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;
275 * This function will change the data written, thus it should only be
276 * used when checking data integrity feature
278 static void bio_integrity_fault_inject(struct bio *bio)
280 struct bio_vec *bvec;
285 bio_for_each_segment_all(bvec, bio, i) {
286 struct page *page = bvec->bv_page;
296 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
297 unsigned int sectors, int tuple_size)
299 __u16 *expected_guard;
303 expected_guard = expected_guard_buf;
304 for (i = 0; i < sectors; i++) {
305 bio_guard = (__u16 *)bio_prot_buf;
306 if (*bio_guard != *expected_guard) {
307 CERROR("unexpected guard tags on sector %d "
308 "expected guard %u, bio guard "
309 "%u, sectors %u, tuple size %d\n",
310 i, *expected_guard, *bio_guard, sectors,
315 bio_prot_buf += tuple_size;
320 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
321 struct osd_iobuf *iobuf, int index)
323 struct blk_integrity *bi = bdev_get_integrity(bdev);
324 struct bio_integrity_payload *bip = bio->bi_integrity;
325 struct niobuf_local *lnb;
326 unsigned short sector_size = blk_integrity_interval(bi);
327 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
328 bip->bip_vec->bv_offset;
330 sector_t sector = bio_start_sector(bio);
331 unsigned int i, sectors, total;
332 __u16 *expected_guard;
336 bio_for_each_segment_all(bv, bio, i) {
337 lnb = iobuf->dr_lnbs[index];
338 expected_guard = lnb->lnb_guards;
339 sectors = bv->bv_len / sector_size;
340 if (lnb->lnb_guard_rpc) {
341 rc = bio_dif_compare(expected_guard, bio_prot_buf,
342 sectors, bi->tuple_size);
348 bio_prot_buf += sectors * bi->tuple_size;
349 total += sectors * bi->tuple_size;
350 LASSERT(total <= bip_size(bio->bi_integrity));
356 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
357 struct osd_iobuf *iobuf,
358 int start_page_idx, bool fault_inject,
359 bool integrity_enabled)
361 struct super_block *sb = osd_sb(osd);
362 integrity_gen_fn *generate_fn = NULL;
363 integrity_vrfy_fn *verify_fn = NULL;
368 if (!integrity_enabled)
371 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
375 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
379 /* Verify and inject fault only when writing */
380 if (iobuf->dr_rw == 1) {
381 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
382 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
388 if (unlikely(fault_inject))
389 bio_integrity_fault_inject(bio);
395 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
396 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
397 static void dio_integrity_complete_routine(struct bio *bio)
399 static void dio_integrity_complete_routine(struct bio *bio, int error)
402 struct osd_bio_private *bio_private = bio->bi_private;
404 bio->bi_private = bio_private->obp_iobuf;
405 osd_dio_complete_routine(bio, error);
407 OBD_FREE_PTR(bio_private);
411 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
412 bool integrity_enabled, int start_page_idx,
413 struct osd_bio_private **pprivate)
419 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
420 if (integrity_enabled) {
421 struct osd_bio_private *bio_private = NULL;
423 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
424 if (bio_private == NULL)
426 bio->bi_end_io = dio_integrity_complete_routine;
427 bio->bi_private = bio_private;
428 bio_private->obp_start_page_idx = start_page_idx;
429 bio_private->obp_iobuf = iobuf;
430 *pprivate = bio_private;
434 bio->bi_end_io = dio_complete_routine;
