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 2008 Sun Microsystems, Inc. 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 atomic_t dr_numreqs; /* number of reqs being processed */
62 wait_queue_head_t dr_wait;
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 atomic_inc(&iobuf->dr_numreqs);
79 if (rw == OBD_BRW_READ) {
80 atomic_inc(&filter->fo_r_in_flight);
81 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
82 atomic_read(&filter->fo_r_in_flight));
83 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
85 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_RPC_HIST],
86 atomic_read(&filter->fo_r_in_flight));
87 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DISK_IOSIZE], size);
89 atomic_inc(&filter->fo_w_in_flight);
90 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
91 atomic_read(&filter->fo_w_in_flight));
92 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
94 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_RPC_HIST],
95 atomic_read(&filter->fo_w_in_flight));
96 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DISK_IOSIZE], size);
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 atomic_dec(&filter->fo_r_in_flight);
110 atomic_dec(&filter->fo_w_in_flight);
112 if (atomic_dec_and_test(&iobuf->dr_numreqs))
113 wake_up(&iobuf->dr_wait);
116 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
118 struct filter_iobuf *iobuf = bio->bi_private;
122 /* CAVEAT EMPTOR: possibly in IRQ context
123 * DO NOT record procfs stats here!!! */
125 if (bio->bi_size) /* Not complete */
128 if (unlikely(iobuf == NULL)) {
129 CERROR("***** bio->bi_private is NULL! This should never "
130 "happen. Normally, I would crash here, but instead I "
131 "will dump the bio contents to the console. Please "
132 "report this to <http://bugzilla.lustre.org/> , along "
133 "with any interesting messages leading up to this point "
134 "(like SCSI errors, perhaps). Because bi_private is "
135 "NULL, I can't wake up the thread that initiated this "
136 "I/O -- so you will probably have to reboot this node.\n");
137 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
138 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
139 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
140 bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
141 bio->bi_end_io, atomic_read(&bio->bi_cnt),
146 /* the check is outside of the cycle for performance reason -bzzz */
147 if (!test_bit(BIO_RW, &bio->bi_rw)) {
148 bio_for_each_segment(bvl, bio, i) {
149 if (likely(error == 0))
150 SetPageUptodate(bvl->bv_page);
151 LASSERT(PageLocked(bvl->bv_page));
152 #ifdef HAVE_PAGE_CONSTANT
153 ClearPageConstant(bvl->bv_page);
156 record_finish_io(iobuf, OBD_BRW_READ, error);
158 #ifdef HAVE_PAGE_CONSTANT
159 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
160 bio_for_each_segment(bvl, bio, i) {
161 ClearPageConstant(bvl->bv_page);
165 record_finish_io(iobuf, OBD_BRW_WRITE, error);
168 /* any real error is good enough -bzzz */
169 if (error != 0 && iobuf->dr_error == 0)
170 iobuf->dr_error = error;
172 /* Completed bios used to be chained off iobuf->dr_bios and freed in
173 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
174 * mempool when serious on-disk fragmentation was encountered,
175 * deadlocking the OST. The bios are now released as soon as complete
176 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
181 static int can_be_merged(struct bio *bio, sector_t sector)
188 size = bio->bi_size >> 9;
189 return bio->bi_sector + size == sector ? 1 : 0;
192 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
193 int rw, int num_pages)
195 struct filter_iobuf *iobuf;
197 LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
199 OBD_ALLOC(iobuf, sizeof(*iobuf));
203 OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
204 if (iobuf->dr_pages == NULL)
207 OBD_ALLOC(iobuf->dr_blocks,
208 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
209 if (iobuf->dr_blocks == NULL)
212 iobuf->dr_filter = filter;
213 init_waitqueue_head(&iobuf->dr_wait);
214 atomic_set(&iobuf->dr_numreqs, 0);
215 iobuf->dr_max_pages = num_pages;
216 iobuf->dr_npages = 0;
222 OBD_FREE(iobuf->dr_pages,
223 num_pages * sizeof(*iobuf->dr_pages));
225 OBD_FREE(iobuf, sizeof(*iobuf));
227 RETURN(ERR_PTR(-ENOMEM));
230 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
232 iobuf->dr_npages = 0;
234 atomic_set(&iobuf->dr_numreqs, 0);
