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) 2012, 2015, 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 * struct OBD_{ALLOC,FREE}*()
56 #include <obd_support.h>
58 #include "osd_internal.h"
61 #include <ldiskfs/ldiskfs_extents.h>
63 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
64 int rw, int line, int pages)
68 LASSERTF(iobuf->dr_elapsed_valid == 0,
69 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
70 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
72 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
74 init_waitqueue_head(&iobuf->dr_wait);
75 atomic_set(&iobuf->dr_numreqs, 0);
80 iobuf->dr_elapsed = 0;
81 /* must be counted before, so assert */
83 iobuf->dr_init_at = line;
85 blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
86 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
87 LASSERT(iobuf->dr_pg_buf.lb_len >=
88 pages * sizeof(iobuf->dr_pages[0]));
92 /* start with 1MB for 4K blocks */
94 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
97 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
98 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
100 blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
101 iobuf->dr_max_pages = 0;
102 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
103 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
105 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
106 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
107 if (unlikely(iobuf->dr_blocks == NULL))
110 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
111 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
112 if (unlikely(iobuf->dr_pages == NULL))
115 iobuf->dr_max_pages = pages;
119 #define osd_init_iobuf(dev, iobuf, rw, pages) \
120 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
122 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
124 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
125 iobuf->dr_pages[iobuf->dr_npages++] = page;
128 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
130 int rw = iobuf->dr_rw;
132 if (iobuf->dr_elapsed_valid) {
133 iobuf->dr_elapsed_valid = 0;
134 LASSERT(iobuf->dr_dev == d);
135 LASSERT(iobuf->dr_frags > 0);
136 lprocfs_oh_tally(&d->od_brw_stats.
137 hist[BRW_R_DIO_FRAGS+rw],
139 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
144 #ifndef REQ_WRITE /* pre-2.6.35 */
145 #define __REQ_WRITE BIO_RW
148 static void dio_complete_routine(struct bio *bio, int error)
150 struct osd_iobuf *iobuf = bio->bi_private;
151 #ifdef HAVE_BVEC_ITER
152 struct bvec_iter iter;
159 /* CAVEAT EMPTOR: possibly in IRQ context
160 * DO NOT record procfs stats here!!! */
162 if (unlikely(iobuf == NULL)) {
163 CERROR("***** bio->bi_private is NULL! This should never "
164 "happen. Normally, I would crash here, but instead I "
165 "will dump the bio contents to the console. Please "
166 "report this to <https://jira.hpdd.intel.com/> , along "
167 "with any interesting messages leading up to this point "
168 "(like SCSI errors, perhaps). Because bi_private is "
169 "NULL, I can't wake up the thread that initiated this "
170 "IO - you will probably have to reboot this node.\n");
171 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
172 "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
173 "bi_private: %p\n", bio->bi_next, bio->bi_flags,
174 bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
175 bio_sectors(bio) << 9, bio->bi_end_io,
176 atomic_read(&bio->bi_cnt), bio->bi_private);
180 /* the check is outside of the cycle for performance reason -bzzz */
181 if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
182 bio_for_each_segment_all(bvl, bio, iter) {
183 if (likely(error == 0))
184 SetPageUptodate(bvl_to_page(bvl));
185 LASSERT(PageLocked(bvl_to_page(bvl)));
187 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
189 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
192 /* any real error is good enough -bzzz */
193 if (error != 0 && iobuf->dr_error == 0)
194 iobuf->dr_error = error;
197 * set dr_elapsed before dr_numreqs turns to 0, otherwise
198 * it's possible that service thread will see dr_numreqs
199 * is zero, but dr_elapsed is not set yet, leading to lost
200 * data in this processing and an assertion in a subsequent
203 if (atomic_read(&iobuf->dr_numreqs) == 1) {
204 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
205 iobuf->dr_elapsed_valid = 1;
207 if (atomic_dec_and_test(&iobuf->dr_numreqs))
208 wake_up(&iobuf->dr_wait);
210 /* Completed bios used to be chained off iobuf->dr_bios and freed in
211 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
212 * mempool when serious on-disk fragmentation was encountered,
213 * deadlocking the OST. The bios are now released as soon as complete
214 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
218 static void record_start_io(struct osd_iobuf *iobuf, int size)
220 struct osd_device *osd = iobuf->dr_dev;
221 struct obd_histogram *h = osd->od_brw_stats.hist;
224 atomic_inc(&iobuf->dr_numreqs);
226 if (iobuf->dr_rw == 0) {
227 atomic_inc(&osd->od_r_in_flight);
228 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
229 atomic_read(&osd->od_r_in_flight));
230 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
