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 #define DEBUG_SUBSYSTEM S_OSD
43 /* prerequisite for linux/xattr.h */
44 #include <linux/types.h>
45 /* prerequisite for linux/xattr.h */
48 #include <linux/pagevec.h>
51 * struct OBD_{ALLOC,FREE}*()
54 #include <obd_support.h>
56 #include "osd_internal.h"
59 #include <ldiskfs/ldiskfs_extents.h>
61 static inline bool osd_use_page_cache(struct osd_device *d)
63 /* do not use pagecache if write and read caching are disabled */
64 if (d->od_writethrough_cache + d->od_read_cache == 0)
66 /* use pagecache by default */
70 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
71 int rw, int line, int pages)
75 LASSERTF(iobuf->dr_elapsed_valid == 0,
76 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
77 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
79 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
81 init_waitqueue_head(&iobuf->dr_wait);
82 atomic_set(&iobuf->dr_numreqs, 0);
87 iobuf->dr_elapsed = ktime_set(0, 0);
88 /* must be counted before, so assert */
90 iobuf->dr_init_at = line;
92 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
93 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
94 LASSERT(iobuf->dr_pg_buf.lb_len >=
95 pages * sizeof(iobuf->dr_pages[0]));
99 /* start with 1MB for 4K blocks */
101 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
104 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
105 (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
107 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
108 iobuf->dr_max_pages = 0;
109 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
110 (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
112 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
113 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
114 if (unlikely(iobuf->dr_blocks == NULL))
117 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
118 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
119 if (unlikely(iobuf->dr_pages == NULL))
122 lu_buf_realloc(&iobuf->dr_lnb_buf,
123 pages * sizeof(iobuf->dr_lnbs[0]));
124 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
125 if (unlikely(iobuf->dr_lnbs == NULL))
128 iobuf->dr_max_pages = pages;
132 #define osd_init_iobuf(dev, iobuf, rw, pages) \
133 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
135 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
136 struct niobuf_local *lnb)
138 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
139 iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
140 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
144 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
146 int rw = iobuf->dr_rw;
148 if (iobuf->dr_elapsed_valid) {
149 iobuf->dr_elapsed_valid = 0;
150 LASSERT(iobuf->dr_dev == d);
151 LASSERT(iobuf->dr_frags > 0);
152 lprocfs_oh_tally(&d->od_brw_stats.hist[BRW_R_DIO_FRAGS+rw],
154 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
155 ktime_to_ms(iobuf->dr_elapsed));
159 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
160 static void dio_complete_routine(struct bio *bio)
162 int error = bio->bi_status;
164 static void dio_complete_routine(struct bio *bio, int error)
167 struct osd_iobuf *iobuf = bio->bi_private;
170 /* CAVEAT EMPTOR: possibly in IRQ context
171 * DO NOT record procfs stats here!!!
174 if (unlikely(iobuf == NULL)) {
175 CERROR("***** bio->bi_private is NULL! This should never happen. Normally, I would crash here, but instead I will dump the bio contents to the console. Please report this to <https://jira.whamcloud.com/> , along with any interesting messages leading up to this point (like SCSI errors, perhaps). Because bi_private is NULL, I can't wake up the thread that initiated this IO - you will probably have to reboot this node.\n");
176 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
177 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
178 bio->bi_next, (unsigned long)bio->bi_flags,
179 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
180 bio_sectors(bio) << 9, bio->bi_end_io,
181 atomic_read(&bio->__bi_cnt),
186 /* the check is outside of the cycle for performance reason -bzzz */
187 if (!bio_data_dir(bio)) {
188 DECLARE_BVEC_ITER_ALL(iter_all);
190 bio_for_each_segment_all(bvl, bio, iter_all) {
191 if (likely(error == 0))
192 SetPageUptodate(bvl_to_page(bvl));
193 LASSERT(PageLocked(bvl_to_page(bvl)));
195 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
197 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
200 /* any real error is good enough -bzzz */
201 if (error != 0 && iobuf->dr_error == 0)
202 iobuf->dr_error = error;
205 * set dr_elapsed before dr_numreqs turns to 0, otherwise
206 * it's possible that service thread will see dr_numreqs
207 * is zero, but dr_elapsed is not set yet, leading to lost
208 * data in this processing and an assertion in a subsequent
211 if (atomic_read(&iobuf->dr_numreqs) == 1) {
212 ktime_t now = ktime_get();
214 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
215 iobuf->dr_elapsed_valid = 1;
217 if (atomic_dec_and_test(&iobuf->dr_numreqs))
218 wake_up(&iobuf->dr_wait);
220 /* Completed bios used to be chained off iobuf->dr_bios and freed in
221 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
222 * mempool when serious on-disk fragmentation was encountered,
223 * deadlocking the OST. The bios are now released as soon as complete
224 * so the pool cannot be exhausted while IOs are competing. b=10076
229 static void record_start_io(struct osd_iobuf *iobuf, int size)
231 struct osd_device *osd = iobuf->dr_dev;
232 struct obd_histogram *h = osd->od_brw_stats.hist;
235 atomic_inc(&iobuf->dr_numreqs);
237 if (iobuf->dr_rw == 0) {
238 atomic_inc(&osd->od_r_in_flight);
239 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
240 atomic_read(&osd->od_r_in_flight));
241 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
242 } else if (iobuf->dr_rw == 1) {
243 atomic_inc(&osd->od_w_in_flight);
244 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
245 atomic_read(&osd->od_w_in_flight));
246 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
252 static void osd_submit_bio(int rw, struct bio *bio)
254 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
255 #ifdef HAVE_SUBMIT_BIO_2ARGS
256 submit_bio(rw ? WRITE : READ, bio);
263 static int can_be_merged(struct bio *bio, sector_t sector)
268 return bio_end_sector(bio) == sector ? 1 : 0;
271 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
273 * This function will change the data written, thus it should only be
274 * used when checking data integrity feature
276 static void bio_integrity_fault_inject(struct bio *bio)
278 struct bio_vec *bvec;
279 DECLARE_BVEC_ITER_ALL(iter_all);
283 bio_for_each_segment_all(bvec, bio, iter_all) {
284 struct page *page = bvec->bv_page;
294 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
295 unsigned int sectors, int tuple_size)
297 __u16 *expected_guard;
301 expected_guard = expected_guard_buf;
302 for (i = 0; i < sectors; i++) {
303 bio_guard = (__u16 *)bio_prot_buf;
304 if (*bio_guard != *expected_guard) {
306 "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
307 i, *expected_guard, *bio_guard, sectors,
312 bio_prot_buf += tuple_size;
317 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
318 struct osd_iobuf *iobuf, int index)
320 struct blk_integrity *bi = bdev_get_integrity(bdev);
321 struct bio_integrity_payload *bip = bio->bi_integrity;
322 struct niobuf_local *lnb;
323 unsigned short sector_size = blk_integrity_interval(bi);
324 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
325 bip->bip_vec->bv_offset;
327 sector_t sector = bio_start_sector(bio);
328 unsigned int sectors, total;
329 DECLARE_BVEC_ITER_ALL(iter_all);
330 __u16 *expected_guard;
334 bio_for_each_segment_all(bv, bio, iter_all) {
335 lnb = iobuf->dr_lnbs[index];
336 expected_guard = lnb->lnb_guards;
337 sectors = bv->bv_len / sector_size;
338 if (lnb->lnb_guard_rpc) {
339 rc = bio_dif_compare(expected_guard, bio_prot_buf,
