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) 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
40 * Author: Nikita Danilov <nikita@clusterfs.com>
41 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
45 /* LUSTRE_VERSION_CODE */
46 #include <lustre_ver.h>
47 /* prerequisite for linux/xattr.h */
48 #include <linux/types.h>
49 /* prerequisite for linux/xattr.h */
53 #include <ldiskfs/ldiskfs.h>
54 #include <ldiskfs/ldiskfs_jbd2.h>
55 #include <ldiskfs/ldiskfs_extents.h>
58 * struct OBD_{ALLOC,FREE}*()
61 #include <obd_support.h>
63 #include "osd_internal.h"
65 #ifndef HAVE_PAGE_CONSTANT
66 #define mapping_cap_page_constant_write(mapping) 0
67 #define SetPageConstant(page) do {} while (0)
68 #define ClearPageConstant(page) do {} while (0)
71 #ifndef HAS_GENERIC_ERROR_REMOVE_PAGE
72 int generic_error_remove_page(struct address_space *mapping, struct page *page)
77 if (mapping != page->mapping)
80 * Only punch for normal data pages for now.
81 * Handling other types like directories would need more auditing.
83 if (!S_ISREG(mapping->host->i_mode))
86 if (page_mapped(page)) {
87 unmap_mapping_range(mapping,
88 (loff_t)page->index << PAGE_CACHE_SHIFT,
91 truncate_complete_page(mapping, page);
96 static void __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
99 LASSERTF(iobuf->dr_elapsed_valid == 0,
100 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
101 cfs_atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
104 cfs_waitq_init(&iobuf->dr_wait);
105 cfs_atomic_set(&iobuf->dr_numreqs, 0);
106 iobuf->dr_max_pages = PTLRPC_MAX_BRW_PAGES;
107 iobuf->dr_npages = 0;
111 iobuf->dr_elapsed = 0;
112 /* must be counted before, so assert */
114 iobuf->dr_init_at = line;
116 #define osd_init_iobuf(dev,iobuf,rw) __osd_init_iobuf(dev, iobuf, rw, __LINE__)
118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
120 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
121 iobuf->dr_pages[iobuf->dr_npages++] = page;
124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
126 int rw = iobuf->dr_rw;
128 if (iobuf->dr_elapsed_valid) {
129 iobuf->dr_elapsed_valid = 0;
130 LASSERT(iobuf->dr_dev == d);
131 LASSERT(iobuf->dr_frags > 0);
132 lprocfs_oh_tally(&d->od_brw_stats.
133 hist[BRW_R_DIO_FRAGS+rw],
135 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
140 #ifdef HAVE_BIO_ENDIO_2ARG
141 #define DIO_RETURN(a)
142 static void dio_complete_routine(struct bio *bio, int error)
144 #define DIO_RETURN(a) return(a)
145 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
148 struct osd_iobuf *iobuf = bio->bi_private;
152 /* CAVEAT EMPTOR: possibly in IRQ context
153 * DO NOT record procfs stats here!!! */
155 if (unlikely(iobuf == NULL)) {
156 CERROR("***** bio->bi_private is NULL! This should never "
157 "happen. Normally, I would crash here, but instead I "
158 "will dump the bio contents to the console. Please "
159 "report this to <http://jira.whamcloud.com/> , along "
160 "with any interesting messages leading up to this point "
161 "(like SCSI errors, perhaps). Because bi_private is "
162 "NULL, I can't wake up the thread that initiated this "
163 "IO - you will probably have to reboot this node.\n");
164 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
165 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
166 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
167 bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
168 bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
173 /* the check is outside of the cycle for performance reason -bzzz */
174 if (!test_bit(BIO_RW, &bio->bi_rw)) {
175 bio_for_each_segment(bvl, bio, i) {
176 if (likely(error == 0))
177 SetPageUptodate(bvl->bv_page);
178 LASSERT(PageLocked(bvl->bv_page));
179 ClearPageConstant(bvl->bv_page);
