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) 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
42 * Author: Nikita Danilov <nikita@clusterfs.com>
43 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
47 /* LUSTRE_VERSION_CODE */
48 #include <lustre_ver.h>
49 /* prerequisite for linux/xattr.h */
50 #include <linux/types.h>
51 /* prerequisite for linux/xattr.h */
55 #include <ldiskfs/ldiskfs.h>
56 #include <ldiskfs/ldiskfs_jbd2.h>
57 #include <ldiskfs/ldiskfs_extents.h>
60 * struct OBD_{ALLOC,FREE}*()
63 #include <obd_support.h>
65 #include "osd_internal.h"
67 #ifndef HAVE_PAGE_CONSTANT
68 #define mapping_cap_page_constant_write(mapping) 0
69 #define SetPageConstant(page) do {} while (0)
70 #define ClearPageConstant(page) do {} while (0)
73 #ifndef HAS_GENERIC_ERROR_REMOVE_PAGE
74 int generic_error_remove_page(struct address_space *mapping, struct page *page)
79 if (mapping != page->mapping)
82 * Only punch for normal data pages for now.
83 * Handling other types like directories would need more auditing.
85 if (!S_ISREG(mapping->host->i_mode))
88 if (page_mapped(page)) {
89 unmap_mapping_range(mapping,
90 (loff_t)page->index << PAGE_CACHE_SHIFT,
93 truncate_complete_page(mapping, page);
98 static void osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,int rw)
100 cfs_waitq_init(&iobuf->dr_wait);
101 cfs_atomic_set(&iobuf->dr_numreqs, 0);
102 iobuf->dr_max_pages = PTLRPC_MAX_BRW_PAGES;
103 iobuf->dr_npages = 0;
107 iobuf->dr_elapsed = 0;
108 /* must be counted before, so assert */
109 LASSERT(iobuf->dr_elapsed_valid == 0);
113 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
115 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
116 iobuf->dr_pages[iobuf->dr_npages++] = page;
119 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
121 int rw = iobuf->dr_rw;
123 if (iobuf->dr_elapsed_valid) {
124 iobuf->dr_elapsed_valid = 0;
125 LASSERT(iobuf->dr_dev == d);
126 LASSERT(iobuf->dr_frags > 0);
127 lprocfs_oh_tally(&d->od_brw_stats.
128 hist[BRW_R_DIO_FRAGS+rw],
130 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
135 #ifdef HAVE_BIO_ENDIO_2ARG
136 #define DIO_RETURN(a)
137 static void dio_complete_routine(struct bio *bio, int error)
139 #define DIO_RETURN(a) return(a)
140 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
143 struct osd_iobuf *iobuf = bio->bi_private;
147 /* CAVEAT EMPTOR: possibly in IRQ context
148 * DO NOT record procfs stats here!!! */
150 if (unlikely(iobuf == NULL)) {
151 CERROR("***** bio->bi_private is NULL! This should never "
152 "happen. Normally, I would crash here, but instead I "
153 "will dump the bio contents to the console. Please "
154 "report this to <http://jira.whamcloud.com/> , along "
155 "with any interesting messages leading up to this point "
156 "(like SCSI errors, perhaps). Because bi_private is "
157 "NULL, I can't wake up the thread that initiated this "
158 "IO - you will probably have to reboot this node.\n");
159 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
160 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
161 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
162 bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
163 bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
168 /* the check is outside of the cycle for performance reason -bzzz */
169 if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
170 bio_for_each_segment(bvl, bio, i) {
171 if (likely(error == 0))
172 SetPageUptodate(bvl->bv_page);
173 LASSERT(PageLocked(bvl->bv_page));
