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, 2016, 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 */
47 * struct OBD_{ALLOC,FREE}*()
50 #include <obd_support.h>
52 #include "osd_internal.h"
55 #include <ldiskfs/ldiskfs_extents.h>
57 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
58 int rw, int line, int pages)
62 LASSERTF(iobuf->dr_elapsed_valid == 0,
63 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
64 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
66 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
68 init_waitqueue_head(&iobuf->dr_wait);
69 atomic_set(&iobuf->dr_numreqs, 0);
74 iobuf->dr_elapsed = 0;
75 /* must be counted before, so assert */
77 iobuf->dr_init_at = line;
79 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
80 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
81 LASSERT(iobuf->dr_pg_buf.lb_len >=
82 pages * sizeof(iobuf->dr_pages[0]));
86 /* start with 1MB for 4K blocks */
88 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
91 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
92 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
94 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
95 iobuf->dr_max_pages = 0;
96 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
97 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
99 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
100 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
101 if (unlikely(iobuf->dr_blocks == NULL))
104 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
105 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
106 if (unlikely(iobuf->dr_pages == NULL))
109 iobuf->dr_max_pages = pages;
113 #define osd_init_iobuf(dev, iobuf, rw, pages) \
114 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
116 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
118 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
119 iobuf->dr_pages[iobuf->dr_npages++] = page;
122 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
124 int rw = iobuf->dr_rw;
126 if (iobuf->dr_elapsed_valid) {
127 iobuf->dr_elapsed_valid = 0;
128 LASSERT(iobuf->dr_dev == d);
129 LASSERT(iobuf->dr_frags > 0);
130 lprocfs_oh_tally(&d->od_brw_stats.
131 hist[BRW_R_DIO_FRAGS+rw],
133 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
138 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
139 static void dio_complete_routine(struct bio *bio)
141 int error = bio->bi_error;
143 static void dio_complete_routine(struct bio *bio, int error)
146 struct osd_iobuf *iobuf = bio->bi_private;
150 /* CAVEAT EMPTOR: possibly in IRQ context
151 * DO NOT record procfs stats here!!! */
153 if (unlikely(iobuf == NULL)) {
154 CERROR("***** bio->bi_private is NULL! This should never "
155 "happen. Normally, I would crash here, but instead I "
156 "will dump the bio contents to the console. Please "
157 "report this to <https://jira.hpdd.intel.com/> , along "
158 "with any interesting messages leading up to this point "
159 "(like SCSI errors, perhaps). Because bi_private is "
160 "NULL, I can't wake up the thread that initiated this "
161 "IO - you will probably have to reboot this node.\n");
162 CERROR("bi_next: %p, bi_flags: %lx, "
168 "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
169 "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
170 (unsigned long)bio->bi_flags,
176 bio->bi_vcnt, bio_idx(bio),
177 bio_sectors(bio) << 9, bio->bi_end_io,
179 atomic_read(&bio->bi_cnt),
181 atomic_read(&bio->__bi_cnt),
187 /* the check is outside of the cycle for performance reason -bzzz */
188 if (!bio_data_dir(bio)) {
189 bio_for_each_segment_all(bvl, bio, iter) {
190 if (likely(error == 0))
191 SetPageUptodate(bvl_to_page(bvl));
192 LASSERT(PageLocked(bvl_to_page(bvl)));
194 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
196 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
199 /* any real error is good enough -bzzz */
200 if (error != 0 && iobuf->dr_error == 0)
201 iobuf->dr_error = error;
204 * set dr_elapsed before dr_numreqs turns to 0, otherwise
205 * it's possible that service thread will see dr_numreqs
206 * is zero, but dr_elapsed is not set yet, leading to lost
207 * data in this processing and an assertion in a subsequent
210 if (atomic_read(&iobuf->dr_numreqs) == 1) {
211 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
212 iobuf->dr_elapsed_valid = 1;
214 if (atomic_dec_and_test(&iobuf->dr_numreqs))
215 wake_up(&iobuf->dr_wait);
217 /* Completed bios used to be chained off iobuf->dr_bios and freed in
218 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
219 * mempool when serious on-disk fragmentation was encountered,
220 * deadlocking the OST. The bios are now released as soon as complete
221 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
225 static void record_start_io(struct osd_iobuf *iobuf, int size)
227 struct osd_device *osd = iobuf->dr_dev;
228 struct obd_histogram *h = osd->od_brw_stats.hist;
231 atomic_inc(&iobuf->dr_numreqs);
233 if (iobuf->dr_rw == 0) {
234 atomic_inc(&osd->od_r_in_flight);
235 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
236 atomic_read(&osd->od_r_in_flight));
237 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
238 } else if (iobuf->dr_rw == 1) {
239 atomic_inc(&osd->od_w_in_flight);
240 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
241 atomic_read(&osd->od_w_in_flight));
242 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
