4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 /* prerequisite for linux/xattr.h */
42 #include <linux/types.h>
43 /* prerequisite for linux/xattr.h */
46 #include <linux/pagevec.h>
49 * struct OBD_{ALLOC,FREE}*()
52 #include <obd_support.h>
54 #include "osd_internal.h"
57 #include <ldiskfs/ldiskfs_extents.h>
59 static inline bool osd_use_page_cache(struct osd_device *d)
61 /* do not use pagecache if write and read caching are disabled */
62 if (d->od_writethrough_cache + d->od_read_cache == 0)
64 /* use pagecache by default */
68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
69 int rw, int line, int pages)
73 LASSERTF(iobuf->dr_elapsed_valid == 0,
74 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
75 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
77 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
79 init_waitqueue_head(&iobuf->dr_wait);
80 atomic_set(&iobuf->dr_numreqs, 0);
85 iobuf->dr_elapsed = ktime_set(0, 0);
86 /* must be counted before, so assert */
88 iobuf->dr_init_at = line;
90 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
91 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
92 LASSERT(iobuf->dr_pg_buf.lb_len >=
93 pages * sizeof(iobuf->dr_pages[0]));
97 /* start with 1MB for 4K blocks */
99 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
102 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
103 (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
105 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
106 iobuf->dr_max_pages = 0;
107 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
108 (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
110 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
111 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
112 if (unlikely(iobuf->dr_blocks == NULL))
115 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
116 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
117 if (unlikely(iobuf->dr_pages == NULL))
120 lu_buf_realloc(&iobuf->dr_lnb_buf,
121 pages * sizeof(iobuf->dr_lnbs[0]));
122 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
123 if (unlikely(iobuf->dr_lnbs == NULL))
126 iobuf->dr_max_pages = pages;
130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
131 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
134 struct niobuf_local *lnb)
136 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
137 iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
138 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
144 int rw = iobuf->dr_rw;
146 if (iobuf->dr_elapsed_valid) {
147 iobuf->dr_elapsed_valid = 0;
148 LASSERT(iobuf->dr_dev == d);
149 LASSERT(iobuf->dr_frags > 0);
150 lprocfs_oh_tally(&d->od_brw_stats.
151 hist[BRW_R_DIO_FRAGS+rw],
153 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
154 ktime_to_ms(iobuf->dr_elapsed));
158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
159 static void dio_complete_routine(struct bio *bio)
161 int error = bio->bi_status;
163 static void dio_complete_routine(struct bio *bio, int error)
166 struct osd_iobuf *iobuf = bio->bi_private;
170 /* CAVEAT EMPTOR: possibly in IRQ context
171 * DO NOT record procfs stats here!!! */
173 if (unlikely(iobuf == NULL)) {
174 CERROR("***** bio->bi_private is NULL! This should never "
175 "happen. Normally, I would crash here, but instead I "
176 "will dump the bio contents to the console. Please "
177 "report this to <https://jira.whamcloud.com/> , along "
178 "with any interesting messages leading up to this point "
179 "(like SCSI errors, perhaps). Because bi_private is "
180 "NULL, I can't wake up the thread that initiated this "
181 "IO - you will probably have to reboot this node.\n");
182 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
183 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
184 bio->bi_next, (unsigned long)bio->bi_flags,
185 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
186 bio_sectors(bio) << 9, bio->bi_end_io,
187 atomic_read(&bio->__bi_cnt),
192 /* the check is outside of the cycle for performance reason -bzzz */
193 if (!bio_data_dir(bio)) {
194 bio_for_each_segment_all(bvl, bio, iter) {
195 if (likely(error == 0))
196 SetPageUptodate(bvl_to_page(bvl));
197 LASSERT(PageLocked(bvl_to_page(bvl)));
199 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
201 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
204 /* any real error is good enough -bzzz */
205 if (error != 0 && iobuf->dr_error == 0)
206 iobuf->dr_error = error;
209 * set dr_elapsed before dr_numreqs turns to 0, otherwise
210 * it's possible that service thread will see dr_numreqs
211 * is zero, but dr_elapsed is not set yet, leading to lost
212 * data in this processing and an assertion in a subsequent
215 if (atomic_read(&iobuf->dr_numreqs) == 1) {
216 ktime_t now = ktime_get();
218 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
219 iobuf->dr_elapsed_valid = 1;
221 if (atomic_dec_and_test(&iobuf->dr_numreqs))
222 wake_up(&iobuf->dr_wait);
224 /* Completed bios used to be chained off iobuf->dr_bios and freed in
225 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
226 * mempool when serious on-disk fragmentation was encountered,
227 * deadlocking the OST. The bios are now released as soon as complete
228 * so the pool cannot be exhausted while IOs are competing. bug 10076 */
232 static void record_start_io(struct osd_iobuf *iobuf, int size)
234 struct osd_device *osd = iobuf->dr_dev;
235 struct obd_histogram *h = osd->od_brw_stats.hist;
238 atomic_inc(&iobuf->dr_numreqs);
240 if (iobuf->dr_rw == 0) {
241 atomic_inc(&osd->od_r_in_flight);
242 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