435 bio->bi_private = iobuf;
441 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
442 struct osd_iobuf *iobuf)
444 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
445 struct page **pages = iobuf->dr_pages;
446 int npages = iobuf->dr_npages;
447 sector_t *blocks = iobuf->dr_blocks;
448 int total_blocks = npages * blocks_per_page;
449 struct super_block *sb = inode->i_sb;
450 int sector_bits = sb->s_blocksize_bits - 9;
451 unsigned int blocksize = sb->s_blocksize;
452 struct block_device *bdev = sb->s_bdev;
453 struct osd_bio_private *bio_private = NULL;
454 struct bio *bio = NULL;
455 int bio_start_page_idx;
457 unsigned int page_offset;
465 bool integrity_enabled;
469 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
470 LASSERT(iobuf->dr_npages == npages);
472 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
474 osd_brw_stats_update(osd, iobuf);
475 iobuf->dr_start_time = ktime_get();
477 blk_start_plug(&plug);
478 for (page_idx = 0, block_idx = 0;
480 page_idx++, block_idx += blocks_per_page) {
482 page = pages[page_idx];
483 LASSERT(block_idx + blocks_per_page <= total_blocks);
485 for (i = 0, page_offset = 0;
487 i += nblocks, page_offset += blocksize * nblocks) {
491 if (blocks[block_idx + i] == 0) { /* hole */
492 LASSERTF(iobuf->dr_rw == 0,
493 "page_idx %u, block_idx %u, i %u\n",
494 page_idx, block_idx, i);
495 memset(kmap(page) + page_offset, 0, blocksize);
500 sector = (sector_t)blocks[block_idx + i] << sector_bits;
502 /* Additional contiguous file blocks? */
503 while (i + nblocks < blocks_per_page &&
504 (sector + (nblocks << sector_bits)) ==
505 ((sector_t)blocks[block_idx + i + nblocks] <<
510 can_be_merged(bio, sector) &&
511 bio_add_page(bio, page,
512 blocksize * nblocks, page_offset) != 0)
513 continue; /* added this frag OK */
516 struct request_queue *q = bio_get_queue(bio);
517 unsigned int bi_size = bio_sectors(bio) << 9;
519 /* Dang! I have to fragment this I/O */
520 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
521 "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
522 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
524 queue_max_sectors(q),
525 bio_phys_segments(q, bio),
526 queue_max_phys_segments(q),
527 0, queue_max_hw_segments(q));
528 rc = osd_bio_integrity_handle(osd, bio,
529 iobuf, bio_start_page_idx,
530 fault_inject, integrity_enabled);
536 record_start_io(iobuf, bi_size);
537 osd_submit_bio(iobuf->dr_rw, bio);
540 bio_start_page_idx = page_idx;
541 /* allocate new bio */
542 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
543 (npages - page_idx) *
546 CERROR("Can't allocate bio %u*%u = %u pages\n",
547 (npages - page_idx), blocks_per_page,
548 (npages - page_idx) * blocks_per_page);
553 bio_set_dev(bio, bdev);
554 bio_set_sector(bio, sector);
555 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
556 rc = osd_bio_init(bio, iobuf, integrity_enabled,
557 bio_start_page_idx, &bio_private);
563 rc = bio_add_page(bio, page,
564 blocksize * nblocks, page_offset);
570 rc = osd_bio_integrity_handle(osd, bio, iobuf,
579 record_start_io(iobuf, bio_sectors(bio) << 9);
580 osd_submit_bio(iobuf->dr_rw, bio);
585 blk_finish_plug(&plug);
587 /* in order to achieve better IO throughput, we don't wait for writes
588 * completion here. instead we proceed with transaction commit in
589 * parallel and wait for IO completion once transaction is stopped
590 * see osd_trans_stop() for more details -bzzz */