174 ClearPageConstant(bvl->bv_page);
176 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
178 struct page *p = iobuf->dr_pages[0];
180 if (mapping_cap_page_constant_write(p->mapping)) {
181 bio_for_each_segment(bvl, bio, i) {
182 ClearPageConstant(bvl->bv_page);
186 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
189 /* any real error is good enough -bzzz */
190 if (error != 0 && iobuf->dr_error == 0)
191 iobuf->dr_error = error;
193 if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs)) {
194 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
195 iobuf->dr_elapsed_valid = 1;
196 cfs_waitq_signal(&iobuf->dr_wait);
199 /* Completed bios used to be chained off iobuf->dr_bios and freed in
200 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
201 * mempool when serious on-disk fragmentation was encountered,
202 * deadlocking the OST. The bios are now released as soon as complete
203 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
208 static void record_start_io(struct osd_iobuf *iobuf, int size)
210 struct osd_device *osd = iobuf->dr_dev;
211 struct obd_histogram *h = osd->od_brw_stats.hist;
214 cfs_atomic_inc(&iobuf->dr_numreqs);
216 if (iobuf->dr_rw == 0) {
217 cfs_atomic_inc(&osd->od_r_in_flight);
218 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
219 cfs_atomic_read(&osd->od_r_in_flight));
220 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
221 } else if (iobuf->dr_rw == 1) {
222 cfs_atomic_inc(&osd->od_w_in_flight);
223 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
224 cfs_atomic_read(&osd->od_w_in_flight));
225 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
231 static void osd_submit_bio(int rw, struct bio *bio)
233 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
235 submit_bio(READ, bio);
237 submit_bio(WRITE, bio);
240 static int can_be_merged(struct bio *bio, sector_t sector)
247 size = bio->bi_size >> 9;
248 return bio->bi_sector + size == sector ? 1 : 0;
251 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
252 struct osd_iobuf *iobuf)
254 int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
255 struct page **pages = iobuf->dr_pages;
256 int npages = iobuf->dr_npages;
257 unsigned long *blocks = iobuf->dr_blocks;
258 int total_blocks = npages * blocks_per_page;
259 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
260 unsigned int blocksize = inode->i_sb->s_blocksize;
261 struct bio *bio = NULL;
263 unsigned int page_offset;
272 LASSERT(iobuf->dr_npages == npages);
274 osd_brw_stats_update(osd, iobuf);
275 iobuf->dr_start_time = cfs_time_current();
277 for (page_idx = 0, block_idx = 0;
279 page_idx++, block_idx += blocks_per_page) {
281 page = pages[page_idx];
282 LASSERT(block_idx + blocks_per_page <= total_blocks);
284 for (i = 0, page_offset = 0;
286 i += nblocks, page_offset += blocksize * nblocks) {
290 if (blocks[block_idx + i] == 0) { /* hole */
291 LASSERTF(iobuf->dr_rw == 0,
292 "page_idx %u, block_idx %u, i %u\n",
293 page_idx, block_idx, i);
294 memset(kmap(page) + page_offset, 0, blocksize);
299 sector = (sector_t)blocks[block_idx + i] << sector_bits;
301 /* Additional contiguous file blocks? */
302 while (i + nblocks < blocks_per_page &&
303 (sector + (nblocks << sector_bits)) ==
304 ((sector_t)blocks[block_idx + i + nblocks] <<
308 /* I only set the page to be constant only if it
309 * is mapped to a contiguous underlying disk block(s).
310 * It will then make sure the corresponding device
311 * cache of raid5 will be overwritten by this page.