231 } else if (iobuf->dr_rw == 1) {
232 atomic_inc(&osd->od_w_in_flight);
233 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
234 atomic_read(&osd->od_w_in_flight));
235 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
241 static void osd_submit_bio(int rw, struct bio *bio)
243 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
245 submit_bio(READ, bio);
247 submit_bio(WRITE, bio);
250 static int can_be_merged(struct bio *bio, sector_t sector)
255 return bio_end_sector(bio) == sector ? 1 : 0;
258 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
259 struct osd_iobuf *iobuf)
261 int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
262 struct page **pages = iobuf->dr_pages;
263 int npages = iobuf->dr_npages;
264 sector_t *blocks = iobuf->dr_blocks;
265 int total_blocks = npages * blocks_per_page;
266 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
267 unsigned int blocksize = inode->i_sb->s_blocksize;
268 struct bio *bio = NULL;
270 unsigned int page_offset;
279 LASSERT(iobuf->dr_npages == npages);
281 osd_brw_stats_update(osd, iobuf);
282 iobuf->dr_start_time = cfs_time_current();
284 for (page_idx = 0, block_idx = 0;
286 page_idx++, block_idx += blocks_per_page) {
288 page = pages[page_idx];
289 LASSERT(block_idx + blocks_per_page <= total_blocks);
291 for (i = 0, page_offset = 0;
293 i += nblocks, page_offset += blocksize * nblocks) {
297 if (blocks[block_idx + i] == 0) { /* hole */
298 LASSERTF(iobuf->dr_rw == 0,
299 "page_idx %u, block_idx %u, i %u\n",
300 page_idx, block_idx, i);
301 memset(kmap(page) + page_offset, 0, blocksize);
306 sector = (sector_t)blocks[block_idx + i] << sector_bits;
308 /* Additional contiguous file blocks? */
309 while (i + nblocks < blocks_per_page &&
310 (sector + (nblocks << sector_bits)) ==
311 ((sector_t)blocks[block_idx + i + nblocks] <<
316 can_be_merged(bio, sector) &&
317 bio_add_page(bio, page,
318 blocksize * nblocks, page_offset) != 0)
319 continue; /* added this frag OK */
322 struct request_queue *q =
323 bdev_get_queue(bio->bi_bdev);
324 unsigned int bi_size = bio_sectors(bio) << 9;
326 /* Dang! I have to fragment this I/O */
327 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
328 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
329 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
331 queue_max_sectors(q),
332 bio_phys_segments(q, bio),
333 queue_max_phys_segments(q),
334 0, queue_max_hw_segments(q));
335 record_start_io(iobuf, bi_size);
336 osd_submit_bio(iobuf->dr_rw, bio);
339 /* allocate new bio */
340 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
341 (npages - page_idx) *
344 CERROR("Can't allocate bio %u*%u = %u pages\n",
345 (npages - page_idx), blocks_per_page,
346 (npages - page_idx) * blocks_per_page);
351 bio->bi_bdev = inode->i_sb->s_bdev;
352 bio_set_sector(bio, sector);
353 bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
354 bio->bi_end_io = dio_complete_routine;
355 bio->bi_private = iobuf;
357 rc = bio_add_page(bio, page,
358 blocksize * nblocks, page_offset);
364 record_start_io(iobuf, bio_sectors(bio) << 9);
365 osd_submit_bio(iobuf->dr_rw, bio);
370 /* in order to achieve better IO throughput, we don't wait for writes
371 * completion here. instead we proceed with transaction commit in
372 * parallel and wait for IO completion once transaction is stopped
373 * see osd_trans_stop() for more details -bzzz */
374 if (iobuf->dr_rw == 0) {
375 wait_event(iobuf->dr_wait,
376 atomic_read(&iobuf->dr_numreqs) == 0);
377 osd_fini_iobuf(osd, iobuf);
381 rc = iobuf->dr_error;
385 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
386 struct niobuf_local *lnb)
393 int poff = offset & (PAGE_CACHE_SIZE - 1);
394 int plen = PAGE_CACHE_SIZE - poff;
398 lnb->lnb_file_offset = offset;
399 lnb->lnb_page_offset = poff;
401 /* lnb->lnb_flags = rnb->rnb_flags; */
403 lnb->lnb_page = NULL;
406 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
417 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
419 struct inode *inode = osd_dt_obj(dt)->oo_inode;
420 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
425 page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
426 GFP_NOFS | __GFP_HIGHMEM);
427 if (unlikely(page == NULL))
428 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
434 * there are following "locks":
445 * - lock pages, unlock
447 * - lock partial page
453 * Unlock and release pages loaded by osd_bufs_get()
455 * Unlock \a npages pages from \a lnb and drop the refcount on them.
457 * \param env thread execution environment
458 * \param dt dt object undergoing IO (OSD object + methods)
459 * \param lnb array of pages undergoing IO
460 * \param npages number of pages in \a lnb
464 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
465 struct niobuf_local *lnb, int npages)
469 for (i = 0; i < npages; i++) {
470 if (lnb[i].lnb_page == NULL)
472 LASSERT(PageLocked(lnb[i].lnb_page));
473 unlock_page(lnb[i].lnb_page);
474 page_cache_release(lnb[i].lnb_page);
475 lu_object_put(env, &dt->do_lu);
476 lnb[i].lnb_page = NULL;
483 * Load and lock pages undergoing IO
485 * Pages as described in the \a lnb array are fetched (from disk or cache)
486 * and locked for IO by the caller.
