4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 /* prerequisite for linux/xattr.h */
42 #include <linux/types.h>
43 /* prerequisite for linux/xattr.h */
46 #include <linux/pagevec.h>
49 * struct OBD_{ALLOC,FREE}*()
52 #include <obd_support.h>
54 #include "osd_internal.h"
57 #include <ldiskfs/ldiskfs_extents.h>
59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
60 int rw, int line, int pages)
64 LASSERTF(iobuf->dr_elapsed_valid == 0,
65 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
66 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
68 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
70 init_waitqueue_head(&iobuf->dr_wait);
71 atomic_set(&iobuf->dr_numreqs, 0);
76 iobuf->dr_elapsed = ktime_set(0, 0);
77 /* must be counted before, so assert */
79 iobuf->dr_init_at = line;
81 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
82 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
83 LASSERT(iobuf->dr_pg_buf.lb_len >=
84 pages * sizeof(iobuf->dr_pages[0]));
88 /* start with 1MB for 4K blocks */
90 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
93 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
94 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
96 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
97 iobuf->dr_max_pages = 0;
98 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
99 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
101 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
102 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
103 if (unlikely(iobuf->dr_blocks == NULL))
106 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
107 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
108 if (unlikely(iobuf->dr_pages == NULL))
111 iobuf->dr_max_pages = pages;
115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
116 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
120 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
121 iobuf->dr_pages[iobuf->dr_npages++] = page;
124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
126 int rw = iobuf->dr_rw;
128 if (iobuf->dr_elapsed_valid) {
129 iobuf->dr_elapsed_valid = 0;
130 LASSERT(iobuf->dr_dev == d);
131 LASSERT(iobuf->dr_frags > 0);
132 lprocfs_oh_tally(&d->od_brw_stats.
133 hist[BRW_R_DIO_FRAGS+rw],
135 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
136 ktime_to_ms(iobuf->dr_elapsed));
140 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
141 static void dio_complete_routine(struct bio *bio)
143 # ifdef HAVE_BI_STATUS
144 int error = bio->bi_status;
146 int error = bio->bi_error;
149 static void dio_complete_routine(struct bio *bio, int error)
152 struct osd_iobuf *iobuf = bio->bi_private;
156 /* CAVEAT EMPTOR: possibly in IRQ context
157 * DO NOT record procfs stats here!!! */
159 if (unlikely(iobuf == NULL)) {
160 CERROR("***** bio->bi_private is NULL! This should never "
161 "happen. Normally, I would crash here, but instead I "
162 "will dump the bio contents to the console. Please "
163 "report this to <https://jira.hpdd.intel.com/> , along "
164 "with any interesting messages leading up to this point "
165 "(like SCSI errors, perhaps). Because bi_private is "
166 "NULL, I can't wake up the thread that initiated this "
167 "IO - you will probably have to reboot this node.\n");
168 CERROR("bi_next: %p, bi_flags: %lx, "
174 "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
175 "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
176 (unsigned long)bio->bi_flags,
182 bio->bi_vcnt, bio_idx(bio),
183 bio_sectors(bio) << 9, bio->bi_end_io,
185 atomic_read(&bio->bi_cnt),
187 atomic_read(&bio->__bi_cnt),
193 /* the check is outside of the cycle for performance reason -bzzz */
194 if (!bio_data_dir(bio)) {
195 bio_for_each_segment_all(bvl, bio, iter) {
196 if (likely(error == 0))
197 SetPageUptodate(bvl_to_page(bvl));
198 LASSERT(PageLocked(bvl_to_page(bvl)));
200 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
202 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
205 /* any real error is good enough -bzzz */
206 if (error != 0 && iobuf->dr_error == 0)
207 iobuf->dr_error = error;
210 * set dr_elapsed before dr_numreqs turns to 0, otherwise
211 * it's possible that service thread will see dr_numreqs
212 * is zero, but dr_elapsed is not set yet, leading to lost
213 * data in this processing and an assertion in a subsequent
216 if (atomic_read(&iobuf->dr_numreqs) == 1) {
217 ktime_t now = ktime_get();
219 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
220 iobuf->dr_elapsed_valid = 1;
222 if (atomic_dec_and_test(&iobuf->dr_numreqs))
223 wake_up(&iobuf->dr_wait);
225 /* Completed bios used to be chained off iobuf->dr_bios and freed in
226 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
227 * mempool when serious on-disk fragmentation was encountered,
228 * deadlocking the OST. The bios are now released as soon as complete
229 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
233 static void record_start_io(struct osd_iobuf *iobuf, int size)
235 struct osd_device *osd = iobuf->dr_dev;
236 struct obd_histogram *h = osd->od_brw_stats.hist;
239 atomic_inc(&iobuf->dr_numreqs);
241 if (iobuf->dr_rw == 0) {
242 atomic_inc(&osd->od_r_in_flight);
243 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
244 atomic_read(&osd->od_r_in_flight));
245 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
246 } else if (iobuf->dr_rw == 1) {