313 if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
314 mapping_cap_page_constant_write(inode->i_mapping))
315 SetPageConstant(page);
318 can_be_merged(bio, sector) &&
319 bio_add_page(bio, page,
320 blocksize * nblocks, page_offset) != 0)
321 continue; /* added this frag OK */
324 struct request_queue *q =
325 bdev_get_queue(bio->bi_bdev);
327 /* Dang! I have to fragment this I/O */
328 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
329 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
331 bio->bi_vcnt, bio->bi_max_vecs,
332 bio->bi_size >> 9, queue_max_sectors(q),
333 bio_phys_segments(q, bio),
334 queue_max_phys_segments(q),
335 bio_hw_segments(q, bio),
336 queue_max_hw_segments(q));
338 record_start_io(iobuf, bio->bi_size);
339 osd_submit_bio(iobuf->dr_rw, bio);
342 /* allocate new bio, limited by max BIO size, b=9945 */
343 bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
344 (npages - page_idx) *
347 CERROR("Can't allocate bio %u*%u = %u pages\n",
348 (npages - page_idx), blocks_per_page,
349 (npages - page_idx) * blocks_per_page);
354 bio->bi_bdev = inode->i_sb->s_bdev;
355 bio->bi_sector = sector;
356 bio->bi_end_io = dio_complete_routine;
357 bio->bi_private = iobuf;
359 rc = bio_add_page(bio, page,
360 blocksize * nblocks, page_offset);
366 record_start_io(iobuf, bio->bi_size);
367 osd_submit_bio(iobuf->dr_rw, bio);
372 /* in order to achieve better IO throughput, we don't wait for writes
373 * completion here. instead we proceed with transaction commit in
374 * parallel and wait for IO completion once transaction is stopped
375 * see osd_trans_stop() for more details -bzzz */
376 if (iobuf->dr_rw == 0) {
377 cfs_wait_event(iobuf->dr_wait,
378 cfs_atomic_read(&iobuf->dr_numreqs) == 0);
382 rc = iobuf->dr_error;
386 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
387 struct niobuf_local *lnb)
394 int poff = offset & (CFS_PAGE_SIZE - 1);
395 int plen = CFS_PAGE_SIZE - poff;
399 lnb->offset = offset;
400 /* lnb->lnb_page_offset = poff; */
402 /* lb->flags = rnb->flags; */
407 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
418 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
420 struct inode *inode = osd_dt_obj(dt)->oo_inode;
421 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
426 page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
427 GFP_NOFS | __GFP_HIGHMEM);
428 if (unlikely(page == NULL))
429 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
435 * there are following "locks":
453 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
454 ssize_t len, struct niobuf_local *lnb, int rw,
455 struct lustre_capa *capa)
457 struct osd_object *obj = osd_dt_obj(d);
458 int npages, i, rc = 0;
460 LASSERT(obj->oo_inode);
462 osd_map_remote_to_local(pos, len, &npages, lnb);
464 for (i = 0; i < npages; i++, lnb++) {
466 /* We still set up for ungranted pages so that granted pages
467 * can be written to disk as they were promised, and portals
468 * needs to keep the pages all aligned properly. */
469 lnb->dentry = (void *) obj;
471 lnb->page = osd_get_page(d, lnb->offset, rw);
472 if (lnb->page == NULL)
473 GOTO(cleanup, rc = -ENOMEM);
475 /* DLM locking protects us from write and truncate competing
476 * for same region, but truncate can leave dirty page in the
477 * cache. it's possible the writeout on a such a page is in
478 * progress when we access it. it's also possible that during
479 * this writeout we put new (partial) data, but then won't
480 * be able to proceed in filter_commitrw_write(). thus let's
481 * just wait for writeout completion, should be rare enough.
483 wait_on_page_writeback(lnb->page);
484 BUG_ON(PageWriteback(lnb->page));
486 lu_object_get(&d->do_lu);
494 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
495 struct niobuf_local *lnb, int npages)
497 struct osd_thread_info *oti = osd_oti_get(env);
498 struct osd_iobuf *iobuf = &oti->oti_iobuf;
499 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
502 /* to do IO stats, notice we do this here because
503 * osd_do_bio() doesn't wait for write to complete */
504 osd_fini_iobuf(d, iobuf);
506 for (i = 0; i < npages; i++) {
507 if (lnb[i].page == NULL)
509 LASSERT(PageLocked(lnb[i].page));
510 unlock_page(lnb[i].page);
511 page_cache_release(lnb[i].page);
512 lu_object_put(env, &dt->do_lu);
