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/
35 * Author: Nikita Danilov <nikita@clusterfs.com>
36 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
40 #define DEBUG_SUBSYSTEM S_OSD
42 /* prerequisite for linux/xattr.h */
43 #include <linux/types.h>
44 /* prerequisite for linux/xattr.h */
47 #include <linux/swap.h>
48 #include <linux/pagevec.h>
51 * struct OBD_{ALLOC,FREE}*()
53 #include <obd_support.h>
54 #include <libcfs/libcfs.h>
56 #include "osd_internal.h"
59 #include <ldiskfs/ldiskfs_extents.h>
60 #include <ldiskfs/ldiskfs.h>
63 #define SECTOR_SHIFT 9
66 struct kmem_cache *biop_cachep;
68 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
69 static void dio_complete_routine(struct bio *bio);
71 static void dio_complete_routine(struct bio *bio, int error);
74 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
77 struct osd_bio_private *bio_private = NULL;
80 OBD_SLAB_ALLOC_GFP(bio_private, biop_cachep, sizeof(*bio_private),
82 if (bio_private == NULL)
85 bio->bi_end_io = dio_complete_routine;
86 bio->bi_private = bio_private;
87 bio_private->obp_start_page_idx = start_page_idx;
88 bio_private->obp_iobuf = iobuf;
93 static void osd_bio_fini(struct bio *bio)
95 struct osd_bio_private *bio_private;
99 bio_private = bio->bi_private;
101 OBD_SLAB_FREE(bio_private, biop_cachep, sizeof(*bio_private));
104 static inline bool osd_use_page_cache(struct osd_device *d)
106 /* do not use pagecache if write and read caching are disabled */
107 if (d->od_writethrough_cache + d->od_read_cache == 0)
109 /* use pagecache by default */
113 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
115 int rw, const short line, int pages)
119 LASSERTF(iobuf->dr_elapsed_valid == 0,
120 "iobuf %px, reqs %d, rw %d, line %d\n", iobuf,
121 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
123 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
125 init_waitqueue_head(&iobuf->dr_wait);
126 atomic_set(&iobuf->dr_numreqs, 0);
127 iobuf->dr_npages = 0;
128 iobuf->dr_lextents = 0;
129 iobuf->dr_pextents = 0;
134 iobuf->dr_elapsed = ktime_set(0, 0);
135 /* must be counted before, so assert */
137 iobuf->dr_init_at = line;
138 iobuf->dr_inode = inode;
140 /* Init dr_start_pg_wblks to 0 for osd_read/write_prep().
141 * For osd_write_commit() need to keep the value assigned in
142 * osd_ldiskfs_map_inode_pages() during retries, and before it ,
143 * init dr_start_pg_wblks to 0 in osd_write_prep() is sufficient.
146 iobuf->dr_start_pg_wblks = 0;
148 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
149 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
153 /* start with 1MB for 4K blocks */
155 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
158 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
159 (unsigned int)(pages * sizeof(iobuf->dr_lnbs[0])), i, pages);
161 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
162 iobuf->dr_max_pages = 0;
163 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
164 (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
166 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
167 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
168 if (unlikely(iobuf->dr_blocks == NULL))
171 lu_buf_realloc(&iobuf->dr_lnb_buf,
172 pages * sizeof(iobuf->dr_lnbs[0]));
173 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
174 if (unlikely(iobuf->dr_lnbs == NULL))
177 iobuf->dr_max_pages = pages;
182 #define osd_init_iobuf(dev, iobuf, inode, rw, pages) \
185 BUILD_BUG_ON(__LINE__ >= (1 << 16)); \
186 __r = __osd_init_iobuf(dev, iobuf, inode, rw, __LINE__, pages); \
190 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
191 struct niobuf_local *lnb)
193 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
194 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
198 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
200 int rw = iobuf->dr_rw;
202 if (iobuf->dr_elapsed_valid) {
203 struct brw_stats *h = &d->od_brw_stats;
205 iobuf->dr_elapsed_valid = 0;
206 LASSERT(iobuf->dr_dev == d);
207 LASSERT(iobuf->dr_frags > 0);
208 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_DIO_FRAGS+rw],
210 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_IO_TIME+rw],
211 ktime_to_ms(iobuf->dr_elapsed));
217 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
218 static void dio_complete_routine(struct bio *bio)
220 int error = blk_status_to_errno(bio->bi_status);
222 static void dio_complete_routine(struct bio *bio, int error)
225 struct osd_bio_private *bio_private = bio->bi_private;
226 struct osd_iobuf *iobuf = bio_private->obp_iobuf;
230 /* CAVEAT EMPTOR: possibly in IRQ context
231 * DO NOT record procfs stats here!!!
233 if (unlikely(iobuf == NULL)) {
234 CERROR("***** bio->bi_private is NULL! Dump the bio contents to the console. Please report this to <https://jira.whamcloud.com/>, and probably have to reboot this node.\n");
235 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
236 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
237 bio->bi_next, (unsigned long)bio->bi_flags,
238 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
239 bio_sectors(bio) << 9, bio->bi_end_io,
240 atomic_read(&bio->__bi_cnt),
245 /* the check is outside of the cycle for performance reason -bzzz */
246 if (!bio_data_dir(bio)) {
247 DECLARE_BVEC_ITER_ALL(iter_all);
249 bio_for_each_segment_all(bvl, bio, iter_all) {
250 if (likely(error == 0))
251 SetPageUptodate(bvl_to_page(bvl));
252 LASSERT(PageLocked(bvl_to_page(bvl)));
254 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
256 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
259 /* any real error is good enough -bzzz */
260 if (error != 0 && iobuf->dr_error == 0)
261 iobuf->dr_error = error;
264 * set dr_elapsed before dr_numreqs turns to 0, otherwise
265 * it's possible that service thread will see dr_numreqs
266 * is zero, but dr_elapsed is not set yet, leading to lost
267 * data in this processing and an assertion in a subsequent
270 if (atomic_read(&iobuf->dr_numreqs) == 1) {
271 ktime_t now = ktime_get();
273 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
274 iobuf->dr_elapsed_valid = 1;
276 if (atomic_dec_and_test(&iobuf->dr_numreqs))
277 wake_up(&iobuf->dr_wait);
279 /* Completed bios used to be chained off iobuf->dr_bios and freed in
280 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
281 * mempool when serious on-disk fragmentation was encountered,
282 * deadlocking the OST. The bios are now released as soon as complete
283 * so the pool cannot be exhausted while IOs are competing. b=10076
288 static void record_start_io(struct osd_iobuf *iobuf, int size)
290 struct osd_device *osd = iobuf->dr_dev;
291 struct brw_stats *h = &osd->od_brw_stats;
294 atomic_inc(&iobuf->dr_numreqs);
296 if (iobuf->dr_rw == 0) {
297 atomic_inc(&osd->od_r_in_flight);
298 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_RPC_HIST],
299 atomic_read(&osd->od_r_in_flight));
300 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_DISK_IOSIZE],
302 } else if (iobuf->dr_rw == 1) {
303 atomic_inc(&osd->od_w_in_flight);
304 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_W_RPC_HIST],
305 atomic_read(&osd->od_w_in_flight));
306 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_W_DISK_IOSIZE],
313 static int osd_submit_bio(struct osd_device *osd,
314 struct osd_iobuf *iobuf,
317 struct request_queue *q;
318 unsigned int bi_size;
324 q = bio_get_queue(bio);
325 bi_size = bio_sectors(bio) << SECTOR_SHIFT;
326 /* Dang! I have to fragment this I/O */
328 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
329 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
331 queue_max_sectors(q),
332 osd_bio_nr_segs(bio),
333 queue_max_segments(q));
335 rc = osd_bio_integrity_handle(osd, bio, iobuf);
339 record_start_io(iobuf, bi_size);
341 #ifdef HAVE_SUBMIT_BIO_2ARGS
342 submit_bio(iobuf->dr_rw ? WRITE : READ, bio);
344 bio->bi_opf |= iobuf->dr_rw;
351 static int can_be_merged(struct bio *bio, sector_t sector)
354 return bio_end_sector(bio) == sector ? 1 : 0;
358 static void osd_mark_page_io_done(struct osd_iobuf *iobuf,
360 sector_t start_blocks,
363 struct niobuf_local **lnbs = iobuf->dr_lnbs;
364 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
367 i = start_blocks / blocks_per_page;
368 end = (start_blocks + count) / blocks_per_page;
369 for ( ; i < end; i++)
370 lnbs[i]->lnb_flags |= OBD_BRW_DONE;
374 * Linux v5.12-rc1-20-ga8affc03a9b3
375 * block: rename BIO_MAX_PAGES to BIO_MAX_VECS
378 #define BIO_MAX_VECS BIO_MAX_PAGES
381 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
382 struct osd_iobuf *iobuf, sector_t start_blocks,
385 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
386 struct niobuf_local **lnbs = iobuf->dr_lnbs;
387 int npages = iobuf->dr_npages;
388 sector_t *blocks = iobuf->dr_blocks;
389 struct super_block *sb = inode->i_sb;
390 int sector_bits = sb->s_blocksize_bits - SECTOR_SHIFT;
391 unsigned int blocksize = sb->s_blocksize;
392 struct block_device *bdev = sb->s_bdev;
393 struct bio *bio = NULL;
394 int bio_start_page_idx = 0;
396 unsigned int page_offset;
399 int block_idx, block_idx_end;
400 int page_idx, page_idx_start;
403 bool integrity_enabled;
404 struct blk_plug plug;
405 int blocks_left_page;
409 LASSERT(iobuf->dr_npages == npages);
410 iobuf->dr_start_time = ktime_get();
411 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
414 count = npages * blocks_per_page;
415 block_idx_end = start_blocks + count;
417 blk_start_plug(&plug);
419 page_idx_start = start_blocks / blocks_per_page;
420 for (page_idx = page_idx_start, block_idx = start_blocks;
421 block_idx < block_idx_end; page_idx++,
422 block_idx += blocks_left_page) {
423 /* For cases where the filesystems blocksize is not the
424 * same as PAGE_SIZE (e.g. ARM with PAGE_SIZE=64KB and
425 * blocksize=4KB), there will be multiple blocks to
426 * read/write per page. Also, the start and end block may
427 * not be aligned to the start and end of the page, so the
428 * first page may skip some blocks at the start ("i != 0",
429 * "blocks_left_page" is reduced), and the last page may
430 * skip some blocks at the end (limited by "count").