181 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
183 struct page *p = iobuf->dr_pages[0];
185 if (mapping_cap_page_constant_write(p->mapping)) {
186 bio_for_each_segment(bvl, bio, i) {
187 ClearPageConstant(bvl->bv_page);
191 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
194 /* any real error is good enough -bzzz */
195 if (error != 0 && iobuf->dr_error == 0)
196 iobuf->dr_error = error;
199 * set dr_elapsed before dr_numreqs turns to 0, otherwise
200 * it's possible that service thread will see dr_numreqs
201 * is zero, but dr_elapsed is not set yet, leading to lost
202 * data in this processing and an assertion in a subsequent
205 if (cfs_atomic_read(&iobuf->dr_numreqs) == 1) {
206 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
207 iobuf->dr_elapsed_valid = 1;
209 if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
210 cfs_waitq_signal(&iobuf->dr_wait);
212 /* Completed bios used to be chained off iobuf->dr_bios and freed in
213 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
214 * mempool when serious on-disk fragmentation was encountered,
215 * deadlocking the OST. The bios are now released as soon as complete
216 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
221 static void record_start_io(struct osd_iobuf *iobuf, int size)
223 struct osd_device *osd = iobuf->dr_dev;
224 struct obd_histogram *h = osd->od_brw_stats.hist;
227 cfs_atomic_inc(&iobuf->dr_numreqs);
229 if (iobuf->dr_rw == 0) {
230 cfs_atomic_inc(&osd->od_r_in_flight);
231 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
232 cfs_atomic_read(&osd->od_r_in_flight));
233 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
234 } else if (iobuf->dr_rw == 1) {
235 cfs_atomic_inc(&osd->od_w_in_flight);
236 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
237 cfs_atomic_read(&osd->od_w_in_flight));
238 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
244 static void osd_submit_bio(int rw, struct bio *bio)
246 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
248 submit_bio(READ, bio);
250 submit_bio(WRITE, bio);
253 static int can_be_merged(struct bio *bio, sector_t sector)
260 size = bio->bi_size >> 9;
261 return bio->bi_sector + size == sector ? 1 : 0;
264 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
265 struct osd_iobuf *iobuf)
267 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
268 struct page **pages = iobuf->dr_pages;
269 int npages = iobuf->dr_npages;
270 unsigned long *blocks = iobuf->dr_blocks;
271 int total_blocks = npages * blocks_per_page;
272 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
273 unsigned int blocksize = inode->i_sb->s_blocksize;
274 struct bio *bio = NULL;
276 unsigned int page_offset;
285 LASSERT(iobuf->dr_npages == npages);
287 osd_brw_stats_update(osd, iobuf);
288 iobuf->dr_start_time = cfs_time_current();
290 for (page_idx = 0, block_idx = 0;
292 page_idx++, block_idx += blocks_per_page) {
294 page = pages[page_idx];
295 LASSERT(block_idx + blocks_per_page <= total_blocks);
297 for (i = 0, page_offset = 0;
299 i += nblocks, page_offset += blocksize * nblocks) {
303 if (blocks[block_idx + i] == 0) { /* hole */
304 LASSERTF(iobuf->dr_rw == 0,
305 "page_idx %u, block_idx %u, i %u\n",
306 page_idx, block_idx, i);
307 memset(kmap(page) + page_offset, 0, blocksize);
312 sector = (sector_t)blocks[block_idx + i] << sector_bits;
314 /* Additional contiguous file blocks? */
315 while (i + nblocks < blocks_per_page &&
316 (sector + (nblocks << sector_bits)) ==
317 ((sector_t)blocks[block_idx + i + nblocks] <<
321 /* I only set the page to be constant only if it
322 * is mapped to a contiguous underlying disk block(s).
323 * It will then make sure the corresponding device
324 * cache of raid5 will be overwritten by this page.