340 sectors, bi->tuple_size);
346 bio_prot_buf += sectors * bi->tuple_size;
347 total += sectors * bi->tuple_size;
348 LASSERT(total <= bip_size(bio->bi_integrity));
354 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
355 struct osd_iobuf *iobuf,
356 int start_page_idx, bool fault_inject,
357 bool integrity_enabled)
359 struct super_block *sb = osd_sb(osd);
360 integrity_gen_fn *generate_fn = NULL;
361 integrity_vrfy_fn *verify_fn = NULL;
366 if (!integrity_enabled)
369 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
373 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
377 /* Verify and inject fault only when writing */
378 if (iobuf->dr_rw == 1) {
379 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
380 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
386 if (unlikely(fault_inject))
387 bio_integrity_fault_inject(bio);
393 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
394 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
395 static void dio_integrity_complete_routine(struct bio *bio)
397 static void dio_integrity_complete_routine(struct bio *bio, int error)
400 struct osd_bio_private *bio_private = bio->bi_private;
402 bio->bi_private = bio_private->obp_iobuf;
403 osd_dio_complete_routine(bio, error);
405 OBD_FREE_PTR(bio_private);
407 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
408 #else /* !CONFIG_BLK_DEV_INTEGRITY */
409 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
410 fault_inject, integrity_enabled) 0
411 #endif /* CONFIG_BLK_DEV_INTEGRITY */
413 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
414 bool integrity_enabled, int start_page_idx,
415 struct osd_bio_private **pprivate)
421 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
422 if (integrity_enabled) {
423 struct osd_bio_private *bio_private = NULL;
425 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
426 if (bio_private == NULL)
428 bio->bi_end_io = dio_integrity_complete_routine;
429 bio->bi_private = bio_private;
430 bio_private->obp_start_page_idx = start_page_idx;
431 bio_private->obp_iobuf = iobuf;
432 *pprivate = bio_private;
436 bio->bi_end_io = dio_complete_routine;
437 bio->bi_private = iobuf;
443 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
444 struct osd_iobuf *iobuf, sector_t start_blocks,
447 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
448 struct page **pages = iobuf->dr_pages;
449 int npages = iobuf->dr_npages;
450 sector_t *blocks = iobuf->dr_blocks;
451 struct super_block *sb = inode->i_sb;
452 int sector_bits = sb->s_blocksize_bits - 9;
453 unsigned int blocksize = sb->s_blocksize;
454 struct block_device *bdev = sb->s_bdev;
455 struct osd_bio_private *bio_private = NULL;
456 struct bio *bio = NULL;
457 int bio_start_page_idx;
459 unsigned int page_offset;
462 int block_idx, block_idx_end;
463 int page_idx, page_idx_start;
467 bool integrity_enabled;
468 struct blk_plug plug;
469 int blocks_left_page;
473 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
474 LASSERT(iobuf->dr_npages == npages);
476 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
478 osd_brw_stats_update(osd, iobuf);
479 iobuf->dr_start_time = ktime_get();
482 count = npages * blocks_per_page;
483 block_idx_end = start_blocks + count;
485 blk_start_plug(&plug);
487 page_idx_start = start_blocks / blocks_per_page;
488 for (page_idx = page_idx_start, block_idx = start_blocks;
489 block_idx < block_idx_end; page_idx++,
490 block_idx += blocks_left_page) {
491 page = pages[page_idx];
492 LASSERT(page_idx < iobuf->dr_npages);
494 i = block_idx % blocks_per_page;
495 blocks_left_page = blocks_per_page - i;
496 for (page_offset = i * blocksize; i < blocks_left_page;
497 i += nblocks, page_offset += blocksize * nblocks) {
500 if (blocks[block_idx + i] == 0) { /* hole */
501 LASSERTF(iobuf->dr_rw == 0,
502 "page_idx %u, block_idx %u, i %u,"
503 "start_blocks: %llu, count: %llu, npages: %d\n",
504 page_idx, block_idx, i,
505 (unsigned long long)start_blocks,
506 (unsigned long long)count, npages);
507 memset(kmap(page) + page_offset, 0, blocksize);
512 sector = (sector_t)blocks[block_idx + i] << sector_bits;
514 /* Additional contiguous file blocks? */
515 while (i + nblocks < blocks_left_page &&
516 (sector + (nblocks << sector_bits)) ==
517 ((sector_t)blocks[block_idx + i + nblocks] <<
521 if (bio && can_be_merged(bio, sector) &&
522 bio_add_page(bio, page, blocksize * nblocks,
524 continue; /* added this frag OK */
527 struct request_queue *q = bio_get_queue(bio);
528 unsigned int bi_size = bio_sectors(bio) << 9;
530 /* Dang! I have to fragment this I/O */
532 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
533 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
535 queue_max_sectors(q),
536 osd_bio_nr_segs(bio),
537 queue_max_segments(q));
538 rc = osd_bio_integrity_handle(osd, bio,
539 iobuf, bio_start_page_idx,
540 fault_inject, integrity_enabled);
546 record_start_io(iobuf, bi_size);
547 osd_submit_bio(iobuf->dr_rw, bio);
550 bio_start_page_idx = page_idx;
551 /* allocate new bio */
552 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
553 (block_idx_end - block_idx +
554 blocks_left_page - 1)));
556 CERROR("Can't allocate bio %u pages\n",
557 block_idx_end - block_idx +
558 blocks_left_page - 1);
563 bio_set_dev(bio, bdev);
564 bio_set_sector(bio, sector);
565 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
566 rc = osd_bio_init(bio, iobuf, integrity_enabled,
567 bio_start_page_idx, &bio_private);
573 rc = bio_add_page(bio, page,
574 blocksize * nblocks, page_offset);
580 rc = osd_bio_integrity_handle(osd, bio, iobuf,
589 record_start_io(iobuf, bio_sectors(bio) << 9);
590 osd_submit_bio(iobuf->dr_rw, bio);
595 blk_finish_plug(&plug);
597 /* in order to achieve better IO throughput, we don't wait for writes
598 * completion here. instead we proceed with transaction commit in
599 * parallel and wait for IO completion once transaction is stopped
600 * see osd_trans_stop() for more details -bzzz
602 if (iobuf->dr_rw == 0 || fault_inject) {
603 wait_event(iobuf->dr_wait,
604 atomic_read(&iobuf->dr_numreqs) == 0);
605 osd_fini_iobuf(osd, iobuf);
609 rc = iobuf->dr_error;
612 OBD_FREE_PTR(bio_private);
618 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
619 struct niobuf_local *lnb, int maxlnb)
627 int poff = offset & (PAGE_SIZE - 1);
628 int plen = PAGE_SIZE - poff;
630 if (*nrpages >= maxlnb) {
637 lnb->lnb_file_offset = offset;
638 lnb->lnb_page_offset = poff;
640 /* lnb->lnb_flags = rnb->rnb_flags; */
642 lnb->lnb_page = NULL;
644 lnb->lnb_guard_rpc = 0;
645 lnb->lnb_guard_disk = 0;
648 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
659 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
660 loff_t offset, gfp_t gfp_mask, bool cache)
662 struct osd_thread_info *oti = osd_oti_get(env);
663 struct inode *inode = osd_dt_obj(dt)->oo_inode;
664 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
671 page = find_or_create_page(inode->i_mapping,
672 offset >> PAGE_SHIFT, gfp_mask);
675 LASSERT(!PagePrivate2(page));
676 wait_on_page_writeback(page);
678 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
684 if (inode->i_mapping->nrpages) {
685 /* consult with pagecache, but do not create new pages */
686 /* this is normally used once */
687 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
689 wait_on_page_writeback(page);
694 LASSERT(oti->oti_dio_pages);
695 cur = oti->oti_dio_pages_used;
696 page = oti->oti_dio_pages[cur];
698 if (unlikely(!page)) {
699 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
700 page = alloc_page(gfp_mask);
703 oti->oti_dio_pages[cur] = page;
704 SetPagePrivate2(page);
708 ClearPageUptodate(page);
709 page->index = offset >> PAGE_SHIFT;
710 oti->oti_dio_pages_used++;
716 * there are following "locks":
727 * - lock pages, unlock
729 * - lock partial page
735 * Unlock and release pages loaded by osd_bufs_get()
737 * Unlock \a npages pages from \a lnb and drop the refcount on them.