247 atomic_inc(&osd->od_w_in_flight);
248 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
249 atomic_read(&osd->od_w_in_flight));
250 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
256 static void osd_submit_bio(int rw, struct bio *bio)
258 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
259 #ifdef HAVE_SUBMIT_BIO_2ARGS
261 submit_bio(READ, bio);
263 submit_bio(WRITE, bio);
270 static int can_be_merged(struct bio *bio, sector_t sector)
275 return bio_end_sector(bio) == sector ? 1 : 0;
278 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
279 struct osd_iobuf *iobuf)
281 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
282 struct page **pages = iobuf->dr_pages;
283 int npages = iobuf->dr_npages;
284 sector_t *blocks = iobuf->dr_blocks;
285 int total_blocks = npages * blocks_per_page;
286 int sector_bits = inode->i_sb->s_blocksize_bits - 9;
287 unsigned int blocksize = inode->i_sb->s_blocksize;
288 struct bio *bio = NULL;
290 unsigned int page_offset;
300 LASSERT(iobuf->dr_npages == npages);
302 osd_brw_stats_update(osd, iobuf);
303 iobuf->dr_start_time = ktime_get();
305 blk_start_plug(&plug);
306 for (page_idx = 0, block_idx = 0;
308 page_idx++, block_idx += blocks_per_page) {
310 page = pages[page_idx];
311 LASSERT(block_idx + blocks_per_page <= total_blocks);
313 for (i = 0, page_offset = 0;
315 i += nblocks, page_offset += blocksize * nblocks) {
319 if (blocks[block_idx + i] == 0) { /* hole */
320 LASSERTF(iobuf->dr_rw == 0,
321 "page_idx %u, block_idx %u, i %u\n",
322 page_idx, block_idx, i);
323 memset(kmap(page) + page_offset, 0, blocksize);
328 sector = (sector_t)blocks[block_idx + i] << sector_bits;
330 /* Additional contiguous file blocks? */
331 while (i + nblocks < blocks_per_page &&
332 (sector + (nblocks << sector_bits)) ==
333 ((sector_t)blocks[block_idx + i + nblocks] <<
338 can_be_merged(bio, sector) &&
339 bio_add_page(bio, page,
340 blocksize * nblocks, page_offset) != 0)
341 continue; /* added this frag OK */
344 struct request_queue *q =
345 bdev_get_queue(bio->bi_bdev);
346 unsigned int bi_size = bio_sectors(bio) << 9;
348 /* Dang! I have to fragment this I/O */
349 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
350 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
351 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
353 queue_max_sectors(q),
354 bio_phys_segments(q, bio),
355 queue_max_phys_segments(q),
356 0, queue_max_hw_segments(q));
357 record_start_io(iobuf, bi_size);
358 osd_submit_bio(iobuf->dr_rw, bio);
361 /* allocate new bio */
362 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
363 (npages - page_idx) *
366 CERROR("Can't allocate bio %u*%u = %u pages\n",
367 (npages - page_idx), blocks_per_page,
368 (npages - page_idx) * blocks_per_page);
373 bio->bi_bdev = inode->i_sb->s_bdev;
374 bio_set_sector(bio, sector);
376 bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
378 bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
380 bio->bi_end_io = dio_complete_routine;
381 bio->bi_private = iobuf;
383 rc = bio_add_page(bio, page,
384 blocksize * nblocks, page_offset);
390 record_start_io(iobuf, bio_sectors(bio) << 9);
391 osd_submit_bio(iobuf->dr_rw, bio);
396 blk_finish_plug(&plug);
398 /* in order to achieve better IO throughput, we don't wait for writes
399 * completion here. instead we proceed with transaction commit in
400 * parallel and wait for IO completion once transaction is stopped
401 * see osd_trans_stop() for more details -bzzz */
402 if (iobuf->dr_rw == 0) {
403 wait_event(iobuf->dr_wait,
404 atomic_read(&iobuf->dr_numreqs) == 0);
405 osd_fini_iobuf(osd, iobuf);
409 rc = iobuf->dr_error;
413 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
414 struct niobuf_local *lnb)
421 int poff = offset & (PAGE_SIZE - 1);
422 int plen = PAGE_SIZE - poff;
426 lnb->lnb_file_offset = offset;
427 lnb->lnb_page_offset = poff;
429 /* lnb->lnb_flags = rnb->rnb_flags; */
431 lnb->lnb_page = NULL;
434 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
445 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
448 struct inode *inode = osd_dt_obj(dt)->oo_inode;
449 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
454 page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
457 if (unlikely(page == NULL))
458 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
464 * there are following "locks":
475 * - lock pages, unlock
477 * - lock partial page
483 * Unlock and release pages loaded by osd_bufs_get()
485 * Unlock \a npages pages from \a lnb and drop the refcount on them.
487 * \param env thread execution environment
488 * \param dt dt object undergoing IO (OSD object + methods)
489 * \param lnb array of pages undergoing IO
490 * \param npages number of pages in \a lnb
494 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
495 struct niobuf_local *lnb, int npages)
500 #ifdef HAVE_PAGEVEC_INIT_ONE_PARAM
503 pagevec_init(&pvec, 0);
506 for (i = 0; i < npages; i++) {
507 if (lnb[i].lnb_page == NULL)
509 LASSERT(PageLocked(lnb[i].lnb_page));
510 unlock_page(lnb[i].lnb_page);
511 if (pagevec_add(&pvec, lnb[i].lnb_page) == 0)
512 pagevec_release(&pvec);
513 dt_object_put(env, dt);
514 lnb[i].lnb_page = NULL;
517 /* Release any partial pagevec */
518 pagevec_release(&pvec);
524 * Load and lock pages undergoing IO
526 * Pages as described in the \a lnb array are fetched (from disk or cache)
527 * and locked for IO by the caller.