488 * DLM locking protects us from write and truncate competing for same region,
489 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
490 * It's possible the writeout on a such a page is in progress when we access
491 * it. It's also possible that during this writeout we put new (partial) data
492 * into the page, but won't be able to proceed in filter_commitrw_write().
493 * Therefore, just wait for writeout completion as it should be rare enough.
495 * \param env thread execution environment
496 * \param dt dt object undergoing IO (OSD object + methods)
497 * \param pos byte offset of IO start
498 * \param len number of bytes of IO
499 * \param lnb array of extents undergoing IO
500 * \param rw read or write operation?
501 * \param capa capabilities
503 * \retval pages (zero or more) loaded successfully
504 * \retval -ENOMEM on memory/page allocation error
506 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
507 loff_t pos, ssize_t len, struct niobuf_local *lnb,
510 struct osd_object *obj = osd_dt_obj(dt);
511 int npages, i, rc = 0;
513 LASSERT(obj->oo_inode);
515 osd_map_remote_to_local(pos, len, &npages, lnb);
517 for (i = 0; i < npages; i++, lnb++) {
518 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
519 if (lnb->lnb_page == NULL)
520 GOTO(cleanup, rc = -ENOMEM);
522 wait_on_page_writeback(lnb->lnb_page);
523 BUG_ON(PageWriteback(lnb->lnb_page));
525 lu_object_get(&dt->do_lu);
532 osd_bufs_put(env, dt, lnb - i, i);
536 #ifndef HAVE_LDISKFS_MAP_BLOCKS
538 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
539 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
550 static long ldiskfs_ext_find_goal(struct inode *inode,
551 struct ldiskfs_ext_path *path,
552 unsigned long block, int *aflags)
554 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
555 unsigned long bg_start;
556 unsigned long colour;
560 struct ldiskfs_extent *ex;
561 depth = path->p_depth;
563 /* try to predict block placement */
564 if ((ex = path[depth].p_ext))
565 return ldiskfs_ext_pblock(ex) +
566 (block - le32_to_cpu(ex->ee_block));
568 /* it looks index is empty
569 * try to find starting from index itself */
570 if (path[depth].p_bh)
571 return path[depth].p_bh->b_blocknr;
574 /* OK. use inode's group */
575 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
576 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
577 colour = (current->pid % 16) *
578 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
579 return bg_start + colour + block;
582 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
583 struct ldiskfs_ext_path *path,
584 unsigned long block, unsigned long *count,
587 struct ldiskfs_allocation_request ar;
588 unsigned long pblock;
591 /* find neighbour allocated blocks */
593 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
597 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
601 /* allocate new block */
602 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
606 ar.flags = LDISKFS_MB_HINT_DATA;
607 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
612 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
613 struct ldiskfs_ext_path *path,
614 struct ldiskfs_ext_cache *cex,
615 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
616 struct ldiskfs_extent *ex,
620 struct bpointers *bp = cbdata;
621 struct ldiskfs_extent nex;
622 unsigned long pblock = 0;
628 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
629 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
631 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
637 if (bp->create == 0) {
639 if (cex->ec_block < bp->start)
640 i = bp->start - cex->ec_block;
641 if (i >= cex->ec_len)
642 CERROR("nothing to do?! i = %d, e_num = %u\n",
644 for (; i < cex->ec_len && bp->num; i++) {
654 tgen = LDISKFS_I(inode)->i_ext_generation;
655 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
657 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
658 count + LDISKFS_ALLOC_NEEDED + 1);
659 if (IS_ERR(handle)) {
660 return PTR_ERR(handle);
663 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
664 /* the tree has changed. so path can be invalid at moment */
665 ldiskfs_journal_stop(handle);
669 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
670 * protected by i_data_sem as whole. so we patch it to store
671 * generation to path and now verify the tree hasn't changed */
672 down_write((&LDISKFS_I(inode)->i_data_sem));
674 /* validate extent, make sure the extent tree does not changed */
675 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
676 /* cex is invalid, try again */
677 up_write(&LDISKFS_I(inode)->i_data_sem);
678 ldiskfs_journal_stop(handle);
683 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
686 BUG_ON(count > cex->ec_len);
688 /* insert new extent */
689 nex.ee_block = cpu_to_le32(cex->ec_block);
690 ldiskfs_ext_store_pblock(&nex, pblock);
691 nex.ee_len = cpu_to_le16(count);
692 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
694 /* free data blocks we just allocated */
695 /* not a good idea to call discard here directly,
696 * but otherwise we'd need to call it every free() */
697 ldiskfs_discard_preallocations(inode);
698 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
699 ldiskfs_free_blocks(handle, inode, NULL,
700 ldiskfs_ext_pblock(&nex),
701 le16_to_cpu(nex.ee_len), 0);
703 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
704 le16_to_cpu(nex.ee_len), 0);
710 * Putting len of the actual extent we just inserted,
711 * we are asking ldiskfs_ext_walk_space() to continue
712 * scaning after that block