739 * \param env thread execution environment
740 * \param dt dt object undergoing IO (OSD object + methods)
741 * \param lnb array of pages undergoing IO
742 * \param npages number of pages in \a lnb
746 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
747 struct niobuf_local *lnb, int npages)
749 struct osd_thread_info *oti = osd_oti_get(env);
753 ll_pagevec_init(&pvec, 0);
755 for (i = 0; i < npages; i++) {
756 struct page *page = lnb[i].lnb_page;
761 /* if the page isn't cached, then reset uptodate
764 if (PagePrivate2(page)) {
765 oti->oti_dio_pages_used--;
767 if (lnb[i].lnb_locked)
769 if (pagevec_add(&pvec, page) == 0)
770 pagevec_release(&pvec);
773 lnb[i].lnb_page = NULL;
776 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
778 /* Release any partial pagevec */
779 pagevec_release(&pvec);
785 * Load and lock pages undergoing IO
787 * Pages as described in the \a lnb array are fetched (from disk or cache)
788 * and locked for IO by the caller.
790 * DLM locking protects us from write and truncate competing for same region,
791 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
792 * It's possible the writeout on a such a page is in progress when we access
793 * it. It's also possible that during this writeout we put new (partial) data
794 * into the page, but won't be able to proceed in filter_commitrw_write().
795 * Therefore, just wait for writeout completion as it should be rare enough.
797 * \param env thread execution environment
798 * \param dt dt object undergoing IO (OSD object + methods)
799 * \param pos byte offset of IO start
800 * \param len number of bytes of IO
801 * \param lnb array of extents undergoing IO
802 * \param rw read or write operation, and other flags
803 * \param capa capabilities
805 * \retval pages (zero or more) loaded successfully
806 * \retval -ENOMEM on memory/page allocation error
808 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
809 loff_t pos, ssize_t len, struct niobuf_local *lnb,
810 int maxlnb, enum dt_bufs_type rw)
812 struct osd_thread_info *oti = osd_oti_get(env);
813 struct osd_object *obj = osd_dt_obj(dt);
814 struct osd_device *osd = osd_obj2dev(obj);
815 int npages, i, iosize, rc = 0;
820 LASSERT(obj->oo_inode);
822 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
826 write = rw & DT_BUFS_TYPE_WRITE;
828 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
829 iosize = fsize - lnb[0].lnb_file_offset;
830 fsize = max(fsize, i_size_read(obj->oo_inode));
832 cache = rw & DT_BUFS_TYPE_READAHEAD;
836 cache = osd_use_page_cache(osd);
839 if (!osd->od_writethrough_cache) {
843 if (iosize > osd->od_writethrough_max_iosize) {
848 if (!osd->od_read_cache) {
852 if (iosize > osd->od_readcache_max_iosize) {
857 /* don't use cache on large files */
858 if (osd->od_readcache_max_filesize &&
859 fsize > osd->od_readcache_max_filesize)
865 if (!cache && unlikely(!oti->oti_dio_pages)) {
866 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
867 PTLRPC_MAX_BRW_PAGES);
868 if (!oti->oti_dio_pages)
872 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
873 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
875 for (i = 0; i < npages; i++, lnb++) {
876 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
878 if (lnb->lnb_page == NULL)
879 GOTO(cleanup, rc = -ENOMEM);
885 /* XXX: this version doesn't invalidate cached pages, but use them */
886 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
887 /* do not allow data aliasing, invalidate pagecache */
888 /* XXX: can be quite expensive in mixed case */
889 invalidate_mapping_pages(obj->oo_inode->i_mapping,
890 lnb[0].lnb_file_offset >> PAGE_SHIFT,
891 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
899 osd_bufs_put(env, dt, lnb - i, i);
902 /* Borrow @ext4_chunk_trans_blocks */
903 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
905 ldiskfs_group_t groups;
911 depth = ext_depth(inode);
912 idxblocks = depth * 2;
915 * Now let's see how many group bitmaps and group descriptors need
918 groups = idxblocks + 1;
920 if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
921 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
922 if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
923 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
925 /* bitmaps and block group descriptor blocks */
926 ret = idxblocks + groups + gdpblocks;
928 /* Blocks for super block, inode, quota and xattr blocks */
929 ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
934 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
935 static int osd_extend_trans(handle_t *handle, int needed,
938 return __ldiskfs_journal_ensure_credits(handle, needed, needed,
939 ldiskfs_trans_default_revoke_credits(inode->i_sb));
942 static int osd_extend_restart_trans(handle_t *handle, int needed,
945 return ldiskfs_journal_ensure_credits(handle, needed,
946 ldiskfs_trans_default_revoke_credits(inode->i_sb));
949 static int osd_extend_trans(handle_t *handle, int needed,
952 if (ldiskfs_handle_has_enough_credits(handle, needed))
955 return ldiskfs_journal_extend(handle,
956 needed - handle->h_buffer_credits);
959 static int osd_extend_restart_trans(handle_t *handle, int needed,
963 int rc = osd_extend_trans(handle, needed, inode);
968 return ldiskfs_journal_restart(handle, needed);
970 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
972 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
973 struct osd_device *osd, sector_t start_blocks,
974 sector_t count, loff_t *disk_size,
977 /* if file has grown, take user_size into account */
978 if (user_size && *disk_size > user_size)
979 *disk_size = user_size;
981 spin_lock(&inode->i_lock);
982 if (*disk_size > i_size_read(inode)) {
983 i_size_write(inode, *disk_size);
984 LDISKFS_I(inode)->i_disksize = *disk_size;
985 spin_unlock(&inode->i_lock);
986 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
988 spin_unlock(&inode->i_lock);
992 * We don't do stats here as in read path because
993 * write is async: we'll do this in osd_put_bufs()
995 return osd_do_bio(osd, inode, iobuf, start_blocks, count);
999 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
1000 struct osd_iobuf *iobuf,
1001 struct osd_device *osd,
1002 int create, __u64 user_size,
1005 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1006 int rc = 0, i = 0, mapped_index = 0;
1007 struct page *fp = NULL;
1009 pgoff_t max_page_index;
1010 handle_t *handle = NULL;
1012 sector_t start_blocks = 0, count = 0;
1013 loff_t disk_size = 0;
1014 struct page **page = iobuf->dr_pages;
1015 int pages = iobuf->dr_npages;
1016 sector_t *blocks = iobuf->dr_blocks;
1017 struct niobuf_local *lnb1, *lnb2;
1018 loff_t size1, size2;
1020 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1022 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1023 inode->i_ino, pages, (*page)->index);
1026 create = LDISKFS_GET_BLOCKS_CREATE;
1027 handle = ldiskfs_journal_current_handle();
1028 LASSERT(handle != NULL);
1029 rc = osd_attach_jinode(inode);
1032 disk_size = i_size_read(inode);
1033 /* if disk_size is already bigger than specified user_size,
1036 if (disk_size > user_size)
1039 /* pages are sorted already. so, we just have to find
1040 * contig. space and process them properly
1043 long blen, total = 0, previous_total = 0;
1044 struct ldiskfs_map_blocks map = { 0 };
1046 if (fp == NULL) { /* start new extent */
1051 } else if (fp->index + clen == (*page)->index) {
1052 /* continue the extent */
1058 if (fp->index + clen >= max_page_index)
1059 GOTO(cleanup, rc = -EFBIG);
1060 /* process found extent */
1061 map.m_lblk = fp->index * blocks_per_page;
1062 map.m_len = blen = clen * blocks_per_page;
1065 * We might restart transaction for block allocations,
1066 * in order to make sure data ordered mode, issue IO, disk
1067 * size update and block allocations need be within same
1068 * transaction to make sure consistency.