529 * DLM locking protects us from write and truncate competing for same region,
530 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
531 * It's possible the writeout on a such a page is in progress when we access
532 * it. It's also possible that during this writeout we put new (partial) data
533 * into the page, but won't be able to proceed in filter_commitrw_write().
534 * Therefore, just wait for writeout completion as it should be rare enough.
536 * \param env thread execution environment
537 * \param dt dt object undergoing IO (OSD object + methods)
538 * \param pos byte offset of IO start
539 * \param len number of bytes of IO
540 * \param lnb array of extents undergoing IO
541 * \param rw read or write operation, and other flags
542 * \param capa capabilities
544 * \retval pages (zero or more) loaded successfully
545 * \retval -ENOMEM on memory/page allocation error
547 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
548 loff_t pos, ssize_t len, struct niobuf_local *lnb,
549 enum dt_bufs_type rw)
551 struct osd_object *obj = osd_dt_obj(dt);
552 int npages, i, rc = 0;
555 LASSERT(obj->oo_inode);
557 osd_map_remote_to_local(pos, len, &npages, lnb);
559 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
560 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
562 for (i = 0; i < npages; i++, lnb++) {
563 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
565 if (lnb->lnb_page == NULL)
566 GOTO(cleanup, rc = -ENOMEM);
568 wait_on_page_writeback(lnb->lnb_page);
569 BUG_ON(PageWriteback(lnb->lnb_page));
571 lu_object_get(&dt->do_lu);
578 osd_bufs_put(env, dt, lnb - i, i);
582 #ifndef HAVE_LDISKFS_MAP_BLOCKS
584 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
585 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
596 static long ldiskfs_ext_find_goal(struct inode *inode,
597 struct ldiskfs_ext_path *path,
598 unsigned long block, int *aflags)
600 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
601 unsigned long bg_start;
602 unsigned long colour;
606 struct ldiskfs_extent *ex;
607 depth = path->p_depth;
609 /* try to predict block placement */
610 if ((ex = path[depth].p_ext))
611 return ldiskfs_ext_pblock(ex) +
612 (block - le32_to_cpu(ex->ee_block));
614 /* it looks index is empty
615 * try to find starting from index itself */
616 if (path[depth].p_bh)
617 return path[depth].p_bh->b_blocknr;
620 /* OK. use inode's group */
621 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
622 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
623 colour = (current->pid % 16) *
624 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
625 return bg_start + colour + block;
628 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
629 struct ldiskfs_ext_path *path,
630 unsigned long block, unsigned long *count,
633 struct ldiskfs_allocation_request ar;
634 unsigned long pblock;
637 /* find neighbour allocated blocks */
639 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
643 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
647 /* allocate new block */
648 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
652 ar.flags = LDISKFS_MB_HINT_DATA;
653 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
658 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
659 struct ldiskfs_ext_path *path,
660 struct ldiskfs_ext_cache *cex,
661 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
662 struct ldiskfs_extent *ex,
666 struct bpointers *bp = cbdata;
667 struct ldiskfs_extent nex;
668 unsigned long pblock = 0;
674 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
675 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
677 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
683 if (bp->create == 0) {
685 if (cex->ec_block < bp->start)
686 i = bp->start - cex->ec_block;
687 if (i >= cex->ec_len)
688 CERROR("nothing to do?! i = %d, e_num = %u\n",
690 for (; i < cex->ec_len && bp->num; i++) {
700 tgen = LDISKFS_I(inode)->i_ext_generation;
701 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
703 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
704 count + LDISKFS_ALLOC_NEEDED + 1);
705 if (IS_ERR(handle)) {
706 return PTR_ERR(handle);
709 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
710 /* the tree has changed. so path can be invalid at moment */
711 ldiskfs_journal_stop(handle);
715 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
716 * protected by i_data_sem as whole. so we patch it to store
717 * generation to path and now verify the tree hasn't changed */
718 down_write((&LDISKFS_I(inode)->i_data_sem));
720 /* validate extent, make sure the extent tree does not changed */
721 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
722 /* cex is invalid, try again */
723 up_write(&LDISKFS_I(inode)->i_data_sem);
724 ldiskfs_journal_stop(handle);
729 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
732 BUG_ON(count > cex->ec_len);
734 /* insert new extent */
735 nex.ee_block = cpu_to_le32(cex->ec_block);
736 ldiskfs_ext_store_pblock(&nex, pblock);
737 nex.ee_len = cpu_to_le16(count);
738 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
740 /* free data blocks we just allocated */
741 /* not a good idea to call discard here directly,
742 * but otherwise we'd need to call it every free() */
743 ldiskfs_discard_preallocations(inode);
744 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
745 ldiskfs_free_blocks(handle, inode, NULL,
746 ldiskfs_ext_pblock(&nex),
747 le16_to_cpu(nex.ee_len), 0);
749 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
750 le16_to_cpu(nex.ee_len), 0);
756 * Putting len of the actual extent we just inserted,
757 * we are asking ldiskfs_ext_walk_space() to continue
