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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 /* prerequisite for linux/xattr.h */
42 #include <linux/types.h>
43 /* prerequisite for linux/xattr.h */
46 #include <linux/pagevec.h>
49 * struct OBD_{ALLOC,FREE}*()
52 #include <obd_support.h>
54 #include "osd_internal.h"
57 #include <ldiskfs/ldiskfs_extents.h>
59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
60 int rw, int line, int pages)
64 LASSERTF(iobuf->dr_elapsed_valid == 0,
65 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
66 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
68 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
70 init_waitqueue_head(&iobuf->dr_wait);
71 atomic_set(&iobuf->dr_numreqs, 0);
76 iobuf->dr_elapsed = ktime_set(0, 0);
77 /* must be counted before, so assert */
79 iobuf->dr_init_at = line;
81 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
82 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
83 LASSERT(iobuf->dr_pg_buf.lb_len >=
84 pages * sizeof(iobuf->dr_pages[0]));
88 /* start with 1MB for 4K blocks */
90 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
93 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
94 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
96 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
97 iobuf->dr_max_pages = 0;
98 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
99 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
101 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
102 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
103 if (unlikely(iobuf->dr_blocks == NULL))
106 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
107 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
108 if (unlikely(iobuf->dr_pages == NULL))
111 iobuf->dr_max_pages = pages;
115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
116 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
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],
136 ktime_to_ms(iobuf->dr_elapsed));
140 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
141 static void dio_complete_routine(struct bio *bio)
143 int error = bio->bi_error;
145 static void dio_complete_routine(struct bio *bio, 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 <https://jira.hpdd.intel.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, "
170 "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
171 "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
172 (unsigned long)bio->bi_flags,
178 bio->bi_vcnt, bio_idx(bio),
179 bio_sectors(bio) << 9, bio->bi_end_io,
181 atomic_read(&bio->bi_cnt),
183 atomic_read(&bio->__bi_cnt),
189 /* the check is outside of the cycle for performance reason -bzzz */
190 if (!bio_data_dir(bio)) {
191 bio_for_each_segment_all(bvl, bio, iter) {
192 if (likely(error == 0))
193 SetPageUptodate(bvl_to_page(bvl));
194 LASSERT(PageLocked(bvl_to_page(bvl)));
196 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
198 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
201 /* any real error is good enough -bzzz */
202 if (error != 0 && iobuf->dr_error == 0)
203 iobuf->dr_error = error;
206 * set dr_elapsed before dr_numreqs turns to 0, otherwise
207 * it's possible that service thread will see dr_numreqs
208 * is zero, but dr_elapsed is not set yet, leading to lost
209 * data in this processing and an assertion in a subsequent
212 if (atomic_read(&iobuf->dr_numreqs) == 1) {
213 ktime_t now = ktime_get();
215 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
216 iobuf->dr_elapsed_valid = 1;
218 if (atomic_dec_and_test(&iobuf->dr_numreqs))
219 wake_up(&iobuf->dr_wait);
221 /* Completed bios used to be chained off iobuf->dr_bios and freed in
222 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
223 * mempool when serious on-disk fragmentation was encountered,
224 * deadlocking the OST. The bios are now released as soon as complete
225 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
229 static void record_start_io(struct osd_iobuf *iobuf, int size)
231 struct osd_device *osd = iobuf->dr_dev;
232 struct obd_histogram *h = osd->od_brw_stats.hist;
235 atomic_inc(&iobuf->dr_numreqs);
237 if (iobuf->dr_rw == 0) {
238 atomic_inc(&osd->od_r_in_flight);
239 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
240 atomic_read(&osd->od_r_in_flight));
241 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
242 } else if (iobuf->dr_rw == 1) {
243 atomic_inc(&osd->od_w_in_flight);
244 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
245 atomic_read(&osd->od_w_in_flight));
246 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
252 static void osd_submit_bio(int rw, struct bio *bio)
254 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
255 #ifdef HAVE_SUBMIT_BIO_2ARGS
257 submit_bio(READ, bio);
259 submit_bio(WRITE, bio);
266 static int can_be_merged(struct bio *bio, sector_t sector)
271 return bio_end_sector(bio) == sector ? 1 : 0;
274 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