714 cex->ec_len = le16_to_cpu(nex.ee_len);
715 cex->ec_start = ldiskfs_ext_pblock(&nex);
716 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
717 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
720 up_write((&LDISKFS_I(inode)->i_data_sem));
721 ldiskfs_journal_stop(handle);
726 CERROR("hmm. why do we find this extent?\n");
727 CERROR("initial space: %lu:%u\n",
728 bp->start, bp->init_num);
729 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
730 CERROR("current extent: %u/%u/%llu %d\n",
731 cex->ec_block, cex->ec_len,
732 (unsigned long long)cex->ec_start,
735 CERROR("current extent: %u/%u/%llu\n",
736 cex->ec_block, cex->ec_len,
737 (unsigned long long)cex->ec_start);
741 if (cex->ec_block < bp->start)
742 i = bp->start - cex->ec_block;
743 if (i >= cex->ec_len)
744 CERROR("nothing to do?! i = %d, e_num = %u\n",
746 for (; i < cex->ec_len && bp->num; i++) {
747 *(bp->blocks) = cex->ec_start + i;
749 /* unmap any possible underlying metadata from
750 * the block device mapping. bug 6998. */
751 unmap_underlying_metadata(inode->i_sb->s_bdev,
762 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
763 int clen, sector_t *blocks, int create)
765 int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
769 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
773 bp.start = index * blocks_per_page;
774 bp.init_num = bp.num = clen * blocks_per_page;
777 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
778 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
780 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
781 ldiskfs_ext_new_extent_cb, &bp);
782 ldiskfs_ext_invalidate_cache(inode);
787 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
788 struct page **page, int pages,
789 sector_t *blocks, int create)
791 int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
792 pgoff_t bitmap_max_page_index;
796 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
798 for (i = 0, b = blocks; i < pages; i++, page++) {
799 if ((*page)->index + 1 >= bitmap_max_page_index) {
803 rc = ldiskfs_map_inode_page(inode, *page, b, create);
805 CERROR("ino %lu, blk %llu create %d: rc %d\n",
807 (unsigned long long)*b, create, rc);
810 b += blocks_per_page;
815 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
817 int pages, sector_t *blocks,
820 int rc = 0, i = 0, clen = 0;
821 struct page *fp = NULL;
823 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
824 inode->i_ino, pages, (*page)->index);
826 /* pages are sorted already. so, we just have to find
827 * contig. space and process them properly */
830 /* start new extent */
835 } else if (fp->index + clen == (*page)->index) {
836 /* continue the extent */
843 /* process found extent */
844 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
849 /* look for next extent */
851 blocks += clen * (PAGE_CACHE_SIZE >> inode->i_blkbits);
855 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
862 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
863 int pages, sector_t *blocks,
868 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
869 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
873 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
878 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
879 int pages, sector_t *blocks,
882 int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
884 struct page *fp = NULL;
886 pgoff_t max_page_index;
888 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
890 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
891 inode->i_ino, pages, (*page)->index);
893 /* pages are sorted already. so, we just have to find
894 * contig. space and process them properly */
896 long blen, total = 0;
897 handle_t *handle = NULL;
898 struct ldiskfs_map_blocks map = { 0 };
900 if (fp == NULL) { /* start new extent */
905 } else if (fp->index + clen == (*page)->index) {
906 /* continue the extent */
912 if (fp->index + clen >= max_page_index)
913 GOTO(cleanup, rc = -EFBIG);
914 /* process found extent */
915 map.m_lblk = fp->index * blocks_per_page;
916 map.m_len = blen = clen * blocks_per_page;
918 create = LDISKFS_GET_BLOCKS_CREATE;
919 handle = ldiskfs_journal_current_handle();
920 LASSERT(handle != NULL);
923 rc = ldiskfs_map_blocks(handle, inode, &map, create);
926 for (; total < blen && c < map.m_len; c++, total++) {
928 *(blocks + total) = 0;
932 *(blocks + total) = map.m_pblk + c;
933 /* unmap any possible underlying
934 * metadata from the block device
935 * mapping. bug 6998. */
936 if ((map.m_flags & LDISKFS_MAP_NEW) &&
938 unmap_underlying_metadata(
945 if (rc == 0 && total < blen) {
946 map.m_lblk = fp->index * blocks_per_page + total;
947 map.m_len = blen - total;
953 /* look for next extent */
955 blocks += blocks_per_page * clen;
960 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
962 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
963 struct niobuf_local *lnb, int npages)
965 struct osd_thread_info *oti = osd_oti_get(env);
966 struct osd_iobuf *iobuf = &oti->oti_iobuf;
967 struct inode *inode = osd_dt_obj(dt)->oo_inode;
968 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
969 struct timeval start;
971 unsigned long timediff;
980 rc = osd_init_iobuf(osd, iobuf, 0, npages);
981 if (unlikely(rc != 0))
984 isize = i_size_read(inode);
985 maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
987 if (osd->od_writethrough_cache)
989 if (isize > osd->od_readcache_max_filesize)
992 do_gettimeofday(&start);
993 for (i = 0; i < npages; i++) {
996 generic_error_remove_page(inode->i_mapping,
1000 * till commit the content of the page is undefined