326 if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
327 mapping_cap_page_constant_write(inode->i_mapping))
328 SetPageConstant(page);
331 can_be_merged(bio, sector) &&
332 bio_add_page(bio, page,
333 blocksize * nblocks, page_offset) != 0)
334 continue; /* added this frag OK */
337 struct request_queue *q =
338 bdev_get_queue(bio->bi_bdev);
340 /* Dang! I have to fragment this I/O */
341 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
342 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
344 bio->bi_vcnt, bio->bi_max_vecs,
345 bio->bi_size >> 9, queue_max_sectors(q),
346 bio_phys_segments(q, bio),
347 queue_max_phys_segments(q),
348 bio_hw_segments(q, bio),
349 queue_max_hw_segments(q));
351 record_start_io(iobuf, bio->bi_size);
352 osd_submit_bio(iobuf->dr_rw, bio);
355 /* allocate new bio */
356 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
357 (npages - page_idx) *
360 CERROR("Can't allocate bio %u*%u = %u pages\n",
361 (npages - page_idx), blocks_per_page,
362 (npages - page_idx) * blocks_per_page);
367 bio->bi_bdev = inode->i_sb->s_bdev;
368 bio->bi_sector = sector;
369 bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
370 bio->bi_end_io = dio_complete_routine;
371 bio->bi_private = iobuf;
373 rc = bio_add_page(bio, page,
374 blocksize * nblocks, page_offset);
380 record_start_io(iobuf, bio->bi_size);
381 osd_submit_bio(iobuf->dr_rw, bio);
386 /* in order to achieve better IO throughput, we don't wait for writes
387 * completion here. instead we proceed with transaction commit in
388 * parallel and wait for IO completion once transaction is stopped
389 * see osd_trans_stop() for more details -bzzz */
390 if (iobuf->dr_rw == 0) {
391 cfs_wait_event(iobuf->dr_wait,
392 cfs_atomic_read(&iobuf->dr_numreqs) == 0);
396 rc = iobuf->dr_error;
400 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
401 struct niobuf_local *lnb)
408 int poff = offset & (CFS_PAGE_SIZE - 1);
409 int plen = CFS_PAGE_SIZE - poff;
413 lnb->lnb_file_offset = offset;
414 lnb->lnb_page_offset = poff;
416 /* lb->flags = rnb->flags; */
421 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
432 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
434 struct inode *inode = osd_dt_obj(dt)->oo_inode;
435 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
440 page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
441 GFP_NOFS | __GFP_HIGHMEM);
442 if (unlikely(page == NULL))
443 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
449 * there are following "locks":
466 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
467 ssize_t len, struct niobuf_local *lnb, int rw,
468 struct lustre_capa *capa)
470 struct osd_object *obj = osd_dt_obj(d);
471 int npages, i, rc = 0;
473 LASSERT(obj->oo_inode);
475 osd_map_remote_to_local(pos, len, &npages, lnb);
477 for (i = 0; i < npages; i++, lnb++) {
479 /* We still set up for ungranted pages so that granted pages
480 * can be written to disk as they were promised, and portals
481 * needs to keep the pages all aligned properly. */
482 lnb->dentry = (void *) obj;
484 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
485 if (lnb->page == NULL)
486 GOTO(cleanup, rc = -ENOMEM);
488 /* DLM locking protects us from write and truncate competing
489 * for same region, but truncate can leave dirty page in the
490 * cache. it's possible the writeout on a such a page is in
491 * progress when we access it. it's also possible that during
492 * this writeout we put new (partial) data, but then won't
493 * be able to proceed in filter_commitrw_write(). thus let's
494 * just wait for writeout completion, should be rare enough.
496 wait_on_page_writeback(lnb->page);
497 BUG_ON(PageWriteback(lnb->page));
499 lu_object_get(&d->do_lu);
507 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
508 struct niobuf_local *lnb, int npages)
510 struct osd_thread_info *oti = osd_oti_get(env);
511 struct osd_iobuf *iobuf = &oti->oti_iobuf;
512 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
515 /* to do IO stats, notice we do this here because
516 * osd_do_bio() doesn't wait for write to complete */
517 osd_fini_iobuf(d, iobuf);
519 for (i = 0; i < npages; i++) {
520 if (lnb[i].page == NULL)
522 LASSERT(PageLocked(lnb[i].page));
523 unlock_page(lnb[i].page);
524 page_cache_release(lnb[i].page);
525 lu_object_put(env, &dt->do_lu);
531 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
