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;
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 */
525 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
526 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
527 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
529 queue_max_sectors(q),
530 bio_phys_segments(q, bio),
531 queue_max_phys_segments(q),
532 0, queue_max_hw_segments(q));
533 rc = osd_bio_integrity_handle(osd, bio,
534 iobuf, bio_start_page_idx,
535 fault_inject, integrity_enabled);
541 record_start_io(iobuf, bi_size);
542 osd_submit_bio(iobuf->dr_rw, bio);
545 bio_start_page_idx = page_idx;
546 /* allocate new bio */
547 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
548 (npages - page_idx) *
551 CERROR("Can't allocate bio %u*%u = %u pages\n",
552 (npages - page_idx), blocks_per_page,
553 (npages - page_idx) * blocks_per_page);
558 bio_set_dev(bio, bdev);
559 bio_set_sector(bio, sector);
560 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
561 rc = osd_bio_init(bio, iobuf, integrity_enabled,
562 bio_start_page_idx, &bio_private);
568 rc = bio_add_page(bio, page,
569 blocksize * nblocks, page_offset);
575 rc = osd_bio_integrity_handle(osd, bio, iobuf,
584 record_start_io(iobuf, bio_sectors(bio) << 9);
585 osd_submit_bio(iobuf->dr_rw, bio);
590 blk_finish_plug(&plug);
592 /* in order to achieve better IO throughput, we don't wait for writes
593 * completion here. instead we proceed with transaction commit in
594 * parallel and wait for IO completion once transaction is stopped
595 * see osd_trans_stop() for more details -bzzz */
596 if (iobuf->dr_rw == 0 || fault_inject) {
597 wait_event(iobuf->dr_wait,
598 atomic_read(&iobuf->dr_numreqs) == 0);
599 osd_fini_iobuf(osd, iobuf);
603 rc = iobuf->dr_error;
606 OBD_FREE_PTR(bio_private);
612 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
613 struct niobuf_local *lnb)
620 int poff = offset & (PAGE_SIZE - 1);
621 int plen = PAGE_SIZE - poff;
625 lnb->lnb_file_offset = offset;
626 lnb->lnb_page_offset = poff;
628 /* lnb->lnb_flags = rnb->rnb_flags; */
630 lnb->lnb_page = NULL;
632 lnb->lnb_guard_rpc = 0;
633 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;
741 LASSERT(PageLocked(page));
743 /* if the page isn't cached, then reset uptodate
744 * to prevent reuse */
745 if (test_bit(PG_private_2, &page->flags)) {
746 clear_bit(PG_private_2, &page->flags);
747 ClearPageUptodate(page);
749 oti->oti_dio_pages_used--;
752 if (pagevec_add(&pvec, page) == 0)
753 pagevec_release(&pvec);
755 dt_object_put(env, dt);
757 lnb[i].lnb_page = NULL;
760 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
762 /* Release any partial pagevec */
763 pagevec_release(&pvec);
769 * Load and lock pages undergoing IO
771 * Pages as described in the \a lnb array are fetched (from disk or cache)
772 * and locked for IO by the caller.
774 * DLM locking protects us from write and truncate competing for same region,
775 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
776 * It's possible the writeout on a such a page is in progress when we access
777 * it. It's also possible that during this writeout we put new (partial) data
778 * into the page, but won't be able to proceed in filter_commitrw_write().
779 * Therefore, just wait for writeout completion as it should be rare enough.