518 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
519 struct niobuf_local *lnb, int npages)
521 struct osd_thread_info *oti = osd_oti_get(env);
522 struct osd_iobuf *iobuf = &oti->oti_iobuf;
523 struct inode *inode = osd_dt_obj(dt)->oo_inode;
524 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
525 struct timeval start;
527 unsigned long timediff;
536 osd_init_iobuf(osd, iobuf, 0);
538 isize = i_size_read(inode);
539 maxidx = ((isize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
541 if (osd->od_writethrough_cache)
543 if (isize > osd->od_readcache_max_filesize)
546 cfs_gettimeofday(&start);
547 for (i = 0; i < npages; i++) {
550 generic_error_remove_page(inode->i_mapping,
554 * till commit the content of the page is undefined
555 * we'll set it uptodate once bulk is done. otherwise
556 * subsequent reads can access non-stable data
558 ClearPageUptodate(lnb[i].page);
560 if (lnb[i].len == CFS_PAGE_SIZE)
563 if (maxidx >= lnb[i].page->index) {
564 osd_iobuf_add_page(iobuf, lnb[i].page);
567 char *p = kmap(lnb[i].page);
572 off = lnb[i].offset + lnb[i].len;
573 off &= ~CFS_PAGE_MASK;
575 memset(p + off, 0, CFS_PAGE_SIZE - off);
579 cfs_gettimeofday(&end);
580 timediff = cfs_timeval_sub(&end, &start, NULL);
581 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
583 if (iobuf->dr_npages) {
584 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
589 if (likely(rc == 0)) {
590 rc = osd_do_bio(osd, inode, iobuf);
591 /* do IO stats for preparation reads */
592 osd_fini_iobuf(osd, iobuf);
598 static int osd_declare_write_commit(const struct lu_env *env,
599 struct dt_object *dt,
600 struct niobuf_local *lnb, int npages,
601 struct thandle *handle)
603 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
604 struct inode *inode = osd_dt_obj(dt)->oo_inode;
605 struct osd_thandle *oh;
612 LASSERT(handle != NULL);
613 oh = container_of0(handle, struct osd_thandle, ot_super);
614 LASSERT(oh->ot_handle == NULL);
616 old = oh->ot_credits;
619 /* calculate number of extents (probably better to pass nb) */
620 for (i = 1; i < npages; i++)
622 lnb[i - 1].offset + lnb[i - 1].len)
626 * each extent can go into new leaf causing a split
627 * 5 is max tree depth: inode + 4 index blocks
628 * with blockmaps, depth is 3 at most
630 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
632 * many concurrent threads may grow tree by the time
633 * our transaction starts. so, consider 2 is a min depth
635 depth = ext_depth(inode);
636 depth = max(depth, 1) + 1;
638 oh->ot_credits++; /* inode */
639 oh->ot_credits += depth * 2 * extents;
643 oh->ot_credits++; /* inode */
644 oh->ot_credits += depth * extents;
647 /* each new block can go in different group (bitmap + gd) */
649 /* we can't dirty more bitmap blocks than exist */
650 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
651 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
653 oh->ot_credits += newblocks;
655 /* we can't dirty more gd blocks than exist */
656 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
657 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
659 oh->ot_credits += newblocks;
664 /* Check if a block is allocated or not */
665 static int osd_is_mapped(struct inode *inode, obd_size offset)
667 sector_t (*fs_bmap)(struct address_space *, sector_t);
669 fs_bmap = inode->i_mapping->a_ops->bmap;
671 /* We can't know if we are overwriting or not */
675 if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
681 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
682 struct niobuf_local *lnb, int npages,
683 struct thandle *thandle)
685 struct osd_thread_info *oti = osd_oti_get(env);
686 struct osd_iobuf *iobuf = &oti->oti_iobuf;
687 struct inode *inode = osd_dt_obj(dt)->oo_inode;
688 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
694 osd_init_iobuf(osd, iobuf, 1);
695 isize = i_size_read(inode);
697 for (i = 0; i < npages; i++) {
698 if (lnb[i].rc == -ENOSPC &&
699 osd_is_mapped(inode, lnb[i].offset)) {
700 /* Allow the write to proceed if overwriting an
705 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
706 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
708 LASSERT(lnb[i].page);
709 generic_error_remove_page(inode->i_mapping,lnb[i].page);
713 LASSERT(PageLocked(lnb[i].page));
714 LASSERT(!PageWriteback(lnb[i].page));