532 struct niobuf_local *lnb, int npages)
534 struct osd_thread_info *oti = osd_oti_get(env);
535 struct osd_iobuf *iobuf = &oti->oti_iobuf;
536 struct inode *inode = osd_dt_obj(dt)->oo_inode;
537 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
538 struct timeval start;
540 unsigned long timediff;
549 osd_init_iobuf(osd, iobuf, 0);
551 isize = i_size_read(inode);
552 maxidx = ((isize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
554 if (osd->od_writethrough_cache)
556 if (isize > osd->od_readcache_max_filesize)
559 cfs_gettimeofday(&start);
560 for (i = 0; i < npages; i++) {
563 generic_error_remove_page(inode->i_mapping,
567 * till commit the content of the page is undefined
568 * we'll set it uptodate once bulk is done. otherwise
569 * subsequent reads can access non-stable data
571 ClearPageUptodate(lnb[i].page);
573 if (lnb[i].len == CFS_PAGE_SIZE)
576 if (maxidx >= lnb[i].page->index) {
577 osd_iobuf_add_page(iobuf, lnb[i].page);
580 char *p = kmap(lnb[i].page);
582 off = lnb[i].lnb_page_offset;
585 off = (lnb[i].lnb_page_offset + lnb[i].len) &
588 memset(p + off, 0, CFS_PAGE_SIZE - off);
592 cfs_gettimeofday(&end);
593 timediff = cfs_timeval_sub(&end, &start, NULL);
594 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
596 if (iobuf->dr_npages) {
597 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
602 if (likely(rc == 0)) {
603 rc = osd_do_bio(osd, inode, iobuf);
604 /* do IO stats for preparation reads */
605 osd_fini_iobuf(osd, iobuf);
611 /* Check if a block is allocated or not */
612 static int osd_is_mapped(struct inode *inode, obd_size offset)
614 sector_t (*fs_bmap)(struct address_space *, sector_t);
616 fs_bmap = inode->i_mapping->a_ops->bmap;
618 /* We can't know if we are overwriting or not */
619 if (unlikely(fs_bmap == NULL))
622 if (i_size_read(inode) == 0)
625 /* Beyond EOF, must not be mapped */
626 if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
627 (offset >> inode->i_blkbits))
630 if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
636 static int osd_declare_write_commit(const struct lu_env *env,
637 struct dt_object *dt,
638 struct niobuf_local *lnb, int npages,
639 struct thandle *handle)
641 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
642 struct inode *inode = osd_dt_obj(dt)->oo_inode;
643 struct osd_thandle *oh;
650 bool ignore_quota = false;
651 long long quota_space = 0;
654 LASSERT(handle != NULL);
655 oh = container_of0(handle, struct osd_thandle, ot_super);
656 LASSERT(oh->ot_handle == NULL);
660 /* calculate number of extents (probably better to pass nb) */
661 for (i = 0; i < npages; i++) {
662 if (i && lnb[i].lnb_file_offset !=
663 lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
666 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
667 quota_space += CFS_PAGE_SIZE;
669 /* ignore quota for the whole request if any page is from
670 * client cache or written by root.
672 * XXX we could handle this on per-lnb basis as done by
674 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
675 !(lnb[i].flags & OBD_BRW_SYNC))
680 * each extent can go into new leaf causing a split
681 * 5 is max tree depth: inode + 4 index blocks
682 * with blockmaps, depth is 3 at most
684 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
686 * many concurrent threads may grow tree by the time
687 * our transaction starts. so, consider 2 is a min depth
689 depth = ext_depth(inode);
690 depth = max(depth, 1) + 1;
692 oh->ot_credits++; /* inode */
693 oh->ot_credits += depth * 2 * extents;
697 oh->ot_credits++; /* inode */
698 oh->ot_credits += depth * extents;
701 /* quota space for metadata blocks */
702 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
704 /* quota space should be reported in 1K blocks */
705 quota_space = toqb(quota_space);
707 /* each new block can go in different group (bitmap + gd) */
709 /* we can't dirty more bitmap blocks than exist */
710 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
711 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
713 oh->ot_credits += newblocks;
715 /* we can't dirty more gd blocks than exist */
716 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
717 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
719 oh->ot_credits += newblocks;
721 /* make sure the over quota flags were not set */
722 lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