1070 if (handle && check_credits) {
1072 * credits to insert 1 extent into extent tree.
1074 credits = osd_chunk_trans_blocks(inode, blen);
1075 rc = osd_extend_trans(handle, credits, inode);
1079 * only issue IO if restart transaction needed,
1080 * as update disk size need hold inode lock, we
1081 * want to avoid that as much as possible.
1084 WARN_ON_ONCE(start_blocks == 0);
1085 rc = osd_ldiskfs_map_write(inode,
1086 iobuf, osd, start_blocks,
1087 count, &disk_size, user_size);
1090 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
1091 rc = ldiskfs_journal_restart(handle, credits,
1092 ldiskfs_trans_default_revoke_credits(inode->i_sb));
1094 rc = ldiskfs_journal_restart(handle, credits);
1098 start_blocks += count;
1099 /* reset IO block count */
1103 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1107 for (; total < blen && c < map.m_len; c++, total++) {
1109 *(blocks + total) = 0;
1113 *(blocks + total) = map.m_pblk + c;
1114 /* unmap any possible underlying
1115 * metadata from the block device
1118 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1120 clean_bdev_aliases(inode->i_sb->s_bdev,
1126 if (rc == 0 && create) {
1127 count += (total - previous_total);
1128 mapped_index = (count + blocks_per_page -
1129 1) / blocks_per_page - 1;
1130 lnb1 = iobuf->dr_lnbs[i - clen];
1131 lnb2 = iobuf->dr_lnbs[mapped_index];
1132 size1 = lnb1->lnb_file_offset -
1133 (lnb1->lnb_file_offset % PAGE_SIZE) +
1134 (total << inode->i_blkbits);
1135 size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1139 if (size1 > disk_size)
1143 if (rc == 0 && total < blen) {
1144 map.m_lblk = fp->index * blocks_per_page + total;
1145 map.m_len = blen - total;
1146 previous_total = total;
1152 /* look for next extent */
1154 blocks += blocks_per_page * clen;
1157 if (rc == 0 && create &&
1158 start_blocks < pages * blocks_per_page) {
1159 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1160 count, &disk_size, user_size);
1161 LASSERT(start_blocks + count == pages * blocks_per_page);
1166 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1167 struct niobuf_local *lnb, int npages)
1169 struct osd_thread_info *oti = osd_oti_get(env);
1170 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1171 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1172 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1181 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1182 if (unlikely(rc != 0))
1185 isize = i_size_read(inode);
1186 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1188 start = ktime_get();
1189 for (i = 0; i < npages; i++) {
1192 * till commit the content of the page is undefined
1193 * we'll set it uptodate once bulk is done. otherwise
1194 * subsequent reads can access non-stable data
1196 ClearPageUptodate(lnb[i].lnb_page);
1198 if (lnb[i].lnb_len == PAGE_SIZE)
1201 if (maxidx >= lnb[i].lnb_page->index) {
1202 osd_iobuf_add_page(iobuf, &lnb[i]);
1205 char *p = kmap(lnb[i].lnb_page);
1207 off = lnb[i].lnb_page_offset;
1210 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1213 memset(p + off, 0, PAGE_SIZE - off);
1214 kunmap(lnb[i].lnb_page);
1218 timediff = ktime_us_delta(end, start);
1219 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1221 if (iobuf->dr_npages) {
1222 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1224 if (likely(rc == 0)) {
1225 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1226 /* do IO stats for preparation reads */
1227 osd_fini_iobuf(osd, iobuf);
1233 struct osd_fextent {
1236 unsigned int mapped:1;
1239 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1240 struct osd_fextent *cached_extent)
1242 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1243 sector_t block = offset >> inode->i_blkbits;
1245 struct fiemap_extent_info fei = { 0 };
1246 struct fiemap_extent fe = { 0 };
1247 mm_segment_t saved_fs;
1250 if (block >= cached_extent->start && block < cached_extent->end)
1251 return cached_extent->mapped;
1253 if (i_size_read(inode) == 0)
1256 /* Beyond EOF, must not be mapped */
1257 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1260 fei.fi_extents_max = 1;
1261 fei.fi_extents_start = &fe;
1263 saved_fs = get_fs();
1265 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1270 start = fe.fe_logical >> inode->i_blkbits;
1272 if (start > block) {
1273 cached_extent->start = block;
1274 cached_extent->end = start;
1275 cached_extent->mapped = 0;
1277 cached_extent->start = start;
1278 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1280 cached_extent->mapped = 1;
1283 return cached_extent->mapped;
1286 static int osd_declare_write_commit(const struct lu_env *env,
1287 struct dt_object *dt,
1288 struct niobuf_local *lnb, int npages,
1289 struct thandle *handle)
1291 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1292 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1293 struct osd_thandle *oh;
1300 long long quota_space = 0;
1301 struct osd_fextent mapped = { 0 }, extent = { 0 };
1302 enum osd_quota_local_flags local_flags = 0;
1303 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1306 LASSERT(handle != NULL);
1307 oh = container_of(handle, struct osd_thandle, ot_super);
1308 LASSERT(oh->ot_handle == NULL);
1310 /* calculate number of extents (probably better to pass nb) */
1311 for (i = 0; i < npages; i++) {
1312 /* ignore quota for the whole request if any page is from
1313 * client cache or written by root.