716 if (lnb[i].offset + lnb[i].len > isize)
717 isize = lnb[i].offset + lnb[i].len;
720 * Since write and truncate are serialized by oo_sem, even
721 * partial-page truncate should not leave dirty pages in the
724 LASSERT(!PageDirty(lnb[i].page));
726 SetPageUptodate(lnb[i].page);
728 osd_iobuf_add_page(iobuf, lnb[i].page);
731 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
733 } else if (iobuf->dr_npages > 0) {
734 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
740 /* no pages to write, no transno is needed */
741 thandle->th_local = 1;
744 if (likely(rc == 0)) {
745 if (isize > i_size_read(inode)) {
746 i_size_write(inode, isize);
747 LDISKFS_I(inode)->i_disksize = isize;
748 inode->i_sb->s_op->dirty_inode(inode);
751 rc = osd_do_bio(osd, inode, iobuf);
752 /* we don't do stats here as in read path because
753 * write is async: we'll do this in osd_put_bufs() */
756 if (unlikely(rc != 0)) {
757 /* if write fails, we should drop pages from the cache */
758 for (i = 0; i < npages; i++) {
759 if (lnb[i].page == NULL)
761 LASSERT(PageLocked(lnb[i].page));
762 generic_error_remove_page(inode->i_mapping,lnb[i].page);
769 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
770 struct niobuf_local *lnb, int npages)
772 struct osd_thread_info *oti = osd_oti_get(env);
773 struct osd_iobuf *iobuf = &oti->oti_iobuf;
774 struct inode *inode = osd_dt_obj(dt)->oo_inode;
775 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
776 struct timeval start, end;
777 unsigned long timediff;
778 int rc = 0, i, m = 0, cache = 0;
782 osd_init_iobuf(osd, iobuf, 0);
784 if (osd->od_read_cache)
786 if (i_size_read(inode) > osd->od_readcache_max_filesize)
789 cfs_gettimeofday(&start);
790 for (i = 0; i < npages; i++) {
792 if (i_size_read(inode) <= lnb[i].offset)
793 /* If there's no more data, abort early.
794 * lnb->rc == 0, so it's easy to detect later. */
797 if (i_size_read(inode) <
798 lnb[i].offset + lnb[i].len - 1)
799 lnb[i].rc = i_size_read(inode) - lnb[i].offset;
801 lnb[i].rc = lnb[i].len;
804 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
805 if (PageUptodate(lnb[i].page)) {
806 lprocfs_counter_add(osd->od_stats,
807 LPROC_OSD_CACHE_HIT, 1);
809 lprocfs_counter_add(osd->od_stats,
810 LPROC_OSD_CACHE_MISS, 1);
811 osd_iobuf_add_page(iobuf, lnb[i].page);
814 generic_error_remove_page(inode->i_mapping,lnb[i].page);
816 cfs_gettimeofday(&end);
817 timediff = cfs_timeval_sub(&end, &start, NULL);
818 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
820 if (iobuf->dr_npages) {
821 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
826 rc = osd_do_bio(osd, inode, iobuf);
828 /* IO stats will be done in osd_bufs_put() */
835 * XXX: Another layering violation for now.
837 * We don't want to use ->f_op->read methods, because generic file write
839 * - serializes on ->i_sem, and
841 * - does a lot of extra work like balance_dirty_pages(),
843 * which doesn't work for globally shared files like /last_rcvd.
845 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
847 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
849 memcpy(buffer, (char *)ei->i_data, buflen);
854 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
856 struct buffer_head *bh;
864 /* prevent reading after eof */
865 cfs_spin_lock(&inode->i_lock);
866 if (i_size_read(inode) < *offs + size) {
867 size = i_size_read(inode) - *offs;
868 cfs_spin_unlock(&inode->i_lock);
870 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
871 i_size_read(inode), *offs);
873 } else if (size == 0) {
877 cfs_spin_unlock(&inode->i_lock);
880 blocksize = 1 << inode->i_blkbits;
883 block = *offs >> inode->i_blkbits;
884 boffs = *offs & (blocksize - 1);
885 csize = min(blocksize - boffs, size);
886 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
888 CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
889 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
890 csize, *offs, inode->i_ino, err);
894 memcpy(buf, bh->b_data + boffs, csize);
904 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
905 struct lu_buf *buf, loff_t *pos,
906 struct lustre_capa *capa)
908 struct inode *inode = osd_dt_obj(dt)->oo_inode;
911 if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
914 /* Read small symlink from inode body as we need to maintain correct
915 * on-disk symlinks for ldiskfs.