758 * scaning after that block
760 cex->ec_len = le16_to_cpu(nex.ee_len);
761 cex->ec_start = ldiskfs_ext_pblock(&nex);
762 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
763 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
766 up_write((&LDISKFS_I(inode)->i_data_sem));
767 ldiskfs_journal_stop(handle);
772 CERROR("hmm. why do we find this extent?\n");
773 CERROR("initial space: %lu:%u\n",
774 bp->start, bp->init_num);
775 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
776 CERROR("current extent: %u/%u/%llu %d\n",
777 cex->ec_block, cex->ec_len,
778 (unsigned long long)cex->ec_start,
781 CERROR("current extent: %u/%u/%llu\n",
782 cex->ec_block, cex->ec_len,
783 (unsigned long long)cex->ec_start);
787 if (cex->ec_block < bp->start)
788 i = bp->start - cex->ec_block;
789 if (i >= cex->ec_len)
790 CERROR("nothing to do?! i = %d, e_num = %u\n",
792 for (; i < cex->ec_len && bp->num; i++) {
793 *(bp->blocks) = cex->ec_start + i;
795 /* unmap any possible underlying metadata from
796 * the block device mapping. bug 6998. */
797 #ifndef HAVE_CLEAN_BDEV_ALIASES
798 unmap_underlying_metadata(inode->i_sb->s_bdev,
801 clean_bdev_aliases(inode->i_sb->s_bdev,
813 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
814 int clen, sector_t *blocks, int create)
816 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
820 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
824 bp.start = index * blocks_per_page;
825 bp.init_num = bp.num = clen * blocks_per_page;
828 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
829 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
831 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
832 ldiskfs_ext_new_extent_cb, &bp);
833 ldiskfs_ext_invalidate_cache(inode);
838 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
839 struct page **page, int pages,
840 sector_t *blocks, int create)
842 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
843 pgoff_t bitmap_max_page_index;
847 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
849 for (i = 0, b = blocks; i < pages; i++, page++) {
850 if ((*page)->index + 1 >= bitmap_max_page_index) {
854 rc = ldiskfs_map_inode_page(inode, *page, b, create);
856 CERROR("ino %lu, blk %llu create %d: rc %d\n",
858 (unsigned long long)*b, create, rc);
861 b += blocks_per_page;
866 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
868 int pages, sector_t *blocks,
871 int rc = 0, i = 0, clen = 0;
872 struct page *fp = NULL;
874 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
875 inode->i_ino, pages, (*page)->index);
877 /* pages are sorted already. so, we just have to find
878 * contig. space and process them properly */
881 /* start new extent */
886 } else if (fp->index + clen == (*page)->index) {
887 /* continue the extent */
894 /* process found extent */
895 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
900 /* look for next extent */
902 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
906 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
913 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
914 int pages, sector_t *blocks,
919 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
920 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
924 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
929 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
930 int pages, sector_t *blocks,
933 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
935 struct page *fp = NULL;
937 pgoff_t max_page_index;
938 handle_t *handle = NULL;
940 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
942 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
943 inode->i_ino, pages, (*page)->index);
946 create = LDISKFS_GET_BLOCKS_CREATE;
947 handle = ldiskfs_journal_current_handle();
948 LASSERT(handle != NULL);
949 rc = osd_attach_jinode(inode);
953 /* pages are sorted already. so, we just have to find
954 * contig. space and process them properly */
956 long blen, total = 0;
957 struct ldiskfs_map_blocks map = { 0 };
959 if (fp == NULL) { /* start new extent */
964 } else if (fp->index + clen == (*page)->index) {
965 /* continue the extent */
971 if (fp->index + clen >= max_page_index)
972 GOTO(cleanup, rc = -EFBIG);
973 /* process found extent */
974 map.m_lblk = fp->index * blocks_per_page;
975 map.m_len = blen = clen * blocks_per_page;
977 rc = ldiskfs_map_blocks(handle, inode, &map, create);
980 for (; total < blen && c < map.m_len; c++, total++) {
982 *(blocks + total) = 0;
986 *(blocks + total) = map.m_pblk + c;
987 /* unmap any possible underlying
988 * metadata from the block device
989 * mapping. bug 6998. */
990 if ((map.m_flags & LDISKFS_MAP_NEW) &&
992 #ifndef HAVE_CLEAN_BDEV_ALIASES
993 unmap_underlying_metadata(
1005 if (rc == 0 && total < blen) {
1006 map.m_lblk = fp->index * blocks_per_page + total;
1007 map.m_len = blen - total;
1013 /* look for next extent */
1015 blocks += blocks_per_page * clen;
1020 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1022 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1023 struct niobuf_local *lnb, int npages)
1025 struct osd_thread_info *oti = osd_oti_get(env);
1026 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1027 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1028 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1040 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1041 if (unlikely(rc != 0))
1044 isize = i_size_read(inode);
1045 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1047 if (osd->od_writethrough_cache)
1049 if (isize > osd->od_readcache_max_filesize)
1052 start = ktime_get();
1053 for (i = 0; i < npages; i++) {
1056 generic_error_remove_page(inode->i_mapping,