275 struct osd_iobuf *iobuf)
277 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
278 struct page **pages = iobuf->dr_pages;
279 int npages = iobuf->dr_npages;
280 sector_t *blocks = iobuf->dr_blocks;
281 int total_blocks = npages * blocks_per_page;
282 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
283 unsigned int blocksize = inode->i_sb->s_blocksize;
284 struct bio *bio = NULL;
286 unsigned int page_offset;
296 LASSERT(iobuf->dr_npages == npages);
298 osd_brw_stats_update(osd, iobuf);
299 iobuf->dr_start_time = ktime_get();
301 blk_start_plug(&plug);
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 LASSERTF(iobuf->dr_rw == 0,
317 "page_idx %u, block_idx %u, i %u\n",
318 page_idx, block_idx, i);
319 memset(kmap(page) + page_offset, 0, blocksize);
324 sector = (sector_t)blocks[block_idx + i] << sector_bits;
326 /* Additional contiguous file blocks? */
327 while (i + nblocks < blocks_per_page &&
328 (sector + (nblocks << sector_bits)) ==
329 ((sector_t)blocks[block_idx + i + nblocks] <<
334 can_be_merged(bio, sector) &&
335 bio_add_page(bio, page,
336 blocksize * nblocks, page_offset) != 0)
337 continue; /* added this frag OK */
340 struct request_queue *q =
341 bdev_get_queue(bio->bi_bdev);
342 unsigned int bi_size = bio_sectors(bio) << 9;
344 /* Dang! I have to fragment this I/O */
345 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
346 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
347 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
349 queue_max_sectors(q),
350 bio_phys_segments(q, bio),
351 queue_max_phys_segments(q),
352 0, queue_max_hw_segments(q));
353 record_start_io(iobuf, bi_size);
354 osd_submit_bio(iobuf->dr_rw, bio);
357 /* allocate new bio */
358 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
359 (npages - page_idx) *
362 CERROR("Can't allocate bio %u*%u = %u pages\n",
363 (npages - page_idx), blocks_per_page,
364 (npages - page_idx) * blocks_per_page);
369 bio->bi_bdev = inode->i_sb->s_bdev;
370 bio_set_sector(bio, sector);
372 bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
374 bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
376 bio->bi_end_io = dio_complete_routine;
377 bio->bi_private = iobuf;
379 rc = bio_add_page(bio, page,
380 blocksize * nblocks, page_offset);
386 record_start_io(iobuf, bio_sectors(bio) << 9);
387 osd_submit_bio(iobuf->dr_rw, bio);
392 blk_finish_plug(&plug);
394 /* in order to achieve better IO throughput, we don't wait for writes
395 * completion here. instead we proceed with transaction commit in
396 * parallel and wait for IO completion once transaction is stopped
397 * see osd_trans_stop() for more details -bzzz */
398 if (iobuf->dr_rw == 0) {
399 wait_event(iobuf->dr_wait,
400 atomic_read(&iobuf->dr_numreqs) == 0);
401 osd_fini_iobuf(osd, iobuf);
405 rc = iobuf->dr_error;
409 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
410 struct niobuf_local *lnb)
417 int poff = offset & (PAGE_SIZE - 1);
418 int plen = PAGE_SIZE - poff;
422 lnb->lnb_file_offset = offset;
423 lnb->lnb_page_offset = poff;
425 /* lnb->lnb_flags = rnb->rnb_flags; */
427 lnb->lnb_page = NULL;
430 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
441 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
444 struct inode *inode = osd_dt_obj(dt)->oo_inode;
445 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
450 page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
453 if (unlikely(page == NULL))
454 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
460 * there are following "locks":
471 * - lock pages, unlock
473 * - lock partial page
479 * Unlock and release pages loaded by osd_bufs_get()
481 * Unlock \a npages pages from \a lnb and drop the refcount on them.
483 * \param env thread execution environment
484 * \param dt dt object undergoing IO (OSD object + methods)
485 * \param lnb array of pages undergoing IO
486 * \param npages number of pages in \a lnb
490 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
491 struct niobuf_local *lnb, int npages)
496 pagevec_init(&pvec, 0);
498 for (i = 0; i < npages; i++) {
499 if (lnb[i].lnb_page == NULL)
501 LASSERT(PageLocked(lnb[i].lnb_page));
502 unlock_page(lnb[i].lnb_page);
503 if (pagevec_add(&pvec, lnb[i].lnb_page) == 0)
504 pagevec_release(&pvec);
505 dt_object_put(env, dt);
506 lnb[i].lnb_page = NULL;
509 /* Release any partial pagevec */
510 pagevec_release(&pvec);
516 * Load and lock pages undergoing IO
518 * Pages as described in the \a lnb array are fetched (from disk or cache)
519 * and locked for IO by the caller.
521 * DLM locking protects us from write and truncate competing for same region,
522 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
523 * It's possible the writeout on a such a page is in progress when we access
524 * it. It's also possible that during this writeout we put new (partial) data
525 * into the page, but won't be able to proceed in filter_commitrw_write().
526 * Therefore, just wait for writeout completion as it should be rare enough.