917 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
918 (buf->lb_len <= sizeof(LDISKFS_I(inode)->i_data)))
919 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
921 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
926 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
927 const loff_t size, loff_t pos,
928 struct thandle *handle)
930 struct osd_thandle *oh;
933 LASSERT(handle != NULL);
935 oh = container_of0(handle, struct osd_thandle, ot_super);
936 LASSERT(oh->ot_handle == NULL);
938 /* XXX: size == 0 or INT_MAX indicating a catalog header update or
939 * llog write, see comment in mdd_declare_llog_record().
941 * This hack will be removed with llog over OSD landing
943 if (size == DECLARE_LLOG_REWRITE)
945 else if (size == DECLARE_LLOG_WRITE)
948 credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
950 OSD_DECLARE_OP(oh, write);
951 oh->ot_credits += credits;
953 if (osd_dt_obj(dt)->oo_inode == NULL)
956 osd_declare_qid(dt, oh, USRQUOTA, osd_dt_obj(dt)->oo_inode->i_uid,
957 osd_dt_obj(dt)->oo_inode);
958 osd_declare_qid(dt, oh, GRPQUOTA, osd_dt_obj(dt)->oo_inode->i_gid,
959 osd_dt_obj(dt)->oo_inode);
963 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
966 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
967 LDISKFS_I(inode)->i_disksize = buflen;
968 i_size_write(inode, buflen);
969 inode->i_sb->s_op->dirty_inode(inode);
974 static int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
975 loff_t *offs, handle_t *handle)
977 struct buffer_head *bh = NULL;
978 loff_t offset = *offs;
979 loff_t new_size = i_size_read(inode);
981 int blocksize = 1 << inode->i_blkbits;
987 while (bufsize > 0) {
991 block = offset >> inode->i_blkbits;
992 boffs = offset & (blocksize - 1);
993 size = min(blocksize - boffs, bufsize);
994 bh = ldiskfs_bread(handle, inode, block, 1, &err);
996 CERROR("%s: error reading offset %llu (block %lu): "
998 inode->i_sb->s_id, offset, block, err);
1002 err = ldiskfs_journal_get_write_access(handle, bh);
1004 CERROR("journal_get_write_access() returned error %d\n",
1008 LASSERTF(boffs + size <= bh->b_size,
1009 "boffs %d size %d bh->b_size %lu",
1010 boffs, size, (unsigned long)bh->b_size);
1011 memcpy(bh->b_data + boffs, buf, size);
1012 err = ldiskfs_journal_dirty_metadata(handle, bh);
1016 if (offset + size > new_size)
1017 new_size = offset + size;
1025 /* correct in-core and on-disk sizes */
1026 if (new_size > i_size_read(inode)) {
1027 cfs_spin_lock(&inode->i_lock);
1028 if (new_size > i_size_read(inode))
1029 i_size_write(inode, new_size);
1030 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1031 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1034 cfs_spin_unlock(&inode->i_lock);
1036 inode->i_sb->s_op->dirty_inode(inode);
1044 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1045 const struct lu_buf *buf, loff_t *pos,
1046 struct thandle *handle, struct lustre_capa *capa,
1049 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1050 struct osd_thandle *oh;
1052 #ifdef HAVE_QUOTA_SUPPORT
1053 cfs_cap_t save = cfs_curproc_cap_pack();
1056 LASSERT(dt_object_exists(dt));
1058 if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1061 LASSERT(handle != NULL);
1063 /* XXX: don't check: one declared chunk can be used many times */
1064 /* OSD_EXEC_OP(handle, write); */
1066 oh = container_of(handle, struct osd_thandle, ot_super);
1067 LASSERT(oh->ot_handle->h_transaction != NULL);
1068 #ifdef HAVE_QUOTA_SUPPORT
1070 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1072 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1074 /* Write small symlink to inode body as we need to maintain correct
1075 * on-disk symlinks for ldiskfs.