1060 * till commit the content of the page is undefined
1061 * we'll set it uptodate once bulk is done. otherwise
1062 * subsequent reads can access non-stable data
1064 ClearPageUptodate(lnb[i].lnb_page);
1066 if (lnb[i].lnb_len == PAGE_SIZE)
1069 if (maxidx >= lnb[i].lnb_page->index) {
1070 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1073 char *p = kmap(lnb[i].lnb_page);
1075 off = lnb[i].lnb_page_offset;
1078 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1081 memset(p + off, 0, PAGE_SIZE - off);
1082 kunmap(lnb[i].lnb_page);
1086 timediff = ktime_us_delta(end, start);
1087 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1089 if (iobuf->dr_npages) {
1090 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1092 iobuf->dr_blocks, 0);
1093 if (likely(rc == 0)) {
1094 rc = osd_do_bio(osd, inode, iobuf);
1095 /* do IO stats for preparation reads */
1096 osd_fini_iobuf(osd, iobuf);
1102 struct osd_fextent {
1105 unsigned int mapped:1;
1108 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1109 struct osd_fextent *cached_extent)
1111 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1112 sector_t block = offset >> inode->i_blkbits;
1114 struct fiemap_extent_info fei = { 0 };
1115 struct fiemap_extent fe = { 0 };
1116 mm_segment_t saved_fs;
1119 if (block >= cached_extent->start && block < cached_extent->end)
1120 return cached_extent->mapped;
1122 if (i_size_read(inode) == 0)
1125 /* Beyond EOF, must not be mapped */
1126 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1129 fei.fi_extents_max = 1;
1130 fei.fi_extents_start = &fe;
1132 saved_fs = get_fs();
1134 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1139 start = fe.fe_logical >> inode->i_blkbits;
1141 if (start > block) {
1142 cached_extent->start = block;
1143 cached_extent->end = start;
1144 cached_extent->mapped = 0;
1146 cached_extent->start = start;
1147 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1149 cached_extent->mapped = 1;
1152 return cached_extent->mapped;
1155 static int osd_declare_write_commit(const struct lu_env *env,
1156 struct dt_object *dt,
1157 struct niobuf_local *lnb, int npages,
1158 struct thandle *handle)
1160 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1161 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1162 struct osd_thandle *oh;
1170 long long quota_space = 0;
1171 struct osd_fextent extent = { 0 };
1172 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1175 LASSERT(handle != NULL);
1176 oh = container_of0(handle, struct osd_thandle, ot_super);
1177 LASSERT(oh->ot_handle == NULL);
1181 /* calculate number of extents (probably better to pass nb) */
1182 for (i = 0; i < npages; i++) {
1183 if (i && lnb[i].lnb_file_offset !=
1184 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1187 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1188 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1190 quota_space += PAGE_SIZE;
1192 /* ignore quota for the whole request if any page is from
1193 * client cache or written by root.
1195 * XXX once we drop the 1.8 client support, the checking
1196 * for whether page is from cache can be simplified as:
1197 * !(lnb[i].flags & OBD_BRW_SYNC)
1199 * XXX we could handle this on per-lnb basis as done by
1201 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1202 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1204 declare_flags |= OSD_QID_FORCE;
1208 * each extent can go into new leaf causing a split
1209 * 5 is max tree depth: inode + 4 index blocks
1210 * with blockmaps, depth is 3 at most
1212 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1214 * many concurrent threads may grow tree by the time
1215 * our transaction starts. so, consider 2 is a min depth
1217 depth = ext_depth(inode);
1218 depth = max(depth, 1) + 1;
1220 credits++; /* inode */
1221 credits += depth * 2 * extents;
1225 credits++; /* inode */
1226 credits += depth * extents;
1229 /* quota space for metadata blocks */
1230 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1232 /* quota space should be reported in 1K blocks */
1233 quota_space = toqb(quota_space);
1235 /* each new block can go in different group (bitmap + gd) */
1237 /* we can't dirty more bitmap blocks than exist */
1238 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1239 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1241 credits += newblocks;
1243 /* we can't dirty more gd blocks than exist */
1244 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1245 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1247 credits += newblocks;
1249 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1251 /* make sure the over quota flags were not set */
1252 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1254 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1255 i_projid_read(inode), quota_space, oh,
1256 osd_dt_obj(dt), &flags, declare_flags);
1258 /* we need only to store the overquota flags in the first lnb for
1259 * now, once we support multiple objects BRW, this code needs be
1261 if (flags & QUOTA_FL_OVER_USRQUOTA)
1262 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1263 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1264 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1265 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1266 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1271 /* Check if a block is allocated or not */
1272 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1273 struct niobuf_local *lnb, int npages,
1274 struct thandle *thandle)
1276 struct osd_thread_info *oti = osd_oti_get(env);