781 * \param env thread execution environment
782 * \param dt dt object undergoing IO (OSD object + methods)
783 * \param pos byte offset of IO start
784 * \param len number of bytes of IO
785 * \param lnb array of extents undergoing IO
786 * \param rw read or write operation, and other flags
787 * \param capa capabilities
789 * \retval pages (zero or more) loaded successfully
790 * \retval -ENOMEM on memory/page allocation error
792 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
793 loff_t pos, ssize_t len, struct niobuf_local *lnb,
794 enum dt_bufs_type rw)
796 struct osd_thread_info *oti = osd_oti_get(env);
797 struct osd_object *obj = osd_dt_obj(dt);
798 int npages, i, rc = 0;
801 LASSERT(obj->oo_inode);
803 if (!osd_use_page_cache(osd_obj2dev(obj))) {
804 if (unlikely(!oti->oti_dio_pages)) {
805 OBD_ALLOC(oti->oti_dio_pages,
806 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
807 if (!oti->oti_dio_pages)
812 osd_map_remote_to_local(pos, len, &npages, lnb);
814 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
815 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
817 for (i = 0; i < npages; i++, lnb++) {
818 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
820 if (lnb->lnb_page == NULL)
821 GOTO(cleanup, rc = -ENOMEM);
823 wait_on_page_writeback(lnb->lnb_page);
824 BUG_ON(PageWriteback(lnb->lnb_page));
826 lu_object_get(&dt->do_lu);
833 osd_bufs_put(env, dt, lnb - i, i);
837 #ifndef HAVE_LDISKFS_MAP_BLOCKS
839 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
840 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
851 static long ldiskfs_ext_find_goal(struct inode *inode,
852 struct ldiskfs_ext_path *path,
853 unsigned long block, int *aflags)
855 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
856 unsigned long bg_start;
857 unsigned long colour;
861 struct ldiskfs_extent *ex;
862 depth = path->p_depth;
864 /* try to predict block placement */
865 if ((ex = path[depth].p_ext))
866 return ldiskfs_ext_pblock(ex) +
867 (block - le32_to_cpu(ex->ee_block));
869 /* it looks index is empty
870 * try to find starting from index itself */
871 if (path[depth].p_bh)
872 return path[depth].p_bh->b_blocknr;
875 /* OK. use inode's group */
876 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
877 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
878 colour = (current->pid % 16) *
879 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
880 return bg_start + colour + block;
883 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
884 struct ldiskfs_ext_path *path,
885 unsigned long block, unsigned long *count,
888 struct ldiskfs_allocation_request ar;
889 unsigned long pblock;
892 /* find neighbour allocated blocks */
894 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
898 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
902 /* allocate new block */
903 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
907 ar.flags = LDISKFS_MB_HINT_DATA;
908 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
913 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
914 struct ldiskfs_ext_path *path,
915 struct ldiskfs_ext_cache *cex,
916 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
917 struct ldiskfs_extent *ex,
921 struct bpointers *bp = cbdata;
922 struct ldiskfs_extent nex;
923 unsigned long pblock = 0;
929 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
930 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
932 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
938 if (bp->create == 0) {
940 if (cex->ec_block < bp->start)
941 i = bp->start - cex->ec_block;
942 if (i >= cex->ec_len)
943 CERROR("nothing to do?! i = %d, e_num = %u\n",
945 for (; i < cex->ec_len && bp->num; i++) {
955 tgen = LDISKFS_I(inode)->i_ext_generation;
956 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
958 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
959 count + LDISKFS_ALLOC_NEEDED + 1);
960 if (IS_ERR(handle)) {
961 return PTR_ERR(handle);
964 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
965 /* the tree has changed. so path can be invalid at moment */
966 ldiskfs_journal_stop(handle);
970 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
971 * protected by i_data_sem as whole. so we patch it to store
972 * generation to path and now verify the tree hasn't changed */
973 down_write((&LDISKFS_I(inode)->i_data_sem));
975 /* validate extent, make sure the extent tree does not changed */
976 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
977 /* cex is invalid, try again */
978 up_write(&LDISKFS_I(inode)->i_data_sem);
979 ldiskfs_journal_stop(handle);
984 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
987 BUG_ON(count > cex->ec_len);
989 /* insert new extent */
990 nex.ee_block = cpu_to_le32(cex->ec_block);
991 ldiskfs_ext_store_pblock(&nex, pblock);
992 nex.ee_len = cpu_to_le16(count);
993 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
995 /* free data blocks we just allocated */
996 /* not a good idea to call discard here directly,
997 * but otherwise we'd need to call it every free() */
998 ldiskfs_discard_preallocations(inode);
999 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
1000 ldiskfs_free_blocks(handle, inode, NULL,
1001 ldiskfs_ext_pblock(&nex),
1002 le16_to_cpu(nex.ee_len), 0);
1004 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
1005 le16_to_cpu(nex.ee_len), 0);
1011 * Putting len of the actual extent we just inserted,
1012 * we are asking ldiskfs_ext_walk_space() to continue
1013 * scaning after that block
1015 cex->ec_len = le16_to_cpu(nex.ee_len);
1016 cex->ec_start = ldiskfs_ext_pblock(&nex);
1017 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
1018 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
1021 up_write((&LDISKFS_I(inode)->i_data_sem));
1022 ldiskfs_journal_stop(handle);
1027 CERROR("hmm. why do we find this extent?\n");
1028 CERROR("initial space: %lu:%u\n",
1029 bp->start, bp->init_num);
1030 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
1031 CERROR("current extent: %u/%u/%llu %d\n",
1032 cex->ec_block, cex->ec_len,
1033 (unsigned long long)cex->ec_start,
1036 CERROR("current extent: %u/%u/%llu\n",
1037 cex->ec_block, cex->ec_len,
1038 (unsigned long long)cex->ec_start);
1042 if (cex->ec_block < bp->start)
1043 i = bp->start - cex->ec_block;
1044 if (i >= cex->ec_len)
1045 CERROR("nothing to do?! i = %d, e_num = %u\n",
1047 for (; i < cex->ec_len && bp->num; i++) {
1048 *(bp->blocks) = cex->ec_start + i;
1049 /* unmap any underlying metadata from
1050 * the block device mapping. b=6998.