432 page = lnbs[page_idx]->lnb_page;
433 LASSERT(page_idx < iobuf->dr_npages);
435 i = block_idx % blocks_per_page;
436 blocks_left_page = blocks_per_page - i;
437 if (block_idx + blocks_left_page > block_idx_end)
438 blocks_left_page = block_idx_end - block_idx;
439 page_offset = i * blocksize;
440 for (i = 0; i < blocks_left_page;
441 i += nblocks, page_offset += blocksize * nblocks) {
444 if (blocks[block_idx + i] == 0) { /* hole */
445 LASSERTF(iobuf->dr_rw == 0,
446 "page_idx %u, block_idx %u, i %u,"
447 "start_blocks: %llu, count: %llu, npages: %d\n",
448 page_idx, block_idx, i,
449 (unsigned long long)start_blocks,
450 (unsigned long long)count, npages);
451 memset(kmap(page) + page_offset, 0, blocksize);
456 sector = (sector_t)blocks[block_idx + i] << sector_bits;
458 /* Additional contiguous file blocks? */
459 while (i + nblocks < blocks_left_page &&
460 (sector + (nblocks << sector_bits)) ==
461 ((sector_t)blocks[block_idx + i + nblocks] <<
465 if (bio && can_be_merged(bio, sector) &&
466 bio_add_page(bio, page, blocksize * nblocks,
468 continue; /* added this frag OK */
470 rc = osd_submit_bio(osd, iobuf, bio);
474 bio_start_page_idx = page_idx;
475 /* allocate new bio */
476 bio = cfs_bio_alloc(bdev,
477 min_t(unsigned short, BIO_MAX_VECS,
478 (block_idx_end - block_idx +
479 blocks_left_page - 1)),
480 iobuf->dr_rw ? REQ_OP_WRITE
484 CERROR("Can't allocate bio %u pages\n",
485 block_idx_end - block_idx +
486 blocks_left_page - 1);
490 bio_set_sector(bio, sector);
491 rc = osd_bio_init(bio, iobuf, bio_start_page_idx);
495 rc = bio_add_page(bio, page,
496 blocksize * nblocks, page_offset);
500 rc = osd_submit_bio(osd, iobuf, bio);
504 blk_finish_plug(&plug);
506 /* in order to achieve better IO throughput, we don't wait for writes
507 * completion here. instead we proceed with transaction commit in
508 * parallel and wait for IO completion once transaction is stopped
509 * see osd_trans_stop() for more details -bzzz
511 if (iobuf->dr_rw == 0 || CFS_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT)) {
512 wait_event(iobuf->dr_wait,
513 atomic_read(&iobuf->dr_numreqs) == 0);
517 rc = iobuf->dr_error;
521 if (iobuf->dr_rw == 0 || CFS_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT))
522 osd_fini_iobuf(osd, iobuf);
525 if (rc == 0 && iobuf->dr_rw)
526 osd_mark_page_io_done(iobuf, inode,
527 start_blocks, count);
532 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
533 struct niobuf_local *lnb, int maxlnb)
541 int poff = offset & (PAGE_SIZE - 1);
542 int plen = PAGE_SIZE - poff;
544 if (*nrpages >= maxlnb) {
551 lnb->lnb_file_offset = offset;
552 lnb->lnb_page_offset = poff;
554 /* lnb->lnb_flags = rnb->rnb_flags; */
556 lnb->lnb_page = NULL;
558 lnb->lnb_guard_rpc = 0;
559 lnb->lnb_guard_disk = 0;
562 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
573 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
574 loff_t offset, gfp_t gfp_mask, bool cache)
576 struct osd_thread_info *oti = osd_oti_get(env);
577 struct inode *inode = osd_dt_obj(dt)->oo_inode;
578 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
585 page = find_or_create_page(inode->i_mapping,
586 offset >> PAGE_SHIFT, gfp_mask);
589 LASSERT(!PagePrivate2(page));
590 wait_on_page_writeback(page);
592 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
598 if (inode->i_mapping->nrpages) {
599 /* consult with pagecache, but do not create new pages */
600 /* this is normally used once */
601 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
603 wait_on_page_writeback(page);
608 LASSERT(oti->oti_dio_pages);
609 cur = oti->oti_dio_pages_used;
610 page = oti->oti_dio_pages[cur];
612 if (unlikely(!page)) {
613 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
614 page = alloc_page(gfp_mask);
617 oti->oti_dio_pages[cur] = page;
618 SetPagePrivate2(page);
622 ClearPageUptodate(page);
623 page->index = offset >> PAGE_SHIFT;
624 oti->oti_dio_pages_used++;
630 * there are following "locks":
641 * - lock pages, unlock
643 * - lock partial page
649 * Unlock and release pages loaded by osd_bufs_get()
651 * Unlock \a npages pages from \a lnb and drop the refcount on them.
653 * \param env thread execution environment
654 * \param dt dt object undergoing IO (OSD object + methods)
655 * \param lnb array of pages undergoing IO
656 * \param npages number of pages in \a lnb
660 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
661 struct niobuf_local *lnb, int npages)
663 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
664 struct osd_thread_info *oti = osd_oti_get(env);
665 struct osd_iobuf *iobuf = &oti->oti_iobuf;
669 osd_brw_stats_update(osd, iobuf);
670 ll_pagevec_init(&pvec, 0);
672 for (i = 0; i < npages; i++) {
673 struct page *page = lnb[i].lnb_page;
678 /* if the page isn't cached, then reset uptodate
681 if (PagePrivate2(page)) {
682 oti->oti_dio_pages_used--;
684 if (lnb[i].lnb_locked)
686 if (pagevec_add(&pvec, page) == 0)
687 pagevec_release(&pvec);
690 lnb[i].lnb_page = NULL;
693 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
695 /* Release any partial pagevec */
696 pagevec_release(&pvec);
702 * Load and lock pages undergoing IO
704 * Pages as described in the \a lnb array are fetched (from disk or cache)
705 * and locked for IO by the caller.