1077 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1078 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1079 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1081 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1084 #ifdef HAVE_QUOTA_SUPPORT
1085 cfs_curproc_cap_unpack(save);
1088 result = buf->lb_len;
1092 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1093 __u64 start, __u64 end, struct thandle *th)
1095 struct osd_thandle *oh;
1099 oh = container_of(th, struct osd_thandle, ot_super);
1101 OSD_DECLARE_OP(oh, punch);
1104 * we don't need to reserve credits for whole truncate
1105 * it's not possible as truncate may need to free too many
1106 * blocks and that won't fit a single transaction. instead
1107 * we reserve credits to change i_size and put inode onto
1108 * orphan list. if needed truncate will extend or restart
1111 oh->ot_credits += osd_dto_credits_noquota[DTO_ATTR_SET_BASE];
1112 oh->ot_credits += 3;
1117 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1118 __u64 start, __u64 end, struct thandle *th,
1119 struct lustre_capa *capa)
1121 struct osd_thandle *oh;
1122 struct osd_object *obj = osd_dt_obj(dt);
1123 struct inode *inode = obj->oo_inode;
1129 LASSERT(end == OBD_OBJECT_EOF);
1130 LASSERT(dt_object_exists(dt));
1131 LASSERT(osd_invariant(obj));
1134 oh = container_of(th, struct osd_thandle, ot_super);
1135 LASSERT(oh->ot_handle->h_transaction != NULL);
1137 OSD_EXEC_OP(th, punch);
1139 tid = oh->ot_handle->h_transaction->t_tid;
1141 rc = vmtruncate(inode, start);
1144 * For a partial-page truncate, flush the page to disk immediately to
1145 * avoid data corruption during direct disk write. b=17397
1147 if (rc == 0 && (start & ~CFS_PAGE_MASK) != 0)
1148 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1150 h = journal_current_handle();
1152 LASSERT(h == oh->ot_handle);
1154 if (tid != h->h_transaction->t_tid) {
1155 int credits = oh->ot_credits;
1157 * transaction has changed during truncate
1158 * we need to restart the handle with our credits
1160 if (h->h_buffer_credits < credits) {
1161 if (ldiskfs_journal_extend(h, credits))
1162 rc2 = ldiskfs_journal_restart(h, credits);
1166 RETURN(rc == 0 ? rc2 : rc);
1169 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1170 struct ll_user_fiemap *fm)
1172 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1173 struct osd_thread_info *info = osd_oti_get(env);
1174 struct dentry *dentry = &info->oti_obj_dentry;
1175 struct file *file = &info->oti_file;
1176 mm_segment_t saved_fs;
1180 dentry->d_inode = inode;
1181 file->f_dentry = dentry;
1182 file->f_mapping = inode->i_mapping;
1183 file->f_op = inode->i_fop;
1185 saved_fs = get_fs();
1187 /* ldiskfs_ioctl does not have a inode argument */
1188 if (inode->i_fop->unlocked_ioctl)
1189 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1198 * in some cases we may need declare methods for objects being created
1199 * e.g., when we create symlink
1201 const struct dt_body_operations osd_body_ops_new = {
1202 .dbo_declare_write = osd_declare_write,
1205 const struct dt_body_operations osd_body_ops = {
1206 .dbo_read = osd_read,
1207 .dbo_declare_write = osd_declare_write,
1208 .dbo_write = osd_write,
1209 .dbo_bufs_get = osd_bufs_get,
1210 .dbo_bufs_put = osd_bufs_put,
1211 .dbo_write_prep = osd_write_prep,
1212 .dbo_declare_write_commit = osd_declare_write_commit,
1213 .dbo_write_commit = osd_write_commit,
1214 .dbo_read_prep = osd_read_prep,
1215 .do_declare_punch = osd_declare_punch,
1216 .do_punch = osd_punch,
1217 .dbo_fiemap_get = osd_fiemap_get,