1053 clean_bdev_aliases(inode->i_sb->s_bdev,
1063 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
1064 int clen, sector_t *blocks, int create)
1066 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1067 struct bpointers bp;
1070 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
1074 bp.start = index * blocks_per_page;
1075 bp.init_num = bp.num = clen * blocks_per_page;
1078 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
1079 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
1081 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
1082 ldiskfs_ext_new_extent_cb, &bp);
1083 ldiskfs_ext_invalidate_cache(inode);
1088 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
1089 struct page **page, int pages,
1090 sector_t *blocks, int create)
1092 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1093 pgoff_t bitmap_max_page_index;
1097 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
1099 for (i = 0, b = blocks; i < pages; i++, page++) {
1100 if ((*page)->index + 1 >= bitmap_max_page_index) {
1104 rc = ldiskfs_map_inode_page(inode, *page, b, create);
1106 CERROR("ino %lu, blk %llu create %d: rc %d\n",
1108 (unsigned long long)*b, create, rc);
1111 b += blocks_per_page;
1116 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
1118 int pages, sector_t *blocks,
1121 int rc = 0, i = 0, clen = 0;
1122 struct page *fp = NULL;
1124 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1125 inode->i_ino, pages, (*page)->index);
1127 /* pages are sorted already. so, we just have to find
1128 * contig. space and process them properly */
1131 /* start new extent */
1136 } else if (fp->index + clen == (*page)->index) {
1137 /* continue the extent */
1144 /* process found extent */
1145 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1150 /* look for next extent */
1152 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1156 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1163 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1164 int pages, sector_t *blocks,
1169 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1170 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1174 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1179 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1180 int pages, sector_t *blocks,
1183 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1185 struct page *fp = NULL;
1187 pgoff_t max_page_index;
1188 handle_t *handle = NULL;
1190 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1192 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1193 inode->i_ino, pages, (*page)->index);
1196 create = LDISKFS_GET_BLOCKS_CREATE;
1197 handle = ldiskfs_journal_current_handle();
1198 LASSERT(handle != NULL);
1199 rc = osd_attach_jinode(inode);
1203 /* pages are sorted already. so, we just have to find
1204 * contig. space and process them properly */
1206 long blen, total = 0;
1207 struct ldiskfs_map_blocks map = { 0 };
1209 if (fp == NULL) { /* start new extent */
1214 } else if (fp->index + clen == (*page)->index) {
1215 /* continue the extent */
1221 if (fp->index + clen >= max_page_index)
1222 GOTO(cleanup, rc = -EFBIG);
1223 /* process found extent */
1224 map.m_lblk = fp->index * blocks_per_page;
1225 map.m_len = blen = clen * blocks_per_page;
1227 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1230 for (; total < blen && c < map.m_len; c++, total++) {
1232 *(blocks + total) = 0;
1236 *(blocks + total) = map.m_pblk + c;
1237 /* unmap any possible underlying
1238 * metadata from the block device
1239 * mapping. bug 6998. */
1240 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1243 inode->i_sb->s_bdev,
1249 if (rc == 0 && total < blen) {
1250 map.m_lblk = fp->index * blocks_per_page + total;
1251 map.m_len = blen - total;
1257 /* look for next extent */
1259 blocks += blocks_per_page * clen;
1264 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1266 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1267 struct niobuf_local *lnb, int npages)
1269 struct osd_thread_info *oti = osd_oti_get(env);
1270 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1271 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1272 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1284 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1285 if (unlikely(rc != 0))
1288 isize = i_size_read(inode);
1289 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1291 if (osd->od_writethrough_cache)
1293 if (isize > osd->od_readcache_max_filesize)
1296 start = ktime_get();
1297 for (i = 0; i < npages; i++) {
1300 generic_error_remove_page(inode->i_mapping,
1304 * till commit the content of the page is undefined
1305 * we'll set it uptodate once bulk is done. otherwise
1306 * subsequent reads can access non-stable data
1308 ClearPageUptodate(lnb[i].lnb_page);
1310 if (lnb[i].lnb_len == PAGE_SIZE)
1313 if (maxidx >= lnb[i].lnb_page->index) {
1314 osd_iobuf_add_page(iobuf, &lnb[i]);
1317 char *p = kmap(lnb[i].lnb_page);
1319 off = lnb[i].lnb_page_offset;
1322 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1325 memset(p + off, 0, PAGE_SIZE - off);
1326 kunmap(lnb[i].lnb_page);
1330 timediff = ktime_us_delta(end, start);
1331 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1333 if (iobuf->dr_npages) {
1334 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1336 iobuf->dr_blocks, 0);