528 * \param env thread execution environment
529 * \param dt dt object undergoing IO (OSD object + methods)
530 * \param pos byte offset of IO start
531 * \param len number of bytes of IO
532 * \param lnb array of extents undergoing IO
533 * \param rw read or write operation, and other flags
534 * \param capa capabilities
536 * \retval pages (zero or more) loaded successfully
537 * \retval -ENOMEM on memory/page allocation error
539 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
540 loff_t pos, ssize_t len, struct niobuf_local *lnb,
541 enum dt_bufs_type rw)
543 struct osd_object *obj = osd_dt_obj(dt);
544 int npages, i, rc = 0;
547 LASSERT(obj->oo_inode);
549 osd_map_remote_to_local(pos, len, &npages, lnb);
551 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
552 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
554 for (i = 0; i < npages; i++, lnb++) {
555 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
557 if (lnb->lnb_page == NULL)
558 GOTO(cleanup, rc = -ENOMEM);
560 wait_on_page_writeback(lnb->lnb_page);
561 BUG_ON(PageWriteback(lnb->lnb_page));
563 lu_object_get(&dt->do_lu);
570 osd_bufs_put(env, dt, lnb - i, i);
574 #ifndef HAVE_LDISKFS_MAP_BLOCKS
576 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
577 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
588 static long ldiskfs_ext_find_goal(struct inode *inode,
589 struct ldiskfs_ext_path *path,
590 unsigned long block, int *aflags)
592 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
593 unsigned long bg_start;
594 unsigned long colour;
598 struct ldiskfs_extent *ex;
599 depth = path->p_depth;
601 /* try to predict block placement */
602 if ((ex = path[depth].p_ext))
603 return ldiskfs_ext_pblock(ex) +
604 (block - le32_to_cpu(ex->ee_block));
606 /* it looks index is empty
607 * try to find starting from index itself */
608 if (path[depth].p_bh)
609 return path[depth].p_bh->b_blocknr;
612 /* OK. use inode's group */
613 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
614 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
615 colour = (current->pid % 16) *
616 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
617 return bg_start + colour + block;
620 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
621 struct ldiskfs_ext_path *path,
622 unsigned long block, unsigned long *count,
625 struct ldiskfs_allocation_request ar;
626 unsigned long pblock;
629 /* find neighbour allocated blocks */
631 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
635 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
639 /* allocate new block */
640 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
644 ar.flags = LDISKFS_MB_HINT_DATA;
645 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
650 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
651 struct ldiskfs_ext_path *path,
652 struct ldiskfs_ext_cache *cex,
653 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
654 struct ldiskfs_extent *ex,
658 struct bpointers *bp = cbdata;
659 struct ldiskfs_extent nex;
660 unsigned long pblock = 0;
666 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
667 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
669 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
675 if (bp->create == 0) {
677 if (cex->ec_block < bp->start)
678 i = bp->start - cex->ec_block;
679 if (i >= cex->ec_len)
680 CERROR("nothing to do?! i = %d, e_num = %u\n",
682 for (; i < cex->ec_len && bp->num; i++) {
692 tgen = LDISKFS_I(inode)->i_ext_generation;
693 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
695 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
696 count + LDISKFS_ALLOC_NEEDED + 1);
697 if (IS_ERR(handle)) {
698 return PTR_ERR(handle);
701 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
702 /* the tree has changed. so path can be invalid at moment */
703 ldiskfs_journal_stop(handle);
707 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
708 * protected by i_data_sem as whole. so we patch it to store
709 * generation to path and now verify the tree hasn't changed */
710 down_write((&LDISKFS_I(inode)->i_data_sem));
712 /* validate extent, make sure the extent tree does not changed */
713 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
714 /* cex is invalid, try again */
715 up_write(&LDISKFS_I(inode)->i_data_sem);
716 ldiskfs_journal_stop(handle);
721 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
724 BUG_ON(count > cex->ec_len);
726 /* insert new extent */
727 nex.ee_block = cpu_to_le32(cex->ec_block);
728 ldiskfs_ext_store_pblock(&nex, pblock);
729 nex.ee_len = cpu_to_le16(count);
730 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
732 /* free data blocks we just allocated */
733 /* not a good idea to call discard here directly,
734 * but otherwise we'd need to call it every free() */
735 ldiskfs_discard_preallocations(inode);
736 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
737 ldiskfs_free_blocks(handle, inode, NULL,
738 ldiskfs_ext_pblock(&nex),
739 le16_to_cpu(nex.ee_len), 0);
741 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
742 le16_to_cpu(nex.ee_len), 0);
748 * Putting len of the actual extent we just inserted,
749 * we are asking ldiskfs_ext_walk_space() to continue
750 * scaning after that block
752 cex->ec_len = le16_to_cpu(nex.ee_len);
753 cex->ec_start = ldiskfs_ext_pblock(&nex);
754 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
755 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
758 up_write((&LDISKFS_I(inode)->i_data_sem));
759 ldiskfs_journal_stop(handle);
764 CERROR("hmm. why do we find this extent?\n");
765 CERROR("initial space: %lu:%u\n",
766 bp->start, bp->init_num);
767 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
768 CERROR("current extent: %u/%u/%llu %d\n",
769 cex->ec_block, cex->ec_len,
770 (unsigned long long)cex->ec_start,
773 CERROR("current extent: %u/%u/%llu\n",
774 cex->ec_block, cex->ec_len,
775 (unsigned long long)cex->ec_start);
779 if (cex->ec_block < bp->start)
780 i = bp->start - cex->ec_block;
781 if (i >= cex->ec_len)
782 CERROR("nothing to do?! i = %d, e_num = %u\n",
784 for (; i < cex->ec_len && bp->num; i++) {
785 *(bp->blocks) = cex->ec_start + i;
787 /* unmap any possible underlying metadata from
788 * the block device mapping. bug 6998. */
789 #ifndef HAVE_CLEAN_BDEV_ALIASES
790 unmap_underlying_metadata(inode->i_sb->s_bdev,
793 clean_bdev_aliases(inode->i_sb->s_bdev,
805 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
806 int clen, sector_t *blocks, int create)
