 secs ? */
805 if (unlikely(oti->oti_transno > obd->obd_last_committed))
806 CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
807 oti->oti_transno, obd->obd_last_committed);
810 fsfilt_check_slow(obd, now, "commitrw commit");
813 lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
816 filter_grant_commit(exp, niocount, res);
818 switch (cleanup_phase) {
820 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
821 LASSERT(current->journal_info == NULL);
823 filter_iobuf_put(&obd->u.filter, iobuf, oti);
826 * lnb->page automatically returns back into per-thread page
832 /* trigger quota pre-acquire */
833 err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
835 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
836 "(rc:%d, uid:%u, gid:%u)\n",
837 err, qcids[USRQUOTA], qcids[GRPQUOTA]);
838 if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
839 qcids[USRQUOTA] = oa->o_uid;
840 qcids[GRPQUOTA] = oa->o_gid;
841 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
842 NULL, rc, FSFILT_OP_CREATE);
843 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
844 "(rc:%d, uid:%u, gid:%u)\n",
845 err, qcids[USRQUOTA], qcids[GRPQUOTA]);
848 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
849 if (lnb->page == NULL)
853 /* If the write has failed, the page cache may
854 * not be consitent with what is on disk, so
855 * force pages to be reread next time it is
857 ClearPageUptodate(lnb->page);
859 LASSERT(PageLocked(lnb->page));
860 unlock_page(lnb->page);
862 page_cache_release(lnb->page);
868 if (fo->fo_writethrough_cache == 0 ||
869 i_size_read(inode) > fo->fo_readcache_max_filesize)
870 filter_release_cache(obd, obj, nb, inode);
871 up_read(&inode->i_alloc_sem);