1337 if (likely(rc == 0)) {
1338 rc = osd_do_bio(osd, inode, iobuf);
1339 /* do IO stats for preparation reads */
1340 osd_fini_iobuf(osd, iobuf);
1346 struct osd_fextent {
1349 unsigned int mapped:1;
1352 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1353 struct osd_fextent *cached_extent)
1355 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1356 sector_t block = offset >> inode->i_blkbits;
1358 struct fiemap_extent_info fei = { 0 };
1359 struct fiemap_extent fe = { 0 };
1360 mm_segment_t saved_fs;
1363 if (block >= cached_extent->start && block < cached_extent->end)
1364 return cached_extent->mapped;
1366 if (i_size_read(inode) == 0)
1369 /* Beyond EOF, must not be mapped */
1370 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1373 fei.fi_extents_max = 1;
1374 fei.fi_extents_start = &fe;
1376 saved_fs = get_fs();
1378 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1383 start = fe.fe_logical >> inode->i_blkbits;
1385 if (start > block) {
1386 cached_extent->start = block;
1387 cached_extent->end = start;
1388 cached_extent->mapped = 0;
1390 cached_extent->start = start;
1391 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1393 cached_extent->mapped = 1;
1396 return cached_extent->mapped;
1399 static int osd_declare_write_commit(const struct lu_env *env,
1400 struct dt_object *dt,
1401 struct niobuf_local *lnb, int npages,
1402 struct thandle *handle)
1404 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1405 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1406 struct osd_thandle *oh;
1414 long long quota_space = 0;
1415 struct osd_fextent extent = { 0 };
1416 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1419 LASSERT(handle != NULL);
1420 oh = container_of0(handle, struct osd_thandle, ot_super);
1421 LASSERT(oh->ot_handle == NULL);
1425 /* calculate number of extents (probably better to pass nb) */
1426 for (i = 0; i < npages; i++) {
1427 if (i && lnb[i].lnb_file_offset !=
1428 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1431 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1432 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1434 quota_space += PAGE_SIZE;
1436 /* ignore quota for the whole request if any page is from
1437 * client cache or written by root.
1439 * XXX once we drop the 1.8 client support, the checking
1440 * for whether page is from cache can be simplified as:
1441 * !(lnb[i].flags & OBD_BRW_SYNC)
1443 * XXX we could handle this on per-lnb basis as done by
1445 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1446 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1448 declare_flags |= OSD_QID_FORCE;
1452 * each extent can go into new leaf causing a split
1453 * 5 is max tree depth: inode + 4 index blocks
1454 * with blockmaps, depth is 3 at most
1456 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1458 * many concurrent threads may grow tree by the time
1459 * our transaction starts. so, consider 2 is a min depth
1461 depth = ext_depth(inode);
1462 depth = max(depth, 1) + 1;
1464 credits++; /* inode */
1465 credits += depth * 2 * extents;
1469 credits++; /* inode */
1470 credits += depth * extents;
1473 /* quota space for metadata blocks */
1474 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1476 /* quota space should be reported in 1K blocks */
1477 quota_space = toqb(quota_space);
1479 /* each new block can go in different group (bitmap + gd) */
1481 /* we can't dirty more bitmap blocks than exist */
1482 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1483 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1485 credits += newblocks;
1487 /* we can't dirty more gd blocks than exist */
1488 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1489 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1491 credits += newblocks;
1493 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1495 /* make sure the over quota flags were not set */
1496 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1498 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1499 i_projid_read(inode), quota_space, oh,
1500 osd_dt_obj(dt), &flags, declare_flags);
1502 /* we need only to store the overquota flags in the first lnb for
1503 * now, once we support multiple objects BRW, this code needs be
1505 if (flags & QUOTA_FL_OVER_USRQUOTA)
1506 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1507 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1508 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1509 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1510 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1513 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1518 /* Check if a block is allocated or not */
1519 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1520 struct niobuf_local *lnb, int npages,
1521 struct thandle *thandle)
1523 struct osd_thread_info *oti = osd_oti_get(env);
1524 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1525 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1526 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1532 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1533 if (unlikely(rc != 0))
1536 isize = i_size_read(inode);
1537 ll_vfs_dq_init(inode);
1539 for (i = 0; i < npages; i++) {
1540 if (lnb[i].lnb_rc == -ENOSPC &&