1001 * we'll set it uptodate once bulk is done. otherwise
1002 * subsequent reads can access non-stable data
1004 ClearPageUptodate(lnb[i].lnb_page);
1006 if (lnb[i].lnb_len == PAGE_CACHE_SIZE)
1009 if (maxidx >= lnb[i].lnb_page->index) {
1010 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1013 char *p = kmap(lnb[i].lnb_page);
1015 off = lnb[i].lnb_page_offset;
1018 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1021 memset(p + off, 0, PAGE_CACHE_SIZE - off);
1022 kunmap(lnb[i].lnb_page);
1025 do_gettimeofday(&end);
1026 timediff = cfs_timeval_sub(&end, &start, NULL);
1027 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1029 if (iobuf->dr_npages) {
1030 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1032 iobuf->dr_blocks, 0);
1033 if (likely(rc == 0)) {
1034 rc = osd_do_bio(osd, inode, iobuf);
1035 /* do IO stats for preparation reads */
1036 osd_fini_iobuf(osd, iobuf);
1042 struct osd_fextent {
1045 unsigned int mapped:1;
1048 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1049 struct osd_fextent *cached_extent)
1051 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1052 sector_t block = offset >> inode->i_blkbits;
1054 struct fiemap_extent_info fei = { 0 };
1055 struct fiemap_extent fe = { 0 };
1056 mm_segment_t saved_fs;
1059 if (block >= cached_extent->start && block < cached_extent->end)
1060 return cached_extent->mapped;
1062 if (i_size_read(inode) == 0)
1065 /* Beyond EOF, must not be mapped */
1066 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1069 fei.fi_extents_max = 1;
1070 fei.fi_extents_start = &fe;
1072 saved_fs = get_fs();
1074 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1079 start = fe.fe_logical >> inode->i_blkbits;
1081 if (start > block) {
1082 cached_extent->start = block;
1083 cached_extent->end = start;
1084 cached_extent->mapped = 0;
1086 cached_extent->start = start;
1087 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1089 cached_extent->mapped = 1;
1092 return cached_extent->mapped;
1095 static int osd_declare_write_commit(const struct lu_env *env,
1096 struct dt_object *dt,
1097 struct niobuf_local *lnb, int npages,
1098 struct thandle *handle)
1100 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1101 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1102 struct osd_thandle *oh;
1110 bool ignore_quota = false;
1111 long long quota_space = 0;
1112 struct osd_fextent extent = { 0 };
1115 LASSERT(handle != NULL);
1116 oh = container_of0(handle, struct osd_thandle, ot_super);
1117 LASSERT(oh->ot_handle == NULL);
1121 /* calculate number of extents (probably better to pass nb) */
1122 for (i = 0; i < npages; i++) {
1123 if (i && lnb[i].lnb_file_offset !=
1124 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1127 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1128 quota_space += PAGE_CACHE_SIZE;
1130 /* ignore quota for the whole request if any page is from
1131 * client cache or written by root.
1133 * XXX once we drop the 1.8 client support, the checking
1134 * for whether page is from cache can be simplified as:
1135 * !(lnb[i].flags & OBD_BRW_SYNC)
1137 * XXX we could handle this on per-lnb basis as done by
1139 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1140 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1142 ignore_quota = true;
1146 * each extent can go into new leaf causing a split
1147 * 5 is max tree depth: inode + 4 index blocks
1148 * with blockmaps, depth is 3 at most
1150 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1152 * many concurrent threads may grow tree by the time
1153 * our transaction starts. so, consider 2 is a min depth
1155 depth = ext_depth(inode);
1156 depth = max(depth, 1) + 1;
1158 credits++; /* inode */
1159 credits += depth * 2 * extents;
1163 credits++; /* inode */
1164 credits += depth * extents;
1167 /* quota space for metadata blocks */
1168 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1170 /* quota space should be reported in 1K blocks */
1171 quota_space = toqb(quota_space);
1173 /* each new block can go in different group (bitmap + gd) */
1175 /* we can't dirty more bitmap blocks than exist */
1176 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1177 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1179 credits += newblocks;
1181 /* we can't dirty more gd blocks than exist */
1182 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1183 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1185 credits += newblocks;
1187 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1189 /* make sure the over quota flags were not set */
1190 lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1192 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1193 quota_space, oh, osd_dt_obj(dt), true,
1194 &flags, ignore_quota);
1196 /* we need only to store the overquota flags in the first lnb for
1197 * now, once we support multiple objects BRW, this code needs be
1199 if (flags & QUOTA_FL_OVER_USRQUOTA)
1200 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1201 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1202 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1207 /* Check if a block is allocated or not */
1208 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1209 struct niobuf_local *lnb, int npages,
1210 struct thandle *thandle)
1212 struct osd_thread_info *oti = osd_oti_get(env);
1213 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1214 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1215 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1218 struct osd_fextent extent = { 0 };