591 if (iobuf->dr_rw == 0 || fault_inject) {
592 wait_event(iobuf->dr_wait,
593 atomic_read(&iobuf->dr_numreqs) == 0);
594 osd_fini_iobuf(osd, iobuf);
598 rc = iobuf->dr_error;
601 OBD_FREE_PTR(bio_private);
607 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
608 struct niobuf_local *lnb)
615 int poff = offset & (PAGE_SIZE - 1);
616 int plen = PAGE_SIZE - poff;
620 lnb->lnb_file_offset = offset;
621 lnb->lnb_page_offset = poff;
623 /* lnb->lnb_flags = rnb->rnb_flags; */
625 lnb->lnb_page = NULL;
627 lnb->lnb_guard_rpc = 0;
628 lnb->lnb_guard_disk = 0;
630 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
641 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
642 loff_t offset, gfp_t gfp_mask)
644 struct osd_thread_info *oti = osd_oti_get(env);
645 struct inode *inode = osd_dt_obj(dt)->oo_inode;
646 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
648 int cur = oti->oti_dio_pages_used;
652 if (osd_use_page_cache(d)) {
653 page = find_or_create_page(inode->i_mapping,
654 offset >> PAGE_SHIFT,
658 LASSERT(!test_bit(PG_private_2, &page->flags));
660 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
663 LASSERT(oti->oti_dio_pages);
665 if (unlikely(!oti->oti_dio_pages[cur])) {
666 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
667 page = alloc_page(gfp_mask);
670 oti->oti_dio_pages[cur] = page;
673 page = oti->oti_dio_pages[cur];
674 LASSERT(!test_bit(PG_private_2, &page->flags));
675 set_bit(PG_private_2, &page->flags);
676 oti->oti_dio_pages_used++;
678 LASSERT(!PageLocked(page));
681 LASSERT(!page->mapping);
682 LASSERT(!PageWriteback(page));
683 ClearPageUptodate(page);
685 page->index = offset >> PAGE_SHIFT;
692 * there are following "locks":
703 * - lock pages, unlock
705 * - lock partial page
711 * Unlock and release pages loaded by osd_bufs_get()
713 * Unlock \a npages pages from \a lnb and drop the refcount on them.
715 * \param env thread execution environment
716 * \param dt dt object undergoing IO (OSD object + methods)
717 * \param lnb array of pages undergoing IO
718 * \param npages number of pages in \a lnb
722 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
723 struct niobuf_local *lnb, int npages)
725 struct osd_thread_info *oti = osd_oti_get(env);
729 ll_pagevec_init(&pvec, 0);
731 for (i = 0; i < npages; i++) {
732 struct page *page = lnb[i].lnb_page;
736 LASSERT(PageLocked(page));
738 /* if the page isn't cached, then reset uptodate
739 * to prevent reuse */
740 if (test_bit(PG_private_2, &page->flags)) {
741 clear_bit(PG_private_2, &page->flags);
742 ClearPageUptodate(page);
744 oti->oti_dio_pages_used--;
747 if (pagevec_add(&pvec, page) == 0)
748 pagevec_release(&pvec);
750 dt_object_put(env, dt);
752 lnb[i].lnb_page = NULL;
755 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
757 /* Release any partial pagevec */
758 pagevec_release(&pvec);
764 * Load and lock pages undergoing IO
766 * Pages as described in the \a lnb array are fetched (from disk or cache)
767 * and locked for IO by the caller.
769 * DLM locking protects us from write and truncate competing for same region,
770 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
771 * It's possible the writeout on a such a page is in progress when we access
772 * it. It's also possible that during this writeout we put new (partial) data
773 * into the page, but won't be able to proceed in filter_commitrw_write().
774 * Therefore, just wait for writeout completion as it should be rare enough.