1541 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1542 /* Allow the write to proceed if overwriting an
1547 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1548 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1550 LASSERT(lnb[i].lnb_page);
1551 generic_error_remove_page(inode->i_mapping,
1556 LASSERT(PageLocked(lnb[i].lnb_page));
1557 LASSERT(!PageWriteback(lnb[i].lnb_page));
1559 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1560 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1563 * Since write and truncate are serialized by oo_sem, even
1564 * partial-page truncate should not leave dirty pages in the
1567 LASSERT(!PageDirty(lnb[i].lnb_page));
1569 SetPageUptodate(lnb[i].lnb_page);
1571 osd_iobuf_add_page(iobuf, &lnb[i]);
1574 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1576 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1578 } else if (iobuf->dr_npages > 0) {
1579 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1581 iobuf->dr_blocks, 1);
1583 /* no pages to write, no transno is needed */
1584 thandle->th_local = 1;
1587 if (likely(rc == 0)) {
1588 spin_lock(&inode->i_lock);
1589 if (isize > i_size_read(inode)) {
1590 i_size_write(inode, isize);
1591 LDISKFS_I(inode)->i_disksize = isize;
1592 spin_unlock(&inode->i_lock);
1593 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1595 spin_unlock(&inode->i_lock);
1598 rc = osd_do_bio(osd, inode, iobuf);
1599 /* we don't do stats here as in read path because
1600 * write is async: we'll do this in osd_put_bufs() */
1602 osd_fini_iobuf(osd, iobuf);
1605 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1607 if (unlikely(rc != 0)) {
1608 /* if write fails, we should drop pages from the cache */
1609 for (i = 0; i < npages; i++) {
1610 if (lnb[i].lnb_page == NULL)
1612 LASSERT(PageLocked(lnb[i].lnb_page));
1613 generic_error_remove_page(inode->i_mapping,
1621 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1622 struct niobuf_local *lnb, int npages)
1624 struct osd_thread_info *oti = osd_oti_get(env);
1625 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1626 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1627 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1628 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1635 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1636 if (unlikely(rc != 0))
1639 isize = i_size_read(inode);
1641 if (osd->od_read_cache)
1643 if (isize > osd->od_readcache_max_filesize)
1646 start = ktime_get();
1647 for (i = 0; i < npages; i++) {
1649 if (isize <= lnb[i].lnb_file_offset)
1650 /* If there's no more data, abort early.
1651 * lnb->lnb_rc == 0, so it's easy to detect later. */
1654 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1655 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1657 lnb[i].lnb_rc = lnb[i].lnb_len;
1659 /* Bypass disk read if fail_loc is set properly */
1660 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1661 SetPageUptodate(lnb[i].lnb_page);
1663 if (PageUptodate(lnb[i].lnb_page)) {
1667 osd_iobuf_add_page(iobuf, &lnb[i]);
1671 generic_error_remove_page(inode->i_mapping,
1675 timediff = ktime_us_delta(end, start);
1676 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1678 if (cache_hits != 0)
1679 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1681 if (cache_misses != 0)
1682 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1684 if (cache_hits + cache_misses != 0)
1685 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1686 cache_hits + cache_misses);
1688 if (iobuf->dr_npages) {
1689 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1691 iobuf->dr_blocks, 0);
1692 rc = osd_do_bio(osd, inode, iobuf);
1694 /* IO stats will be done in osd_bufs_put() */
1701 * XXX: Another layering violation for now.
1703 * We don't want to use ->f_op->read methods, because generic file write
1705 * - serializes on ->i_sem, and
1707 * - does a lot of extra work like balance_dirty_pages(),
1709 * which doesn't work for globally shared files like /last_rcvd.
1711 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1713 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1715 memcpy(buffer, (char *)ei->i_data, buflen);
1720 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1722 struct buffer_head *bh;
1723 unsigned long block;
1729 /* prevent reading after eof */
1730 spin_lock(&inode->i_lock);
1731 if (i_size_read(inode) < *offs + size) {
1732 loff_t diff = i_size_read(inode) - *offs;
1733 spin_unlock(&inode->i_lock);
1735 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1736 i_size_read(inode), *offs);
1738 } else if (diff == 0) {
1744 spin_unlock(&inode->i_lock);
1747 blocksize = 1 << inode->i_blkbits;
1750 block = *offs >> inode->i_blkbits;
1751 boffs = *offs & (blocksize - 1);
1752 csize = min(blocksize - boffs, size);
1753 bh = __ldiskfs_bread(NULL, inode, block, 0);
1755 CERROR("%s: can't read %u@%llu on ino %lu: "
1756 "rc = %ld\n", osd_ino2name(inode),
1757 csize, *offs, inode->i_ino,
1763 memcpy(buf, bh->b_data + boffs, csize);
1766 memset(buf, 0, csize);
1776 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1777 struct lu_buf *buf, loff_t *pos)
1779 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1782 /* Read small symlink from inode body as we need to maintain correct
1783 * on-disk symlinks for ldiskfs.