724 rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
725 quota_space, oh, true, true, &flags,
728 /* we need only to store the overquota flags in the first lnb for
729 * now, once we support multiple objects BRW, this code needs be
731 if (flags & QUOTA_FL_OVER_USRQUOTA)
732 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
733 if (flags & QUOTA_FL_OVER_GRPQUOTA)
734 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
739 /* Check if a block is allocated or not */
740 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
741 struct niobuf_local *lnb, int npages,
742 struct thandle *thandle)
744 struct osd_thread_info *oti = osd_oti_get(env);
745 struct osd_iobuf *iobuf = &oti->oti_iobuf;
746 struct inode *inode = osd_dt_obj(dt)->oo_inode;
747 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
753 osd_init_iobuf(osd, iobuf, 1);
754 isize = i_size_read(inode);
755 ll_vfs_dq_init(inode);
757 for (i = 0; i < npages; i++) {
758 if (lnb[i].rc == -ENOSPC &&
759 osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
760 /* Allow the write to proceed if overwriting an
765 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
766 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
768 LASSERT(lnb[i].page);
769 generic_error_remove_page(inode->i_mapping,lnb[i].page);
773 LASSERT(PageLocked(lnb[i].page));
774 LASSERT(!PageWriteback(lnb[i].page));
776 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
777 isize = lnb[i].lnb_file_offset + lnb[i].len;
780 * Since write and truncate are serialized by oo_sem, even
781 * partial-page truncate should not leave dirty pages in the
784 LASSERT(!PageDirty(lnb[i].page));
786 SetPageUptodate(lnb[i].page);
788 osd_iobuf_add_page(iobuf, lnb[i].page);
791 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
793 } else if (iobuf->dr_npages > 0) {
794 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
800 /* no pages to write, no transno is needed */
801 thandle->th_local = 1;
804 if (likely(rc == 0)) {
805 if (isize > i_size_read(inode)) {
806 i_size_write(inode, isize);
807 LDISKFS_I(inode)->i_disksize = isize;
808 inode->i_sb->s_op->dirty_inode(inode);
811 rc = osd_do_bio(osd, inode, iobuf);
812 /* we don't do stats here as in read path because
813 * write is async: we'll do this in osd_put_bufs() */
816 if (unlikely(rc != 0)) {
817 /* if write fails, we should drop pages from the cache */
818 for (i = 0; i < npages; i++) {
819 if (lnb[i].page == NULL)
821 LASSERT(PageLocked(lnb[i].page));
822 generic_error_remove_page(inode->i_mapping,lnb[i].page);
829 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
830 struct niobuf_local *lnb, int npages)
832 struct osd_thread_info *oti = osd_oti_get(env);
833 struct osd_iobuf *iobuf = &oti->oti_iobuf;
834 struct inode *inode = osd_dt_obj(dt)->oo_inode;
835 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
836 struct timeval start, end;
837 unsigned long timediff;
838 int rc = 0, i, m = 0, cache = 0;
842 osd_init_iobuf(osd, iobuf, 0);
844 if (osd->od_read_cache)
846 if (i_size_read(inode) > osd->od_readcache_max_filesize)
849 cfs_gettimeofday(&start);
850 for (i = 0; i < npages; i++) {
852 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
853 /* If there's no more data, abort early.
854 * lnb->rc == 0, so it's easy to detect later. */
857 if (i_size_read(inode) <
858 lnb[i].lnb_file_offset + lnb[i].len - 1)
859 lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
861 lnb[i].rc = lnb[i].len;
864 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
865 if (PageUptodate(lnb[i].page)) {
866 lprocfs_counter_add(osd->od_stats,
867 LPROC_OSD_CACHE_HIT, 1);
869 lprocfs_counter_add(osd->od_stats,
870 LPROC_OSD_CACHE_MISS, 1);
871 osd_iobuf_add_page(iobuf, lnb[i].page);
874 generic_error_remove_page(inode->i_mapping,lnb[i].page);
876 cfs_gettimeofday(&end);
877 timediff = cfs_timeval_sub(&end, &start, NULL);
878 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
880 if (iobuf->dr_npages) {
881 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
886 rc = osd_do_bio(osd, inode, iobuf);
888 /* IO stats will be done in osd_bufs_put() */
895 * XXX: Another layering violation for now.
897 * We don't want to use ->f_op->read methods, because generic file write
899 * - serializes on ->i_sem, and
901 * - does a lot of extra work like balance_dirty_pages(),
903 * which doesn't work for globally shared files like /last_rcvd.