1315 * XXX once we drop the 1.8 client support, the checking
1316 * for whether page is from cache can be simplified as:
1317 * !(lnb[i].flags & OBD_BRW_SYNC)
1319 * XXX we could handle this on per-lnb basis as done by
1322 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1323 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1325 declare_flags |= OSD_QID_FORCE;
1327 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1328 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1332 /* count only unmapped changes */
1334 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1336 extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1338 extent.end += lnb[i].lnb_len;
1341 quota_space += PAGE_SIZE;
1344 credits++; /* inode */
1346 * overwrite case, no need to modify tree and
1352 * each extent can go into new leaf causing a split
1353 * 5 is max tree depth: inode + 4 index blocks
1354 * with blockmaps, depth is 3 at most
1356 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1358 * many concurrent threads may grow tree by the time
1359 * our transaction starts. so, consider 2 is a min depth
1361 depth = ext_depth(inode);
1362 depth = max(depth, 1) + 1;
1364 credits += depth * 2 * extents;
1368 credits += depth * extents;
1372 * try a bit more extents to avoid restart
1373 * as much as possible in normal case.
1375 if (npages > 1 && extents)
1378 /* quota space for metadata blocks */
1379 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1381 /* quota space should be reported in 1K blocks */
1382 quota_space = toqb(quota_space);
1384 /* each new block can go in different group (bitmap + gd) */
1386 /* we can't dirty more bitmap blocks than exist */
1387 if (extents > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1388 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1392 /* we can't dirty more gd blocks than exist */
1393 if (extents > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1394 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1399 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1401 /* make sure the over quota flags were not set */
1402 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1404 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1405 i_projid_read(inode), quota_space, oh,
1406 osd_dt_obj(dt), &local_flags, declare_flags);
1408 /* we need only to store the overquota flags in the first lnb for
1409 * now, once we support multiple objects BRW, this code needs be
1412 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1413 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1414 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1415 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1416 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1417 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1420 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1425 /* Check if a block is allocated or not */
1426 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1427 struct niobuf_local *lnb, int npages,
1428 struct thandle *thandle, __u64 user_size)
1430 struct osd_thread_info *oti = osd_oti_get(env);
1431 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1432 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1433 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1434 int rc = 0, i, check_credits = 0;
1435 struct osd_thandle *oh = container_of(thandle,
1436 struct osd_thandle, ot_super);
1437 unsigned int save_credits = oh->ot_credits;
1441 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1442 if (unlikely(rc != 0))
1445 dquot_initialize(inode);
1447 for (i = 0; i < npages; i++) {
1448 if (lnb[i].lnb_rc == -ENOSPC &&
1449 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1450 /* Allow the write to proceed if overwriting an
1456 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1457 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1459 LASSERT(lnb[i].lnb_page);
1460 generic_error_remove_page(inode->i_mapping,
1465 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1468 LASSERT(PageLocked(lnb[i].lnb_page));
1469 LASSERT(!PageWriteback(lnb[i].lnb_page));
1472 * Since write and truncate are serialized by oo_sem, even
1473 * partial-page truncate should not leave dirty pages in the
1476 LASSERT(!PageDirty(lnb[i].lnb_page));
1478 SetPageUptodate(lnb[i].lnb_page);
1480 osd_iobuf_add_page(iobuf, &lnb[i]);
1483 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1485 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1487 } else if (iobuf->dr_npages > 0) {
1488 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1492 * Write might restart transaction, extend credits
1493 * if needed for operations such as attribute set.
1496 handle_t *handle = ldiskfs_journal_current_handle();
1498 LASSERT(handle != NULL);
1499 rc = osd_extend_restart_trans(handle, save_credits,
1503 /* no pages to write, no transno is needed */
1504 thandle->th_local = 1;
1508 osd_fini_iobuf(osd, iobuf);
1510 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1512 if (unlikely(rc != 0)) {
1513 /* if write fails, we should drop pages from the cache */
1514 for (i = 0; i < npages; i++) {
1515 if (lnb[i].lnb_page == NULL)
1517 if (!PagePrivate2(lnb[i].lnb_page)) {
1518 LASSERT(PageLocked(lnb[i].lnb_page));
1519 generic_error_remove_page(inode->i_mapping,
1528 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1529 struct niobuf_local *lnb, int npages)
1531 struct osd_thread_info *oti = osd_oti_get(env);
1532 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1533 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1534 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1535 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1542 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1543 if (unlikely(rc != 0))
1546 isize = i_size_read(inode);
1548 start = ktime_get();
1549 for (i = 0; i < npages; i++) {
1551 if (isize <= lnb[i].lnb_file_offset)
1552 /* If there's no more data, abort early.
1553 * lnb->lnb_rc == 0, so it's easy to detect later.
1557 /* instead of looking if we go beyong isize, send complete
1558 * pages all the time
1560 lnb[i].lnb_rc = lnb[i].lnb_len;
1562 /* Bypass disk read if fail_loc is set properly */
1563 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1564 SetPageUptodate(lnb[i].lnb_page);
1566 if (PageUptodate(lnb[i].lnb_page)) {
1568 unlock_page(lnb[i].lnb_page);
1571 osd_iobuf_add_page(iobuf, &lnb[i]);
1573 /* no need to unlock in osd_bufs_put(), the sooner page is
1574 * unlocked, the earlier another client can access it.
1575 * notice real unlock_page() can be called few lines
1576 * below after osd_do_bio(). lnb is a per-thread, so it's
1577 * fine to have PG_locked and lnb_locked inconsistent here
1579 lnb[i].lnb_locked = 0;
1582 timediff = ktime_us_delta(end, start);
1583 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1585 if (cache_hits != 0)
1586 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1588 if (cache_misses != 0)
1589 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1591 if (cache_hits + cache_misses != 0)
1592 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1593 cache_hits + cache_misses);
1595 if (iobuf->dr_npages) {
1596 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1599 rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1601 /* IO stats will be done in osd_bufs_put() */
1603 /* early release to let others read data during the bulk */
1604 for (i = 0; i < iobuf->dr_npages; i++) {
1605 LASSERT(PageLocked(iobuf->dr_pages[i]));
1606 if (!PagePrivate2(iobuf->dr_pages[i]))
1607 unlock_page(iobuf->dr_pages[i]);
1615 * XXX: Another layering violation for now.
1617 * We don't want to use ->f_op->read methods, because generic file write
1619 * - serializes on ->i_sem, and
1621 * - does a lot of extra work like balance_dirty_pages(),
1623 * which doesn't work for globally shared files like /last_rcvd.
1625 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1627 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1629 memcpy(buffer, (char *)ei->i_data, buflen);
1634 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1636 struct buffer_head *bh;
1637 unsigned long block;
1643 /* prevent reading after eof */
1644 spin_lock(&inode->i_lock);
1645 if (i_size_read(inode) < *offs + size) {
1646 loff_t diff = i_size_read(inode) - *offs;
1648 spin_unlock(&inode->i_lock);
1651 "size %llu is too short to read @%llu\n",
1652 i_size_read(inode), *offs);
1654 } else if (diff == 0) {
1660 spin_unlock(&inode->i_lock);
1663 blocksize = 1 << inode->i_blkbits;
1666 block = *offs >> inode->i_blkbits;
1667 boffs = *offs & (blocksize - 1);
1668 csize = min(blocksize - boffs, size);
1669 bh = __ldiskfs_bread(NULL, inode, block, 0);
1671 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1672 osd_ino2name(inode), csize, *offs, inode->i_ino,
1678 memcpy(buf, bh->b_data + boffs, csize);
1681 memset(buf, 0, csize);
1691 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1692 struct lu_buf *buf, loff_t *pos)
1694 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1697 /* Read small symlink from inode body as we need to maintain correct
1698 * on-disk symlinks for ldiskfs.