248 static void osd_submit_bio(int rw, struct bio *bio)
250 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
251 #ifdef HAVE_SUBMIT_BIO_2ARGS
253 submit_bio(READ, bio);
255 submit_bio(WRITE, bio);
262 static int can_be_merged(struct bio *bio, sector_t sector)
267 return bio_end_sector(bio) == sector ? 1 : 0;
270 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
271 struct osd_iobuf *iobuf)
273 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
274 struct page **pages = iobuf->dr_pages;
275 int npages = iobuf->dr_npages;
276 sector_t *blocks = iobuf->dr_blocks;
277 int total_blocks = npages * blocks_per_page;
278 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
279 unsigned int blocksize = inode->i_sb->s_blocksize;
280 struct bio *bio = NULL;
282 unsigned int page_offset;
292 LASSERT(iobuf->dr_npages == npages);
294 osd_brw_stats_update(osd, iobuf);
295 iobuf->dr_start_time = cfs_time_current();
297 blk_start_plug(&plug);
298 for (page_idx = 0, block_idx = 0;
300 page_idx++, block_idx += blocks_per_page) {
302 page = pages[page_idx];
303 LASSERT(block_idx + blocks_per_page <= total_blocks);
305 for (i = 0, page_offset = 0;
307 i += nblocks, page_offset += blocksize * nblocks) {
311 if (blocks[block_idx + i] == 0) { /* hole */
312 LASSERTF(iobuf->dr_rw == 0,
313 "page_idx %u, block_idx %u, i %u\n",
314 page_idx, block_idx, i);
315 memset(kmap(page) + page_offset, 0, blocksize);
320 sector = (sector_t)blocks[block_idx + i] << sector_bits;
322 /* Additional contiguous file blocks? */
323 while (i + nblocks < blocks_per_page &&
324 (sector + (nblocks << sector_bits)) ==
325 ((sector_t)blocks[block_idx + i + nblocks] <<
330 can_be_merged(bio, sector) &&
331 bio_add_page(bio, page,
332 blocksize * nblocks, page_offset) != 0)
333 continue; /* added this frag OK */
336 struct request_queue *q =
337 bdev_get_queue(bio->bi_bdev);
338 unsigned int bi_size = bio_sectors(bio) << 9;
340 /* Dang! I have to fragment this I/O */
341 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
342 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
343 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
345 queue_max_sectors(q),
346 bio_phys_segments(q, bio),
347 queue_max_phys_segments(q),
348 0, queue_max_hw_segments(q));
349 record_start_io(iobuf, bi_size);
350 osd_submit_bio(iobuf->dr_rw, bio);
353 /* allocate new bio */
354 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
355 (npages - page_idx) *
358 CERROR("Can't allocate bio %u*%u = %u pages\n",
359 (npages - page_idx), blocks_per_page,
360 (npages - page_idx) * blocks_per_page);
365 bio->bi_bdev = inode->i_sb->s_bdev;
366 bio_set_sector(bio, sector);
368 bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
370 bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
372 bio->bi_end_io = dio_complete_routine;
373 bio->bi_private = iobuf;
375 rc = bio_add_page(bio, page,
376 blocksize * nblocks, page_offset);
382 record_start_io(iobuf, bio_sectors(bio) << 9);
383 osd_submit_bio(iobuf->dr_rw, bio);
388 blk_finish_plug(&plug);
390 /* in order to achieve better IO throughput, we don't wait for writes
391 * completion here. instead we proceed with transaction commit in
392 * parallel and wait for IO completion once transaction is stopped
393 * see osd_trans_stop() for more details -bzzz */
394 if (iobuf->dr_rw == 0) {
395 wait_event(iobuf->dr_wait,
396 atomic_read(&iobuf->dr_numreqs) == 0);
397 osd_fini_iobuf(osd, iobuf);
401 rc = iobuf->dr_error;
405 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
406 struct niobuf_local *lnb)
413 int poff = offset & (PAGE_SIZE - 1);
414 int plen = PAGE_SIZE - poff;
418 lnb->lnb_file_offset = offset;
419 lnb->lnb_page_offset = poff;
421 /* lnb->lnb_flags = rnb->rnb_flags; */
423 lnb->lnb_page = NULL;
426 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
437 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
440 struct inode *inode = osd_dt_obj(dt)->oo_inode;
441 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
446 page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
449 if (unlikely(page == NULL))
450 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
456 * there are following "locks":
467 * - lock pages, unlock
469 * - lock partial page
475 * Unlock and release pages loaded by osd_bufs_get()
477 * Unlock \a npages pages from \a lnb and drop the refcount on them.
479 * \param env thread execution environment
480 * \param dt dt object undergoing IO (OSD object + methods)
481 * \param lnb array of pages undergoing IO
482 * \param npages number of pages in \a lnb
486 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
487 struct niobuf_local *lnb, int npages)
491 for (i = 0; i < npages; i++) {
492 if (lnb[i].lnb_page == NULL)
494 LASSERT(PageLocked(lnb[i].lnb_page));
495 unlock_page(lnb[i].lnb_page);
496 put_page(lnb[i].lnb_page);
497 dt_object_put(env, dt);
498 lnb[i].lnb_page = NULL;
505 * Load and lock pages undergoing IO
507 * Pages as described in the \a lnb array are fetched (from disk or cache)
508 * and locked for IO by the caller.
510 * DLM locking protects us from write and truncate competing for same region,
511 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
512 * It's possible the writeout on a such a page is in progress when we access
513 * it. It's also possible that during this writeout we put new (partial) data
514 * into the page, but won't be able to proceed in filter_commitrw_write().