905 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
907 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
909 memcpy(buffer, (char *)ei->i_data, buflen);
914 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
916 struct buffer_head *bh;
924 /* prevent reading after eof */
925 spin_lock(&inode->i_lock);
926 if (i_size_read(inode) < *offs + size) {
927 loff_t diff = i_size_read(inode) - *offs;
928 spin_unlock(&inode->i_lock);
930 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
931 i_size_read(inode), *offs);
933 } else if (diff == 0) {
939 spin_unlock(&inode->i_lock);
942 blocksize = 1 << inode->i_blkbits;
945 block = *offs >> inode->i_blkbits;
946 boffs = *offs & (blocksize - 1);
947 csize = min(blocksize - boffs, size);
948 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
950 CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
951 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
952 csize, *offs, inode->i_ino, err);
956 memcpy(buf, bh->b_data + boffs, csize);
966 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
967 struct lu_buf *buf, loff_t *pos,
968 struct lustre_capa *capa)
970 struct inode *inode = osd_dt_obj(dt)->oo_inode;
973 if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
976 /* Read small symlink from inode body as we need to maintain correct
977 * on-disk symlinks for ldiskfs.
979 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
980 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
981 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
983 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
988 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
989 const loff_t size, loff_t pos,
990 struct thandle *handle)
992 struct osd_thandle *oh;
998 LASSERT(handle != NULL);
1000 oh = container_of0(handle, struct osd_thandle, ot_super);
1001 LASSERT(oh->ot_handle == NULL);
1003 /* XXX: size == 0 or INT_MAX indicating a catalog header update or
1004 * llog write, see comment in mdd_declare_llog_record().
1006 * This hack will be removed with llog over OSD landing
1008 if (size == DECLARE_LLOG_REWRITE)
1010 else if (size == DECLARE_LLOG_WRITE)
1013 credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1015 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1017 inode = osd_dt_obj(dt)->oo_inode;
1019 /* we may declare write to non-exist llog */
1023 /* dt_declare_write() is usually called for system objects, such
1024 * as llog or last_rcvd files. We needn't enforce quota on those
1025 * objects, so always set the lqi_space as 0. */
1026 rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1027 true, true, NULL, false);
1031 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1033 /* LU-2634: clear the extent format for fast symlink */
1034 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1036 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1037 LDISKFS_I(inode)->i_disksize = buflen;
1038 i_size_write(inode, buflen);
1039 inode->i_sb->s_op->dirty_inode(inode);
1044 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1045 int write_NUL, loff_t *offs, handle_t *handle)
1047 struct buffer_head *bh = NULL;
1048 loff_t offset = *offs;
1049 loff_t new_size = i_size_read(inode);
1050 unsigned long block;
1051 int blocksize = 1 << inode->i_blkbits;
1055 int dirty_inode = 0;
1059 * long symlink write does not count the NUL terminator in
1060 * bufsize, we write it, and the inode's file size does not
1061 * count the NUL terminator as well.
1063 ((char *)buf)[bufsize] = '\0';
1066 while (bufsize > 0) {
1070 block = offset >> inode->i_blkbits;
1071 boffs = offset & (blocksize - 1);
1072 size = min(blocksize - boffs, bufsize);
1073 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1075 CERROR("%s: error reading offset %llu (block %lu): "
1077 inode->i_sb->s_id, offset, block, err);
1081 err = ldiskfs_journal_get_write_access(handle, bh);
1083 CERROR("journal_get_write_access() returned error %d\n",
1087 LASSERTF(boffs + size <= bh->b_size,
1088 "boffs %d size %d bh->b_size %lu",
1089 boffs, size, (unsigned long)bh->b_size);
1090 memcpy(bh->b_data + boffs, buf, size);
1091 err = ldiskfs_journal_dirty_metadata(handle, bh);
1095 if (offset + size > new_size)
1096 new_size = offset + size;
1106 /* correct in-core and on-disk sizes */
1107 if (new_size > i_size_read(inode)) {
1108 spin_lock(&inode->i_lock);
1109 if (new_size > i_size_read(inode))
1110 i_size_write(inode, new_size);
1111 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1112 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1115 spin_unlock(&inode->i_lock);
1117 inode->i_sb->s_op->dirty_inode(inode);
1125 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1126 const struct lu_buf *buf, loff_t *pos,
1127 struct thandle *handle, struct lustre_capa *capa,
1130 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1131 struct osd_thandle *oh;
1135 LASSERT(dt_object_exists(dt));
1137 if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1140 LASSERT(handle != NULL);
1141 LASSERT(inode != NULL);
1142 ll_vfs_dq_init(inode);
1144 /* XXX: don't check: one declared chunk can be used many times */
1145 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1147 oh = container_of(handle, struct osd_thandle, ot_super);
1148 LASSERT(oh->ot_handle->h_transaction != NULL);
1149 /* Write small symlink to inode body as we need to maintain correct
1150 * on-disk symlinks for ldiskfs.