808 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
812 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
816 bp.start = index * blocks_per_page;
817 bp.init_num = bp.num = clen * blocks_per_page;
820 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
821 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
823 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
824 ldiskfs_ext_new_extent_cb, &bp);
825 ldiskfs_ext_invalidate_cache(inode);
830 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
831 struct page **page, int pages,
832 sector_t *blocks, int create)
834 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
835 pgoff_t bitmap_max_page_index;
839 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
841 for (i = 0, b = blocks; i < pages; i++, page++) {
842 if ((*page)->index + 1 >= bitmap_max_page_index) {
846 rc = ldiskfs_map_inode_page(inode, *page, b, create);
848 CERROR("ino %lu, blk %llu create %d: rc %d\n",
850 (unsigned long long)*b, create, rc);
853 b += blocks_per_page;
858 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
860 int pages, sector_t *blocks,
863 int rc = 0, i = 0, clen = 0;
864 struct page *fp = NULL;
866 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
867 inode->i_ino, pages, (*page)->index);
869 /* pages are sorted already. so, we just have to find
870 * contig. space and process them properly */
873 /* start new extent */
878 } else if (fp->index + clen == (*page)->index) {
879 /* continue the extent */
886 /* process found extent */
887 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
892 /* look for next extent */
894 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
898 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
905 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
906 int pages, sector_t *blocks,
911 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
912 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
916 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
921 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
922 int pages, sector_t *blocks,
925 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
927 struct page *fp = NULL;
929 pgoff_t max_page_index;
930 handle_t *handle = NULL;
932 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
934 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
935 inode->i_ino, pages, (*page)->index);
938 create = LDISKFS_GET_BLOCKS_CREATE;
939 handle = ldiskfs_journal_current_handle();
940 LASSERT(handle != NULL);
941 rc = osd_attach_jinode(inode);
945 /* pages are sorted already. so, we just have to find
946 * contig. space and process them properly */
948 long blen, total = 0;
949 struct ldiskfs_map_blocks map = { 0 };
951 if (fp == NULL) { /* start new extent */
956 } else if (fp->index + clen == (*page)->index) {
957 /* continue the extent */
963 if (fp->index + clen >= max_page_index)
964 GOTO(cleanup, rc = -EFBIG);
965 /* process found extent */
966 map.m_lblk = fp->index * blocks_per_page;
967 map.m_len = blen = clen * blocks_per_page;
969 rc = ldiskfs_map_blocks(handle, inode, &map, create);
972 for (; total < blen && c < map.m_len; c++, total++) {
974 *(blocks + total) = 0;
978 *(blocks + total) = map.m_pblk + c;
979 /* unmap any possible underlying
980 * metadata from the block device
981 * mapping. bug 6998. */
982 if ((map.m_flags & LDISKFS_MAP_NEW) &&
984 #ifndef HAVE_CLEAN_BDEV_ALIASES
985 unmap_underlying_metadata(
997 if (rc == 0 && total < blen) {
998 map.m_lblk = fp->index * blocks_per_page + total;
999 map.m_len = blen - total;
1005 /* look for next extent */
1007 blocks += blocks_per_page * clen;
1012 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1014 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1015 struct niobuf_local *lnb, int npages)
1017 struct osd_thread_info *oti = osd_oti_get(env);
1018 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1019 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1020 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1032 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1033 if (unlikely(rc != 0))
1036 isize = i_size_read(inode);
1037 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1039 if (osd->od_writethrough_cache)
1041 if (isize > osd->od_readcache_max_filesize)
1044 start = ktime_get();
1045 for (i = 0; i < npages; i++) {
1048 generic_error_remove_page(inode->i_mapping,
1052 * till commit the content of the page is undefined
1053 * we'll set it uptodate once bulk is done. otherwise
1054 * subsequent reads can access non-stable data
1056 ClearPageUptodate(lnb[i].lnb_page);
1058 if (lnb[i].lnb_len == PAGE_SIZE)
1061 if (maxidx >= lnb[i].lnb_page->index) {
1062 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1065 char *p = kmap(lnb[i].lnb_page);
1067 off = lnb[i].lnb_page_offset;
1070 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1073 memset(p + off, 0, PAGE_SIZE - off);
1074 kunmap(lnb[i].lnb_page);
1078 timediff = ktime_us_delta(end, start);
1079 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1081 if (iobuf->dr_npages) {
1082 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1084 iobuf->dr_blocks, 0);
1085 if (likely(rc == 0)) {
1086 rc = osd_do_bio(osd, inode, iobuf);
1087 /* do IO stats for preparation reads */
1088 osd_fini_iobuf(osd, iobuf);
1094 struct osd_fextent {
1097 unsigned int mapped:1;
1100 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1101 struct osd_fextent *cached_extent)
1103 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1104 sector_t block = offset >> inode->i_blkbits;
1106 struct fiemap_extent_info fei = { 0 };
1107 struct fiemap_extent fe = { 0 };
1108 mm_segment_t saved_fs;
1111 if (block >= cached_extent->start && block < cached_extent->end)
1112 return cached_extent->mapped;
1114 if (i_size_read(inode) == 0)
1117 /* Beyond EOF, must not be mapped */
1118 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1121 fei.fi_extents_max = 1;
1122 fei.fi_extents_start = &fe;
1124 saved_fs = get_fs();
1126 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1131 start = fe.fe_logical >> inode->i_blkbits;
1133 if (start > block) {
1134 cached_extent->start = block;
1135 cached_extent->end = start;
1136 cached_extent->mapped = 0;
1138 cached_extent->start = start;
1139 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1141 cached_extent->mapped = 1;
1144 return cached_extent->mapped;
1147 static int osd_declare_write_commit(const struct lu_env *env,
1148 struct dt_object *dt,