1785 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1786 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1787 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1789 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1794 static inline int osd_extents_enabled(struct super_block *sb,
1795 struct inode *inode)
1797 if (inode != NULL) {
1798 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1800 } else if (ldiskfs_has_feature_extents(sb)) {
1806 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1807 const loff_t size, const loff_t pos,
1810 int credits, bits, bs, i;
1812 bits = sb->s_blocksize_bits;
1815 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1816 * we do not expect blockmaps on the large files,
1817 * so let's shrink it to 2 levels (4GB files) */
1819 /* this is default reservation: 2 levels */
1820 credits = (blocks + 2) * 3;
1822 /* actual offset is unknown, hard to optimize */
1826 /* now check for few specific cases to optimize */
1827 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1830 /* allocate if not allocated */
1831 if (inode == NULL) {
1832 credits += blocks * 2;
1835 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1836 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1837 if (LDISKFS_I(inode)->i_data[i] == 0)
1840 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1841 /* single indirect */
1842 credits = blocks * 3;
1843 if (inode == NULL ||
1844 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1847 /* The indirect block may be modified. */
1854 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1855 const struct lu_buf *buf, loff_t _pos,
1856 struct thandle *handle)
1858 struct osd_object *obj = osd_dt_obj(dt);
1859 struct inode *inode = obj->oo_inode;
1860 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1861 struct osd_thandle *oh;
1862 int rc = 0, est = 0, credits, blocks, allocated = 0;
1868 LASSERT(buf != NULL);
1869 LASSERT(handle != NULL);
1871 oh = container_of0(handle, struct osd_thandle, ot_super);
1872 LASSERT(oh->ot_handle == NULL);
1875 bits = sb->s_blocksize_bits;
1879 /* if this is an append, then we
1880 * should expect cross-block record */
1886 /* blocks to modify */
1887 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1888 LASSERT(blocks > 0);
1890 if (inode != NULL && _pos != -1) {
1891 /* object size in blocks */
1892 est = (i_size_read(inode) + bs - 1) >> bits;
1893 allocated = inode->i_blocks >> (bits - 9);
1894 if (pos + size <= i_size_read(inode) && est <= allocated) {
1895 /* looks like an overwrite, no need to modify tree */
1897 /* no need to modify i_size */
1902 if (osd_extents_enabled(sb, inode)) {
1904 * many concurrent threads may grow tree by the time
1905 * our transaction starts. so, consider 2 is a min depth
1906 * for every level we may need to allocate a new block
1907 * and take some entries from the old one. so, 3 blocks
1908 * to allocate (bitmap, gd, itself) + old block - 4 per
1911 depth = inode != NULL ? ext_depth(inode) : 0;
1912 depth = max(depth, 1) + 1;
1914 /* if not append, then split may need to modify
1915 * existing blocks moving entries into the new ones */
1918 /* blocks to store data: bitmap,gd,itself */
1919 credits += blocks * 3;
1921 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1923 /* if inode is created as part of the transaction,
1924 * then it's counted already by the creation method */
1930 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1932 /* dt_declare_write() is usually called for system objects, such
1933 * as llog or last_rcvd files. We needn't enforce quota on those
1934 * objects, so always set the lqi_space as 0. */
1936 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1938 i_projid_read(inode), 0,
1939 oh, obj, NULL, OSD_QID_BLK);
1942 rc = osd_trunc_lock(obj, oh, true);
1947 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1949 /* LU-2634: clear the extent format for fast symlink */
1950 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1952 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1953 spin_lock(&inode->i_lock);
1954 LDISKFS_I(inode)->i_disksize = buflen;
1955 i_size_write(inode, buflen);
1956 spin_unlock(&inode->i_lock);
1957 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1962 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1963 int write_NUL, loff_t *offs, handle_t *handle)
1965 struct buffer_head *bh = NULL;
1966 loff_t offset = *offs;
1967 loff_t new_size = i_size_read(inode);
1968 unsigned long block;
1969 int blocksize = 1 << inode->i_blkbits;
1973 int dirty_inode = 0;
1977 * long symlink write does not count the NUL terminator in
1978 * bufsize, we write it, and the inode's file size does not
1979 * count the NUL terminator as well.