1151 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1152 * does not count it in.
1154 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1155 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1156 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1158 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1159 buf->lb_len, is_link, pos,
1162 result = buf->lb_len;
1166 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1167 __u64 start, __u64 end, struct thandle *th)
1169 struct osd_thandle *oh;
1170 struct inode *inode;
1175 oh = container_of(th, struct osd_thandle, ot_super);
1178 * we don't need to reserve credits for whole truncate
1179 * it's not possible as truncate may need to free too many
1180 * blocks and that won't fit a single transaction. instead
1181 * we reserve credits to change i_size and put inode onto
1182 * orphan list. if needed truncate will extend or restart
1185 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1186 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1188 inode = osd_dt_obj(dt)->oo_inode;
1191 rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1192 true, true, NULL, false);
1196 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1197 __u64 start, __u64 end, struct thandle *th,
1198 struct lustre_capa *capa)
1200 struct osd_thandle *oh;
1201 struct osd_object *obj = osd_dt_obj(dt);
1202 struct inode *inode = obj->oo_inode;
1208 LASSERT(end == OBD_OBJECT_EOF);
1209 LASSERT(dt_object_exists(dt));
1210 LASSERT(osd_invariant(obj));
1211 LASSERT(inode != NULL);
1212 ll_vfs_dq_init(inode);
1215 oh = container_of(th, struct osd_thandle, ot_super);
1216 LASSERT(oh->ot_handle->h_transaction != NULL);
1218 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1220 tid = oh->ot_handle->h_transaction->t_tid;
1222 rc = vmtruncate(inode, start);
1225 * For a partial-page truncate, flush the page to disk immediately to
1226 * avoid data corruption during direct disk write. b=17397
1228 if (rc == 0 && (start & ~CFS_PAGE_MASK) != 0)
1229 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1231 h = journal_current_handle();
1233 LASSERT(h == oh->ot_handle);
1235 if (tid != h->h_transaction->t_tid) {
1236 int credits = oh->ot_credits;
1238 * transaction has changed during truncate
1239 * we need to restart the handle with our credits
1241 if (h->h_buffer_credits < credits) {
1242 if (ldiskfs_journal_extend(h, credits))
1243 rc2 = ldiskfs_journal_restart(h, credits);
1247 RETURN(rc == 0 ? rc2 : rc);
1250 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1251 struct ll_user_fiemap *fm)
1253 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1254 struct osd_thread_info *info = osd_oti_get(env);
1255 struct dentry *dentry = &info->oti_obj_dentry;
1256 struct file *file = &info->oti_file;
1257 mm_segment_t saved_fs;
1261 dentry->d_inode = inode;
1262 file->f_dentry = dentry;
1263 file->f_mapping = inode->i_mapping;
1264 file->f_op = inode->i_fop;
1266 saved_fs = get_fs();
1268 /* ldiskfs_ioctl does not have a inode argument */
1269 if (inode->i_fop->unlocked_ioctl)
1270 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1279 * in some cases we may need declare methods for objects being created
1280 * e.g., when we create symlink
1282 const struct dt_body_operations osd_body_ops_new = {
1283 .dbo_declare_write = osd_declare_write,
1286 const struct dt_body_operations osd_body_ops = {
1287 .dbo_read = osd_read,
1288 .dbo_declare_write = osd_declare_write,
1289 .dbo_write = osd_write,
1290 .dbo_bufs_get = osd_bufs_get,
1291 .dbo_bufs_put = osd_bufs_put,
1292 .dbo_write_prep = osd_write_prep,
1293 .dbo_declare_write_commit = osd_declare_write_commit,
1294 .dbo_write_commit = osd_write_commit,
1295 .dbo_read_prep = osd_read_prep,
1296 .do_declare_punch = osd_declare_punch,
1297 .do_punch = osd_punch,
1298 .dbo_fiemap_get = osd_fiemap_get,