1149 struct niobuf_local *lnb, int npages,
1150 struct thandle *handle)
1152 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1153 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1154 struct osd_thandle *oh;
1162 long long quota_space = 0;
1163 struct osd_fextent extent = { 0 };
1164 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1167 LASSERT(handle != NULL);
1168 oh = container_of0(handle, struct osd_thandle, ot_super);
1169 LASSERT(oh->ot_handle == NULL);
1173 /* calculate number of extents (probably better to pass nb) */
1174 for (i = 0; i < npages; i++) {
1175 if (i && lnb[i].lnb_file_offset !=
1176 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1179 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1180 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1182 quota_space += PAGE_SIZE;
1184 /* ignore quota for the whole request if any page is from
1185 * client cache or written by root.
1187 * XXX once we drop the 1.8 client support, the checking
1188 * for whether page is from cache can be simplified as:
1189 * !(lnb[i].flags & OBD_BRW_SYNC)
1191 * XXX we could handle this on per-lnb basis as done by
1193 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1194 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1196 declare_flags |= OSD_QID_FORCE;
1200 * each extent can go into new leaf causing a split
1201 * 5 is max tree depth: inode + 4 index blocks
1202 * with blockmaps, depth is 3 at most
1204 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1206 * many concurrent threads may grow tree by the time
1207 * our transaction starts. so, consider 2 is a min depth
1209 depth = ext_depth(inode);
1210 depth = max(depth, 1) + 1;
1212 credits++; /* inode */
1213 credits += depth * 2 * extents;
1217 credits++; /* inode */
1218 credits += depth * extents;
1221 /* quota space for metadata blocks */
1222 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1224 /* quota space should be reported in 1K blocks */
1225 quota_space = toqb(quota_space);
1227 /* each new block can go in different group (bitmap + gd) */
1229 /* we can't dirty more bitmap blocks than exist */
1230 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1231 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1233 credits += newblocks;
1235 /* we can't dirty more gd blocks than exist */
1236 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1237 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1239 credits += newblocks;
1241 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1243 /* make sure the over quota flags were not set */
1244 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1246 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1247 i_projid_read(inode), quota_space, oh,
1248 osd_dt_obj(dt), &flags, declare_flags);
1250 /* we need only to store the overquota flags in the first lnb for
1251 * now, once we support multiple objects BRW, this code needs be
1253 if (flags & QUOTA_FL_OVER_USRQUOTA)
1254 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1255 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1256 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1257 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1258 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1263 /* Check if a block is allocated or not */
1264 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1265 struct niobuf_local *lnb, int npages,
1266 struct thandle *thandle)
1268 struct osd_thread_info *oti = osd_oti_get(env);
1269 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1270 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1271 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1277 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1278 if (unlikely(rc != 0))
1281 isize = i_size_read(inode);
1282 ll_vfs_dq_init(inode);
1284 for (i = 0; i < npages; i++) {
1285 if (lnb[i].lnb_rc == -ENOSPC &&
1286 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1287 /* Allow the write to proceed if overwriting an
1292 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1293 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1295 LASSERT(lnb[i].lnb_page);
1296 generic_error_remove_page(inode->i_mapping,
1301 LASSERT(PageLocked(lnb[i].lnb_page));
1302 LASSERT(!PageWriteback(lnb[i].lnb_page));
1304 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1305 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1308 * Since write and truncate are serialized by oo_sem, even
1309 * partial-page truncate should not leave dirty pages in the
1312 LASSERT(!PageDirty(lnb[i].lnb_page));
1314 SetPageUptodate(lnb[i].lnb_page);
1316 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1319 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1321 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1323 } else if (iobuf->dr_npages > 0) {
1324 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1326 iobuf->dr_blocks, 1);
1328 /* no pages to write, no transno is needed */
1329 thandle->th_local = 1;
1332 if (likely(rc == 0)) {
1333 spin_lock(&inode->i_lock);
1334 if (isize > i_size_read(inode)) {
1335 i_size_write(inode, isize);
1336 LDISKFS_I(inode)->i_disksize = isize;
1337 spin_unlock(&inode->i_lock);
1338 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1340 spin_unlock(&inode->i_lock);
1343 rc = osd_do_bio(osd, inode, iobuf);
1344 /* we don't do stats here as in read path because
1345 * write is async: we'll do this in osd_put_bufs() */
1347 osd_fini_iobuf(osd, iobuf);
1350 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1352 if (unlikely(rc != 0)) {
1353 /* if write fails, we should drop pages from the cache */
1354 for (i = 0; i < npages; i++) {
1355 if (lnb[i].lnb_page == NULL)
1357 LASSERT(PageLocked(lnb[i].lnb_page));
1358 generic_error_remove_page(inode->i_mapping,
1366 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1367 struct niobuf_local *lnb, int npages)
1369 struct osd_thread_info *oti = osd_oti_get(env);
1370 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1371 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1372 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1373 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1380 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1381 if (unlikely(rc != 0))
1384 isize = i_size_read(inode);
1386 if (osd->od_read_cache)
1388 if (isize > osd->od_readcache_max_filesize)
1391 start = ktime_get();
1392 for (i = 0; i < npages; i++) {
1394 if (isize <= lnb[i].lnb_file_offset)
1395 /* If there's no more data, abort early.