1700 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1701 loff_t size = i_size_read(inode);
1703 if (buf->lb_len < size)
1706 if (size < sizeof(LDISKFS_I(inode)->i_data))
1707 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1709 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1711 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1717 static inline int osd_extents_enabled(struct super_block *sb,
1718 struct inode *inode)
1720 if (inode != NULL) {
1721 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1723 } else if (ldiskfs_has_feature_extents(sb)) {
1729 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1730 const loff_t size, const loff_t pos,
1733 int credits, bits, bs, i;
1735 bits = sb->s_blocksize_bits;
1738 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1739 * we do not expect blockmaps on the large files,
1740 * so let's shrink it to 2 levels (4GB files)
1743 /* this is default reservation: 2 levels */
1744 credits = (blocks + 2) * 3;
1746 /* actual offset is unknown, hard to optimize */
1750 /* now check for few specific cases to optimize */
1751 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1754 /* allocate if not allocated */
1755 if (inode == NULL) {
1756 credits += blocks * 2;
1759 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1760 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1761 if (LDISKFS_I(inode)->i_data[i] == 0)
1764 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1765 /* single indirect */
1766 credits = blocks * 3;
1767 if (inode == NULL ||
1768 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1771 /* The indirect block may be modified. */
1778 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1779 const struct lu_buf *buf, loff_t _pos,
1780 struct thandle *handle)
1782 struct osd_object *obj = osd_dt_obj(dt);
1783 struct inode *inode = obj->oo_inode;
1784 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1785 struct osd_thandle *oh;
1786 int rc = 0, est = 0, credits, blocks, allocated = 0;
1792 LASSERT(buf != NULL);
1793 LASSERT(handle != NULL);
1795 oh = container_of(handle, struct osd_thandle, ot_super);
1796 LASSERT(oh->ot_handle == NULL);
1799 bits = sb->s_blocksize_bits;
1803 /* if this is an append, then we
1804 * should expect cross-block record
1811 /* blocks to modify */
1812 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1813 LASSERT(blocks > 0);
1815 if (inode != NULL && _pos != -1) {
1816 /* object size in blocks */
1817 est = (i_size_read(inode) + bs - 1) >> bits;
1818 allocated = inode->i_blocks >> (bits - 9);
1819 if (pos + size <= i_size_read(inode) && est <= allocated) {
1820 /* looks like an overwrite, no need to modify tree */
1822 /* no need to modify i_size */
1827 if (osd_extents_enabled(sb, inode)) {
1829 * many concurrent threads may grow tree by the time
1830 * our transaction starts. so, consider 2 is a min depth
1831 * for every level we may need to allocate a new block
1832 * and take some entries from the old one. so, 3 blocks
1833 * to allocate (bitmap, gd, itself) + old block - 4 per
1836 depth = inode != NULL ? ext_depth(inode) : 0;
1837 depth = max(depth, 1) + 1;
1839 /* if not append, then split may need to modify
1840 * existing blocks moving entries into the new ones
1844 /* blocks to store data: bitmap,gd,itself */
1845 credits += blocks * 3;
1847 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1849 /* if inode is created as part of the transaction,
1850 * then it's counted already by the creation method
1857 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1859 /* dt_declare_write() is usually called for system objects, such
1860 * as llog or last_rcvd files. We needn't enforce quota on those
1861 * objects, so always set the lqi_space as 0.
1864 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1866 i_projid_read(inode), 0,
1867 oh, obj, NULL, OSD_QID_BLK);
1870 rc = osd_trunc_lock(obj, oh, true);
1875 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1877 /* LU-2634: clear the extent format for fast symlink */
1878 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1880 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1881 spin_lock(&inode->i_lock);
1882 LDISKFS_I(inode)->i_disksize = buflen;
1883 i_size_write(inode, buflen);
1884 spin_unlock(&inode->i_lock);
1885 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1890 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1891 int bufsize, int write_NUL, loff_t *offs,
1894 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1895 struct buffer_head *bh = NULL;
1896 loff_t offset = *offs;
1897 loff_t new_size = i_size_read(inode);
1898 unsigned long block;
1899 int blocksize = 1 << inode->i_blkbits;
1900 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1904 int dirty_inode = 0;
1905 bool create, sparse, sync = false;
1909 * long symlink write does not count the NUL terminator in
1910 * bufsize, we write it, and the inode's file size does not
1911 * count the NUL terminator as well.
1913 ((char *)buf)[bufsize] = '\0';
1917 dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1920 /* sparse checking is racy, but sparse is very rare case, leave as is */
1921 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1922 ((new_size - 1) >> inode->i_blkbits) + 1);
1924 while (bufsize > 0) {
1925 int credits = handle->h_buffer_credits;
1926 unsigned long last_block = (new_size == 0) ? 0 :
1927 (new_size - 1) >> inode->i_blkbits;
1932 block = offset >> inode->i_blkbits;
1933 boffs = offset & (blocksize - 1);
1934 size = min(blocksize - boffs, bufsize);
1935 sync = (block > last_block || new_size == 0 || sparse);
1938 down(&ei->i_append_sem);
1940 bh = __ldiskfs_bread(handle, inode, block, 0);
1942 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1944 "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1945 osd_ino2name(inode),
1946 offset, block, bufsize, *offs);
1948 if (IS_ERR_OR_NULL(bh)) {
1949 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1950 int flags = LDISKFS_GET_BLOCKS_CREATE;
1952 /* while the file system is being mounted, avoid
1953 * preallocation otherwise mount can take a long
1954 * time as mballoc cache is cold.
1955 * XXX: this is a workaround until we have a proper
1957 * XXX: works with extent-based files only */
1958 if (!osd->od_cl_seq)
1959 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1960 bh = __ldiskfs_bread(handle, inode, block, flags);
1964 up(&ei->i_append_sem);
1969 if (IS_ERR_OR_NULL(bh)) {
1978 "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1979 osd_ino2name(inode), offset, block, bufsize,
1980 *offs, credits, handle->h_buffer_credits, err);
1984 err = ldiskfs_journal_get_write_access(handle, bh);
1986 CERROR("journal_get_write_access() returned error %d\n",
1990 LASSERTF(boffs + size <= bh->b_size,
1991 "boffs %d size %d bh->b_size %lu\n",
1992 boffs, size, (unsigned long)bh->b_size);
1994 memset(bh->b_data, 0, bh->b_size);
1996 up(&ei->i_append_sem);
2000 memcpy(bh->b_data + boffs, buf, size);
2001 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2005 if (offset + size > new_size)
2006 new_size = offset + size;
2012 up(&ei->i_append_sem);
2019 /* correct in-core and on-disk sizes */
2020 if (new_size > i_size_read(inode)) {
2021 spin_lock(&inode->i_lock);
2022 if (new_size > i_size_read(inode))
2023 i_size_write(inode, new_size);
2024 if (i_size_read(inode) > ei->i_disksize) {
2025 ei->i_disksize = i_size_read(inode);
2028 spin_unlock(&inode->i_lock);
2031 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2038 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2039 const struct lu_buf *buf, loff_t *pos,
2040 struct thandle *handle)
2042 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2043 struct osd_thandle *oh;
2047 LASSERT(dt_object_exists(dt));
2049 LASSERT(handle != NULL);
2050 LASSERT(inode != NULL);
2051 dquot_initialize(inode);
2053 /* XXX: don't check: one declared chunk can be used many times */
2054 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2056 oh = container_of(handle, struct osd_thandle, ot_super);
2057 LASSERT(oh->ot_handle->h_transaction != NULL);
2058 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2060 /* Write small symlink to inode body as we need to maintain correct
2061 * on-disk symlinks for ldiskfs.