707 * DLM locking protects us from write and truncate competing for same region,
708 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
709 * It's possible the writeout on a such a page is in progress when we access
710 * it. It's also possible that during this writeout we put new (partial) data
711 * into the page, but won't be able to proceed in filter_commitrw_write().
712 * Therefore, just wait for writeout completion as it should be rare enough.
714 * \param env thread execution environment
715 * \param dt dt object undergoing IO (OSD object + methods)
716 * \param pos byte offset of IO start
717 * \param len number of bytes of IO
718 * \param lnb array of extents undergoing IO
719 * \param rw read or write operation, and other flags
720 * \param capa capabilities
722 * \retval pages (zero or more) loaded successfully
723 * \retval -ENOMEM on memory/page allocation error
725 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
726 loff_t pos, ssize_t len, struct niobuf_local *lnb,
727 int maxlnb, enum dt_bufs_type rw)
729 struct osd_thread_info *oti = osd_oti_get(env);
730 struct osd_object *obj = osd_dt_obj(dt);
731 struct osd_device *osd = osd_obj2dev(obj);
732 int npages, i, iosize, rc = 0;
737 LASSERT(obj->oo_inode);
739 if (unlikely(obj->oo_destroyed))
742 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
746 write = rw & DT_BUFS_TYPE_WRITE;
748 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
749 iosize = fsize - lnb[0].lnb_file_offset;
750 fsize = max(fsize, i_size_read(obj->oo_inode));
752 cache = rw & DT_BUFS_TYPE_READAHEAD;
756 cache = osd_use_page_cache(osd);
759 if (!osd->od_writethrough_cache) {
763 if (iosize >= osd->od_writethrough_max_iosize) {
768 if (!osd->od_read_cache) {
772 if (iosize >= osd->od_readcache_max_iosize) {
777 /* don't use cache on large files */
778 if (osd->od_readcache_max_filesize &&
779 fsize > osd->od_readcache_max_filesize)
785 if (!cache && unlikely(!oti->oti_dio_pages)) {
786 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
787 PTLRPC_MAX_BRW_PAGES);
788 if (!oti->oti_dio_pages)
792 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
793 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
795 for (i = 0; i < npages; i++, lnb++) {
796 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
798 if (lnb->lnb_page == NULL)
799 GOTO(cleanup, rc = -ENOMEM);
803 mark_page_accessed(lnb->lnb_page);
807 /* XXX: this version doesn't invalidate cached pages, but use them */
808 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
809 /* do not allow data aliasing, invalidate pagecache */
810 /* XXX: can be quite expensive in mixed case */
811 invalidate_mapping_pages(obj->oo_inode->i_mapping,
812 lnb[0].lnb_file_offset >> PAGE_SHIFT,
813 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
821 osd_bufs_put(env, dt, lnb - i, i);
825 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
826 static int osd_extend_restart_trans(handle_t *handle, int needed,
831 rc = ldiskfs_journal_ensure_credits(handle, needed,
832 ldiskfs_trans_default_revoke_credits(inode->i_sb));
833 /* this means journal has been restarted */
840 static int osd_extend_restart_trans(handle_t *handle, int needed,
845 if (ldiskfs_handle_has_enough_credits(handle, needed))
847 rc = ldiskfs_journal_extend(handle,
848 needed - handle->h_buffer_credits);
852 return ldiskfs_journal_restart(handle, needed);
854 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
856 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
857 struct osd_device *osd, sector_t start_blocks,
858 sector_t count, loff_t *disk_size,
861 /* if file has grown, take user_size into account */
862 if (user_size && *disk_size > user_size)
863 *disk_size = user_size;
865 spin_lock(&inode->i_lock);
866 if (*disk_size > i_size_read(inode)) {
867 i_size_write(inode, *disk_size);
868 LDISKFS_I(inode)->i_disksize = *disk_size;
869 spin_unlock(&inode->i_lock);
870 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
872 spin_unlock(&inode->i_lock);
876 * We don't do stats here as in read path because
877 * write is async: we'll do this in osd_put_bufs()
879 return osd_do_bio(osd, inode, iobuf, start_blocks, count);
882 static unsigned int osd_extent_bytes(const struct osd_device *o)
884 unsigned int *extent_bytes_ptr =
885 raw_cpu_ptr(o->od_extent_bytes_percpu);
887 if (likely(*extent_bytes_ptr))
888 return *extent_bytes_ptr;
890 /* initialize on first access or CPU hotplug */
891 if (!ldiskfs_has_feature_extents(osd_sb(o)))
892 *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
894 *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
896 return *extent_bytes_ptr;
899 #define EXTENT_BYTES_DECAY 64
900 static void osd_decay_extent_bytes(struct osd_device *osd,
901 unsigned int new_bytes)
903 unsigned int old_bytes;
905 if (!ldiskfs_has_feature_extents(osd_sb(osd)))
908 old_bytes = osd_extent_bytes(osd);
909 *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
910 (old_bytes * (EXTENT_BYTES_DECAY - 1) +
911 min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
912 EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
915 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
916 struct osd_iobuf *iobuf,
917 struct osd_device *osd,
918 int create, __u64 user_size,
920 struct thandle *thandle)
922 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
923 int rc = 0, i = 0, mapped_index = 0;
924 struct page *fp = NULL;
926 pgoff_t max_page_index;
927 handle_t *handle = NULL;
928 sector_t start_blocks = 0, count = 0;
929 loff_t disk_size = 0;
930 struct niobuf_local **lnbs = iobuf->dr_lnbs;
931 int pages = iobuf->dr_npages;
932 sector_t *blocks = iobuf->dr_blocks;
933 struct niobuf_local *lnb1, *lnb2;
936 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
938 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
939 inode->i_ino, pages, (*lnbs)->lnb_page->index);
942 create = LDISKFS_GET_BLOCKS_CREATE;
943 handle = ldiskfs_journal_current_handle();
944 LASSERT(handle != NULL);
945 rc = osd_attach_jinode(inode);
948 disk_size = i_size_read(inode);
949 /* if disk_size is already bigger than specified user_size,
952 if (disk_size > user_size)
955 /* pages are sorted already. so, we just have to find
956 * contig. space and process them properly
959 long blen, total = 0, previous_total = 0;
960 struct ldiskfs_map_blocks map = { 0 };
963 if (fp == NULL) { /* start new extent */
964 fp = (*lnbs)->lnb_page;
967 iobuf->dr_lextents++;
970 } else if (fp->index + clen == (*lnbs)->lnb_page->index) {
971 /* continue the extent */
977 if (fp->index + clen >= max_page_index)
978 GOTO(cleanup, rc = -EFBIG);
979 /* process found extent */
980 map.m_lblk = fp->index * blocks_per_page;
981 map.m_len = blen = clen * blocks_per_page;
984 * Skip already written blocks of the start page.
985 * Note that this branch will not go into for 4K PAGE_SIZE.
986 * Because dr_start_pg_wblks is always 0 for 4K PAGE_SIZE.
987 * iobuf->dr_start_pg_wblks = (start_blocks + count) %
990 if (iobuf->dr_start_pg_wblks > 0) {
991 total = previous_total = start_blocks =
992 iobuf->dr_start_pg_wblks;
993 map.m_lblk = fp->index * blocks_per_page +
995 map.m_len = blen - total;
996 iobuf->dr_start_pg_wblks = 0;
1001 * We might restart transaction for block allocations,
1002 * in order to make sure data ordered mode, issue IO, disk
1003 * size update and block allocations need be within same
1004 * transaction to make sure consistency.
1006 if (handle && check_credits) {
1007 struct osd_thandle *oh;
1009 LASSERT(thandle != NULL);
1010 oh = container_of(thandle, struct osd_thandle,
1013 * only issue IO if restart transaction needed,
1014 * as update disk size need hold inode lock, we
1015 * want to avoid that as much as possible.
1017 if (oh->oh_declared_ext <= 0) {
1018 rc = osd_ldiskfs_map_write(inode,
1019 iobuf, osd, start_blocks,
1020 count, &disk_size, user_size);
1023 thandle->th_restart_tran = 1;
1024 iobuf->dr_start_pg_wblks = (start_blocks +
1025 count) % blocks_per_page;
1026 GOTO(cleanup, rc = -EAGAIN);
1029 if (CFS_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
1030 oh->oh_declared_ext = 0;
1032 oh->oh_declared_ext--;
1036 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1037 time = ktime_sub(ktime_get(), time);
1040 struct brw_stats *h = &osd->od_brw_stats;
1043 iobuf->dr_pextents++;
1045 idx = map.m_flags & LDISKFS_MAP_NEW ?