1981 ((char *)buf)[bufsize] = '\0';
1985 while (bufsize > 0) {
1986 int credits = handle->h_buffer_credits;
1991 block = offset >> inode->i_blkbits;
1992 boffs = offset & (blocksize - 1);
1993 size = min(blocksize - boffs, bufsize);
1994 bh = __ldiskfs_bread(handle, inode, block, 1);
1995 if (IS_ERR_OR_NULL(bh)) {
2003 CERROR("%s: error reading offset %llu (block %lu, "
2004 "size %d, offs %llu), credits %d/%d: rc = %d\n",
2005 inode->i_sb->s_id, offset, block, bufsize, *offs,
2006 credits, handle->h_buffer_credits, err);
2010 err = ldiskfs_journal_get_write_access(handle, bh);
2012 CERROR("journal_get_write_access() returned error %d\n",
2016 LASSERTF(boffs + size <= bh->b_size,
2017 "boffs %d size %d bh->b_size %lu\n",
2018 boffs, size, (unsigned long)bh->b_size);
2019 memcpy(bh->b_data + boffs, buf, size);
2020 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2024 if (offset + size > new_size)
2025 new_size = offset + size;
2035 /* correct in-core and on-disk sizes */
2036 if (new_size > i_size_read(inode)) {
2037 spin_lock(&inode->i_lock);
2038 if (new_size > i_size_read(inode))
2039 i_size_write(inode, new_size);
2040 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
2041 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
2044 spin_unlock(&inode->i_lock);
2046 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2054 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2055 const struct lu_buf *buf, loff_t *pos,
2056 struct thandle *handle)
2058 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2059 struct osd_thandle *oh;
2063 LASSERT(dt_object_exists(dt));
2065 LASSERT(handle != NULL);
2066 LASSERT(inode != NULL);
2067 ll_vfs_dq_init(inode);
2069 /* XXX: don't check: one declared chunk can be used many times */
2070 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2072 oh = container_of(handle, struct osd_thandle, ot_super);
2073 LASSERT(oh->ot_handle->h_transaction != NULL);
2074 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2076 /* Write small symlink to inode body as we need to maintain correct
2077 * on-disk symlinks for ldiskfs.
2078 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2079 * does not count it in.
2081 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2082 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2083 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2085 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
2086 buf->lb_len, is_link, pos,
2089 result = buf->lb_len;
2091 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2096 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2097 __u64 start, __u64 end, struct thandle *th)
2099 struct osd_thandle *oh;
2100 struct inode *inode;
2105 oh = container_of(th, struct osd_thandle, ot_super);
2108 * we don't need to reserve credits for whole truncate
2109 * it's not possible as truncate may need to free too many
2110 * blocks and that won't fit a single transaction. instead
2111 * we reserve credits to change i_size and put inode onto
2112 * orphan list. if needed truncate will extend or restart
2115 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2116 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2118 inode = osd_dt_obj(dt)->oo_inode;
2121 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2122 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2126 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2131 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2132 __u64 start, __u64 end, struct thandle *th)
2134 struct osd_object *obj = osd_dt_obj(dt);
2135 struct osd_device *osd = osd_obj2dev(obj);
2136 struct inode *inode = obj->oo_inode;
2137 struct osd_access_lock *al;
2138 struct osd_thandle *oh;
2139 int rc = 0, found = 0;
2143 LASSERT(end == OBD_OBJECT_EOF);
2144 LASSERT(dt_object_exists(dt));
2145 LASSERT(osd_invariant(obj));
2146 LASSERT(inode != NULL);
2147 ll_vfs_dq_init(inode);
2150 oh = container_of(th, struct osd_thandle, ot_super);
2151 LASSERT(oh->ot_handle->h_transaction != NULL);
2153 /* we used to skip truncate to current size to
2154 * optimize truncates on OST. with DoM we can
2155 * get attr_set to set specific size (MDS_REINT)
2156 * and then get truncate RPC which essentially
2157 * would be skipped. this is bad.. so, disable
2158 * this optimization on MDS till the client stop
2159 * to sent MDS_REINT (LU-11033) -bzzz */
2160 if (osd->od_is_ost && i_size_read(inode) == start)
2163 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2165 spin_lock(&inode->i_lock);
2166 if (i_size_read(inode) < start)
2168 i_size_write(inode, start);
2169 spin_unlock(&inode->i_lock);
2170 ll_truncate_pagecache(inode, start);
2172 /* optimize grow case */
2174 osd_execute_truncate(obj);
2178 /* add to orphan list to ensure truncate completion
2179 * if this transaction succeed. ldiskfs_truncate()
2180 * will take the inode out of the list */
2181 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2185 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2186 if (obj != al->tl_obj)
2188 LASSERT(al->tl_shared == 0);
2190 /* do actual truncate in osd_trans_stop() */
2191 al->tl_truncate = 1;
2200 static int fiemap_check_ranges(struct inode *inode,
2201 u64 start, u64 len, u64 *new_len)
2210 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2211 maxbytes = inode->i_sb->s_maxbytes;
2213 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2215 if (start > maxbytes)
2219 * Shrink request scope to what the fs can actually handle.