1396 * lnb->lnb_rc == 0, so it's easy to detect later. */
1399 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1400 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1402 lnb[i].lnb_rc = lnb[i].lnb_len;
1404 /* Bypass disk read if fail_loc is set properly */
1405 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1406 SetPageUptodate(lnb[i].lnb_page);
1408 if (PageUptodate(lnb[i].lnb_page)) {
1412 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1416 generic_error_remove_page(inode->i_mapping,
1420 timediff = ktime_us_delta(end, start);
1421 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1423 if (cache_hits != 0)
1424 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1426 if (cache_misses != 0)
1427 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1429 if (cache_hits + cache_misses != 0)
1430 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1431 cache_hits + cache_misses);
1433 if (iobuf->dr_npages) {
1434 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1436 iobuf->dr_blocks, 0);
1437 rc = osd_do_bio(osd, inode, iobuf);
1439 /* IO stats will be done in osd_bufs_put() */
1446 * XXX: Another layering violation for now.
1448 * We don't want to use ->f_op->read methods, because generic file write
1450 * - serializes on ->i_sem, and
1452 * - does a lot of extra work like balance_dirty_pages(),
1454 * which doesn't work for globally shared files like /last_rcvd.
1456 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1458 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1460 memcpy(buffer, (char *)ei->i_data, buflen);
1465 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1467 struct buffer_head *bh;
1468 unsigned long block;
1474 /* prevent reading after eof */
1475 spin_lock(&inode->i_lock);
1476 if (i_size_read(inode) < *offs + size) {
1477 loff_t diff = i_size_read(inode) - *offs;
1478 spin_unlock(&inode->i_lock);
1480 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1481 i_size_read(inode), *offs);
1483 } else if (diff == 0) {
1489 spin_unlock(&inode->i_lock);
1492 blocksize = 1 << inode->i_blkbits;
1495 block = *offs >> inode->i_blkbits;
1496 boffs = *offs & (blocksize - 1);
1497 csize = min(blocksize - boffs, size);
1498 bh = __ldiskfs_bread(NULL, inode, block, 0);
1500 CERROR("%s: can't read %u@%llu on ino %lu: "
1501 "rc = %ld\n", osd_ino2name(inode),
1502 csize, *offs, inode->i_ino,
1508 memcpy(buf, bh->b_data + boffs, csize);
1511 memset(buf, 0, csize);
1521 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1522 struct lu_buf *buf, loff_t *pos)
1524 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1527 /* Read small symlink from inode body as we need to maintain correct
1528 * on-disk symlinks for ldiskfs.