237 void filter_free_iobuf(struct filter_iobuf *iobuf)
239 int num_pages = iobuf->dr_max_pages;
241 filter_clear_iobuf(iobuf);
243 OBD_FREE(iobuf->dr_blocks,
244 MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
245 OBD_FREE(iobuf->dr_pages,
246 num_pages * sizeof(*iobuf->dr_pages));
250 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
251 struct obd_trans_info *oti)
253 int thread_id = oti ? oti->oti_thread_id : -1;
255 if (unlikely(thread_id < 0)) {
256 filter_free_iobuf(iobuf);
260 LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
261 "iobuf mismatch for thread %d: pool %p iobuf %p\n",
262 thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
263 filter_clear_iobuf(iobuf);
266 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
267 struct inode *inode, struct page *page)
269 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
270 iobuf->dr_pages[iobuf->dr_npages++] = page;
275 int filter_do_bio(struct obd_export *exp, struct inode *inode,
276 struct filter_iobuf *iobuf, int rw)
278 struct obd_device *obd = exp->exp_obd;
279 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
280 struct page **pages = iobuf->dr_pages;
281 int npages = iobuf->dr_npages;
282 unsigned long *blocks = iobuf->dr_blocks;
283 int total_blocks = npages * blocks_per_page;
284 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
285 unsigned int blocksize = inode->i_sb->s_blocksize;
286 struct bio *bio = NULL;
288 unsigned long start_time = jiffies;
290 unsigned int page_offset;
299 LASSERT(iobuf->dr_npages == npages);
300 LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
302 for (page_idx = 0, block_idx = 0;
304 page_idx++, block_idx += blocks_per_page) {
306 page = pages[page_idx];
307 LASSERT (block_idx + blocks_per_page <= total_blocks);
309 for (i = 0, page_offset = 0;
311 i += nblocks, page_offset += blocksize * nblocks) {
315 if (blocks[block_idx + i] == 0) { /* hole */
316 LASSERT(rw == OBD_BRW_READ);
317 memset(kmap(page) + page_offset, 0, blocksize);
322 sector = (sector_t)blocks[block_idx + i] << sector_bits;
324 /* Additional contiguous file blocks? */
325 while (i + nblocks < blocks_per_page &&
326 (sector + (nblocks << sector_bits)) ==
327 ((sector_t)blocks[block_idx + i + nblocks] <<
331 #ifdef HAVE_PAGE_CONSTANT
332 /* I only set the page to be constant only if it
333 * is mapped to a contiguous underlying disk block(s).
334 * It will then make sure the corresponding device
335 * cache of raid5 will be overwritten by this page.
337 if ((rw == OBD_BRW_WRITE) &&
338 (nblocks == blocks_per_page) &&
339 mapping_cap_page_constant_write(inode->i_mapping))
340 SetPageConstant(page);
344 can_be_merged(bio, sector) &&
345 bio_add_page(bio, page,
346 blocksize * nblocks, page_offset) != 0)
347 continue; /* added this frag OK */
351 bdev_get_queue(bio->bi_bdev);
353 /* Dang! I have to fragment this I/O */
354 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
355 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
357 bio->bi_vcnt, bio->bi_max_vecs,
358 bio->bi_size >> 9, q->max_sectors,
359 bio_phys_segments(q, bio),
360 q->max_phys_segments,
361 bio_hw_segments(q, bio),
364 record_start_io(iobuf, rw, bio->bi_size, exp);
365 rc = fsfilt_send_bio(rw, obd, inode, bio);
367 CERROR("Can't send bio: %d\n", rc);
368 record_finish_io(iobuf, rw, rc);
374 /* allocate new bio */
375 bio = bio_alloc(GFP_NOIO,
376 (npages - page_idx) * blocks_per_page);
378 CERROR("Can't allocate bio %u*%u = %u pages\n",
379 (npages - page_idx), blocks_per_page,
380 (npages - page_idx) * blocks_per_page);
385 bio->bi_bdev = inode->i_sb->s_bdev;
386 bio->bi_sector = sector;
387 bio->bi_end_io = dio_complete_routine;
388 bio->bi_private = iobuf;
390 rc = bio_add_page(bio, page,
391 blocksize * nblocks, page_offset);
397 record_start_io(iobuf, rw, bio->bi_size, exp);
398 rc = fsfilt_send_bio(rw, obd, inode, bio);
403 CERROR("Can't send bio: %d\n", rc);
404 record_finish_io(iobuf, rw, rc);
409 wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
411 if (rw == OBD_BRW_READ) {
412 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS],
414 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
416 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
417 jiffies - start_time);
418 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time);
419 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
420 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_DIO_FRAGS],