2221 if (len > maxbytes || (maxbytes - len) < start)
2222 *new_len = maxbytes - start;
2227 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2228 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2230 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2233 struct fiemap_extent_info fieinfo = {0, };
2234 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2237 mm_segment_t cur_fs;
2240 if (inode->i_op->fiemap == NULL)
2243 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2246 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2250 fieinfo.fi_flags = fm->fm_flags;
2251 fieinfo.fi_extents_max = fm->fm_extent_count;
2252 fieinfo.fi_extents_start = fm->fm_extents;
2254 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2255 filemap_write_and_wait(inode->i_mapping);
2257 /* Save previous value address limit */
2259 /* Set the address limit of the kernel */
2262 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2263 fm->fm_flags = fieinfo.fi_flags;
2264 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2266 /* Restore the previous address limt */
2272 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2273 __u64 start, __u64 end, enum lu_ladvise_type advice)
2276 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2280 case LU_LADVISE_DONTNEED:
2283 invalidate_mapping_pages(inode->i_mapping,
2284 start >> PAGE_SHIFT,
2285 (end - 1) >> PAGE_SHIFT);
2296 * in some cases we may need declare methods for objects being created
2297 * e.g., when we create symlink
2299 const struct dt_body_operations osd_body_ops_new = {
2300 .dbo_declare_write = osd_declare_write,
2303 const struct dt_body_operations osd_body_ops = {
2304 .dbo_read = osd_read,
2305 .dbo_declare_write = osd_declare_write,
2306 .dbo_write = osd_write,
2307 .dbo_bufs_get = osd_bufs_get,
2308 .dbo_bufs_put = osd_bufs_put,
2309 .dbo_write_prep = osd_write_prep,
2310 .dbo_declare_write_commit = osd_declare_write_commit,
2311 .dbo_write_commit = osd_write_commit,
2312 .dbo_read_prep = osd_read_prep,
2313 .dbo_declare_punch = osd_declare_punch,
2314 .dbo_punch = osd_punch,
2315 .dbo_fiemap_get = osd_fiemap_get,
2316 .dbo_ladvise = osd_ladvise,
2320 * Get a truncate lock
2322 * In order to take multi-transaction truncate out of main transaction we let
2323 * the caller grab a lock on the object passed. the lock can be shared (for
2324 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2325 * and write in the same transaction handle (do not confuse with big ldiskfs
2326 * transaction containing lots of handles).
2327 * The lock must be taken at declaration.
2329 * \param obj object to lock
2331 * \shared shared or exclusive
2333 * \retval 0 lock is granted
2334 * \retval -NOMEM no memory to allocate lock
2336 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2338 struct osd_access_lock *al, *tmp;
2343 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2344 if (tmp->tl_obj != obj)
2346 LASSERT(tmp->tl_shared == shared);
2347 /* found same lock */
2352 if (unlikely(al == NULL))
2355 al->tl_truncate = false;
2357 down_read(&obj->oo_ext_idx_sem);
2359 down_write(&obj->oo_ext_idx_sem);
2360 al->tl_shared = shared;
2362 list_add(&al->tl_list, &oh->ot_trunc_locks);
2367 void osd_trunc_unlock_all(struct list_head *list)
2369 struct osd_access_lock *al, *tmp;
2370 list_for_each_entry_safe(al, tmp, list, tl_list) {
2372 up_read(&al->tl_obj->oo_ext_idx_sem);
2374 up_write(&al->tl_obj->oo_ext_idx_sem);
2375 list_del(&al->tl_list);
2380 void osd_execute_truncate(struct osd_object *obj)
2382 struct osd_device *d = osd_obj2dev(obj);
2383 struct inode *inode = obj->oo_inode;
2386 /* simulate crash before (in the middle) of delayed truncate */
2387 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2388 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2389 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2391 mutex_lock(&sbi->s_orphan_lock);
2392 list_del_init(&ei->i_orphan);
2393 mutex_unlock(&sbi->s_orphan_lock);
2397 ldiskfs_truncate(inode);
2400 * For a partial-page truncate, flush the page to disk immediately to
2401 * avoid data corruption during direct disk write. b=17397
2403 size = i_size_read(inode);
2404 if ((size & ~PAGE_MASK) == 0)
2406 if (osd_use_page_cache(d)) {
2407 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2409 /* Notice we use "wait" version to ensure I/O is complete */
2410 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2411 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2412 size >> PAGE_SHIFT);
2416 void osd_process_truncates(struct list_head *list)
2418 struct osd_access_lock *al;
2420 LASSERT(journal_current_handle() == NULL);
2422 list_for_each_entry(al, list, tl_list) {
2425 if (!al->tl_truncate)
2427 osd_execute_truncate(al->tl_obj);