1530 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1531 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1532 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1534 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1539 static inline int osd_extents_enabled(struct super_block *sb,
1540 struct inode *inode)
1542 if (inode != NULL) {
1543 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1545 } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1546 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1552 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1553 const loff_t size, const loff_t pos,
1556 int credits, bits, bs, i;
1558 bits = sb->s_blocksize_bits;
1561 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1562 * we do not expect blockmaps on the large files,
1563 * so let's shrink it to 2 levels (4GB files) */
1565 /* this is default reservation: 2 levels */
1566 credits = (blocks + 2) * 3;
1568 /* actual offset is unknown, hard to optimize */
1572 /* now check for few specific cases to optimize */
1573 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1576 /* allocate if not allocated */
1577 if (inode == NULL) {
1578 credits += blocks * 2;
1581 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1582 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1583 if (LDISKFS_I(inode)->i_data[i] == 0)
1586 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1587 /* single indirect */
1588 credits = blocks * 3;
1589 if (inode == NULL ||
1590 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1593 /* The indirect block may be modified. */
1600 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1601 const struct lu_buf *buf, loff_t _pos,
1602 struct thandle *handle)
1604 struct osd_object *obj = osd_dt_obj(dt);
1605 struct inode *inode = obj->oo_inode;
1606 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1607 struct osd_thandle *oh;
1608 int rc = 0, est = 0, credits, blocks, allocated = 0;
1614 LASSERT(buf != NULL);
1615 LASSERT(handle != NULL);
1617 oh = container_of0(handle, struct osd_thandle, ot_super);
1618 LASSERT(oh->ot_handle == NULL);
1621 bits = sb->s_blocksize_bits;
1625 /* if this is an append, then we
1626 * should expect cross-block record */
1632 /* blocks to modify */
1633 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1634 LASSERT(blocks > 0);
1636 if (inode != NULL && _pos != -1) {
1637 /* object size in blocks */
1638 est = (i_size_read(inode) + bs - 1) >> bits;
1639 allocated = inode->i_blocks >> (bits - 9);
1640 if (pos + size <= i_size_read(inode) && est <= allocated) {
1641 /* looks like an overwrite, no need to modify tree */
1643 /* no need to modify i_size */
1648 if (osd_extents_enabled(sb, inode)) {
1650 * many concurrent threads may grow tree by the time
1651 * our transaction starts. so, consider 2 is a min depth
1652 * for every level we may need to allocate a new block
1653 * and take some entries from the old one. so, 3 blocks
1654 * to allocate (bitmap, gd, itself) + old block - 4 per
1657 depth = inode != NULL ? ext_depth(inode) : 0;
1658 depth = max(depth, 1) + 1;
1660 /* if not append, then split may need to modify
1661 * existing blocks moving entries into the new ones */
1664 /* blocks to store data: bitmap,gd,itself */
1665 credits += blocks * 3;
1667 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1669 /* if inode is created as part of the transaction,
1670 * then it's counted already by the creation method */
1676 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1678 /* dt_declare_write() is usually called for system objects, such
1679 * as llog or last_rcvd files. We needn't enforce quota on those
1680 * objects, so always set the lqi_space as 0. */
1682 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1684 i_projid_read(inode), 0,
1685 oh, obj, NULL, OSD_QID_BLK);
1689 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1691 /* LU-2634: clear the extent format for fast symlink */
1692 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1694 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1695 spin_lock(&inode->i_lock);
1696 LDISKFS_I(inode)->i_disksize = buflen;
1697 i_size_write(inode, buflen);
1698 spin_unlock(&inode->i_lock);
1699 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1704 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1705 int write_NUL, loff_t *offs, handle_t *handle)
1707 struct buffer_head *bh = NULL;
1708 loff_t offset = *offs;
1709 loff_t new_size = i_size_read(inode);
1710 unsigned long block;
1711 int blocksize = 1 << inode->i_blkbits;
1715 int dirty_inode = 0;
1719 * long symlink write does not count the NUL terminator in
1720 * bufsize, we write it, and the inode's file size does not
1721 * count the NUL terminator as well.
1723 ((char *)buf)[bufsize] = '\0';
1727 while (bufsize > 0) {
1728 int credits = handle->h_buffer_credits;
1733 block = offset >> inode->i_blkbits;
1734 boffs = offset & (blocksize - 1);
1735 size = min(blocksize - boffs, bufsize);
1736 bh = __ldiskfs_bread(handle, inode, block, 1);
1737 if (IS_ERR_OR_NULL(bh)) {
1745 CERROR("%s: error reading offset %llu (block %lu, "
1746 "size %d, offs %llu), credits %d/%d: rc = %d\n",
1747 inode->i_sb->s_id, offset, block, bufsize, *offs,
1748 credits, handle->h_buffer_credits, err);
1752 err = ldiskfs_journal_get_write_access(handle, bh);
1754 CERROR("journal_get_write_access() returned error %d\n",
1758 LASSERTF(boffs + size <= bh->b_size,
1759 "boffs %d size %d bh->b_size %lu\n",
1760 boffs, size, (unsigned long)bh->b_size);
1761 memcpy(bh->b_data + boffs, buf, size);
1762 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1766 if (offset + size > new_size)
1767 new_size = offset + size;
1777 /* correct in-core and on-disk sizes */
1778 if (new_size > i_size_read(inode)) {
1779 spin_lock(&inode->i_lock);
1780 if (new_size > i_size_read(inode))
1781 i_size_write(inode, new_size);
1782 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1783 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1786 spin_unlock(&inode->i_lock);
1788 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1796 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1797 const struct lu_buf *buf, loff_t *pos,
1798 struct thandle *handle, int ignore_quota)
1800 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1801 struct osd_thandle *oh;
1805 LASSERT(dt_object_exists(dt));
1807 LASSERT(handle != NULL);
1808 LASSERT(inode != NULL);
1809 ll_vfs_dq_init(inode);
1811 /* XXX: don't check: one declared chunk can be used many times */
1812 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1814 oh = container_of(handle, struct osd_thandle, ot_super);
1815 LASSERT(oh->ot_handle->h_transaction != NULL);
1816 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1818 /* Write small symlink to inode body as we need to maintain correct
1819 * on-disk symlinks for ldiskfs.