243 atomic_read(&osd->od_r_in_flight));
244 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
245 } else if (iobuf->dr_rw == 1) {
246 atomic_inc(&osd->od_w_in_flight);
247 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
248 atomic_read(&osd->od_w_in_flight));
249 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
255 static void osd_submit_bio(int rw, struct bio *bio)
257 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
258 #ifdef HAVE_SUBMIT_BIO_2ARGS
259 submit_bio(rw ? WRITE : READ, bio);
266 static int can_be_merged(struct bio *bio, sector_t sector)
271 return bio_end_sector(bio) == sector ? 1 : 0;
274 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
276 * This function will change the data written, thus it should only be
277 * used when checking data integrity feature
279 static void bio_integrity_fault_inject(struct bio *bio)
281 struct bio_vec *bvec;
286 bio_for_each_segment_all(bvec, bio, i) {
287 struct page *page = bvec->bv_page;
297 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
298 unsigned int sectors, int tuple_size)
300 __u16 *expected_guard;
304 expected_guard = expected_guard_buf;
305 for (i = 0; i < sectors; i++) {
306 bio_guard = (__u16 *)bio_prot_buf;
307 if (*bio_guard != *expected_guard) {
308 CERROR("unexpected guard tags on sector %d "
309 "expected guard %u, bio guard "
310 "%u, sectors %u, tuple size %d\n",
311 i, *expected_guard, *bio_guard, sectors,
316 bio_prot_buf += tuple_size;
321 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
322 struct osd_iobuf *iobuf, int index)
324 struct blk_integrity *bi = bdev_get_integrity(bdev);
325 struct bio_integrity_payload *bip = bio->bi_integrity;
326 struct niobuf_local *lnb;
327 unsigned short sector_size = blk_integrity_interval(bi);
328 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
329 bip->bip_vec->bv_offset;
331 sector_t sector = bio_start_sector(bio);
332 unsigned int i, sectors, total;
333 __u16 *expected_guard;
337 bio_for_each_segment_all(bv, bio, i) {
338 lnb = iobuf->dr_lnbs[index];
339 expected_guard = lnb->lnb_guards;
340 sectors = bv->bv_len / sector_size;
341 if (lnb->lnb_guard_rpc) {
342 rc = bio_dif_compare(expected_guard, bio_prot_buf,
343 sectors, bi->tuple_size);
349 bio_prot_buf += sectors * bi->tuple_size;
350 total += sectors * bi->tuple_size;
351 LASSERT(total <= bip_size(bio->bi_integrity));
357 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
358 struct osd_iobuf *iobuf,
359 int start_page_idx, bool fault_inject,
360 bool integrity_enabled)
362 struct super_block *sb = osd_sb(osd);
363 integrity_gen_fn *generate_fn = NULL;
364 integrity_vrfy_fn *verify_fn = NULL;
369 if (!integrity_enabled)
372 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
376 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
380 /* Verify and inject fault only when writing */
381 if (iobuf->dr_rw == 1) {
382 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
383 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
389 if (unlikely(fault_inject))
390 bio_integrity_fault_inject(bio);
396 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
397 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
398 static void dio_integrity_complete_routine(struct bio *bio)
400 static void dio_integrity_complete_routine(struct bio *bio, int error)
403 struct osd_bio_private *bio_private = bio->bi_private;
405 bio->bi_private = bio_private->obp_iobuf;
406 osd_dio_complete_routine(bio, error);
408 OBD_FREE_PTR(bio_private);
410 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
411 #else /* !CONFIG_BLK_DEV_INTEGRITY */
412 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
413 fault_inject, integrity_enabled) 0
414 #endif /* CONFIG_BLK_DEV_INTEGRITY */
416 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
417 bool integrity_enabled, int start_page_idx,
418 struct osd_bio_private **pprivate)
424 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
425 if (integrity_enabled) {
426 struct osd_bio_private *bio_private = NULL;
428 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
429 if (bio_private == NULL)
431 bio->bi_end_io = dio_integrity_complete_routine;
432 bio->bi_private = bio_private;
433 bio_private->obp_start_page_idx = start_page_idx;
434 bio_private->obp_iobuf = iobuf;
435 *pprivate = bio_private;
439 bio->bi_end_io = dio_complete_routine;
440 bio->bi_private = iobuf;
446 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
447 struct osd_iobuf *iobuf)
449 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
450 struct page **pages = iobuf->dr_pages;
451 int npages = iobuf->dr_npages;
452 sector_t *blocks = iobuf->dr_blocks;
453 int total_blocks = npages * blocks_per_page;
454 struct super_block *sb = inode->i_sb;
455 int sector_bits = sb->s_blocksize_bits - 9;
456 unsigned int blocksize = sb->s_blocksize;
457 struct block_device *bdev = sb->s_bdev;
458 struct osd_bio_private *bio_private = NULL;
459 struct bio *bio = NULL;
460 int bio_start_page_idx;
462 unsigned int page_offset;
470 bool integrity_enabled;
471 struct blk_plug plug;
474 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
475 LASSERT(iobuf->dr_npages == npages);
477 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
479 osd_brw_stats_update(osd, iobuf);
480 iobuf->dr_start_time = ktime_get();
482 blk_start_plug(&plug);
483 for (page_idx = 0, block_idx = 0;
485 page_idx++, block_idx += blocks_per_page) {
487 page = pages[page_idx];
488 LASSERT(block_idx + blocks_per_page <= total_blocks);
490 for (i = 0, page_offset = 0;
492 i += nblocks, page_offset += blocksize * nblocks) {