1046 BRW_ALLOC_TIME : BRW_MAP_TIME;
1047 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[idx],
1050 for (; total < blen && c < map.m_len; c++, total++) {
1052 *(blocks + total) = 0;
1056 if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
1058 /* don't try to read allocated, but
1059 * unwritten blocks, instead fill the
1060 * patches with zeros in osd_do_bio() */
1061 *(blocks + total) = 0;
1064 *(blocks + total) = map.m_pblk + c;
1065 /* unmap any possible underlying
1066 * metadata from the block device
1069 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1071 clean_bdev_aliases(inode->i_sb->s_bdev,
1077 if (rc == 0 && create) {
1078 count += (total - previous_total);
1079 mapped_index = (start_blocks + count + blocks_per_page -
1080 1) / blocks_per_page - 1;
1081 lnb1 = iobuf->dr_lnbs[i - clen];
1082 lnb2 = iobuf->dr_lnbs[mapped_index];
1083 size1 = lnb1->lnb_file_offset -
1084 (lnb1->lnb_file_offset % PAGE_SIZE) +
1085 (total << inode->i_blkbits);
1086 size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1090 if (size1 > disk_size)
1094 if (rc == 0 && total < blen) {
1096 * decay extent blocks if we could not
1097 * allocate extent once.
1099 osd_decay_extent_bytes(osd,
1100 (total - previous_total) << inode->i_blkbits);
1101 map.m_lblk = fp->index * blocks_per_page + total;
1102 map.m_len = blen - total;
1103 previous_total = total;
1109 * decay extent blocks if we could allocate
1110 * good large extent.
1112 if (total - previous_total >=
1113 osd_extent_bytes(osd) >> inode->i_blkbits)
1114 osd_decay_extent_bytes(osd,
1115 (total - previous_total) << inode->i_blkbits);
1116 /* look for next extent */
1118 blocks += blocks_per_page * clen;
1121 if (rc == 0 && create &&
1122 start_blocks < pages * blocks_per_page) {
1123 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1124 count, &disk_size, user_size);
1125 LASSERT(start_blocks + count == pages * blocks_per_page);
1130 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1131 struct niobuf_local *lnb, int npages)
1133 struct osd_thread_info *oti = osd_oti_get(env);
1134 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1135 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1136 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1145 rc = osd_init_iobuf(osd, iobuf, inode, 0, npages);
1146 if (unlikely(rc != 0))
1149 isize = i_size_read(inode);
1150 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1152 start = ktime_get();
1153 for (i = 0; i < npages; i++) {
1156 * till commit the content of the page is undefined
1157 * we'll set it uptodate once bulk is done. otherwise
1158 * subsequent reads can access non-stable data
1160 ClearPageUptodate(lnb[i].lnb_page);
1162 if (lnb[i].lnb_len == PAGE_SIZE)
1165 if (maxidx >= lnb[i].lnb_page->index) {
1166 osd_iobuf_add_page(iobuf, &lnb[i]);
1169 char *p = kmap(lnb[i].lnb_page);
1171 off = lnb[i].lnb_page_offset;
1174 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1177 memset(p + off, 0, PAGE_SIZE - off);
1178 kunmap(lnb[i].lnb_page);
1182 timediff = ktime_us_delta(end, start);
1183 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1185 if (iobuf->dr_npages) {
1186 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1188 if (likely(rc == 0)) {
1189 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1190 /* do IO stats for preparation reads */
1191 osd_fini_iobuf(osd, iobuf);
1198 #define DECLARE_MM_SEGMENT_T(name) mm_segment_t name
1199 #define access_set_kernel(saved_fs, fei) \
1201 saved_fs = get_fs(); \
1202 set_fs(KERNEL_DS); \
1204 #define access_unset_kernel(saved_fs, fei) set_fs((saved_fs))
1206 #define DECLARE_MM_SEGMENT_T(name)
1207 #define access_set_kernel(saved_fs, fei) \
1208 (fei)->fi_flags |= LDISKFS_FIEMAP_FLAG_MEMCPY
1209 #define access_unset_kernel(saved_fs, fei) \
1210 (fei)->fi_flags &= ~(LDISKFS_FIEMAP_FLAG_MEMCPY)
1211 #endif /* KERNEL_DS */
1213 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1214 struct ldiskfs_map_blocks *map)
1216 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1218 sector_t block = osd_i_blocks(inode, offset);
1221 if (i_size_read(inode) == 0)
1224 /* Beyond EOF, must not be mapped */
1225 if ((i_size_read(inode) - 1) < offset)
1228 end = map->m_lblk + map->m_len;
1229 if (block >= map->m_lblk && block < end)
1230 return map->m_flags & LDISKFS_MAP_MAPPED;
1232 map->m_lblk = block;
1233 map->m_len = INT_MAX;
1235 mapped = ldiskfs_map_blocks(NULL, inode, map, 0);
1241 return map->m_flags & LDISKFS_MAP_MAPPED;
1244 #define MAX_EXTENTS_PER_WRITE 100
1245 static int osd_declare_write_commit(const struct lu_env *env,
1246 struct dt_object *dt,
1247 struct niobuf_local *lnb, int npages,
1248 struct thandle *handle)
1250 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1251 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1252 struct osd_thandle *oh;
1253 int extents = 0, new_meta = 0;
1254 int depth, new_blocks = 0;
1256 int dirty_groups = 0;
1259 long long quota_space = 0;
1260 struct ldiskfs_map_blocks map;
1261 enum osd_quota_local_flags local_flags = 0;
1262 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1263 unsigned int extent_bytes;
1264 loff_t extent_start = 0;
1265 loff_t extent_end = 0;
1268 LASSERT(handle != NULL);
1269 oh = container_of(handle, struct osd_thandle, ot_super);
1270 LASSERT(oh->ot_handle == NULL);
1273 * We track a decaying average extent blocks per filesystem,
1274 * for most of time, it will be 1M, with filesystem becoming
1275 * heavily-fragmented, it will be reduced to 4K at the worst.
1277 extent_bytes = osd_extent_bytes(osd);
1278 LASSERT(extent_bytes >= osd_sb(osd)->s_blocksize);
1280 /* calculate number of extents (probably better to pass nb) */
1281 for (i = 0; i < npages; i++) {
1282 /* ignore quota for the whole request if any page is from
1283 * client cache or written by root.
1285 * XXX we could handle this on per-lnb basis as done by
1288 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1289 (lnb[i].lnb_flags & OBD_BRW_SYS_RESOURCE) ||
1290 !(lnb[i].lnb_flags & OBD_BRW_SYNC))
1291 declare_flags |= OSD_QID_FORCE;
1294 * Convert unwritten extent might need split extents, could
1297 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &map) &&
1298 !(map.m_flags & LDISKFS_MAP_UNWRITTEN)) {
1299 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1303 if (lnb[i].lnb_flags & OBD_BRW_DONE) {
1304 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1308 /* count only unmapped changes */
1310 if (lnb[i].lnb_file_offset != extent_end || extent_end == 0) {
1311 if (extent_end != 0)
1312 extents += (extent_end - extent_start +
1313 extent_bytes - 1) / extent_bytes;
1314 extent_start = lnb[i].lnb_file_offset;
1315 extent_end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1317 extent_end += lnb[i].lnb_len;
1320 quota_space += PAGE_SIZE;
1323 credits++; /* inode */
1325 * overwrite case, no need to modify tree and
1331 extents += (extent_end - extent_start +
1332 extent_bytes - 1) / extent_bytes;
1334 * with system space usage growing up, mballoc codes won't
1335 * try best to scan block group to align best free extent as
1336 * we can. So extent bytes per extent could be decayed to a
1337 * very small value, this could make us reserve too many credits.
1338 * We could be more optimistic in the credit reservations, even
1339 * in a case where the filesystem is nearly full, it is extremely
1340 * unlikely that the worst case would ever be hit.