2062 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2063 * does not count it in.
2065 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2066 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2067 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2069 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2070 is_link, pos, oh->ot_handle);
2072 result = buf->lb_len;
2074 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2079 static int osd_declare_fallocate(const struct lu_env *env,
2080 struct dt_object *dt, __u64 start, __u64 end,
2081 int mode, struct thandle *th)
2083 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2084 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2085 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2086 long long quota_space = 0;
2087 /* 5 is max tree depth. (inode + 4 index blocks) */
2094 * Only mode == 0 (which is standard prealloc) is supported now.
2095 * Rest of mode options is not supported yet.
2097 if (mode & ~FALLOC_FL_KEEP_SIZE)
2098 RETURN(-EOPNOTSUPP);
2103 /* quota space for metadata blocks
2104 * approximate metadata estimate should be good enough.
2106 quota_space += PAGE_SIZE;
2107 quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2109 /* quota space should be reported in 1K blocks */
2110 quota_space = toqb(quota_space) + toqb(end - start) +
2111 LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2113 /* We don't need to reserve credits for whole fallocate here.
2114 * We reserve space only for metadata. Fallocate credits are
2115 * extended as required
2117 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2118 i_projid_read(inode), quota_space, oh,
2119 osd_dt_obj(dt), NULL, OSD_QID_BLK);
2123 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2124 __u64 start, __u64 end, int mode, struct thandle *th)
2126 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2127 handle_t *handle = ldiskfs_journal_current_handle();
2128 unsigned int save_credits = oh->ot_credits;
2129 struct osd_object *obj = osd_dt_obj(dt);
2130 struct inode *inode = obj->oo_inode;
2131 struct ldiskfs_map_blocks map;
2132 unsigned int credits;
2133 ldiskfs_lblk_t blen;
2134 ldiskfs_lblk_t boff;
2135 loff_t new_size = 0;
2142 LASSERT(dt_object_exists(dt));
2143 LASSERT(osd_invariant(obj));
2144 LASSERT(inode != NULL);
2146 CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2147 inode->i_ino, start, end, mode);
2149 dquot_initialize(inode);
2153 boff = start >> inode->i_blkbits;
2154 blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2156 flags = LDISKFS_GET_BLOCKS_CREATE;
2157 if (mode & FALLOC_FL_KEEP_SIZE)
2158 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2163 * We only support preallocation for extent-based file only.
2165 if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)))
2166 GOTO(out, rc = -EOPNOTSUPP);
2168 if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2169 end > LDISKFS_I(inode)->i_disksize)) {
2171 rc = inode_newsize_ok(inode, new_size);
2176 inode_dio_wait(inode);
2181 /* Don't normalize the request if it can fit in one extent so
2182 * that it doesn't get unnecessarily split into multiple extents.
2184 if (blen <= EXT_UNWRITTEN_MAX_LEN)
2185 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2188 * credits to insert 1 extent into extent tree.
2190 credits = osd_chunk_trans_blocks(inode, blen);
2191 depth = ext_depth(inode);
2193 while (rc >= 0 && blen) {
2197 * Recalculate credits when extent tree depth changes.
2199 if (depth != ext_depth(inode)) {
2200 credits = osd_chunk_trans_blocks(inode, blen);
2201 depth = ext_depth(inode);
2204 /* TODO: quota check */
2205 rc = osd_extend_restart_trans(handle, credits, inode);
2209 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2212 "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2213 inode->i_ino, map.m_lblk, map.m_len, rc);
2214 ldiskfs_mark_inode_dirty(handle, inode);
2219 map.m_len = blen = blen - rc;
2220 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2221 inode->i_ctime = current_time(inode);
2225 if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2226 inode->i_mtime = inode->i_ctime;
2228 if (epos > inode->i_size)
2229 ldiskfs_set_inode_flag(inode,
2230 LDISKFS_INODE_EOFBLOCKS);
2233 ldiskfs_mark_inode_dirty(handle, inode);
2237 /* extand credits if needed for operations such as attribute set */
2239 rc = osd_extend_restart_trans(handle, save_credits, inode);
2241 inode_unlock(inode);
2246 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2247 __u64 start, __u64 end, struct thandle *th)
2249 struct osd_thandle *oh;
2250 struct inode *inode;
2255 oh = container_of(th, struct osd_thandle, ot_super);
2258 * we don't need to reserve credits for whole truncate
2259 * it's not possible as truncate may need to free too many
2260 * blocks and that won't fit a single transaction. instead
2261 * we reserve credits to change i_size and put inode onto
2262 * orphan list. if needed truncate will extend or restart
2265 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2266 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2268 inode = osd_dt_obj(dt)->oo_inode;
2271 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2272 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2276 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2281 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2282 __u64 start, __u64 end, struct thandle *th)
2284 struct osd_object *obj = osd_dt_obj(dt);
2285 struct osd_device *osd = osd_obj2dev(obj);
2286 struct inode *inode = obj->oo_inode;
2287 struct osd_access_lock *al;
2288 struct osd_thandle *oh;
2289 int rc = 0, found = 0;
2293 LASSERT(dt_object_exists(dt));
2294 LASSERT(osd_invariant(obj));
2295 LASSERT(inode != NULL);
2296 dquot_initialize(inode);
2299 oh = container_of(th, struct osd_thandle, ot_super);
2300 LASSERT(oh->ot_handle->h_transaction != NULL);
2302 /* we used to skip truncate to current size to
2303 * optimize truncates on OST. with DoM we can
2304 * get attr_set to set specific size (MDS_REINT)
2305 * and then get truncate RPC which essentially
2306 * would be skipped. this is bad.. so, disable
2307 * this optimization on MDS till the client stop
2308 * to sent MDS_REINT (LU-11033) -bzzz
2310 if (osd->od_is_ost && i_size_read(inode) == start)
2313 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2315 spin_lock(&inode->i_lock);
2316 if (i_size_read(inode) < start)
2318 i_size_write(inode, start);
2319 spin_unlock(&inode->i_lock);
2320 /* if object holds encrypted content, we need to make sure we truncate
2321 * on an encryption unit boundary, or subsequent reads will get
2324 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2325 start & ~LUSTRE_ENCRYPTION_MASK)
2326 start = (start & LUSTRE_ENCRYPTION_MASK) +
2327 LUSTRE_ENCRYPTION_UNIT_SIZE;
2328 ll_truncate_pagecache(inode, start);
2330 /* optimize grow case */
2332 osd_execute_truncate(obj);
2337 /* add to orphan list to ensure truncate completion
2338 * if this transaction succeed. ldiskfs_truncate()
2339 * will take the inode out of the list
2341 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2342 inode_unlock(inode);
2346 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2347 if (obj != al->tl_obj)
2349 LASSERT(al->tl_shared == 0);
2351 /* do actual truncate in osd_trans_stop() */
2352 al->tl_truncate = 1;
2361 static int fiemap_check_ranges(struct inode *inode,
2362 u64 start, u64 len, u64 *new_len)
2371 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2372 maxbytes = inode->i_sb->s_maxbytes;
2374 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2376 if (start > maxbytes)
2380 * Shrink request scope to what the fs can actually handle.