1277 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1278 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1279 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1285 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1286 if (unlikely(rc != 0))
1289 isize = i_size_read(inode);
1290 ll_vfs_dq_init(inode);
1292 for (i = 0; i < npages; i++) {
1293 if (lnb[i].lnb_rc == -ENOSPC &&
1294 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1295 /* Allow the write to proceed if overwriting an
1300 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1301 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1303 LASSERT(lnb[i].lnb_page);
1304 generic_error_remove_page(inode->i_mapping,
1309 LASSERT(PageLocked(lnb[i].lnb_page));
1310 LASSERT(!PageWriteback(lnb[i].lnb_page));
1312 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1313 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1316 * Since write and truncate are serialized by oo_sem, even
1317 * partial-page truncate should not leave dirty pages in the
1320 LASSERT(!PageDirty(lnb[i].lnb_page));
1322 SetPageUptodate(lnb[i].lnb_page);
1324 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1327 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1329 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1331 } else if (iobuf->dr_npages > 0) {
1332 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1334 iobuf->dr_blocks, 1);
1336 /* no pages to write, no transno is needed */
1337 thandle->th_local = 1;
1340 if (likely(rc == 0)) {
1341 spin_lock(&inode->i_lock);
1342 if (isize > i_size_read(inode)) {
1343 i_size_write(inode, isize);
1344 LDISKFS_I(inode)->i_disksize = isize;
1345 spin_unlock(&inode->i_lock);
1346 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1348 spin_unlock(&inode->i_lock);
1351 rc = osd_do_bio(osd, inode, iobuf);
1352 /* we don't do stats here as in read path because
1353 * write is async: we'll do this in osd_put_bufs() */
1355 osd_fini_iobuf(osd, iobuf);
1358 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1360 if (unlikely(rc != 0)) {
1361 /* if write fails, we should drop pages from the cache */
1362 for (i = 0; i < npages; i++) {
1363 if (lnb[i].lnb_page == NULL)
1365 LASSERT(PageLocked(lnb[i].lnb_page));
1366 generic_error_remove_page(inode->i_mapping,
1374 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1375 struct niobuf_local *lnb, int npages)
1377 struct osd_thread_info *oti = osd_oti_get(env);
1378 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1379 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1380 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1381 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1388 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1389 if (unlikely(rc != 0))
1392 isize = i_size_read(inode);
1394 if (osd->od_read_cache)
1396 if (isize > osd->od_readcache_max_filesize)
1399 start = ktime_get();
1400 for (i = 0; i < npages; i++) {
1402 if (isize <= lnb[i].lnb_file_offset)
1403 /* If there's no more data, abort early.
1404 * lnb->lnb_rc == 0, so it's easy to detect later. */
1407 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1408 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1410 lnb[i].lnb_rc = lnb[i].lnb_len;
1412 /* Bypass disk read if fail_loc is set properly */
1413 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1414 SetPageUptodate(lnb[i].lnb_page);
1416 if (PageUptodate(lnb[i].lnb_page)) {
1420 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1424 generic_error_remove_page(inode->i_mapping,
1428 timediff = ktime_us_delta(end, start);
1429 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1431 if (cache_hits != 0)
1432 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1434 if (cache_misses != 0)
1435 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1437 if (cache_hits + cache_misses != 0)
1438 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1439 cache_hits + cache_misses);
1441 if (iobuf->dr_npages) {
1442 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1444 iobuf->dr_blocks, 0);
1445 rc = osd_do_bio(osd, inode, iobuf);
1447 /* IO stats will be done in osd_bufs_put() */
1454 * XXX: Another layering violation for now.
1456 * We don't want to use ->f_op->read methods, because generic file write
1458 * - serializes on ->i_sem, and
1460 * - does a lot of extra work like balance_dirty_pages(),
1462 * which doesn't work for globally shared files like /last_rcvd.
1464 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1466 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1468 memcpy(buffer, (char *)ei->i_data, buflen);
1473 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1475 struct buffer_head *bh;
1476 unsigned long block;
1482 /* prevent reading after eof */
1483 spin_lock(&inode->i_lock);
1484 if (i_size_read(inode) < *offs + size) {
1485 loff_t diff = i_size_read(inode) - *offs;
1486 spin_unlock(&inode->i_lock);
1488 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1489 i_size_read(inode), *offs);
1491 } else if (diff == 0) {
1497 spin_unlock(&inode->i_lock);
1500 blocksize = 1 << inode->i_blkbits;
1503 block = *offs >> inode->i_blkbits;
1504 boffs = *offs & (blocksize - 1);
1505 csize = min(blocksize - boffs, size);
1506 bh = __ldiskfs_bread(NULL, inode, block, 0);
1508 CERROR("%s: can't read %u@%llu on ino %lu: "
1509 "rc = %ld\n", osd_ino2name(inode),
1510 csize, *offs, inode->i_ino,
1516 memcpy(buf, bh->b_data + boffs, csize);
1519 memset(buf, 0, csize);
1529 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1530 struct lu_buf *buf, loff_t *pos)
1532 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1535 /* Read small symlink from inode body as we need to maintain correct
1536 * on-disk symlinks for ldiskfs.