496 if (blocks[block_idx + i] == 0) { /* hole */
497 LASSERTF(iobuf->dr_rw == 0,
498 "page_idx %u, block_idx %u, i %u\n",
499 page_idx, block_idx, i);
500 memset(kmap(page) + page_offset, 0, blocksize);
505 sector = (sector_t)blocks[block_idx + i] << sector_bits;
507 /* Additional contiguous file blocks? */
508 while (i + nblocks < blocks_per_page &&
509 (sector + (nblocks << sector_bits)) ==
510 ((sector_t)blocks[block_idx + i + nblocks] <<
515 can_be_merged(bio, sector) &&
516 bio_add_page(bio, page,
517 blocksize * nblocks, page_offset) != 0)
518 continue; /* added this frag OK */
521 struct request_queue *q = bio_get_queue(bio);
522 unsigned int bi_size = bio_sectors(bio) << 9;
524 /* Dang! I have to fragment this I/O */
526 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
527 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
529 queue_max_sectors(q),
530 bio->bi_phys_segments,
531 queue_max_segments(q));
532 rc = osd_bio_integrity_handle(osd, bio,
533 iobuf, bio_start_page_idx,
534 fault_inject, integrity_enabled);
540 record_start_io(iobuf, bi_size);
541 osd_submit_bio(iobuf->dr_rw, bio);
544 bio_start_page_idx = page_idx;
545 /* allocate new bio */
546 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
547 (npages - page_idx) *
550 CERROR("Can't allocate bio %u*%u = %u pages\n",
551 (npages - page_idx), blocks_per_page,
552 (npages - page_idx) * blocks_per_page);
557 bio_set_dev(bio, bdev);
558 bio_set_sector(bio, sector);
559 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
560 rc = osd_bio_init(bio, iobuf, integrity_enabled,
561 bio_start_page_idx, &bio_private);
567 rc = bio_add_page(bio, page,
568 blocksize * nblocks, page_offset);
574 rc = osd_bio_integrity_handle(osd, bio, iobuf,
583 record_start_io(iobuf, bio_sectors(bio) << 9);
584 osd_submit_bio(iobuf->dr_rw, bio);
589 blk_finish_plug(&plug);
591 /* in order to achieve better IO throughput, we don't wait for writes
592 * completion here. instead we proceed with transaction commit in
593 * parallel and wait for IO completion once transaction is stopped
594 * see osd_trans_stop() for more details -bzzz */
595 if (iobuf->dr_rw == 0 || fault_inject) {
596 wait_event(iobuf->dr_wait,
597 atomic_read(&iobuf->dr_numreqs) == 0);
598 osd_fini_iobuf(osd, iobuf);
602 rc = iobuf->dr_error;
605 OBD_FREE_PTR(bio_private);
611 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
612 struct niobuf_local *lnb)
619 int poff = offset & (PAGE_SIZE - 1);
620 int plen = PAGE_SIZE - poff;
624 lnb->lnb_file_offset = offset;
625 lnb->lnb_page_offset = poff;
627 /* lnb->lnb_flags = rnb->rnb_flags; */
629 lnb->lnb_page = NULL;
631 lnb->lnb_guard_rpc = 0;
632 lnb->lnb_guard_disk = 0;
635 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
646 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
647 loff_t offset, gfp_t gfp_mask)
649 struct osd_thread_info *oti = osd_oti_get(env);
650 struct inode *inode = osd_dt_obj(dt)->oo_inode;
651 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
653 int cur = oti->oti_dio_pages_used;
657 if (osd_use_page_cache(d)) {
658 page = find_or_create_page(inode->i_mapping,
659 offset >> PAGE_SHIFT,
663 LASSERT(!test_bit(PG_private_2, &page->flags));
665 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
668 LASSERT(oti->oti_dio_pages);
670 if (unlikely(!oti->oti_dio_pages[cur])) {
671 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
672 page = alloc_page(gfp_mask);
675 oti->oti_dio_pages[cur] = page;
678 page = oti->oti_dio_pages[cur];
679 LASSERT(!test_bit(PG_private_2, &page->flags));
680 set_bit(PG_private_2, &page->flags);
681 oti->oti_dio_pages_used++;
683 LASSERT(!PageLocked(page));
686 LASSERT(!page->mapping);
687 LASSERT(!PageWriteback(page));
688 ClearPageUptodate(page);
690 page->index = offset >> PAGE_SHIFT;
697 * there are following "locks":
708 * - lock pages, unlock
710 * - lock partial page
716 * Unlock and release pages loaded by osd_bufs_get()
718 * Unlock \a npages pages from \a lnb and drop the refcount on them.
720 * \param env thread execution environment
721 * \param dt dt object undergoing IO (OSD object + methods)
722 * \param lnb array of pages undergoing IO
723 * \param npages number of pages in \a lnb
727 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
728 struct niobuf_local *lnb, int npages)
730 struct osd_thread_info *oti = osd_oti_get(env);
734 ll_pagevec_init(&pvec, 0);
736 for (i = 0; i < npages; i++) {
737 struct page *page = lnb[i].lnb_page;
742 /* if the page isn't cached, then reset uptodate
743 * to prevent reuse */
744 if (test_bit(PG_private_2, &page->flags)) {
745 clear_bit(PG_private_2, &page->flags);
746 ClearPageUptodate(page);
747 if (lnb[i].lnb_locked)
749 oti->oti_dio_pages_used--;
751 if (lnb[i].lnb_locked)
753 if (pagevec_add(&pvec, page) == 0)
754 pagevec_release(&pvec);
756 dt_object_put(env, dt);
758 lnb[i].lnb_page = NULL;
761 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
763 /* Release any partial pagevec */
764 pagevec_release(&pvec);
770 * Load and lock pages undergoing IO
772 * Pages as described in the \a lnb array are fetched (from disk or cache)
773 * and locked for IO by the caller.
775 * DLM locking protects us from write and truncate competing for same region,
776 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
777 * It's possible the writeout on a such a page is in progress when we access
778 * it. It's also possible that during this writeout we put new (partial) data
779 * into the page, but won't be able to proceed in filter_commitrw_write().
780 * Therefore, just wait for writeout completion as it should be rare enough.