1222 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1223 if (unlikely(rc != 0))
1226 isize = i_size_read(inode);
1227 ll_vfs_dq_init(inode);
1229 for (i = 0; i < npages; i++) {
1230 if (lnb[i].lnb_rc == -ENOSPC &&
1231 osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1232 /* Allow the write to proceed if overwriting an
1237 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1238 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1240 LASSERT(lnb[i].lnb_page);
1241 generic_error_remove_page(inode->i_mapping,
1246 LASSERT(PageLocked(lnb[i].lnb_page));
1247 LASSERT(!PageWriteback(lnb[i].lnb_page));
1249 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1250 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1253 * Since write and truncate are serialized by oo_sem, even
1254 * partial-page truncate should not leave dirty pages in the
1257 LASSERT(!PageDirty(lnb[i].lnb_page));
1259 SetPageUptodate(lnb[i].lnb_page);
1261 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1264 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1266 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1268 } else if (iobuf->dr_npages > 0) {
1269 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1271 iobuf->dr_blocks, 1);
1273 /* no pages to write, no transno is needed */
1274 thandle->th_local = 1;
1277 if (likely(rc == 0)) {
1278 spin_lock(&inode->i_lock);
1279 if (isize > i_size_read(inode)) {
1280 i_size_write(inode, isize);
1281 LDISKFS_I(inode)->i_disksize = isize;
1282 spin_unlock(&inode->i_lock);
1283 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1285 spin_unlock(&inode->i_lock);
1288 rc = osd_do_bio(osd, inode, iobuf);
1289 /* we don't do stats here as in read path because
1290 * write is async: we'll do this in osd_put_bufs() */
1292 osd_fini_iobuf(osd, iobuf);
1295 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1297 if (unlikely(rc != 0)) {
1298 /* if write fails, we should drop pages from the cache */
1299 for (i = 0; i < npages; i++) {
1300 if (lnb[i].lnb_page == NULL)
1302 LASSERT(PageLocked(lnb[i].lnb_page));
1303 generic_error_remove_page(inode->i_mapping,
1311 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1312 struct niobuf_local *lnb, int npages)
1314 struct osd_thread_info *oti = osd_oti_get(env);
1315 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1316 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1317 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1318 struct timeval start, end;
1319 unsigned long timediff;
1320 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1325 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1326 if (unlikely(rc != 0))
1329 isize = i_size_read(inode);
1331 if (osd->od_read_cache)
1333 if (isize > osd->od_readcache_max_filesize)
1336 do_gettimeofday(&start);
1337 for (i = 0; i < npages; i++) {
1339 if (isize <= lnb[i].lnb_file_offset)
1340 /* If there's no more data, abort early.
1341 * lnb->lnb_rc == 0, so it's easy to detect later. */
1344 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1345 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1347 lnb[i].lnb_rc = lnb[i].lnb_len;
1349 if (PageUptodate(lnb[i].lnb_page)) {
1353 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1357 generic_error_remove_page(inode->i_mapping,
1360 do_gettimeofday(&end);
1361 timediff = cfs_timeval_sub(&end, &start, NULL);
1362 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1364 if (cache_hits != 0)
1365 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1367 if (cache_misses != 0)
1368 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1370 if (cache_hits + cache_misses != 0)
1371 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1372 cache_hits + cache_misses);
1374 if (iobuf->dr_npages) {
1375 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1377 iobuf->dr_blocks, 0);
1378 rc = osd_do_bio(osd, inode, iobuf);
1380 /* IO stats will be done in osd_bufs_put() */
1387 * XXX: Another layering violation for now.
1389 * We don't want to use ->f_op->read methods, because generic file write
1391 * - serializes on ->i_sem, and
1393 * - does a lot of extra work like balance_dirty_pages(),
1395 * which doesn't work for globally shared files like /last_rcvd.
1397 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1399 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1401 memcpy(buffer, (char *)ei->i_data, buflen);
1406 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1408 struct buffer_head *bh;
1409 unsigned long block;
1415 /* prevent reading after eof */
1416 spin_lock(&inode->i_lock);
1417 if (i_size_read(inode) < *offs + size) {
1418 loff_t diff = i_size_read(inode) - *offs;
1419 spin_unlock(&inode->i_lock);
1421 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1422 i_size_read(inode), *offs);
1424 } else if (diff == 0) {
1430 spin_unlock(&inode->i_lock);
1433 blocksize = 1 << inode->i_blkbits;
1436 block = *offs >> inode->i_blkbits;
1437 boffs = *offs & (blocksize - 1);
1438 csize = min(blocksize - boffs, size);
1439 bh = __ldiskfs_bread(NULL, inode, block, 0);
1441 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1442 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1443 csize, *offs, inode->i_ino,
1449 memcpy(buf, bh->b_data + boffs, csize);
1452 memset(buf, 0, csize);
1462 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1463 struct lu_buf *buf, loff_t *pos)
1465 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1468 /* Read small symlink from inode body as we need to maintain correct
1469 * on-disk symlinks for ldiskfs.