1538 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1539 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1540 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1542 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1547 static inline int osd_extents_enabled(struct super_block *sb,
1548 struct inode *inode)
1550 if (inode != NULL) {
1551 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1553 } else if (ldiskfs_has_feature_extents(sb)) {
1559 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1560 const loff_t size, const loff_t pos,
1563 int credits, bits, bs, i;
1565 bits = sb->s_blocksize_bits;
1568 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1569 * we do not expect blockmaps on the large files,
1570 * so let's shrink it to 2 levels (4GB files) */
1572 /* this is default reservation: 2 levels */
1573 credits = (blocks + 2) * 3;
1575 /* actual offset is unknown, hard to optimize */
1579 /* now check for few specific cases to optimize */
1580 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1583 /* allocate if not allocated */
1584 if (inode == NULL) {
1585 credits += blocks * 2;
1588 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1589 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1590 if (LDISKFS_I(inode)->i_data[i] == 0)
1593 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1594 /* single indirect */
1595 credits = blocks * 3;
1596 if (inode == NULL ||
1597 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1600 /* The indirect block may be modified. */
1607 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1608 const struct lu_buf *buf, loff_t _pos,
1609 struct thandle *handle)
1611 struct osd_object *obj = osd_dt_obj(dt);
1612 struct inode *inode = obj->oo_inode;
1613 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1614 struct osd_thandle *oh;
1615 int rc = 0, est = 0, credits, blocks, allocated = 0;
1621 LASSERT(buf != NULL);
1622 LASSERT(handle != NULL);
1624 oh = container_of0(handle, struct osd_thandle, ot_super);
1625 LASSERT(oh->ot_handle == NULL);
1628 bits = sb->s_blocksize_bits;
1632 /* if this is an append, then we
1633 * should expect cross-block record */
1639 /* blocks to modify */
1640 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1641 LASSERT(blocks > 0);
1643 if (inode != NULL && _pos != -1) {
1644 /* object size in blocks */
1645 est = (i_size_read(inode) + bs - 1) >> bits;
1646 allocated = inode->i_blocks >> (bits - 9);
1647 if (pos + size <= i_size_read(inode) && est <= allocated) {
1648 /* looks like an overwrite, no need to modify tree */
1650 /* no need to modify i_size */
1655 if (osd_extents_enabled(sb, inode)) {
1657 * many concurrent threads may grow tree by the time
1658 * our transaction starts. so, consider 2 is a min depth
1659 * for every level we may need to allocate a new block
1660 * and take some entries from the old one. so, 3 blocks
1661 * to allocate (bitmap, gd, itself) + old block - 4 per
1664 depth = inode != NULL ? ext_depth(inode) : 0;
1665 depth = max(depth, 1) + 1;
1667 /* if not append, then split may need to modify
1668 * existing blocks moving entries into the new ones */
1671 /* blocks to store data: bitmap,gd,itself */
1672 credits += blocks * 3;
1674 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1676 /* if inode is created as part of the transaction,
1677 * then it's counted already by the creation method */
1683 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1685 /* dt_declare_write() is usually called for system objects, such
1686 * as llog or last_rcvd files. We needn't enforce quota on those
1687 * objects, so always set the lqi_space as 0. */
1689 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1691 i_projid_read(inode), 0,
1692 oh, obj, NULL, OSD_QID_BLK);
1696 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1698 /* LU-2634: clear the extent format for fast symlink */
1699 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1701 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1702 spin_lock(&inode->i_lock);
1703 LDISKFS_I(inode)->i_disksize = buflen;
1704 i_size_write(inode, buflen);
1705 spin_unlock(&inode->i_lock);
1706 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1711 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1712 int write_NUL, loff_t *offs, handle_t *handle)
1714 struct buffer_head *bh = NULL;
1715 loff_t offset = *offs;
1716 loff_t new_size = i_size_read(inode);
1717 unsigned long block;
1718 int blocksize = 1 << inode->i_blkbits;
1722 int dirty_inode = 0;
1726 * long symlink write does not count the NUL terminator in
1727 * bufsize, we write it, and the inode's file size does not
1728 * count the NUL terminator as well.
1730 ((char *)buf)[bufsize] = '\0';
1734 while (bufsize > 0) {
1735 int credits = handle->h_buffer_credits;
1740 block = offset >> inode->i_blkbits;
1741 boffs = offset & (blocksize - 1);
1742 size = min(blocksize - boffs, bufsize);
1743 bh = __ldiskfs_bread(handle, inode, block, 1);
1744 if (IS_ERR_OR_NULL(bh)) {
1752 CERROR("%s: error reading offset %llu (block %lu, "
1753 "size %d, offs %llu), credits %d/%d: rc = %d\n",
1754 inode->i_sb->s_id, offset, block, bufsize, *offs,
1755 credits, handle->h_buffer_credits, err);
1759 err = ldiskfs_journal_get_write_access(handle, bh);
1761 CERROR("journal_get_write_access() returned error %d\n",
1765 LASSERTF(boffs + size <= bh->b_size,
1766 "boffs %d size %d bh->b_size %lu\n",
1767 boffs, size, (unsigned long)bh->b_size);
1768 memcpy(bh->b_data + boffs, buf, size);
1769 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1773 if (offset + size > new_size)
1774 new_size = offset + size;
1784 /* correct in-core and on-disk sizes */
1785 if (new_size > i_size_read(inode)) {
1786 spin_lock(&inode->i_lock);
1787 if (new_size > i_size_read(inode))
1788 i_size_write(inode, new_size);
1789 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1790 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1793 spin_unlock(&inode->i_lock);
1795 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1803 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1804 const struct lu_buf *buf, loff_t *pos,
1805 struct thandle *handle, int ignore_quota)
1807 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1808 struct osd_thandle *oh;
1812 LASSERT(dt_object_exists(dt));
1814 LASSERT(handle != NULL);
1815 LASSERT(inode != NULL);
1816 ll_vfs_dq_init(inode);
1818 /* XXX: don't check: one declared chunk can be used many times */
1819 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1821 oh = container_of(handle, struct osd_thandle, ot_super);
1822 LASSERT(oh->ot_handle->h_transaction != NULL);
1823 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1825 /* Write small symlink to inode body as we need to maintain correct
1826 * on-disk symlinks for ldiskfs.