782 * \param env thread execution environment
783 * \param dt dt object undergoing IO (OSD object + methods)
784 * \param pos byte offset of IO start
785 * \param len number of bytes of IO
786 * \param lnb array of extents undergoing IO
787 * \param rw read or write operation, and other flags
788 * \param capa capabilities
790 * \retval pages (zero or more) loaded successfully
791 * \retval -ENOMEM on memory/page allocation error
793 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
794 loff_t pos, ssize_t len, struct niobuf_local *lnb,
795 enum dt_bufs_type rw)
797 struct osd_thread_info *oti = osd_oti_get(env);
798 struct osd_object *obj = osd_dt_obj(dt);
799 int npages, i, rc = 0;
802 LASSERT(obj->oo_inode);
804 if (!osd_use_page_cache(osd_obj2dev(obj))) {
805 if (unlikely(!oti->oti_dio_pages)) {
806 OBD_ALLOC(oti->oti_dio_pages,
807 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
808 if (!oti->oti_dio_pages)
813 osd_map_remote_to_local(pos, len, &npages, lnb);
815 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
816 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
818 for (i = 0; i < npages; i++, lnb++) {
819 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
821 if (lnb->lnb_page == NULL)
822 GOTO(cleanup, rc = -ENOMEM);
825 wait_on_page_writeback(lnb->lnb_page);
826 BUG_ON(PageWriteback(lnb->lnb_page));
828 lu_object_get(&dt->do_lu);
835 osd_bufs_put(env, dt, lnb - i, i);
839 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
840 int pages, sector_t *blocks,
843 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
845 struct page *fp = NULL;
847 pgoff_t max_page_index;
848 handle_t *handle = NULL;
850 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
852 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
853 inode->i_ino, pages, (*page)->index);
856 create = LDISKFS_GET_BLOCKS_CREATE;
857 handle = ldiskfs_journal_current_handle();
858 LASSERT(handle != NULL);
859 rc = osd_attach_jinode(inode);
863 /* pages are sorted already. so, we just have to find
864 * contig. space and process them properly */
866 long blen, total = 0;
867 struct ldiskfs_map_blocks map = { 0 };
869 if (fp == NULL) { /* start new extent */
874 } else if (fp->index + clen == (*page)->index) {
875 /* continue the extent */
881 if (fp->index + clen >= max_page_index)
882 GOTO(cleanup, rc = -EFBIG);
883 /* process found extent */
884 map.m_lblk = fp->index * blocks_per_page;
885 map.m_len = blen = clen * blocks_per_page;
887 rc = ldiskfs_map_blocks(handle, inode, &map, create);
890 for (; total < blen && c < map.m_len; c++, total++) {
892 *(blocks + total) = 0;
896 *(blocks + total) = map.m_pblk + c;
897 /* unmap any possible underlying
898 * metadata from the block device
899 * mapping. bug 6998. */
900 if ((map.m_flags & LDISKFS_MAP_NEW) &&
909 if (rc == 0 && total < blen) {
910 map.m_lblk = fp->index * blocks_per_page + total;
911 map.m_len = blen - total;
917 /* look for next extent */
919 blocks += blocks_per_page * clen;
925 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
926 struct niobuf_local *lnb, int npages)
928 struct osd_thread_info *oti = osd_oti_get(env);
929 struct osd_iobuf *iobuf = &oti->oti_iobuf;
930 struct inode *inode = osd_dt_obj(dt)->oo_inode;
931 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
943 rc = osd_init_iobuf(osd, iobuf, 0, npages);
944 if (unlikely(rc != 0))
947 isize = i_size_read(inode);
948 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
950 if (osd->od_writethrough_cache)
952 if (isize > osd->od_readcache_max_filesize)
956 for (i = 0; i < npages; i++) {
959 generic_error_remove_page(inode->i_mapping,
963 * till commit the content of the page is undefined
964 * we'll set it uptodate once bulk is done. otherwise
965 * subsequent reads can access non-stable data
967 ClearPageUptodate(lnb[i].lnb_page);
969 if (lnb[i].lnb_len == PAGE_SIZE)
972 if (maxidx >= lnb[i].lnb_page->index) {
973 osd_iobuf_add_page(iobuf, &lnb[i]);
976 char *p = kmap(lnb[i].lnb_page);
978 off = lnb[i].lnb_page_offset;
981 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
984 memset(p + off, 0, PAGE_SIZE - off);
985 kunmap(lnb[i].lnb_page);
989 timediff = ktime_us_delta(end, start);
990 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
992 if (iobuf->dr_npages) {
993 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
995 iobuf->dr_blocks, 0);
996 if (likely(rc == 0)) {
997 rc = osd_do_bio(osd, inode, iobuf);
998 /* do IO stats for preparation reads */
999 osd_fini_iobuf(osd, iobuf);
1005 struct osd_fextent {
1008 unsigned int mapped:1;
1011 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1012 struct osd_fextent *cached_extent)
1014 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1015 sector_t block = offset >> inode->i_blkbits;
1017 struct fiemap_extent_info fei = { 0 };
1018 struct fiemap_extent fe = { 0 };
1019 mm_segment_t saved_fs;
1022 if (block >= cached_extent->start && block < cached_extent->end)
1023 return cached_extent->mapped;
1025 if (i_size_read(inode) == 0)
1028 /* Beyond EOF, must not be mapped */
1029 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1032 fei.fi_extents_max = 1;
1033 fei.fi_extents_start = &fe;
1035 saved_fs = get_fs();
1037 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1042 start = fe.fe_logical >> inode->i_blkbits;
1044 if (start > block) {
1045 cached_extent->start = block;
1046 cached_extent->end = start;
1047 cached_extent->mapped = 0;
1049 cached_extent->start = start;
1050 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1052 cached_extent->mapped = 1;
1055 return cached_extent->mapped;
1058 static int osd_declare_write_commit(const struct lu_env *env,
1059 struct dt_object *dt,
1060 struct niobuf_local *lnb, int npages,
1061 struct thandle *handle)
1063 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1064 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1065 struct osd_thandle *oh;
1073 long long quota_space = 0;
1074 struct osd_fextent extent = { 0 };
1075 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1078 LASSERT(handle != NULL);
1079 oh = container_of0(handle, struct osd_thandle, ot_super);
1080 LASSERT(oh->ot_handle == NULL);
1084 /* calculate number of extents (probably better to pass nb) */
1085 for (i = 0; i < npages; i++) {
1086 if (i && lnb[i].lnb_file_offset !=
1087 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1090 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1091 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1093 quota_space += PAGE_SIZE;
1095 /* ignore quota for the whole request if any page is from
1096 * client cache or written by root.