1471 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1472 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1473 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1475 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1480 static inline int osd_extents_enabled(struct super_block *sb,
1481 struct inode *inode)
1483 if (inode != NULL) {
1484 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1486 } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1487 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1493 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1494 const loff_t size, const loff_t pos,
1497 int credits, bits, bs, i;
1499 bits = sb->s_blocksize_bits;
1502 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1503 * we do not expect blockmaps on the large files,
1504 * so let's shrink it to 2 levels (4GB files) */
1506 /* this is default reservation: 2 levels */
1507 credits = (blocks + 2) * 3;
1509 /* actual offset is unknown, hard to optimize */
1513 /* now check for few specific cases to optimize */
1514 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1517 /* allocate if not allocated */
1518 if (inode == NULL) {
1519 credits += blocks * 2;
1522 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1523 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1524 if (LDISKFS_I(inode)->i_data[i] == 0)
1527 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1528 /* single indirect */
1529 credits = blocks * 3;
1530 if (inode == NULL ||
1531 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1534 /* The indirect block may be modified. */
1541 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1542 const struct lu_buf *buf, loff_t _pos,
1543 struct thandle *handle)
1545 struct osd_object *obj = osd_dt_obj(dt);
1546 struct inode *inode = obj->oo_inode;
1547 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1548 struct osd_thandle *oh;
1549 int rc = 0, est = 0, credits, blocks, allocated = 0;
1555 LASSERT(buf != NULL);
1556 LASSERT(handle != NULL);
1558 oh = container_of0(handle, struct osd_thandle, ot_super);
1559 LASSERT(oh->ot_handle == NULL);
1562 bits = sb->s_blocksize_bits;
1566 /* if this is an append, then we
1567 * should expect cross-block record */
1573 /* blocks to modify */
1574 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1575 LASSERT(blocks > 0);
1577 if (inode != NULL && _pos != -1) {
1578 /* object size in blocks */
1579 est = (i_size_read(inode) + bs - 1) >> bits;
1580 allocated = inode->i_blocks >> (bits - 9);
1581 if (pos + size <= i_size_read(inode) && est <= allocated) {
1582 /* looks like an overwrite, no need to modify tree */
1584 /* no need to modify i_size */
1589 if (osd_extents_enabled(sb, inode)) {
1591 * many concurrent threads may grow tree by the time
1592 * our transaction starts. so, consider 2 is a min depth
1593 * for every level we may need to allocate a new block
1594 * and take some entries from the old one. so, 3 blocks
1595 * to allocate (bitmap, gd, itself) + old block - 4 per
1598 depth = inode != NULL ? ext_depth(inode) : 0;
1599 depth = max(depth, 1) + 1;
1601 /* if not append, then split may need to modify
1602 * existing blocks moving entries into the new ones */
1605 /* blocks to store data: bitmap,gd,itself */
1606 credits += blocks * 3;
1608 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1610 /* if inode is created as part of the transaction,
1611 * then it's counted already by the creation method */
1617 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1619 /* dt_declare_write() is usually called for system objects, such
1620 * as llog or last_rcvd files. We needn't enforce quota on those
1621 * objects, so always set the lqi_space as 0. */
1623 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1624 i_gid_read(inode), 0, oh, obj, true,
1629 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1631 /* LU-2634: clear the extent format for fast symlink */
1632 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1634 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1635 LDISKFS_I(inode)->i_disksize = buflen;
1636 i_size_write(inode, buflen);
1637 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1642 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1643 int write_NUL, loff_t *offs, handle_t *handle)
1645 struct buffer_head *bh = NULL;
1646 loff_t offset = *offs;
1647 loff_t new_size = i_size_read(inode);
1648 unsigned long block;
1649 int blocksize = 1 << inode->i_blkbits;
1653 int dirty_inode = 0;
1657 * long symlink write does not count the NUL terminator in
1658 * bufsize, we write it, and the inode's file size does not
1659 * count the NUL terminator as well.
1661 ((char *)buf)[bufsize] = '\0';
1664 while (bufsize > 0) {
1668 block = offset >> inode->i_blkbits;
1669 boffs = offset & (blocksize - 1);
1670 size = min(blocksize - boffs, bufsize);
1671 bh = __ldiskfs_bread(handle, inode, block, 1);
1672 if (IS_ERR_OR_NULL(bh)) {
1679 CERROR("%s: error reading offset %llu (block %lu): "
1681 inode->i_sb->s_id, offset, block, err);
1685 err = ldiskfs_journal_get_write_access(handle, bh);
1687 CERROR("journal_get_write_access() returned error %d\n",
1691 LASSERTF(boffs + size <= bh->b_size,
1692 "boffs %d size %d bh->b_size %lu\n",
1693 boffs, size, (unsigned long)bh->b_size);
1694 memcpy(bh->b_data + boffs, buf, size);
1695 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1699 if (offset + size > new_size)
1700 new_size = offset + size;
1710 /* correct in-core and on-disk sizes */
1711 if (new_size > i_size_read(inode)) {
1712 spin_lock(&inode->i_lock);
1713 if (new_size > i_size_read(inode))
1714 i_size_write(inode, new_size);
1715 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1716 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1719 spin_unlock(&inode->i_lock);
1721 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1729 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1730 const struct lu_buf *buf, loff_t *pos,
1731 struct thandle *handle, int ignore_quota)
1733 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1734 struct osd_thandle *oh;
1738 LASSERT(dt_object_exists(dt));
1740 LASSERT(handle != NULL);
1741 LASSERT(inode != NULL);
1742 ll_vfs_dq_init(inode);
1744 /* XXX: don't check: one declared chunk can be used many times */
1745 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1747 oh = container_of(handle, struct osd_thandle, ot_super);
1748 LASSERT(oh->ot_handle->h_transaction != NULL);
1749 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1751 /* Write small symlink to inode body as we need to maintain correct
1752 * on-disk symlinks for ldiskfs.