515 * Therefore, just wait for writeout completion as it should be rare enough.
517 * \param env thread execution environment
518 * \param dt dt object undergoing IO (OSD object + methods)
519 * \param pos byte offset of IO start
520 * \param len number of bytes of IO
521 * \param lnb array of extents undergoing IO
522 * \param rw read or write operation, and other flags
523 * \param capa capabilities
525 * \retval pages (zero or more) loaded successfully
526 * \retval -ENOMEM on memory/page allocation error
528 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
529 loff_t pos, ssize_t len, struct niobuf_local *lnb,
530 enum dt_bufs_type rw)
532 struct osd_object *obj = osd_dt_obj(dt);
533 int npages, i, rc = 0;
536 LASSERT(obj->oo_inode);
538 osd_map_remote_to_local(pos, len, &npages, lnb);
540 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
541 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
543 for (i = 0; i < npages; i++, lnb++) {
544 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
546 if (lnb->lnb_page == NULL)
547 GOTO(cleanup, rc = -ENOMEM);
549 wait_on_page_writeback(lnb->lnb_page);
550 BUG_ON(PageWriteback(lnb->lnb_page));
552 lu_object_get(&dt->do_lu);
559 osd_bufs_put(env, dt, lnb - i, i);
563 #ifndef HAVE_LDISKFS_MAP_BLOCKS
565 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
566 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
577 static long ldiskfs_ext_find_goal(struct inode *inode,
578 struct ldiskfs_ext_path *path,
579 unsigned long block, int *aflags)
581 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
582 unsigned long bg_start;
583 unsigned long colour;
587 struct ldiskfs_extent *ex;
588 depth = path->p_depth;
590 /* try to predict block placement */
591 if ((ex = path[depth].p_ext))
592 return ldiskfs_ext_pblock(ex) +
593 (block - le32_to_cpu(ex->ee_block));
595 /* it looks index is empty
596 * try to find starting from index itself */
597 if (path[depth].p_bh)
598 return path[depth].p_bh->b_blocknr;
601 /* OK. use inode's group */
602 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
603 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
604 colour = (current->pid % 16) *
605 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
606 return bg_start + colour + block;
609 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
610 struct ldiskfs_ext_path *path,
611 unsigned long block, unsigned long *count,
614 struct ldiskfs_allocation_request ar;
615 unsigned long pblock;
618 /* find neighbour allocated blocks */
620 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
624 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
628 /* allocate new block */
629 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
633 ar.flags = LDISKFS_MB_HINT_DATA;
634 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
639 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
640 struct ldiskfs_ext_path *path,
641 struct ldiskfs_ext_cache *cex,
642 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
643 struct ldiskfs_extent *ex,
647 struct bpointers *bp = cbdata;
648 struct ldiskfs_extent nex;
649 unsigned long pblock = 0;
655 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
656 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
658 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
664 if (bp->create == 0) {
666 if (cex->ec_block < bp->start)
667 i = bp->start - cex->ec_block;
668 if (i >= cex->ec_len)
669 CERROR("nothing to do?! i = %d, e_num = %u\n",
671 for (; i < cex->ec_len && bp->num; i++) {
681 tgen = LDISKFS_I(inode)->i_ext_generation;
682 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
684 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
685 count + LDISKFS_ALLOC_NEEDED + 1);
686 if (IS_ERR(handle)) {
687 return PTR_ERR(handle);
690 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
691 /* the tree has changed. so path can be invalid at moment */
692 ldiskfs_journal_stop(handle);
696 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
697 * protected by i_data_sem as whole. so we patch it to store
698 * generation to path and now verify the tree hasn't changed */
699 down_write((&LDISKFS_I(inode)->i_data_sem));
701 /* validate extent, make sure the extent tree does not changed */
702 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
703 /* cex is invalid, try again */
704 up_write(&LDISKFS_I(inode)->i_data_sem);
705 ldiskfs_journal_stop(handle);
710 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
713 BUG_ON(count > cex->ec_len);
715 /* insert new extent */
716 nex.ee_block = cpu_to_le32(cex->ec_block);
717 ldiskfs_ext_store_pblock(&nex, pblock);
718 nex.ee_len = cpu_to_le16(count);
719 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
721 /* free data blocks we just allocated */
722 /* not a good idea to call discard here directly,
723 * but otherwise we'd need to call it every free() */
724 ldiskfs_discard_preallocations(inode);
725 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
726 ldiskfs_free_blocks(handle, inode, NULL,
727 ldiskfs_ext_pblock(&nex),
728 le16_to_cpu(nex.ee_len), 0);
730 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
731 le16_to_cpu(nex.ee_len), 0);
737 * Putting len of the actual extent we just inserted,
738 * we are asking ldiskfs_ext_walk_space() to continue
739 * scaning after that block
741 cex->ec_len = le16_to_cpu(nex.ee_len);
742 cex->ec_start = ldiskfs_ext_pblock(&nex);
743 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
744 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
747 up_write((&LDISKFS_I(inode)->i_data_sem));