1098 * XXX once we drop the 1.8 client support, the checking
1099 * for whether page is from cache can be simplified as:
1100 * !(lnb[i].flags & OBD_BRW_SYNC)
1102 * XXX we could handle this on per-lnb basis as done by
1104 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1105 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1107 declare_flags |= OSD_QID_FORCE;
1111 * each extent can go into new leaf causing a split
1112 * 5 is max tree depth: inode + 4 index blocks
1113 * with blockmaps, depth is 3 at most
1115 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1117 * many concurrent threads may grow tree by the time
1118 * our transaction starts. so, consider 2 is a min depth
1120 depth = ext_depth(inode);
1121 depth = max(depth, 1) + 1;
1123 credits++; /* inode */
1124 credits += depth * 2 * extents;
1128 credits++; /* inode */
1129 credits += depth * extents;
1132 /* quota space for metadata blocks */
1133 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1135 /* quota space should be reported in 1K blocks */
1136 quota_space = toqb(quota_space);
1138 /* each new block can go in different group (bitmap + gd) */
1140 /* we can't dirty more bitmap blocks than exist */
1141 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1142 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1144 credits += newblocks;
1146 /* we can't dirty more gd blocks than exist */
1147 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1148 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1150 credits += newblocks;
1152 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1154 /* make sure the over quota flags were not set */
1155 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1157 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1158 i_projid_read(inode), quota_space, oh,
1159 osd_dt_obj(dt), &flags, declare_flags);
1161 /* we need only to store the overquota flags in the first lnb for
1162 * now, once we support multiple objects BRW, this code needs be
1164 if (flags & QUOTA_FL_OVER_USRQUOTA)
1165 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1166 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1167 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1168 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1169 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1172 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1177 /* Check if a block is allocated or not */
1178 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1179 struct niobuf_local *lnb, int npages,
1180 struct thandle *thandle)
1182 struct osd_thread_info *oti = osd_oti_get(env);
1183 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1184 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1185 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1191 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1192 if (unlikely(rc != 0))
1195 isize = i_size_read(inode);
1196 dquot_initialize(inode);
1198 for (i = 0; i < npages; i++) {
1199 if (lnb[i].lnb_rc == -ENOSPC &&
1200 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1201 /* Allow the write to proceed if overwriting an
1206 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1207 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1209 LASSERT(lnb[i].lnb_page);
1210 generic_error_remove_page(inode->i_mapping,
1215 LASSERT(PageLocked(lnb[i].lnb_page));
1216 LASSERT(!PageWriteback(lnb[i].lnb_page));
1218 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1219 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1222 * Since write and truncate are serialized by oo_sem, even
1223 * partial-page truncate should not leave dirty pages in the
1226 LASSERT(!PageDirty(lnb[i].lnb_page));
1228 SetPageUptodate(lnb[i].lnb_page);
1230 osd_iobuf_add_page(iobuf, &lnb[i]);
1233 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1235 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1237 } else if (iobuf->dr_npages > 0) {
1238 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1240 iobuf->dr_blocks, 1);
1242 /* no pages to write, no transno is needed */
1243 thandle->th_local = 1;
1246 if (likely(rc == 0)) {
1247 spin_lock(&inode->i_lock);
1248 if (isize > i_size_read(inode)) {
1249 i_size_write(inode, isize);
1250 LDISKFS_I(inode)->i_disksize = isize;
1251 spin_unlock(&inode->i_lock);
1252 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1254 spin_unlock(&inode->i_lock);
1257 rc = osd_do_bio(osd, inode, iobuf);
1258 /* we don't do stats here as in read path because
1259 * write is async: we'll do this in osd_put_bufs() */
1261 osd_fini_iobuf(osd, iobuf);
1264 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1266 if (unlikely(rc != 0)) {
1267 /* if write fails, we should drop pages from the cache */
1268 for (i = 0; i < npages; i++) {
1269 if (lnb[i].lnb_page == NULL)
1271 LASSERT(PageLocked(lnb[i].lnb_page));
1272 generic_error_remove_page(inode->i_mapping,
1280 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1281 struct niobuf_local *lnb, int npages)
1283 struct osd_thread_info *oti = osd_oti_get(env);
1284 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1285 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1286 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1287 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1294 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1295 if (unlikely(rc != 0))
1298 isize = i_size_read(inode);
1300 if (osd->od_read_cache)
1302 if (isize > osd->od_readcache_max_filesize)
1305 start = ktime_get();
1306 for (i = 0; i < npages; i++) {
1308 if (isize <= lnb[i].lnb_file_offset)
1309 /* If there's no more data, abort early.