1342 if (extents > MAX_EXTENTS_PER_WRITE)
1343 extents = MAX_EXTENTS_PER_WRITE;
1346 * If we add a single extent, then in the worse case, each tree
1347 * level index/leaf need to be changed in case of the tree split.
1348 * If more extents are inserted, they could cause the whole tree
1349 * split more than once, but this is really rare.
1351 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1353 * many concurrent threads may grow tree by the time
1354 * our transaction starts. so, consider 2 is a min depth.
1356 depth = ext_depth(inode);
1357 depth = min(max(depth, 1) + 1, LDISKFS_MAX_EXTENT_DEPTH);
1359 credits += depth * 2 * extents;
1362 credits += depth * 3 * extents;
1363 new_meta = depth * 2 * extents;
1367 * With N contiguous data blocks, we need at most
1368 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
1369 * 2 dindirect blocks, and 1 tindirect block
1371 new_meta = DIV_ROUND_UP(new_blocks,
1372 LDISKFS_ADDR_PER_BLOCK(inode->i_sb)) + 4;
1373 credits += new_meta;
1375 dirty_groups += (extents + new_meta);
1377 oh->oh_declared_ext = extents;
1379 /* quota space for metadata blocks */
1380 quota_space += new_meta * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1382 /* quota space should be reported in 1K blocks */
1383 quota_space = toqb(quota_space);
1385 /* each new block can go in different group (bitmap + gd) */
1387 /* we can't dirty more bitmap blocks than exist */
1388 if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1389 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1391 credits += dirty_groups;
1393 /* we can't dirty more gd blocks than exist */
1394 if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1395 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1397 credits += dirty_groups;
1400 "%s: inode #%lu extent_bytes %u extents %d credits %d\n",
1401 osd_ino2name(inode), inode->i_ino, extent_bytes, extents,
1405 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1407 /* make sure the over quota flags were not set */
1408 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1410 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1411 i_projid_read(inode), quota_space, oh,
1412 osd_dt_obj(dt), &local_flags, declare_flags);
1414 /* we need only to store the overquota flags in the first lnb for
1415 * now, once we support multiple objects BRW, this code needs be
1418 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1419 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1420 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1421 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1422 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1423 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1424 if (local_flags & QUOTA_FL_ROOT_PRJQUOTA)
1425 lnb[0].lnb_flags |= OBD_BRW_ROOT_PRJQUOTA;
1428 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1433 /* Check if a block is allocated or not */
1434 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1435 struct niobuf_local *lnb, int npages,
1436 struct thandle *thandle, __u64 user_size)
1438 struct osd_thread_info *oti = osd_oti_get(env);
1439 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1440 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1441 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1442 int rc = 0, i, check_credits = 0;
1446 rc = osd_init_iobuf(osd, iobuf, inode, 1, npages);
1447 if (unlikely(rc != 0))
1450 dquot_initialize(inode);
1452 for (i = 0; i < npages; i++) {
1453 if (lnb[i].lnb_rc == -ENOSPC &&
1454 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1455 /* Allow the write to proceed if overwriting an
1461 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1462 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1464 LASSERT(lnb[i].lnb_page);
1465 generic_error_remove_page(inode->i_mapping,
1470 if (lnb[i].lnb_flags & OBD_BRW_DONE)
1473 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1476 LASSERT(PageLocked(lnb[i].lnb_page));
1477 LASSERT(!PageWriteback(lnb[i].lnb_page));
1480 * Since write and truncate are serialized by oo_sem, even
1481 * partial-page truncate should not leave dirty pages in the
1484 LASSERT(!PageDirty(lnb[i].lnb_page));
1486 SetPageUptodate(lnb[i].lnb_page);
1488 osd_iobuf_add_page(iobuf, &lnb[i]);
1491 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1493 if (CFS_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1495 } else if (iobuf->dr_npages > 0) {
1496 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1501 /* no pages to write, no transno is needed */
1502 thandle->th_local = 1;
1505 if (rc != 0 && !thandle->th_restart_tran)
1506 osd_fini_iobuf(osd, iobuf);
1508 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1510 if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
1511 /* if write fails, we should drop pages from the cache */
1512 for (i = 0; i < npages; i++) {
1513 if (lnb[i].lnb_page == NULL)
1515 if (!PagePrivate2(lnb[i].lnb_page)) {
1516 LASSERT(PageLocked(lnb[i].lnb_page));
1517 generic_error_remove_page(inode->i_mapping,
1526 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1527 struct niobuf_local *lnb, int npages)
1529 struct osd_thread_info *oti = osd_oti_get(env);
1530 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1531 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1532 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1533 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1540 rc = osd_init_iobuf(osd, iobuf, inode, 0, npages);
1541 if (unlikely(rc != 0))
1544 isize = i_size_read(inode);
1546 start = ktime_get();
1547 for (i = 0; i < npages; i++) {
1549 if (isize <= lnb[i].lnb_file_offset)
1550 /* If there's no more data, abort early.
1551 * lnb->lnb_rc == 0, so it's easy to detect later.
1555 /* instead of looking if we go beyong isize, send complete
1556 * pages all the time
1558 lnb[i].lnb_rc = lnb[i].lnb_len;
1560 /* Bypass disk read if fail_loc is set properly */
1561 if (CFS_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW))
1562 SetPageUptodate(lnb[i].lnb_page);
1564 if (PageUptodate(lnb[i].lnb_page)) {
1566 unlock_page(lnb[i].lnb_page);
1569 osd_iobuf_add_page(iobuf, &lnb[i]);
1571 /* no need to unlock in osd_bufs_put(), the sooner page is
1572 * unlocked, the earlier another client can access it.
1573 * notice real unlock_page() can be called few lines
1574 * below after osd_do_bio(). lnb is a per-thread, so it's
1575 * fine to have PG_locked and lnb_locked inconsistent here
1577 lnb[i].lnb_locked = 0;
1580 timediff = ktime_us_delta(end, start);
1581 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1583 if (cache_hits != 0)
1584 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1586 if (cache_misses != 0)
1587 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1589 if (cache_hits + cache_misses != 0)
1590 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1591 cache_hits + cache_misses);
1593 if (iobuf->dr_npages) {
1594 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1597 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1599 /* IO stats will be done in osd_bufs_put() */
1601 /* early release to let others read data during the bulk */
1602 for (i = 0; i < iobuf->dr_npages; i++) {
1603 struct page *page = iobuf->dr_lnbs[i]->lnb_page;
1604 LASSERT(PageLocked(page));
1605 if (!PagePrivate2(page))
1614 * XXX: Another layering violation for now.
1616 * We don't want to use ->f_op->read methods, because generic file write
1618 * - serializes on ->i_sem, and
1620 * - does a lot of extra work like balance_dirty_pages(),
1622 * which doesn't work for globally shared files like /last_rcvd.
1624 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1626 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1628 memcpy(buffer, (char *)ei->i_data, buflen);
1633 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1635 struct buffer_head *bh;
1636 unsigned long block;
1642 /* prevent reading after eof */
1643 spin_lock(&inode->i_lock);
1644 if (i_size_read(inode) < *offs + size) {
1645 loff_t diff = i_size_read(inode) - *offs;
1647 spin_unlock(&inode->i_lock);
1650 "size %llu is too short to read @%llu\n",
1651 i_size_read(inode), *offs);
1653 } else if (diff == 0) {
1659 spin_unlock(&inode->i_lock);
1662 blocksize = 1 << inode->i_blkbits;
1665 block = *offs >> inode->i_blkbits;
1666 boffs = *offs & (blocksize - 1);
1667 csize = min(blocksize - boffs, size);
1668 bh = __ldiskfs_bread(NULL, inode, block, 0);
1670 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1671 osd_ino2name(inode), csize, *offs, inode->i_ino,
1677 memcpy(buf, bh->b_data + boffs, csize);
1680 memset(buf, 0, csize);
1690 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1691 struct lu_buf *buf, loff_t *pos)
1693 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1696 /* Read small symlink from inode body as we need to maintain correct
1697 * on-disk symlinks for ldiskfs.