748 ldiskfs_journal_stop(handle);
753 CERROR("hmm. why do we find this extent?\n");
754 CERROR("initial space: %lu:%u\n",
755 bp->start, bp->init_num);
756 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
757 CERROR("current extent: %u/%u/%llu %d\n",
758 cex->ec_block, cex->ec_len,
759 (unsigned long long)cex->ec_start,
762 CERROR("current extent: %u/%u/%llu\n",
763 cex->ec_block, cex->ec_len,
764 (unsigned long long)cex->ec_start);
768 if (cex->ec_block < bp->start)
769 i = bp->start - cex->ec_block;
770 if (i >= cex->ec_len)
771 CERROR("nothing to do?! i = %d, e_num = %u\n",
773 for (; i < cex->ec_len && bp->num; i++) {
774 *(bp->blocks) = cex->ec_start + i;
776 /* unmap any possible underlying metadata from
777 * the block device mapping. bug 6998. */
778 #ifndef HAVE_CLEAN_BDEV_ALIASES
779 unmap_underlying_metadata(inode->i_sb->s_bdev,
782 clean_bdev_aliases(inode->i_sb->s_bdev,
794 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
795 int clen, sector_t *blocks, int create)
797 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
801 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
805 bp.start = index * blocks_per_page;
806 bp.init_num = bp.num = clen * blocks_per_page;
809 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
810 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
812 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
813 ldiskfs_ext_new_extent_cb, &bp);
814 ldiskfs_ext_invalidate_cache(inode);
819 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
820 struct page **page, int pages,
821 sector_t *blocks, int create)
823 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
824 pgoff_t bitmap_max_page_index;
828 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
830 for (i = 0, b = blocks; i < pages; i++, page++) {
831 if ((*page)->index + 1 >= bitmap_max_page_index) {
835 rc = ldiskfs_map_inode_page(inode, *page, b, create);
837 CERROR("ino %lu, blk %llu create %d: rc %d\n",
839 (unsigned long long)*b, create, rc);
842 b += blocks_per_page;
847 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
849 int pages, sector_t *blocks,
852 int rc = 0, i = 0, clen = 0;
853 struct page *fp = NULL;
855 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
856 inode->i_ino, pages, (*page)->index);
858 /* pages are sorted already. so, we just have to find
859 * contig. space and process them properly */
862 /* start new extent */
867 } else if (fp->index + clen == (*page)->index) {
868 /* continue the extent */
875 /* process found extent */
876 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
881 /* look for next extent */
883 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
887 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
894 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
895 int pages, sector_t *blocks,
900 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
901 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
905 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
910 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
911 int pages, sector_t *blocks,
914 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
916 struct page *fp = NULL;
918 pgoff_t max_page_index;
919 handle_t *handle = NULL;
921 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
923 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
924 inode->i_ino, pages, (*page)->index);
927 create = LDISKFS_GET_BLOCKS_CREATE;
928 handle = ldiskfs_journal_current_handle();
929 LASSERT(handle != NULL);
930 rc = osd_attach_jinode(inode);
934 /* pages are sorted already. so, we just have to find
935 * contig. space and process them properly */
937 long blen, total = 0;
938 struct ldiskfs_map_blocks map = { 0 };
940 if (fp == NULL) { /* start new extent */
945 } else if (fp->index + clen == (*page)->index) {
946 /* continue the extent */
952 if (fp->index + clen >= max_page_index)
953 GOTO(cleanup, rc = -EFBIG);
954 /* process found extent */
955 map.m_lblk = fp->index * blocks_per_page;
956 map.m_len = blen = clen * blocks_per_page;
958 rc = ldiskfs_map_blocks(handle, inode, &map, create);
961 for (; total < blen && c < map.m_len; c++, total++) {
963 *(blocks + total) = 0;
967 *(blocks + total) = map.m_pblk + c;
968 /* unmap any possible underlying
969 * metadata from the block device
970 * mapping. bug 6998. */
971 if ((map.m_flags & LDISKFS_MAP_NEW) &&
973 #ifndef HAVE_CLEAN_BDEV_ALIASES
974 unmap_underlying_metadata(
986 if (rc == 0 && total < blen) {
987 map.m_lblk = fp->index * blocks_per_page + total;
988 map.m_len = blen - total;
994 /* look for next extent */
996 blocks += blocks_per_page * clen;
1001 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1003 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1004 struct niobuf_local *lnb, int npages)
1006 struct osd_thread_info *oti = osd_oti_get(env);
1007 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1008 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1009 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1021 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1022 if (unlikely(rc != 0))
1025 isize = i_size_read(inode);
1026 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1028 if (osd->od_writethrough_cache)
1030 if (isize > osd->od_readcache_max_filesize)
1033 start = ktime_get();
1034 for (i = 0; i < npages; i++) {
1037 generic_error_remove_page(inode->i_mapping,
1041 * till commit the content of the page is undefined
1042 * we'll set it uptodate once bulk is done. otherwise
1043 * subsequent reads can access non-stable data