1310 * lnb->lnb_rc == 0, so it's easy to detect later. */
1313 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1314 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1316 lnb[i].lnb_rc = lnb[i].lnb_len;
1318 /* Bypass disk read if fail_loc is set properly */
1319 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1320 SetPageUptodate(lnb[i].lnb_page);
1323 generic_error_remove_page(inode->i_mapping,
1326 if (PageUptodate(lnb[i].lnb_page)) {
1328 unlock_page(lnb[i].lnb_page);
1331 osd_iobuf_add_page(iobuf, &lnb[i]);
1333 /* no need to unlock in osd_bufs_put(), the sooner page is
1334 * unlocked, the earlier another client can access it.
1335 * notice real unlock_page() can be called few lines
1336 * below after osd_do_bio(). lnb is a per-thread, so it's
1337 * fine to have PG_locked and lnb_locked inconsistent here */
1338 lnb[i].lnb_locked = 0;
1341 timediff = ktime_us_delta(end, start);
1342 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1344 if (cache_hits != 0)
1345 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1347 if (cache_misses != 0)
1348 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1350 if (cache_hits + cache_misses != 0)
1351 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1352 cache_hits + cache_misses);
1354 if (iobuf->dr_npages) {
1355 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1357 iobuf->dr_blocks, 0);
1358 rc = osd_do_bio(osd, inode, iobuf);
1360 /* IO stats will be done in osd_bufs_put() */
1362 /* early release to let others read data during the bulk */
1363 for (i = 0; i < iobuf->dr_npages; i++) {
1364 LASSERT(PageLocked(iobuf->dr_pages[i]));
1365 unlock_page(iobuf->dr_pages[i]);
1373 * XXX: Another layering violation for now.
1375 * We don't want to use ->f_op->read methods, because generic file write
1377 * - serializes on ->i_sem, and
1379 * - does a lot of extra work like balance_dirty_pages(),
1381 * which doesn't work for globally shared files like /last_rcvd.
1383 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1385 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1387 memcpy(buffer, (char *)ei->i_data, buflen);
1392 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1394 struct buffer_head *bh;
1395 unsigned long block;
1401 /* prevent reading after eof */
1402 spin_lock(&inode->i_lock);
1403 if (i_size_read(inode) < *offs + size) {
1404 loff_t diff = i_size_read(inode) - *offs;
1405 spin_unlock(&inode->i_lock);
1408 "size %llu is too short to read @%llu\n",
1409 i_size_read(inode), *offs);
1411 } else if (diff == 0) {
1417 spin_unlock(&inode->i_lock);
1420 blocksize = 1 << inode->i_blkbits;
1423 block = *offs >> inode->i_blkbits;
1424 boffs = *offs & (blocksize - 1);
1425 csize = min(blocksize - boffs, size);
1426 bh = __ldiskfs_bread(NULL, inode, block, 0);
1428 CERROR("%s: can't read %u@%llu on ino %lu: "
1429 "rc = %ld\n", osd_ino2name(inode),
1430 csize, *offs, inode->i_ino,
1436 memcpy(buf, bh->b_data + boffs, csize);
1439 memset(buf, 0, csize);
1449 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1450 struct lu_buf *buf, loff_t *pos)
1452 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1455 /* Read small symlink from inode body as we need to maintain correct
1456 * on-disk symlinks for ldiskfs.
1458 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1459 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1460 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1462 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1467 static inline int osd_extents_enabled(struct super_block *sb,
1468 struct inode *inode)
1470 if (inode != NULL) {
1471 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1473 } else if (ldiskfs_has_feature_extents(sb)) {
1479 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1480 const loff_t size, const loff_t pos,
1483 int credits, bits, bs, i;
1485 bits = sb->s_blocksize_bits;
1488 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1489 * we do not expect blockmaps on the large files,
1490 * so let's shrink it to 2 levels (4GB files) */
1492 /* this is default reservation: 2 levels */
1493 credits = (blocks + 2) * 3;
1495 /* actual offset is unknown, hard to optimize */
1499 /* now check for few specific cases to optimize */
1500 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1503 /* allocate if not allocated */
1504 if (inode == NULL) {
1505 credits += blocks * 2;
1508 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1509 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1510 if (LDISKFS_I(inode)->i_data[i] == 0)
1513 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1514 /* single indirect */
1515 credits = blocks * 3;
1516 if (inode == NULL ||
1517 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1520 /* The indirect block may be modified. */
1527 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1528 const struct lu_buf *buf, loff_t _pos,
1529 struct thandle *handle)
1531 struct osd_object *obj = osd_dt_obj(dt);
1532 struct inode *inode = obj->oo_inode;
1533 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1534 struct osd_thandle *oh;
1535 int rc = 0, est = 0, credits, blocks, allocated = 0;
1541 LASSERT(buf != NULL);
1542 LASSERT(handle != NULL);
1544 oh = container_of0(handle, struct osd_thandle, ot_super);
1545 LASSERT(oh->ot_handle == NULL);
1548 bits = sb->s_blocksize_bits;
1552 /* if this is an append, then we
1553 * should expect cross-block record */
1559 /* blocks to modify */
1560 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1561 LASSERT(blocks > 0);
1563 if (inode != NULL && _pos != -1) {
1564 /* object size in blocks */
1565 est = (i_size_read(inode) + bs - 1) >> bits;
1566 allocated = inode->i_blocks >> (bits - 9);
1567 if (pos + size <= i_size_read(inode) && est <= allocated) {
1568 /* looks like an overwrite, no need to modify tree */
1570 /* no need to modify i_size */
1575 if (osd_extents_enabled(sb, inode)) {
1577 * many concurrent threads may grow tree by the time
1578 * our transaction starts. so, consider 2 is a min depth
1579 * for every level we may need to allocate a new block
1580 * and take some entries from the old one. so, 3 blocks
1581 * to allocate (bitmap, gd, itself) + old block - 4 per
1584 depth = inode != NULL ? ext_depth(inode) : 0;
1585 depth = max(depth, 1) + 1;
1587 /* if not append, then split may need to modify
1588 * existing blocks moving entries into the new ones */
1591 /* blocks to store data: bitmap,gd,itself */
1592 credits += blocks * 3;
1594 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1596 /* if inode is created as part of the transaction,
1597 * then it's counted already by the creation method */
1603 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1605 /* dt_declare_write() is usually called for system objects, such
1606 * as llog or last_rcvd files. We needn't enforce quota on those
1607 * objects, so always set the lqi_space as 0. */
1609 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1611 i_projid_read(inode), 0,
1612 oh, obj, NULL, OSD_QID_BLK);
1615 rc = osd_trunc_lock(obj, oh, true);
1620 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1622 /* LU-2634: clear the extent format for fast symlink */
1623 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1625 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1626 spin_lock(&inode->i_lock);
1627 LDISKFS_I(inode)->i_disksize = buflen;
1628 i_size_write(inode, buflen);
1629 spin_unlock(&inode->i_lock);
1630 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1635 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1636 int write_NUL, loff_t *offs, handle_t *handle)
1638 struct buffer_head *bh = NULL;
1639 loff_t offset = *offs;
1640 loff_t new_size = i_size_read(inode);
1641 unsigned long block;
1642 int blocksize = 1 << inode->i_blkbits;
1646 int dirty_inode = 0;
1650 * long symlink write does not count the NUL terminator in
1651 * bufsize, we write it, and the inode's file size does not
1652 * count the NUL terminator as well.