1699 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1700 loff_t size = i_size_read(inode);
1702 if (buf->lb_len < size)
1705 if (size < sizeof(LDISKFS_I(inode)->i_data))
1706 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1708 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1710 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1716 static inline int osd_extents_enabled(struct super_block *sb,
1717 struct inode *inode)
1719 if (inode != NULL) {
1720 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1722 } else if (ldiskfs_has_feature_extents(sb)) {
1728 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1729 const loff_t size, const loff_t pos,
1732 int credits, bits, bs, i;
1734 bits = sb->s_blocksize_bits;
1737 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1738 * we do not expect blockmaps on the large files,
1739 * so let's shrink it to 2 levels (4GB files)
1742 /* this is default reservation: 2 levels */
1743 credits = (blocks + 2) * 3;
1745 /* actual offset is unknown, hard to optimize */
1749 /* now check for few specific cases to optimize */
1750 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1753 /* allocate if not allocated */
1754 if (inode == NULL) {
1755 credits += blocks * 2;
1758 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1759 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1760 if (LDISKFS_I(inode)->i_data[i] == 0)
1763 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1764 /* single indirect */
1765 credits = blocks * 3;
1766 if (inode == NULL ||
1767 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1770 /* The indirect block may be modified. */
1777 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1778 const struct lu_buf *buf, loff_t _pos,
1779 struct thandle *handle)
1781 struct osd_object *obj = osd_dt_obj(dt);
1782 struct inode *inode = obj->oo_inode;
1783 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1784 struct osd_thandle *oh;
1785 int rc = 0, est = 0, credits, blocks, allocated = 0;
1791 LASSERT(buf != NULL);
1792 LASSERT(handle != NULL);
1794 oh = container_of(handle, struct osd_thandle, ot_super);
1795 LASSERT(oh->ot_handle == NULL);
1798 bits = sb->s_blocksize_bits;
1801 if (osd_tx_was_declared(env, oh, dt, DTO_WRITE_BASE, _pos))
1805 /* if this is an append, then we
1806 * should expect cross-block record
1813 /* blocks to modify */
1814 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1815 LASSERT(blocks > 0);
1817 if (inode != NULL && _pos != -1) {
1818 /* object size in blocks */
1819 est = (i_size_read(inode) + bs - 1) >> bits;
1820 allocated = inode->i_blocks >> (bits - 9);
1821 if (pos + size <= i_size_read(inode) && est <= allocated) {
1822 /* looks like an overwrite, no need to modify tree */
1824 /* no need to modify i_size */
1829 if (osd_extents_enabled(sb, inode)) {
1831 * many concurrent threads may grow tree by the time
1832 * our transaction starts. so, consider 2 is a min depth
1833 * for every level we may need to allocate a new block
1834 * and take some entries from the old one. so, 3 blocks
1835 * to allocate (bitmap, gd, itself) + old block - 4 per
1838 depth = inode != NULL ? ext_depth(inode) : 0;
1839 depth = min(max(depth, 1) + 3, LDISKFS_MAX_EXTENT_DEPTH);
1841 /* if not append, then split may need to modify
1842 * existing blocks moving entries into the new ones
1846 /* blocks to store data: bitmap,gd,itself */
1847 credits += blocks * 3;
1849 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1851 /* if inode is created as part of the transaction,
1852 * then it's counted already by the creation method
1859 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1861 /* dt_declare_write() is usually called for system objects, such
1862 * as llog or last_rcvd files. We needn't enforce quota on those
1863 * objects, so always set the lqi_space as 0.
1866 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1868 i_projid_read(inode), 0,
1869 oh, obj, NULL, OSD_QID_BLK);
1872 rc = osd_trunc_lock(obj, oh, true);
1877 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1879 /* LU-2634: clear the extent format for fast symlink */
1880 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1882 /* Copying the NUL byte terminating the link target as well */
1883 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen + 1);
1884 spin_lock(&inode->i_lock);
1885 LDISKFS_I(inode)->i_disksize = buflen;
1886 i_size_write(inode, buflen);
1887 spin_unlock(&inode->i_lock);
1888 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1893 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1894 int bufsize, int write_NUL, loff_t *offs,
1897 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1898 struct buffer_head *bh = NULL;
1899 loff_t offset = *offs;
1900 loff_t new_size = i_size_read(inode);
1901 unsigned long block;
1902 int blocksize = 1 << inode->i_blkbits;
1903 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1907 int dirty_inode = 0;
1908 bool create, sparse, sync = false;
1912 * long symlink write does not count the NUL terminator in
1913 * bufsize, we write it, and the inode's file size does not
1914 * count the NUL terminator as well.
1916 ((char *)buf)[bufsize] = '\0';
1920 /* only the first flag-set matters */
1921 dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1924 /* sparse checking is racy, but sparse is very rare case, leave as is */
1925 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1926 ((new_size - 1) >> inode->i_blkbits) + 1);
1928 while (bufsize > 0) {
1929 int credits = handle->h_buffer_credits;
1930 unsigned long last_block = (new_size == 0) ? 0 :
1931 (new_size - 1) >> inode->i_blkbits;
1936 block = offset >> inode->i_blkbits;
1937 boffs = offset & (blocksize - 1);
1938 size = min(blocksize - boffs, bufsize);
1939 sync = (block > last_block || new_size == 0 || sparse);
1942 down(&ei->i_append_sem);
1944 bh = __ldiskfs_bread(handle, inode, block, 0);
1946 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1948 "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1949 osd_ino2name(inode),
1950 offset, block, bufsize, *offs);
1952 if (IS_ERR_OR_NULL(bh)) {
1953 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1954 int flags = LDISKFS_GET_BLOCKS_CREATE;
1956 /* while the file system is being mounted, avoid
1957 * preallocation otherwise mount can take a long
1958 * time as mballoc cache is cold.
1959 * XXX: this is a workaround until we have a proper
1961 * XXX: works with extent-based files only */
1962 if (!osd->od_cl_seq)
1963 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1964 bh = __ldiskfs_bread(handle, inode, block, flags);
1968 up(&ei->i_append_sem);
1973 if (IS_ERR_OR_NULL(bh)) {
1982 "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1983 osd_ino2name(inode), offset, block, bufsize,
1984 *offs, credits, handle->h_buffer_credits, err);
1988 err = osd_ldiskfs_journal_get_write_access(handle, inode->i_sb,
1992 CERROR("journal_get_write_access() returned error %d\n",
1996 LASSERTF(boffs + size <= bh->b_size,
1997 "boffs %d size %d bh->b_size %lu\n",
1998 boffs, size, (unsigned long)bh->b_size);
2000 memset(bh->b_data, 0, bh->b_size);
2002 up(&ei->i_append_sem);
2006 memcpy(bh->b_data + boffs, buf, size);
2007 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2011 if (offset + size > new_size)
2012 new_size = offset + size;
2018 up(&ei->i_append_sem);
2025 /* correct in-core and on-disk sizes */
2026 if (new_size > i_size_read(inode)) {
2027 spin_lock(&inode->i_lock);
2028 if (new_size > i_size_read(inode))
2029 i_size_write(inode, new_size);
2030 if (i_size_read(inode) > ei->i_disksize) {
2031 ei->i_disksize = i_size_read(inode);
2034 spin_unlock(&inode->i_lock);
2037 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2044 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2045 const struct lu_buf *buf, loff_t *pos,
2046 struct thandle *handle)
2048 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2049 struct osd_thandle *oh;
2053 LASSERT(dt_object_exists(dt));
2055 LASSERT(handle != NULL);
2056 LASSERT(inode != NULL);
2057 dquot_initialize(inode);
2059 /* XXX: don't check: one declared chunk can be used many times */
2060 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2062 oh = container_of(handle, struct osd_thandle, ot_super);
2063 LASSERT(oh->ot_handle->h_transaction != NULL);
2064 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2066 /* Write small symlink to inode body as we need to maintain correct
2067 * on-disk symlinks for ldiskfs.
2068 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2069 * does not count it in.
2071 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2072 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2073 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2075 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2076 is_link, pos, oh->ot_handle);
2078 result = buf->lb_len;
2080 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2085 static int osd_declare_fallocate(const struct lu_env *env,
2086 struct dt_object *dt, __u64 start, __u64 end,
2087 int mode, struct thandle *th)
2089 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2090 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2091 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2092 long long quota_space = 0;
2093 /* 5 is max tree depth. (inode + 4 index blocks) */
2100 * mode == 0 (which is standard prealloc) and PUNCH is supported
2101 * Rest of mode options is not supported yet.