1045 ClearPageUptodate(lnb[i].lnb_page);
1047 if (lnb[i].lnb_len == PAGE_SIZE)
1050 if (maxidx >= lnb[i].lnb_page->index) {
1051 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1054 char *p = kmap(lnb[i].lnb_page);
1056 off = lnb[i].lnb_page_offset;
1059 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1062 memset(p + off, 0, PAGE_SIZE - off);
1063 kunmap(lnb[i].lnb_page);
1067 timediff = ktime_us_delta(end, start);
1068 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1070 if (iobuf->dr_npages) {
1071 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1073 iobuf->dr_blocks, 0);
1074 if (likely(rc == 0)) {
1075 rc = osd_do_bio(osd, inode, iobuf);
1076 /* do IO stats for preparation reads */
1077 osd_fini_iobuf(osd, iobuf);
1083 struct osd_fextent {
1086 unsigned int mapped:1;
1089 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1090 struct osd_fextent *cached_extent)
1092 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1093 sector_t block = offset >> inode->i_blkbits;
1095 struct fiemap_extent_info fei = { 0 };
1096 struct fiemap_extent fe = { 0 };
1097 mm_segment_t saved_fs;
1100 if (block >= cached_extent->start && block < cached_extent->end)
1101 return cached_extent->mapped;
1103 if (i_size_read(inode) == 0)
1106 /* Beyond EOF, must not be mapped */
1107 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1110 fei.fi_extents_max = 1;
1111 fei.fi_extents_start = &fe;
1113 saved_fs = get_fs();
1115 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1120 start = fe.fe_logical >> inode->i_blkbits;
1122 if (start > block) {
1123 cached_extent->start = block;
1124 cached_extent->end = start;
1125 cached_extent->mapped = 0;
1127 cached_extent->start = start;
1128 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1130 cached_extent->mapped = 1;
1133 return cached_extent->mapped;
1136 static int osd_declare_write_commit(const struct lu_env *env,
1137 struct dt_object *dt,
1138 struct niobuf_local *lnb, int npages,
1139 struct thandle *handle)
1141 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1142 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1143 struct osd_thandle *oh;
1151 long long quota_space = 0;
1152 struct osd_fextent extent = { 0 };
1153 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1156 LASSERT(handle != NULL);
1157 oh = container_of0(handle, struct osd_thandle, ot_super);
1158 LASSERT(oh->ot_handle == NULL);
1162 /* calculate number of extents (probably better to pass nb) */
1163 for (i = 0; i < npages; i++) {
1164 if (i && lnb[i].lnb_file_offset !=
1165 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1168 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1169 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1171 quota_space += PAGE_SIZE;
1173 /* ignore quota for the whole request if any page is from
1174 * client cache or written by root.
1176 * XXX once we drop the 1.8 client support, the checking
1177 * for whether page is from cache can be simplified as:
1178 * !(lnb[i].flags & OBD_BRW_SYNC)
1180 * XXX we could handle this on per-lnb basis as done by
1182 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1183 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1185 declare_flags |= OSD_QID_FORCE;
1189 * each extent can go into new leaf causing a split
1190 * 5 is max tree depth: inode + 4 index blocks
1191 * with blockmaps, depth is 3 at most
1193 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1195 * many concurrent threads may grow tree by the time
1196 * our transaction starts. so, consider 2 is a min depth
1198 depth = ext_depth(inode);
1199 depth = max(depth, 1) + 1;
1201 credits++; /* inode */
1202 credits += depth * 2 * extents;
1206 credits++; /* inode */
1207 credits += depth * extents;
1210 /* quota space for metadata blocks */
1211 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1213 /* quota space should be reported in 1K blocks */
1214 quota_space = toqb(quota_space);
1216 /* each new block can go in different group (bitmap + gd) */
1218 /* we can't dirty more bitmap blocks than exist */
1219 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1220 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1222 credits += newblocks;
1224 /* we can't dirty more gd blocks than exist */
1225 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1226 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1228 credits += newblocks;
1230 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1232 /* make sure the over quota flags were not set */
1233 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1235 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1236 i_projid_read(inode), quota_space, oh,
1237 osd_dt_obj(dt), &flags, declare_flags);
1239 /* we need only to store the overquota flags in the first lnb for
1240 * now, once we support multiple objects BRW, this code needs be
1242 if (flags & QUOTA_FL_OVER_USRQUOTA)
1243 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1244 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1245 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1246 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1247 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1252 /* Check if a block is allocated or not */
1253 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1254 struct niobuf_local *lnb, int npages,
1255 struct thandle *thandle)
1257 struct osd_thread_info *oti = osd_oti_get(env);
1258 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1259 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1260 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1266 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1267 if (unlikely(rc != 0))