1654 ((char *)buf)[bufsize] = '\0';
1658 while (bufsize > 0) {
1659 int credits = handle->h_buffer_credits;
1664 block = offset >> inode->i_blkbits;
1665 boffs = offset & (blocksize - 1);
1666 size = min(blocksize - boffs, bufsize);
1667 bh = __ldiskfs_bread(handle, inode, block, 1);
1668 if (IS_ERR_OR_NULL(bh)) {
1676 CERROR("%s: error reading offset %llu (block %lu, "
1677 "size %d, offs %llu), credits %d/%d: rc = %d\n",
1678 inode->i_sb->s_id, offset, block, bufsize, *offs,
1679 credits, handle->h_buffer_credits, err);
1683 err = ldiskfs_journal_get_write_access(handle, bh);
1685 CERROR("journal_get_write_access() returned error %d\n",
1689 LASSERTF(boffs + size <= bh->b_size,
1690 "boffs %d size %d bh->b_size %lu\n",
1691 boffs, size, (unsigned long)bh->b_size);
1692 memcpy(bh->b_data + boffs, buf, size);
1693 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1697 if (offset + size > new_size)
1698 new_size = offset + size;
1708 /* correct in-core and on-disk sizes */
1709 if (new_size > i_size_read(inode)) {
1710 spin_lock(&inode->i_lock);
1711 if (new_size > i_size_read(inode))
1712 i_size_write(inode, new_size);
1713 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1714 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1717 spin_unlock(&inode->i_lock);
1719 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1727 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1728 const struct lu_buf *buf, loff_t *pos,
1729 struct thandle *handle)
1731 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1732 struct osd_thandle *oh;
1736 LASSERT(dt_object_exists(dt));
1738 LASSERT(handle != NULL);
1739 LASSERT(inode != NULL);
1740 dquot_initialize(inode);
1742 /* XXX: don't check: one declared chunk can be used many times */
1743 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1745 oh = container_of(handle, struct osd_thandle, ot_super);
1746 LASSERT(oh->ot_handle->h_transaction != NULL);
1747 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1749 /* Write small symlink to inode body as we need to maintain correct
1750 * on-disk symlinks for ldiskfs.
1751 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1752 * does not count it in.
1754 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1755 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1756 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1758 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1759 buf->lb_len, is_link, pos,
1762 result = buf->lb_len;
1764 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1769 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1770 __u64 start, __u64 end, struct thandle *th)
1772 struct osd_thandle *oh;
1773 struct inode *inode;
1778 oh = container_of(th, struct osd_thandle, ot_super);
1781 * we don't need to reserve credits for whole truncate
1782 * it's not possible as truncate may need to free too many
1783 * blocks and that won't fit a single transaction. instead
1784 * we reserve credits to change i_size and put inode onto
1785 * orphan list. if needed truncate will extend or restart
1788 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1789 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1791 inode = osd_dt_obj(dt)->oo_inode;
1794 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1795 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1799 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1804 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1805 __u64 start, __u64 end, struct thandle *th)
1807 struct osd_object *obj = osd_dt_obj(dt);
1808 struct osd_device *osd = osd_obj2dev(obj);
1809 struct inode *inode = obj->oo_inode;
1810 struct osd_access_lock *al;
1811 struct osd_thandle *oh;
1812 int rc = 0, found = 0;
1816 LASSERT(end == OBD_OBJECT_EOF);
1817 LASSERT(dt_object_exists(dt));
1818 LASSERT(osd_invariant(obj));
1819 LASSERT(inode != NULL);
1820 dquot_initialize(inode);
1823 oh = container_of(th, struct osd_thandle, ot_super);
1824 LASSERT(oh->ot_handle->h_transaction != NULL);
1826 /* we used to skip truncate to current size to
1827 * optimize truncates on OST. with DoM we can
1828 * get attr_set to set specific size (MDS_REINT)
1829 * and then get truncate RPC which essentially
1830 * would be skipped. this is bad.. so, disable
1831 * this optimization on MDS till the client stop
1832 * to sent MDS_REINT (LU-11033) -bzzz */
1833 if (osd->od_is_ost && i_size_read(inode) == start)
1836 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1838 spin_lock(&inode->i_lock);
1839 if (i_size_read(inode) < start)
1841 i_size_write(inode, start);
1842 spin_unlock(&inode->i_lock);
1843 ll_truncate_pagecache(inode, start);
1845 /* optimize grow case */
1847 osd_execute_truncate(obj);
1851 /* add to orphan list to ensure truncate completion
1852 * if this transaction succeed. ldiskfs_truncate()
1853 * will take the inode out of the list */
1854 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
1858 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
1859 if (obj != al->tl_obj)
1861 LASSERT(al->tl_shared == 0);
1863 /* do actual truncate in osd_trans_stop() */
1864 al->tl_truncate = 1;
1873 static int fiemap_check_ranges(struct inode *inode,
1874 u64 start, u64 len, u64 *new_len)
1883 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1884 maxbytes = inode->i_sb->s_maxbytes;
1886 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1888 if (start > maxbytes)
1892 * Shrink request scope to what the fs can actually handle.