422 lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_IO_TIME],
423 jiffies - start_time);
426 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS],
428 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
430 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
431 jiffies - start_time);
432 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], jiffies - start_time);
433 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
434 lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_DIO_FRAGS],
436 lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_IO_TIME],
437 jiffies - start_time);
442 rc = iobuf->dr_error;
446 /* Must be called with i_mutex taken for writes; this will drop it */
447 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
448 struct obd_export *exp, struct iattr *attr,
449 struct obd_trans_info *oti, void **wait_handle)
451 struct obd_device *obd = exp->exp_obd;
452 struct inode *inode = dchild->d_inode;
453 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
455 struct semaphore *sem;
458 LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
459 iobuf->dr_npages, iobuf->dr_max_pages);
460 LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
462 if (rw == OBD_BRW_READ) {
463 if (iobuf->dr_npages == 0)
468 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
469 LASSERT(iobuf->dr_npages > 0);
471 sem = &obd->u.filter.fo_alloc_lock;
473 lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota);
476 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
477 iobuf->dr_npages, iobuf->dr_blocks,
478 obdfilter_created_scratchpad, create, sem);
480 if (rw == OBD_BRW_WRITE) {
482 filter_tally(exp, iobuf->dr_pages,
483 iobuf->dr_npages, iobuf->dr_blocks,
485 if (attr->ia_size > i_size_read(inode))
486 attr->ia_valid |= ATTR_SIZE;
487 rc = fsfilt_setattr(obd, dchild,
488 oti->oti_handle, attr, 0);
491 UNLOCK_INODE_MUTEX(inode);
493 rc2 = filter_finish_transno(exp, oti, 0, 0);
495 CERROR("can't close transaction: %d\n", rc2);
500 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
506 } else if (rc == 0) {
507 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
508 iobuf->dr_blocks, blocks_per_page, 0);
511 RETURN(filter_do_bio(exp, inode, iobuf, rw));
514 /* See if there are unallocated parts in given file region */
515 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
517 sector_t (*fs_bmap)(struct address_space *, sector_t) =
518 inode->i_mapping->a_ops->bmap;
521 /* We can't know if we are overwriting or not */
525 offset >>= inode->i_blkbits;
526 len >>= inode->i_blkbits;
528 for (j = 0; j <= len; j++)
529 if (fs_bmap(inode->i_mapping, offset + j) == 0)
536 * interesting use cases on how it interacts with VM:
538 * - vm writeout -- shouldn't see our pages as we don't mark them dirty
539 * though vm can find partial page left dirty by truncate. in this
540 * usual writeout is used unless our write rewrite that page - then we
541 * drop PG_dirty with PG_lock held.
546 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
547 int objcount, struct obd_ioobj *obj,
548 struct niobuf_remote *nb, int niocount,
549 struct niobuf_local *res, struct obd_trans_info *oti,
552 struct niobuf_local *lnb;
553 struct filter_iobuf *iobuf = NULL;
554 struct lvfs_run_ctxt saved;
555 struct fsfilt_objinfo fso;
556 struct iattr iattr = { 0 };
557 struct inode *inode = NULL;
558 unsigned long now = jiffies;
559 int i, err, cleanup_phase = 0;
560 struct obd_device *obd = exp->exp_obd;
561 struct filter_obd *fo = &obd->u.filter;
563 int total_size = 0, rc2;
564 unsigned int qcids[MAXQUOTAS] = {0, 0};
567 LASSERT(oti != NULL);
568 LASSERT(objcount == 1);
569 LASSERT(current->journal_info == NULL);
574 /* Unfortunately, if quota master is too busy to handle the
575 * pre-dqacq in time and quota hash on ost is used up, we
576 * have to wait for the completion of in flight dqacq/dqrel,
578 if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
579 oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
580 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
581 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
590 iobuf = filter_iobuf_get(&obd->u.filter, oti);
592 GOTO(cleanup, rc = PTR_ERR(iobuf));
595 fso.fso_dentry = res->dentry;
596 fso.fso_bufcnt = obj->ioo_bufcnt;
597 inode = res->dentry->d_inode;
599 iobuf->dr_ignore_quota = 0;
600 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
603 /* If overwriting an existing block, we don't need a grant */