1827 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1828 * does not count it in.
1830 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1831 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1832 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1834 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1835 buf->lb_len, is_link, pos,
1838 result = buf->lb_len;
1840 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1845 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1846 __u64 start, __u64 end, struct thandle *th)
1848 struct osd_thandle *oh;
1849 struct inode *inode;
1854 oh = container_of(th, struct osd_thandle, ot_super);
1857 * we don't need to reserve credits for whole truncate
1858 * it's not possible as truncate may need to free too many
1859 * blocks and that won't fit a single transaction. instead
1860 * we reserve credits to change i_size and put inode onto
1861 * orphan list. if needed truncate will extend or restart
1864 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1865 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1867 inode = osd_dt_obj(dt)->oo_inode;
1870 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1871 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1876 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1877 __u64 start, __u64 end, struct thandle *th)
1879 struct osd_thandle *oh;
1880 struct osd_object *obj = osd_dt_obj(dt);
1881 struct inode *inode = obj->oo_inode;
1884 int rc = 0, rc2 = 0;
1887 LASSERT(end == OBD_OBJECT_EOF);
1888 LASSERT(dt_object_exists(dt));
1889 LASSERT(osd_invariant(obj));
1890 LASSERT(inode != NULL);
1891 ll_vfs_dq_init(inode);
1894 oh = container_of(th, struct osd_thandle, ot_super);
1895 LASSERT(oh->ot_handle->h_transaction != NULL);
1897 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1899 tid = oh->ot_handle->h_transaction->t_tid;
1901 spin_lock(&inode->i_lock);
1902 i_size_write(inode, start);
1903 spin_unlock(&inode->i_lock);
1904 ll_truncate_pagecache(inode, start);
1905 #ifdef HAVE_INODEOPS_TRUNCATE
1906 if (inode->i_op->truncate) {
1907 inode->i_op->truncate(inode);
1910 ldiskfs_truncate(inode);
1913 * For a partial-page truncate, flush the page to disk immediately to
1914 * avoid data corruption during direct disk write. b=17397
1916 if ((start & ~PAGE_MASK) != 0)
1917 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1919 h = journal_current_handle();
1921 LASSERT(h == oh->ot_handle);
1923 /* do not check credits with osd_trans_exec_check() as the truncate
1924 * can restart the transaction internally and we restart the
1925 * transaction in this case */
1927 if (tid != h->h_transaction->t_tid) {
1928 int credits = oh->ot_credits;
1930 * transaction has changed during truncate
1931 * we need to restart the handle with our credits
1933 if (h->h_buffer_credits < credits) {
1934 if (ldiskfs_journal_extend(h, credits))
1935 rc2 = ldiskfs_journal_restart(h, credits);
1939 RETURN(rc == 0 ? rc2 : rc);
1942 static int fiemap_check_ranges(struct inode *inode,
1943 u64 start, u64 len, u64 *new_len)
1952 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1953 maxbytes = inode->i_sb->s_maxbytes;
1955 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1957 if (start > maxbytes)
1961 * Shrink request scope to what the fs can actually handle.
1963 if (len > maxbytes || (maxbytes - len) < start)
1964 *new_len = maxbytes - start;
1969 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1970 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
1972 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1975 struct fiemap_extent_info fieinfo = {0, };
1976 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1982 if (inode->i_op->fiemap == NULL)
1985 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1988 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1992 fieinfo.fi_flags = fm->fm_flags;
1993 fieinfo.fi_extents_max = fm->fm_extent_count;
1994 fieinfo.fi_extents_start = fm->fm_extents;
1996 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1997 filemap_write_and_wait(inode->i_mapping);
1999 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2000 fm->fm_flags = fieinfo.fi_flags;
2001 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2006 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2007 __u64 start, __u64 end, enum lu_ladvise_type advice)
2010 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2014 case LU_LADVISE_DONTNEED:
2017 invalidate_mapping_pages(inode->i_mapping,
2018 start >> PAGE_SHIFT,
2019 (end - 1) >> PAGE_SHIFT);
2030 * in some cases we may need declare methods for objects being created
2031 * e.g., when we create symlink
2033 const struct dt_body_operations osd_body_ops_new = {
2034 .dbo_declare_write = osd_declare_write,
2037 const struct dt_body_operations osd_body_ops = {
2038 .dbo_read = osd_read,
2039 .dbo_declare_write = osd_declare_write,
2040 .dbo_write = osd_write,
2041 .dbo_bufs_get = osd_bufs_get,
2042 .dbo_bufs_put = osd_bufs_put,
2043 .dbo_write_prep = osd_write_prep,
2044 .dbo_declare_write_commit = osd_declare_write_commit,
2045 .dbo_write_commit = osd_write_commit,
2046 .dbo_read_prep = osd_read_prep,
2047 .dbo_declare_punch = osd_declare_punch,
2048 .dbo_punch = osd_punch,
2049 .dbo_fiemap_get = osd_fiemap_get,
2050 .dbo_ladvise = osd_ladvise,