1753 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1754 * does not count it in.
1756 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1757 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1758 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1760 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1761 buf->lb_len, is_link, pos,
1764 result = buf->lb_len;
1766 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1771 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1772 __u64 start, __u64 end, struct thandle *th)
1774 struct osd_thandle *oh;
1775 struct inode *inode;
1780 oh = container_of(th, struct osd_thandle, ot_super);
1783 * we don't need to reserve credits for whole truncate
1784 * it's not possible as truncate may need to free too many
1785 * blocks and that won't fit a single transaction. instead
1786 * we reserve credits to change i_size and put inode onto
1787 * orphan list. if needed truncate will extend or restart
1790 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1791 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1793 inode = osd_dt_obj(dt)->oo_inode;
1796 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1797 0, oh, osd_dt_obj(dt), true, NULL, false);
1801 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1802 __u64 start, __u64 end, struct thandle *th)
1804 struct osd_thandle *oh;
1805 struct osd_object *obj = osd_dt_obj(dt);
1806 struct inode *inode = obj->oo_inode;
1809 int rc = 0, rc2 = 0;
1812 LASSERT(end == OBD_OBJECT_EOF);
1813 LASSERT(dt_object_exists(dt));
1814 LASSERT(osd_invariant(obj));
1815 LASSERT(inode != NULL);
1816 ll_vfs_dq_init(inode);
1819 oh = container_of(th, struct osd_thandle, ot_super);
1820 LASSERT(oh->ot_handle->h_transaction != NULL);
1822 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1824 tid = oh->ot_handle->h_transaction->t_tid;
1826 i_size_write(inode, start);
1827 ll_truncate_pagecache(inode, start);
1828 #ifdef HAVE_INODEOPS_TRUNCATE
1829 if (inode->i_op->truncate) {
1830 inode->i_op->truncate(inode);
1833 ldiskfs_truncate(inode);
1836 * For a partial-page truncate, flush the page to disk immediately to
1837 * avoid data corruption during direct disk write. b=17397
1839 if ((start & ~PAGE_MASK) != 0)
1840 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1842 h = journal_current_handle();
1844 LASSERT(h == oh->ot_handle);
1846 /* do not check credits with osd_trans_exec_check() as the truncate
1847 * can restart the transaction internally and we restart the
1848 * transaction in this case */
1850 if (tid != h->h_transaction->t_tid) {
1851 int credits = oh->ot_credits;
1853 * transaction has changed during truncate
1854 * we need to restart the handle with our credits
1856 if (h->h_buffer_credits < credits) {
1857 if (ldiskfs_journal_extend(h, credits))
1858 rc2 = ldiskfs_journal_restart(h, credits);
1862 RETURN(rc == 0 ? rc2 : rc);
1865 static int fiemap_check_ranges(struct inode *inode,
1866 u64 start, u64 len, u64 *new_len)
1875 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1876 maxbytes = inode->i_sb->s_maxbytes;
1878 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1880 if (start > maxbytes)
1884 * Shrink request scope to what the fs can actually handle.
1886 if (len > maxbytes || (maxbytes - len) < start)
1887 *new_len = maxbytes - start;
1892 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1893 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
1895 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1898 struct fiemap_extent_info fieinfo = {0, };
1899 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1905 if (inode->i_op->fiemap == NULL)
1908 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1911 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1915 fieinfo.fi_flags = fm->fm_flags;
1916 fieinfo.fi_extents_max = fm->fm_extent_count;
1917 fieinfo.fi_extents_start = fm->fm_extents;
1919 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1920 filemap_write_and_wait(inode->i_mapping);
1922 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1923 fm->fm_flags = fieinfo.fi_flags;
1924 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1929 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1930 __u64 start, __u64 end, enum lu_ladvise_type advice)
1945 * in some cases we may need declare methods for objects being created
1946 * e.g., when we create symlink
1948 const struct dt_body_operations osd_body_ops_new = {
1949 .dbo_declare_write = osd_declare_write,
1952 const struct dt_body_operations osd_body_ops = {
1953 .dbo_read = osd_read,
1954 .dbo_declare_write = osd_declare_write,
1955 .dbo_write = osd_write,
1956 .dbo_bufs_get = osd_bufs_get,
1957 .dbo_bufs_put = osd_bufs_put,
1958 .dbo_write_prep = osd_write_prep,
1959 .dbo_declare_write_commit = osd_declare_write_commit,
1960 .dbo_write_commit = osd_write_commit,
1961 .dbo_read_prep = osd_read_prep,
1962 .dbo_declare_punch = osd_declare_punch,
1963 .dbo_punch = osd_punch,
1964 .dbo_fiemap_get = osd_fiemap_get,
1965 .dbo_ladvise = osd_ladvise,