1270 isize = i_size_read(inode);
1271 ll_vfs_dq_init(inode);
1273 for (i = 0; i < npages; i++) {
1274 if (lnb[i].lnb_rc == -ENOSPC &&
1275 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1276 /* Allow the write to proceed if overwriting an
1281 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1282 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1284 LASSERT(lnb[i].lnb_page);
1285 generic_error_remove_page(inode->i_mapping,
1290 LASSERT(PageLocked(lnb[i].lnb_page));
1291 LASSERT(!PageWriteback(lnb[i].lnb_page));
1293 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1294 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1297 * Since write and truncate are serialized by oo_sem, even
1298 * partial-page truncate should not leave dirty pages in the
1301 LASSERT(!PageDirty(lnb[i].lnb_page));
1303 SetPageUptodate(lnb[i].lnb_page);
1305 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1308 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1310 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1312 } else if (iobuf->dr_npages > 0) {
1313 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1315 iobuf->dr_blocks, 1);
1317 /* no pages to write, no transno is needed */
1318 thandle->th_local = 1;
1321 if (likely(rc == 0)) {
1322 spin_lock(&inode->i_lock);
1323 if (isize > i_size_read(inode)) {
1324 i_size_write(inode, isize);
1325 LDISKFS_I(inode)->i_disksize = isize;
1326 spin_unlock(&inode->i_lock);
1327 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1329 spin_unlock(&inode->i_lock);
1332 rc = osd_do_bio(osd, inode, iobuf);
1333 /* we don't do stats here as in read path because
1334 * write is async: we'll do this in osd_put_bufs() */
1336 osd_fini_iobuf(osd, iobuf);
1339 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1341 if (unlikely(rc != 0)) {
1342 /* if write fails, we should drop pages from the cache */
1343 for (i = 0; i < npages; i++) {
1344 if (lnb[i].lnb_page == NULL)
1346 LASSERT(PageLocked(lnb[i].lnb_page));
1347 generic_error_remove_page(inode->i_mapping,
1355 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1356 struct niobuf_local *lnb, int npages)
1358 struct osd_thread_info *oti = osd_oti_get(env);
1359 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1360 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1361 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1362 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1369 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1370 if (unlikely(rc != 0))
1373 isize = i_size_read(inode);
1375 if (osd->od_read_cache)
1377 if (isize > osd->od_readcache_max_filesize)
1380 start = ktime_get();
1381 for (i = 0; i < npages; i++) {
1383 if (isize <= lnb[i].lnb_file_offset)
1384 /* If there's no more data, abort early.
1385 * lnb->lnb_rc == 0, so it's easy to detect later. */
1388 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1389 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1391 lnb[i].lnb_rc = lnb[i].lnb_len;
1393 /* Bypass disk read if fail_loc is set properly */
1394 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1395 SetPageUptodate(lnb[i].lnb_page);
1397 if (PageUptodate(lnb[i].lnb_page)) {
1401 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1405 generic_error_remove_page(inode->i_mapping,
1409 timediff = ktime_us_delta(end, start);
1410 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1412 if (cache_hits != 0)
1413 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1415 if (cache_misses != 0)
1416 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1418 if (cache_hits + cache_misses != 0)
1419 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1420 cache_hits + cache_misses);
1422 if (iobuf->dr_npages) {
1423 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1425 iobuf->dr_blocks, 0);
1426 rc = osd_do_bio(osd, inode, iobuf);
1428 /* IO stats will be done in osd_bufs_put() */
1435 * XXX: Another layering violation for now.
1437 * We don't want to use ->f_op->read methods, because generic file write
1439 * - serializes on ->i_sem, and
1441 * - does a lot of extra work like balance_dirty_pages(),
1443 * which doesn't work for globally shared files like /last_rcvd.
1445 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1447 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1449 memcpy(buffer, (char *)ei->i_data, buflen);
1454 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1456 struct buffer_head *bh;
1457 unsigned long block;
1463 /* prevent reading after eof */
1464 spin_lock(&inode->i_lock);
1465 if (i_size_read(inode) < *offs + size) {
1466 loff_t diff = i_size_read(inode) - *offs;
1467 spin_unlock(&inode->i_lock);
1469 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1470 i_size_read(inode), *offs);
1472 } else if (diff == 0) {
1478 spin_unlock(&inode->i_lock);
1481 blocksize = 1 << inode->i_blkbits;
1484 block = *offs >> inode->i_blkbits;
1485 boffs = *offs & (blocksize - 1);
1486 csize = min(blocksize - boffs, size);
1487 bh = __ldiskfs_bread(NULL, inode, block, 0);
1489 CERROR("%s: can't read %u@%llu on ino %lu: "
1490 "rc = %ld\n", osd_ino2name(inode),
1491 csize, *offs, inode->i_ino,
1497 memcpy(buf, bh->b_data + boffs, csize);
1500 memset(buf, 0, csize);
1510 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1511 struct lu_buf *buf, loff_t *pos)
1513 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1516 /* Read small symlink from inode body as we need to maintain correct
1517 * on-disk symlinks for ldiskfs.