2103 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2104 RETURN(-EOPNOTSUPP);
2106 /* disable fallocate completely */
2107 if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
2108 RETURN(-EOPNOTSUPP);
2113 if ((mode & FALLOC_FL_PUNCH_HOLE) == 0) {
2114 /* quota space for metadata blocks
2115 * approximate metadata estimate should be good enough.
2117 quota_space += PAGE_SIZE;
2118 quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2120 /* quota space should be reported in 1K blocks */
2121 quota_space = toqb(quota_space) + toqb(end - start) +
2122 LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2125 * We don't need to reserve credits for whole fallocate here.
2126 * We reserve space only for metadata. Fallocate credits are
2127 * extended as required
2130 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2131 i_projid_read(inode), quota_space, oh,
2132 osd_dt_obj(dt), NULL, OSD_QID_BLK);
2137 * The both hole punch and allocation may need few transactions
2138 * to complete, so we have to avoid concurrent writes/truncates
2139 * as we can't release object lock from within ldiskfs.
2140 * Notice locking order: transaction start, then lock object
2141 * (don't confuse object lock dt_{read|write}_lock() with the
2144 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2149 static int osd_fallocate_preallocate(const struct lu_env *env,
2150 struct dt_object *dt,
2151 __u64 start, __u64 end, int mode,
2154 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2155 handle_t *handle = ldiskfs_journal_current_handle();
2156 unsigned int save_credits = oh->ot_credits;
2157 struct osd_object *obj = osd_dt_obj(dt);
2158 struct inode *inode = obj->oo_inode;
2159 struct ldiskfs_map_blocks map;
2160 unsigned int credits;
2161 ldiskfs_lblk_t blen;
2162 ldiskfs_lblk_t boff;
2163 loff_t new_size = 0;
2170 LASSERT(dt_object_exists(dt));
2171 LASSERT(osd_invariant(obj));
2172 LASSERT(inode != NULL);
2174 CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2175 inode->i_ino, start, end, mode);
2177 dquot_initialize(inode);
2181 boff = osd_i_blocks(inode, start);
2182 blen = osd_i_blocks(inode, ALIGN(end, 1 << inode->i_blkbits)) - boff;
2184 /* Create and mark new extents as either zero or unwritten */
2185 flags = (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ||
2186 !ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) ?
2187 LDISKFS_GET_BLOCKS_CREATE_ZERO :
2188 LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
2189 #ifdef LDISKFS_GET_BLOCKS_KEEP_SIZE
2190 if (mode & FALLOC_FL_KEEP_SIZE)
2191 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2195 if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2196 end > LDISKFS_I(inode)->i_disksize)) {
2198 rc = inode_newsize_ok(inode, new_size);
2203 inode_dio_wait(inode);
2208 /* Don't normalize the request if it can fit in one extent so
2209 * that it doesn't get unnecessarily split into multiple extents.
2211 if (blen <= EXT_UNWRITTEN_MAX_LEN)
2212 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2215 * credits to insert 1 extent into extent tree.
2217 credits = ldiskfs_chunk_trans_blocks(inode, blen);
2218 depth = ext_depth(inode);
2220 while (rc >= 0 && blen) {
2224 * Recalculate credits when extent tree depth changes.
2226 if (depth != ext_depth(inode)) {
2227 credits = ldiskfs_chunk_trans_blocks(inode, blen);
2228 depth = ext_depth(inode);
2231 /* TODO: quota check */
2232 rc = osd_extend_restart_trans(handle, credits, inode);
2236 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2239 "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2240 inode->i_ino, map.m_lblk, map.m_len, rc);
2241 ldiskfs_mark_inode_dirty(handle, inode);
2246 map.m_len = blen = blen - rc;
2247 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2248 inode->i_ctime = current_time(inode);
2252 if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2253 inode->i_mtime = inode->i_ctime;
2254 #ifdef LDISKFS_EOFBLOCKS_FL
2256 if (epos > inode->i_size)
2257 ldiskfs_set_inode_flag(inode,
2258 LDISKFS_INODE_EOFBLOCKS);
2262 ldiskfs_mark_inode_dirty(handle, inode);
2266 /* extand credits if needed for operations such as attribute set */
2268 rc = osd_extend_restart_trans(handle, save_credits, inode);
2270 inode_unlock(inode);
2275 static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
2276 __u64 start, __u64 end, int mode,
2279 struct osd_object *obj = osd_dt_obj(dt);
2280 struct inode *inode = obj->oo_inode;
2281 struct osd_access_lock *al;
2282 struct osd_thandle *oh;
2283 int rc = 0, found = 0;
2287 LASSERT(dt_object_exists(dt));
2288 LASSERT(osd_invariant(obj));
2289 LASSERT(inode != NULL);
2291 dquot_initialize(inode);
2294 oh = container_of(th, struct osd_thandle, ot_super);
2295 LASSERT(oh->ot_handle->h_transaction != NULL);
2297 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2298 if (obj != al->tl_obj)
2300 LASSERT(al->tl_shared == 0);
2302 /* do actual punch in osd_trans_stop() */
2303 al->tl_start = start;
2306 al->tl_punch = true;
2313 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2314 __u64 start, __u64 end, int mode, struct thandle *th)
2320 if (mode & FALLOC_FL_PUNCH_HOLE) {
2322 rc = osd_fallocate_punch(env, dt, start, end, mode, th);
2324 /* standard preallocate */
2325 rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
2330 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2331 __u64 start, __u64 end, struct thandle *th)
2333 struct osd_thandle *oh;
2334 struct osd_object *obj = osd_dt_obj(dt);
2335 struct inode *inode;
2340 oh = container_of(th, struct osd_thandle, ot_super);
2343 * we don't need to reserve credits for whole truncate
2344 * it's not possible as truncate may need to free too many
2345 * blocks and that won't fit a single transaction. instead
2346 * we reserve credits to change i_size and put inode onto
2347 * orphan list. if needed truncate will extend or restart
2350 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2351 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2353 inode = obj->oo_inode;
2356 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2357 i_projid_read(inode), 0, oh, obj,
2360 /* if object holds encrypted content, we need to make sure we truncate
2361 * on an encryption unit boundary, or subsequent reads will get
2365 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2366 start & ~LUSTRE_ENCRYPTION_MASK)
2367 start = (start & LUSTRE_ENCRYPTION_MASK) +
2368 LUSTRE_ENCRYPTION_UNIT_SIZE;
2369 ll_truncate_pagecache(inode, start);
2370 rc = osd_trunc_lock(obj, oh, false);
2376 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2377 __u64 start, __u64 end, struct thandle *th)
2379 struct osd_object *obj = osd_dt_obj(dt);
2380 struct osd_device *osd = osd_obj2dev(obj);
2381 struct inode *inode = obj->oo_inode;
2382 struct osd_access_lock *al;
2383 struct osd_thandle *oh;
2384 int rc = 0, found = 0;
2388 LASSERT(dt_object_exists(dt));
2389 LASSERT(osd_invariant(obj));
2390 LASSERT(inode != NULL);
2391 dquot_initialize(inode);
2394 oh = container_of(th, struct osd_thandle, ot_super);
2395 LASSERT(oh->ot_handle->h_transaction != NULL);
2397 /* we used to skip truncate to current size to
2398 * optimize truncates on OST. with DoM we can
2399 * get attr_set to set specific size (MDS_REINT)
2400 * and then get truncate RPC which essentially
2401 * would be skipped. this is bad.. so, disable
2402 * this optimization on MDS till the client stop
2403 * to sent MDS_REINT (LU-11033) -bzzz
2405 if (osd->od_is_ost && i_size_read(inode) == start)
2408 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2410 spin_lock(&inode->i_lock);
2411 if (i_size_read(inode) < start)
2413 i_size_write(inode, start);
2414 spin_unlock(&inode->i_lock);
2416 /* optimize grow case */
2418 osd_execute_truncate(obj);
2423 /* add to orphan list to ensure truncate completion
2424 * if this transaction succeed. ldiskfs_truncate()
2425 * will take the inode out of the list
2427 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2428 inode_unlock(inode);
2432 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2433 if (obj != al->tl_obj)
2435 LASSERT(al->tl_shared == 0);
2437 /* do actual truncate in osd_trans_stop() */
2438 al->tl_truncate = 1;
2447 static int fiemap_check_ranges(struct inode *inode,
2448 u64 start, u64 len, u64 *new_len)
2457 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2458 maxbytes = inode->i_sb->s_maxbytes;
2460 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2462 if (start > maxbytes)
2466 * Shrink request scope to what the fs can actually handle.