1894 if (len > maxbytes || (maxbytes - len) < start)
1895 *new_len = maxbytes - start;
1900 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1901 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
1903 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1906 struct fiemap_extent_info fieinfo = {0, };
1907 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1910 mm_segment_t cur_fs;
1913 if (inode->i_op->fiemap == NULL)
1916 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1919 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1923 fieinfo.fi_flags = fm->fm_flags;
1924 fieinfo.fi_extents_max = fm->fm_extent_count;
1925 fieinfo.fi_extents_start = fm->fm_extents;
1927 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1928 filemap_write_and_wait(inode->i_mapping);
1930 /* Save previous value address limit */
1932 /* Set the address limit of the kernel */
1935 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1936 fm->fm_flags = fieinfo.fi_flags;
1937 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1939 /* Restore the previous address limt */
1945 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1946 __u64 start, __u64 end, enum lu_ladvise_type advice)
1949 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1953 case LU_LADVISE_DONTNEED:
1956 invalidate_mapping_pages(inode->i_mapping,
1957 start >> PAGE_SHIFT,
1958 (end - 1) >> PAGE_SHIFT);
1969 * in some cases we may need declare methods for objects being created
1970 * e.g., when we create symlink
1972 const struct dt_body_operations osd_body_ops_new = {
1973 .dbo_declare_write = osd_declare_write,
1976 const struct dt_body_operations osd_body_ops = {
1977 .dbo_read = osd_read,
1978 .dbo_declare_write = osd_declare_write,
1979 .dbo_write = osd_write,
1980 .dbo_bufs_get = osd_bufs_get,
1981 .dbo_bufs_put = osd_bufs_put,
1982 .dbo_write_prep = osd_write_prep,
1983 .dbo_declare_write_commit = osd_declare_write_commit,
1984 .dbo_write_commit = osd_write_commit,
1985 .dbo_read_prep = osd_read_prep,
1986 .dbo_declare_punch = osd_declare_punch,
1987 .dbo_punch = osd_punch,
1988 .dbo_fiemap_get = osd_fiemap_get,
1989 .dbo_ladvise = osd_ladvise,
1993 * Get a truncate lock
1995 * In order to take multi-transaction truncate out of main transaction we let
1996 * the caller grab a lock on the object passed. the lock can be shared (for
1997 * writes) and exclusive (for truncate). It's not allowed to mix truncate
1998 * and write in the same transaction handle (do not confuse with big ldiskfs
1999 * transaction containing lots of handles).
2000 * The lock must be taken at declaration.
2002 * \param obj object to lock
2004 * \shared shared or exclusive
2006 * \retval 0 lock is granted
2007 * \retval -NOMEM no memory to allocate lock
2009 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2011 struct osd_access_lock *al, *tmp;
2016 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2017 if (tmp->tl_obj != obj)
2019 LASSERT(tmp->tl_shared == shared);
2020 /* found same lock */
2025 if (unlikely(al == NULL))
2028 al->tl_truncate = false;
2030 down_read(&obj->oo_ext_idx_sem);
2032 down_write(&obj->oo_ext_idx_sem);
2033 al->tl_shared = shared;
2035 list_add(&al->tl_list, &oh->ot_trunc_locks);
2040 void osd_trunc_unlock_all(struct list_head *list)
2042 struct osd_access_lock *al, *tmp;
2043 list_for_each_entry_safe(al, tmp, list, tl_list) {
2045 up_read(&al->tl_obj->oo_ext_idx_sem);
2047 up_write(&al->tl_obj->oo_ext_idx_sem);
2048 list_del(&al->tl_list);
2053 void osd_execute_truncate(struct osd_object *obj)
2055 struct osd_device *d = osd_obj2dev(obj);
2056 struct inode *inode = obj->oo_inode;
2059 /* simulate crash before (in the middle) of delayed truncate */
2060 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2061 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2062 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2064 mutex_lock(&sbi->s_orphan_lock);
2065 list_del_init(&ei->i_orphan);
2066 mutex_unlock(&sbi->s_orphan_lock);
2070 ldiskfs_truncate(inode);
2073 * For a partial-page truncate, flush the page to disk immediately to
2074 * avoid data corruption during direct disk write. b=17397
2076 size = i_size_read(inode);
2077 if ((size & ~PAGE_MASK) == 0)
2079 if (osd_use_page_cache(d)) {
2080 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2082 /* Notice we use "wait" version to ensure I/O is complete */
2083 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2084 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2085 size >> PAGE_SHIFT);
2089 void osd_process_truncates(struct list_head *list)
2091 struct osd_access_lock *al;
2093 LASSERT(journal_current_handle() == NULL);
2095 list_for_each_entry(al, list, tl_list) {
2098 if (!al->tl_truncate)
2100 osd_execute_truncate(al->tl_obj);