1519 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1520 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1521 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1523 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1528 static inline int osd_extents_enabled(struct super_block *sb,
1529 struct inode *inode)
1531 if (inode != NULL) {
1532 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1534 } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1535 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1541 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1542 const loff_t size, const loff_t pos,
1545 int credits, bits, bs, i;
1547 bits = sb->s_blocksize_bits;
1550 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1551 * we do not expect blockmaps on the large files,
1552 * so let's shrink it to 2 levels (4GB files) */
1554 /* this is default reservation: 2 levels */
1555 credits = (blocks + 2) * 3;
1557 /* actual offset is unknown, hard to optimize */
1561 /* now check for few specific cases to optimize */
1562 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1565 /* allocate if not allocated */
1566 if (inode == NULL) {
1567 credits += blocks * 2;
1570 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1571 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1572 if (LDISKFS_I(inode)->i_data[i] == 0)
1575 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1576 /* single indirect */
1577 credits = blocks * 3;
1578 if (inode == NULL ||
1579 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1582 /* The indirect block may be modified. */
1589 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1590 const struct lu_buf *buf, loff_t _pos,
1591 struct thandle *handle)
1593 struct osd_object *obj = osd_dt_obj(dt);
1594 struct inode *inode = obj->oo_inode;
1595 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1596 struct osd_thandle *oh;
1597 int rc = 0, est = 0, credits, blocks, allocated = 0;
1603 LASSERT(buf != NULL);
1604 LASSERT(handle != NULL);
1606 oh = container_of0(handle, struct osd_thandle, ot_super);
1607 LASSERT(oh->ot_handle == NULL);
1610 bits = sb->s_blocksize_bits;
1614 /* if this is an append, then we
1615 * should expect cross-block record */
1621 /* blocks to modify */
1622 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1623 LASSERT(blocks > 0);
1625 if (inode != NULL && _pos != -1) {
1626 /* object size in blocks */
1627 est = (i_size_read(inode) + bs - 1) >> bits;
1628 allocated = inode->i_blocks >> (bits - 9);
1629 if (pos + size <= i_size_read(inode) && est <= allocated) {
1630 /* looks like an overwrite, no need to modify tree */
1632 /* no need to modify i_size */
1637 if (osd_extents_enabled(sb, inode)) {
1639 * many concurrent threads may grow tree by the time
1640 * our transaction starts. so, consider 2 is a min depth
1641 * for every level we may need to allocate a new block
1642 * and take some entries from the old one. so, 3 blocks
1643 * to allocate (bitmap, gd, itself) + old block - 4 per
1646 depth = inode != NULL ? ext_depth(inode) : 0;
1647 depth = max(depth, 1) + 1;
1649 /* if not append, then split may need to modify
1650 * existing blocks moving entries into the new ones */
1653 /* blocks to store data: bitmap,gd,itself */
1654 credits += blocks * 3;
1656 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1658 /* if inode is created as part of the transaction,
1659 * then it's counted already by the creation method */
1665 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1667 /* dt_declare_write() is usually called for system objects, such
1668 * as llog or last_rcvd files. We needn't enforce quota on those
1669 * objects, so always set the lqi_space as 0. */
1671 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1673 i_projid_read(inode), 0,
1674 oh, obj, NULL, OSD_QID_BLK);
1678 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1680 /* LU-2634: clear the extent format for fast symlink */
1681 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1683 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1684 spin_lock(&inode->i_lock);
1685 LDISKFS_I(inode)->i_disksize = buflen;
1686 i_size_write(inode, buflen);
1687 spin_unlock(&inode->i_lock);
1688 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1693 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1694 int write_NUL, loff_t *offs, handle_t *handle)
1696 struct buffer_head *bh = NULL;
1697 loff_t offset = *offs;
1698 loff_t new_size = i_size_read(inode);
1699 unsigned long block;
1700 int blocksize = 1 << inode->i_blkbits;
1704 int dirty_inode = 0;
1708 * long symlink write does not count the NUL terminator in
1709 * bufsize, we write it, and the inode's file size does not
1710 * count the NUL terminator as well.
1712 ((char *)buf)[bufsize] = '\0';
1716 while (bufsize > 0) {
1717 int credits = handle->h_buffer_credits;
1722 block = offset >> inode->i_blkbits;
1723 boffs = offset & (blocksize - 1);
1724 size = min(blocksize - boffs, bufsize);
1725 bh = __ldiskfs_bread(handle, inode, block, 1);
1726 if (IS_ERR_OR_NULL(bh)) {
1734 CERROR("%s: error reading offset %llu (block %lu, "
1735 "size %d, offs %llu), credits %d/%d: rc = %d\n",
1736 inode->i_sb->s_id, offset, block, bufsize, *offs,
1737 credits, handle->h_buffer_credits, err);
1741 err = ldiskfs_journal_get_write_access(handle, bh);
1743 CERROR("journal_get_write_access() returned error %d\n",
1747 LASSERTF(boffs + size <= bh->b_size,
1748 "boffs %d size %d bh->b_size %lu\n",
1749 boffs, size, (unsigned long)bh->b_size);
1750 memcpy(bh->b_data + boffs, buf, size);
1751 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1755 if (offset + size > new_size)
1756 new_size = offset + size;
1766 /* correct in-core and on-disk sizes */
1767 if (new_size > i_size_read(inode)) {
1768 spin_lock(&inode->i_lock);
1769 if (new_size > i_size_read(inode))
1770 i_size_write(inode, new_size);
1771 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1772 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1775 spin_unlock(&inode->i_lock);
1777 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1785 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1786 const struct lu_buf *buf, loff_t *pos,
1787 struct thandle *handle, int ignore_quota)
1789 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1790 struct osd_thandle *oh;
1794 LASSERT(dt_object_exists(dt));
1796 LASSERT(handle != NULL);
1797 LASSERT(inode != NULL);
1798 ll_vfs_dq_init(inode);
1800 /* XXX: don't check: one declared chunk can be used many times */
1801 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1803 oh = container_of(handle, struct osd_thandle, ot_super);
1804 LASSERT(oh->ot_handle->h_transaction != NULL);
1805 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1807 /* Write small symlink to inode body as we need to maintain correct
1808 * on-disk symlinks for ldiskfs.