1820 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1821 * does not count it in.
1823 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1824 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1825 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1827 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1828 buf->lb_len, is_link, pos,
1831 result = buf->lb_len;
1833 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1838 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1839 __u64 start, __u64 end, struct thandle *th)
1841 struct osd_thandle *oh;
1842 struct inode *inode;
1847 oh = container_of(th, struct osd_thandle, ot_super);
1850 * we don't need to reserve credits for whole truncate
1851 * it's not possible as truncate may need to free too many
1852 * blocks and that won't fit a single transaction. instead
1853 * we reserve credits to change i_size and put inode onto
1854 * orphan list. if needed truncate will extend or restart
1857 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1858 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1860 inode = osd_dt_obj(dt)->oo_inode;
1863 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1864 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1869 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1870 __u64 start, __u64 end, struct thandle *th)
1872 struct osd_thandle *oh;
1873 struct osd_object *obj = osd_dt_obj(dt);
1874 struct inode *inode = obj->oo_inode;
1877 int rc = 0, rc2 = 0;
1880 LASSERT(end == OBD_OBJECT_EOF);
1881 LASSERT(dt_object_exists(dt));
1882 LASSERT(osd_invariant(obj));
1883 LASSERT(inode != NULL);
1884 ll_vfs_dq_init(inode);
1887 oh = container_of(th, struct osd_thandle, ot_super);
1888 LASSERT(oh->ot_handle->h_transaction != NULL);
1890 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1892 tid = oh->ot_handle->h_transaction->t_tid;
1894 spin_lock(&inode->i_lock);
1895 i_size_write(inode, start);
1896 spin_unlock(&inode->i_lock);
1897 ll_truncate_pagecache(inode, start);
1898 #ifdef HAVE_INODEOPS_TRUNCATE
1899 if (inode->i_op->truncate) {
1900 inode->i_op->truncate(inode);
1903 ldiskfs_truncate(inode);
1906 * For a partial-page truncate, flush the page to disk immediately to
1907 * avoid data corruption during direct disk write. b=17397
1909 if ((start & ~PAGE_MASK) != 0)
1910 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1912 h = journal_current_handle();
1914 LASSERT(h == oh->ot_handle);
1916 /* do not check credits with osd_trans_exec_check() as the truncate
1917 * can restart the transaction internally and we restart the
1918 * transaction in this case */
1920 if (tid != h->h_transaction->t_tid) {
1921 int credits = oh->ot_credits;
1923 * transaction has changed during truncate
1924 * we need to restart the handle with our credits
1926 if (h->h_buffer_credits < credits) {
1927 if (ldiskfs_journal_extend(h, credits))
1928 rc2 = ldiskfs_journal_restart(h, credits);
1932 RETURN(rc == 0 ? rc2 : rc);
1935 static int fiemap_check_ranges(struct inode *inode,
1936 u64 start, u64 len, u64 *new_len)
1945 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1946 maxbytes = inode->i_sb->s_maxbytes;
1948 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1950 if (start > maxbytes)
1954 * Shrink request scope to what the fs can actually handle.
1956 if (len > maxbytes || (maxbytes - len) < start)
1957 *new_len = maxbytes - start;
1962 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1963 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
1965 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1968 struct fiemap_extent_info fieinfo = {0, };
1969 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1975 if (inode->i_op->fiemap == NULL)
1978 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1981 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1985 fieinfo.fi_flags = fm->fm_flags;
1986 fieinfo.fi_extents_max = fm->fm_extent_count;
1987 fieinfo.fi_extents_start = fm->fm_extents;
1989 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1990 filemap_write_and_wait(inode->i_mapping);
1992 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1993 fm->fm_flags = fieinfo.fi_flags;
1994 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1999 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2000 __u64 start, __u64 end, enum lu_ladvise_type advice)
2003 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2007 case LU_LADVISE_DONTNEED:
2010 invalidate_mapping_pages(inode->i_mapping,
2011 start >> PAGE_CACHE_SHIFT,
2012 (end - 1) >> PAGE_CACHE_SHIFT);
2023 * in some cases we may need declare methods for objects being created
2024 * e.g., when we create symlink
2026 const struct dt_body_operations osd_body_ops_new = {
2027 .dbo_declare_write = osd_declare_write,
2030 const struct dt_body_operations osd_body_ops = {
2031 .dbo_read = osd_read,
2032 .dbo_declare_write = osd_declare_write,
2033 .dbo_write = osd_write,
2034 .dbo_bufs_get = osd_bufs_get,
2035 .dbo_bufs_put = osd_bufs_put,
2036 .dbo_write_prep = osd_write_prep,
2037 .dbo_declare_write_commit = osd_declare_write_commit,
2038 .dbo_write_commit = osd_write_commit,
2039 .dbo_read_prep = osd_read_prep,
2040 .dbo_declare_punch = osd_declare_punch,
2041 .dbo_punch = osd_punch,
2042 .dbo_fiemap_get = osd_fiemap_get,
2043 .dbo_ladvise = osd_ladvise,