2468 if (len > maxbytes || (maxbytes - len) < start)
2469 *new_len = maxbytes - start;
2474 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2475 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2477 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2480 struct fiemap_extent_info fieinfo = {0, };
2481 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2484 DECLARE_MM_SEGMENT_T(saved_fs);
2487 if (inode->i_op->fiemap == NULL)
2490 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2493 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2497 fieinfo.fi_flags = fm->fm_flags;
2498 fieinfo.fi_extents_max = fm->fm_extent_count;
2499 fieinfo.fi_extents_start = fm->fm_extents;
2501 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2502 filemap_write_and_wait(inode->i_mapping);
2504 access_set_kernel(saved_fs, &fieinfo);
2505 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2506 access_unset_kernel(saved_fs, &fieinfo);
2507 fm->fm_flags = fieinfo.fi_flags;
2508 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2513 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2514 __u64 start, __u64 end, enum lu_ladvise_type advice)
2516 struct osd_object *obj = osd_dt_obj(dt);
2521 case LU_LADVISE_DONTNEED:
2523 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2524 start >> PAGE_SHIFT,
2525 (end - 1) >> PAGE_SHIFT);
2535 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2536 loff_t offset, int whence)
2538 struct osd_object *obj = osd_dt_obj(dt);
2539 struct osd_device *dev = osd_obj2dev(obj);
2540 struct inode *inode = obj->oo_inode;
2545 LASSERT(dt_object_exists(dt));
2546 LASSERT(osd_invariant(obj));
2548 LASSERT(offset >= 0);
2550 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
2553 RETURN(PTR_ERR(file));
2555 file->f_mode |= FMODE_64BITHASH;
2556 result = file->f_op->llseek(file, offset, whence);
2560 * If 'offset' is beyond end of object file then treat it as not error
2561 * but valid case for SEEK_HOLE and return 'offset' as result.
2562 * LOV will decide if it is beyond real end of file or not.
2564 if (whence == SEEK_HOLE && result == -ENXIO)
2567 CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2568 "hole" : "data", offset, result);
2573 * in some cases we may need declare methods for objects being created
2574 * e.g., when we create symlink
2576 const struct dt_body_operations osd_body_ops_new = {
2577 .dbo_declare_write = osd_declare_write,
2580 const struct dt_body_operations osd_body_ops = {
2581 .dbo_read = osd_read,
2582 .dbo_declare_write = osd_declare_write,
2583 .dbo_write = osd_write,
2584 .dbo_bufs_get = osd_bufs_get,
2585 .dbo_bufs_put = osd_bufs_put,
2586 .dbo_write_prep = osd_write_prep,
2587 .dbo_declare_write_commit = osd_declare_write_commit,
2588 .dbo_write_commit = osd_write_commit,
2589 .dbo_read_prep = osd_read_prep,
2590 .dbo_declare_punch = osd_declare_punch,
2591 .dbo_punch = osd_punch,
2592 .dbo_fiemap_get = osd_fiemap_get,
2593 .dbo_ladvise = osd_ladvise,
2594 .dbo_declare_fallocate = osd_declare_fallocate,
2595 .dbo_fallocate = osd_fallocate,
2596 .dbo_lseek = osd_lseek,
2600 * Get a truncate lock
2602 * In order to take multi-transaction truncate out of main transaction we let
2603 * the caller grab a lock on the object passed. the lock can be shared (for
2604 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2605 * and write in the same transaction handle (do not confuse with big ldiskfs
2606 * transaction containing lots of handles).
2607 * The lock must be taken at declaration.
2609 * \param obj object to lock
2611 * \shared shared or exclusive
2613 * \retval 0 lock is granted
2614 * \retval -NOMEM no memory to allocate lock
2616 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2618 struct osd_access_lock *al, *tmp;
2623 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2624 if (tmp->tl_obj != obj)
2626 LASSERT(tmp->tl_shared == shared);
2627 /* found same lock */
2632 if (unlikely(al == NULL))
2635 al->tl_truncate = false;
2637 down_read(&obj->oo_ext_idx_sem);
2639 down_write(&obj->oo_ext_idx_sem);
2640 al->tl_shared = shared;
2641 lu_object_get(&obj->oo_dt.do_lu);
2643 list_add(&al->tl_list, &oh->ot_trunc_locks);
2648 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2650 struct osd_access_lock *al, *tmp;
2652 list_for_each_entry_safe(al, tmp, list, tl_list) {
2654 up_read(&al->tl_obj->oo_ext_idx_sem);
2656 up_write(&al->tl_obj->oo_ext_idx_sem);
2657 osd_object_put(env, al->tl_obj);
2658 list_del(&al->tl_list);
2663 /* For a partial-page punch, flush punch range to disk immediately */
2664 static void osd_partial_page_flush_punch(struct osd_device *d,
2665 struct inode *inode, loff_t start,
2668 if (osd_use_page_cache(d)) {
2669 filemap_fdatawrite_range(inode->i_mapping, start, end);
2671 /* Notice we use "wait" version to ensure I/O is complete */
2672 filemap_write_and_wait_range(inode->i_mapping, start,
2674 invalidate_mapping_pages(inode->i_mapping, start >> PAGE_SHIFT,
2680 * For a partial-page truncate, flush the page to disk immediately to
2681 * avoid data corruption during direct disk write. b=17397
2683 static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
2686 if (!(offset & ~PAGE_MASK))
2689 if (osd_use_page_cache(d)) {
2690 filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
2692 /* Notice we use "wait" version to ensure I/O is complete */
2693 filemap_write_and_wait_range(inode->i_mapping, offset,
2695 invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
2696 offset >> PAGE_SHIFT);
2700 void osd_execute_truncate(struct osd_object *obj)
2702 struct osd_device *d = osd_obj2dev(obj);
2703 struct inode *inode = obj->oo_inode;
2706 /* simulate crash before (in the middle) of delayed truncate */
2707 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2708 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2709 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2711 mutex_lock(&sbi->s_orphan_lock);
2712 list_del_init(&ei->i_orphan);
2713 mutex_unlock(&sbi->s_orphan_lock);
2717 size = i_size_read(inode);
2719 /* if object holds encrypted content, we need to make sure we truncate
2720 * on an encryption unit boundary, or block content will get corrupted
2722 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2723 size & ~LUSTRE_ENCRYPTION_MASK)
2724 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2725 LUSTRE_ENCRYPTION_UNIT_SIZE;
2726 ldiskfs_truncate(inode);
2727 inode_unlock(inode);
2728 if (inode->i_size != size) {
2729 spin_lock(&inode->i_lock);
2730 i_size_write(inode, size);
2731 LDISKFS_I(inode)->i_disksize = size;
2732 spin_unlock(&inode->i_lock);
2733 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2735 osd_partial_page_flush(d, inode, size);
2738 static int osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
2739 loff_t start, loff_t end, int mode)
2741 struct osd_device *d = osd_obj2dev(obj);
2742 struct inode *inode = obj->oo_inode;
2746 file = alloc_file_pseudo(inode, d->od_mnt, "/", O_NOATIME,
2749 RETURN(PTR_ERR(file));
2751 file->f_mode |= FMODE_64BITHASH;
2752 rc = file->f_op->fallocate(file, mode, start, end - start);
2756 osd_partial_page_flush_punch(d, inode, start, end - 1);
2760 int osd_process_truncates(const struct lu_env *env, struct list_head *list)
2762 struct osd_access_lock *al;
2765 LASSERT(!journal_current_handle());
2767 list_for_each_entry(al, list, tl_list) {
2770 if (al->tl_truncate)
2771 osd_execute_truncate(al->tl_obj);
2772 else if (al->tl_punch)
2773 rc = osd_execute_punch(env, al->tl_obj, al->tl_start,
2774 al->tl_end, al->tl_mode);