1809 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1810 * does not count it in.
1812 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1813 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1814 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1816 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1817 buf->lb_len, is_link, pos,
1820 result = buf->lb_len;
1822 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1827 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1828 __u64 start, __u64 end, struct thandle *th)
1830 struct osd_thandle *oh;
1831 struct inode *inode;
1836 oh = container_of(th, struct osd_thandle, ot_super);
1839 * we don't need to reserve credits for whole truncate
1840 * it's not possible as truncate may need to free too many
1841 * blocks and that won't fit a single transaction. instead
1842 * we reserve credits to change i_size and put inode onto
1843 * orphan list. if needed truncate will extend or restart
1846 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1847 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1849 inode = osd_dt_obj(dt)->oo_inode;
1852 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1853 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1858 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1859 __u64 start, __u64 end, struct thandle *th)
1861 struct osd_thandle *oh;
1862 struct osd_object *obj = osd_dt_obj(dt);
1863 struct inode *inode = obj->oo_inode;
1866 int rc = 0, rc2 = 0;
1869 LASSERT(end == OBD_OBJECT_EOF);
1870 LASSERT(dt_object_exists(dt));
1871 LASSERT(osd_invariant(obj));
1872 LASSERT(inode != NULL);
1873 ll_vfs_dq_init(inode);
1876 oh = container_of(th, struct osd_thandle, ot_super);
1877 LASSERT(oh->ot_handle->h_transaction != NULL);
1879 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1881 tid = oh->ot_handle->h_transaction->t_tid;
1883 spin_lock(&inode->i_lock);
1884 i_size_write(inode, start);
1885 spin_unlock(&inode->i_lock);
1886 ll_truncate_pagecache(inode, start);
1887 #ifdef HAVE_INODEOPS_TRUNCATE
1888 if (inode->i_op->truncate) {
1889 inode->i_op->truncate(inode);
1892 ldiskfs_truncate(inode);
1895 * For a partial-page truncate, flush the page to disk immediately to
1896 * avoid data corruption during direct disk write. b=17397
1898 if ((start & ~PAGE_MASK) != 0)
1899 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1901 h = journal_current_handle();
1903 LASSERT(h == oh->ot_handle);
1905 /* do not check credits with osd_trans_exec_check() as the truncate
1906 * can restart the transaction internally and we restart the
1907 * transaction in this case */
1909 if (tid != h->h_transaction->t_tid) {
1910 int credits = oh->ot_credits;
1912 * transaction has changed during truncate
1913 * we need to restart the handle with our credits
1915 if (h->h_buffer_credits < credits) {
1916 if (ldiskfs_journal_extend(h, credits))
1917 rc2 = ldiskfs_journal_restart(h, credits);
1921 RETURN(rc == 0 ? rc2 : rc);
1924 static int fiemap_check_ranges(struct inode *inode,
1925 u64 start, u64 len, u64 *new_len)
1934 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1935 maxbytes = inode->i_sb->s_maxbytes;
1937 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1939 if (start > maxbytes)
1943 * Shrink request scope to what the fs can actually handle.
1945 if (len > maxbytes || (maxbytes - len) < start)
1946 *new_len = maxbytes - start;
1951 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1952 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
1954 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1957 struct fiemap_extent_info fieinfo = {0, };
1958 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1964 if (inode->i_op->fiemap == NULL)
1967 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1970 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1974 fieinfo.fi_flags = fm->fm_flags;
1975 fieinfo.fi_extents_max = fm->fm_extent_count;
1976 fieinfo.fi_extents_start = fm->fm_extents;
1978 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1979 filemap_write_and_wait(inode->i_mapping);
1981 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1982 fm->fm_flags = fieinfo.fi_flags;
1983 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1988 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1989 __u64 start, __u64 end, enum lu_ladvise_type advice)
1992 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1996 case LU_LADVISE_DONTNEED:
1999 invalidate_mapping_pages(inode->i_mapping,
2000 start >> PAGE_CACHE_SHIFT,
2001 (end - 1) >> PAGE_CACHE_SHIFT);
2012 * in some cases we may need declare methods for objects being created
2013 * e.g., when we create symlink
2015 const struct dt_body_operations osd_body_ops_new = {
2016 .dbo_declare_write = osd_declare_write,
2019 const struct dt_body_operations osd_body_ops = {
2020 .dbo_read = osd_read,
2021 .dbo_declare_write = osd_declare_write,
2022 .dbo_write = osd_write,
2023 .dbo_bufs_get = osd_bufs_get,
2024 .dbo_bufs_put = osd_bufs_put,
2025 .dbo_write_prep = osd_write_prep,
2026 .dbo_declare_write_commit = osd_declare_write_commit,
2027 .dbo_write_commit = osd_write_commit,
2028 .dbo_read_prep = osd_read_prep,
2029 .dbo_declare_punch = osd_declare_punch,
2030 .dbo_punch = osd_punch,
2031 .dbo_fiemap_get = osd_fiemap_get,
2032 .dbo_ladvise = osd_ladvise,