604 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
605 filter_range_is_mapped(inode, lnb->offset, lnb->len))
608 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
609 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
613 LASSERT(PageLocked(lnb->page));
614 LASSERT(!PageWriteback(lnb->page));
616 /* truncate might leave tail dirty */
617 clear_page_dirty_for_io(lnb->page);
619 SetPageUptodate(lnb->page);
621 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
624 total_size += lnb->len;
626 /* we expect these pages to be in offset order, but we'll
628 this_size = lnb->offset + lnb->len;
629 if (this_size > iattr.ia_size)
630 iattr.ia_size = this_size;
632 /* if one page is a write-back page from client cache, or it's
633 * written by root, then mark the whole io request as ignore
635 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
636 iobuf->dr_ignore_quota = 1;
639 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
644 LOCK_INODE_MUTEX(inode);
645 fsfilt_check_slow(obd, now, "i_mutex");
646 oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
648 if (IS_ERR(oti->oti_handle)) {
649 UNLOCK_INODE_MUTEX(inode);
650 rc = PTR_ERR(oti->oti_handle);
651 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
652 "error starting transaction: rc = %d\n", rc);
653 oti->oti_handle = NULL;
656 /* have to call fsfilt_commit() from this point on */
658 fsfilt_check_slow(obd, now, "brw_start");
660 i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
662 /* If the inode still has SUID+SGID bits set (see filter_precreate())
663 * then we will accept the UID+GID if sent by the client for
664 * initializing the ownership of this inode. We only allow this to
665 * happen once (so clear these bits) and later only allow setattr. */
666 if (inode->i_mode & S_ISUID)
668 if (inode->i_mode & S_ISGID)
671 iattr_from_obdo(&iattr, oa, i);
672 if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
675 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
676 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
678 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
680 iattr.ia_valid |= ATTR_MODE;
681 iattr.ia_mode = inode->i_mode;
682 if (iattr.ia_valid & ATTR_UID)
683 iattr.ia_mode &= ~S_ISUID;
684 if (iattr.ia_valid & ATTR_GID)
685 iattr.ia_mode &= ~S_ISGID;
687 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
689 /* To avoid problems with quotas, UID and GID must be set
690 * in the inode before filter_direct_io() - see bug 10357. */
691 save = iattr.ia_valid;
692 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
693 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
694 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
695 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
696 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
699 /* filter_direct_io drops i_mutex */
700 rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
703 obdo_from_inode(oa, inode,
704 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
706 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
708 lquota_getflag(filter_quota_interface_ref, obd, oa);
710 fsfilt_check_slow(obd, now, "direct_io");
712 err = fsfilt_commit_wait(obd, inode, wait_handle);
714 CERROR("Failure to commit OST transaction (%d)?\n", err);
718 if (obd->obd_replayable && !rc)
719 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
720 "oti_transno "LPU64" last_committed "LPU64"\n",
721 oti->oti_transno, obd->obd_last_committed);
723 fsfilt_check_slow(obd, now, "commitrw commit");
726 filter_grant_commit(exp, niocount, res);
728 switch (cleanup_phase) {
730 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
731 LASSERT(current->journal_info == NULL);
733 filter_iobuf_put(&obd->u.filter, iobuf, oti);
736 * lnb->page automatically returns back into per-thread page
742 /* trigger quota pre-acquire */
743 qcids[USRQUOTA] = oa->o_uid;
744 qcids[GRPQUOTA] = oa->o_gid;
745 err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
747 CDEBUG(err ? D_ERROR : D_QUOTA,
748 "filter adjust qunit! (rc:%d)\n", err);
750 for (i = 0, lnb = res; i < niocount; i++, lnb++) {
751 if (lnb->page == NULL)
754 LASSERT(PageLocked(lnb->page));
755 unlock_page(lnb->page);
757 page_cache_release(lnb->page);
761 if (inode && (fo->fo_writethrough_cache == 0 ||
762 i_size_read(inode) > fo->fo_readcache_max_filesize))
763 filter_invalidate_cache(obd, obj, nb, inode);