2382 if (len > maxbytes || (maxbytes - len) < start)
2383 *new_len = maxbytes - start;
2388 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2389 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2391 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2394 struct fiemap_extent_info fieinfo = {0, };
2395 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2398 mm_segment_t cur_fs;
2401 if (inode->i_op->fiemap == NULL)
2404 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2407 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2411 fieinfo.fi_flags = fm->fm_flags;
2412 fieinfo.fi_extents_max = fm->fm_extent_count;
2413 fieinfo.fi_extents_start = fm->fm_extents;
2415 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2416 filemap_write_and_wait(inode->i_mapping);
2418 /* Save previous value address limit */
2420 /* Set the address limit of the kernel */
2423 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2424 fm->fm_flags = fieinfo.fi_flags;
2425 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2427 /* Restore the previous address limt */
2433 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2434 __u64 start, __u64 end, enum lu_ladvise_type advice)
2436 struct osd_object *obj = osd_dt_obj(dt);
2441 case LU_LADVISE_DONTNEED:
2443 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2444 start >> PAGE_SHIFT,
2445 (end - 1) >> PAGE_SHIFT);
2455 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2456 loff_t offset, int whence)
2458 struct osd_object *obj = osd_dt_obj(dt);
2459 struct inode *inode = obj->oo_inode;
2465 LASSERT(dt_object_exists(dt));
2466 LASSERT(osd_invariant(obj));
2468 LASSERT(offset >= 0);
2470 file = osd_quasi_file(env, inode);
2471 result = file->f_op->llseek(file, offset, whence);
2474 * If 'offset' is beyond end of object file then treat it as not error
2475 * but valid case for SEEK_HOLE and return 'offset' as result.
2476 * LOV will decide if it is beyond real end of file or not.
2478 if (whence == SEEK_HOLE && result == -ENXIO)
2481 CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2482 "hole" : "data", offset, result);
2487 * in some cases we may need declare methods for objects being created
2488 * e.g., when we create symlink
2490 const struct dt_body_operations osd_body_ops_new = {
2491 .dbo_declare_write = osd_declare_write,
2494 const struct dt_body_operations osd_body_ops = {
2495 .dbo_read = osd_read,
2496 .dbo_declare_write = osd_declare_write,
2497 .dbo_write = osd_write,
2498 .dbo_bufs_get = osd_bufs_get,
2499 .dbo_bufs_put = osd_bufs_put,
2500 .dbo_write_prep = osd_write_prep,
2501 .dbo_declare_write_commit = osd_declare_write_commit,
2502 .dbo_write_commit = osd_write_commit,
2503 .dbo_read_prep = osd_read_prep,
2504 .dbo_declare_punch = osd_declare_punch,
2505 .dbo_punch = osd_punch,
2506 .dbo_fiemap_get = osd_fiemap_get,
2507 .dbo_ladvise = osd_ladvise,
2508 .dbo_declare_fallocate = osd_declare_fallocate,
2509 .dbo_fallocate = osd_fallocate,
2510 .dbo_lseek = osd_lseek,
2514 * Get a truncate lock
2516 * In order to take multi-transaction truncate out of main transaction we let
2517 * the caller grab a lock on the object passed. the lock can be shared (for
2518 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2519 * and write in the same transaction handle (do not confuse with big ldiskfs
2520 * transaction containing lots of handles).
2521 * The lock must be taken at declaration.
2523 * \param obj object to lock
2525 * \shared shared or exclusive
2527 * \retval 0 lock is granted
2528 * \retval -NOMEM no memory to allocate lock
2530 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2532 struct osd_access_lock *al, *tmp;
2537 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2538 if (tmp->tl_obj != obj)
2540 LASSERT(tmp->tl_shared == shared);
2541 /* found same lock */
2546 if (unlikely(al == NULL))
2549 al->tl_truncate = false;
2551 down_read(&obj->oo_ext_idx_sem);
2553 down_write(&obj->oo_ext_idx_sem);
2554 al->tl_shared = shared;
2555 lu_object_get(&obj->oo_dt.do_lu);
2557 list_add(&al->tl_list, &oh->ot_trunc_locks);
2562 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2564 struct osd_access_lock *al, *tmp;
2566 list_for_each_entry_safe(al, tmp, list, tl_list) {
2568 up_read(&al->tl_obj->oo_ext_idx_sem);
2570 up_write(&al->tl_obj->oo_ext_idx_sem);
2571 osd_object_put(env, al->tl_obj);
2572 list_del(&al->tl_list);
2577 void osd_execute_truncate(struct osd_object *obj)
2579 struct osd_device *d = osd_obj2dev(obj);
2580 struct inode *inode = obj->oo_inode;
2583 /* simulate crash before (in the middle) of delayed truncate */
2584 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2585 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2586 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2588 mutex_lock(&sbi->s_orphan_lock);
2589 list_del_init(&ei->i_orphan);
2590 mutex_unlock(&sbi->s_orphan_lock);
2594 size = i_size_read(inode);
2596 /* if object holds encrypted content, we need to make sure we truncate
2597 * on an encryption unit boundary, or block content will get corrupted
2599 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2600 size & ~LUSTRE_ENCRYPTION_MASK)
2601 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2602 LUSTRE_ENCRYPTION_UNIT_SIZE;
2603 ldiskfs_truncate(inode);
2604 inode_unlock(inode);
2605 if (inode->i_size != size) {
2606 spin_lock(&inode->i_lock);
2607 i_size_write(inode, size);
2608 LDISKFS_I(inode)->i_disksize = size;
2609 spin_unlock(&inode->i_lock);
2610 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2614 * For a partial-page truncate, flush the page to disk immediately to
2615 * avoid data corruption during direct disk write. b=17397
2617 if ((size & ~PAGE_MASK) == 0)
2619 if (osd_use_page_cache(d)) {
2620 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2622 /* Notice we use "wait" version to ensure I/O is complete */
2623 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2624 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2625 size >> PAGE_SHIFT);
2629 void osd_process_truncates(struct list_head *list)
2631 struct osd_access_lock *al;
2633 LASSERT(journal_current_handle() == NULL);
2635 list_for_each_entry(al, list, tl_list) {
2638 if (!al->tl_truncate)
2640 osd_execute_truncate(al->tl_obj);