776 * \param env thread execution environment
777 * \param dt dt object undergoing IO (OSD object + methods)
778 * \param pos byte offset of IO start
779 * \param len number of bytes of IO
780 * \param lnb array of extents undergoing IO
781 * \param rw read or write operation, and other flags
782 * \param capa capabilities
784 * \retval pages (zero or more) loaded successfully
785 * \retval -ENOMEM on memory/page allocation error
787 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
788 loff_t pos, ssize_t len, struct niobuf_local *lnb,
789 enum dt_bufs_type rw)
791 struct osd_thread_info *oti = osd_oti_get(env);
792 struct osd_object *obj = osd_dt_obj(dt);
793 int npages, i, rc = 0;
796 LASSERT(obj->oo_inode);
798 if (!osd_use_page_cache(osd_obj2dev(obj))) {
799 if (unlikely(!oti->oti_dio_pages)) {
800 OBD_ALLOC(oti->oti_dio_pages,
801 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
802 if (!oti->oti_dio_pages)
807 osd_map_remote_to_local(pos, len, &npages, lnb);
809 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
810 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
812 for (i = 0; i < npages; i++, lnb++) {
813 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
815 if (lnb->lnb_page == NULL)
816 GOTO(cleanup, rc = -ENOMEM);
818 wait_on_page_writeback(lnb->lnb_page);
819 BUG_ON(PageWriteback(lnb->lnb_page));
821 lu_object_get(&dt->do_lu);
828 osd_bufs_put(env, dt, lnb - i, i);
832 #ifndef HAVE_LDISKFS_MAP_BLOCKS
834 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
835 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
846 static long ldiskfs_ext_find_goal(struct inode *inode,
847 struct ldiskfs_ext_path *path,
848 unsigned long block, int *aflags)
850 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
851 unsigned long bg_start;
852 unsigned long colour;
856 struct ldiskfs_extent *ex;
857 depth = path->p_depth;
859 /* try to predict block placement */
860 if ((ex = path[depth].p_ext))
861 return ldiskfs_ext_pblock(ex) +
862 (block - le32_to_cpu(ex->ee_block));
864 /* it looks index is empty
865 * try to find starting from index itself */
866 if (path[depth].p_bh)
867 return path[depth].p_bh->b_blocknr;
870 /* OK. use inode's group */
871 bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
872 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
873 colour = (current->pid % 16) *
874 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
875 return bg_start + colour + block;
878 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
879 struct ldiskfs_ext_path *path,
880 unsigned long block, unsigned long *count,
883 struct ldiskfs_allocation_request ar;
884 unsigned long pblock;
887 /* find neighbour allocated blocks */
889 *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
893 *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
897 /* allocate new block */
898 ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
902 ar.flags = LDISKFS_MB_HINT_DATA;
903 pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
908 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
909 struct ldiskfs_ext_path *path,
910 struct ldiskfs_ext_cache *cex,
911 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
912 struct ldiskfs_extent *ex,
916 struct bpointers *bp = cbdata;
917 struct ldiskfs_extent nex;
918 unsigned long pblock = 0;
924 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
925 if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
927 if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
933 if (bp->create == 0) {
935 if (cex->ec_block < bp->start)
936 i = bp->start - cex->ec_block;
937 if (i >= cex->ec_len)
938 CERROR("nothing to do?! i = %d, e_num = %u\n",
940 for (; i < cex->ec_len && bp->num; i++) {
950 tgen = LDISKFS_I(inode)->i_ext_generation;
951 count = ldiskfs_ext_calc_credits_for_insert(inode, path);
953 handle = osd_journal_start(inode, LDISKFS_HT_MISC,
954 count + LDISKFS_ALLOC_NEEDED + 1);
955 if (IS_ERR(handle)) {
956 return PTR_ERR(handle);
959 if (tgen != LDISKFS_I(inode)->i_ext_generation) {
960 /* the tree has changed. so path can be invalid at moment */
961 ldiskfs_journal_stop(handle);
965 /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
966 * protected by i_data_sem as whole. so we patch it to store
967 * generation to path and now verify the tree hasn't changed */
968 down_write((&LDISKFS_I(inode)->i_data_sem));
970 /* validate extent, make sure the extent tree does not changed */
971 if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
972 /* cex is invalid, try again */
973 up_write(&LDISKFS_I(inode)->i_data_sem);
974 ldiskfs_journal_stop(handle);
979 pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
982 BUG_ON(count > cex->ec_len);
984 /* insert new extent */
985 nex.ee_block = cpu_to_le32(cex->ec_block);
986 ldiskfs_ext_store_pblock(&nex, pblock);
987 nex.ee_len = cpu_to_le16(count);
988 err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
990 /* free data blocks we just allocated */
991 /* not a good idea to call discard here directly,
992 * but otherwise we'd need to call it every free() */
993 ldiskfs_discard_preallocations(inode);
994 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
995 ldiskfs_free_blocks(handle, inode, NULL,
996 ldiskfs_ext_pblock(&nex),
997 le16_to_cpu(nex.ee_len), 0);
999 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
1000 le16_to_cpu(nex.ee_len), 0);
1006 * Putting len of the actual extent we just inserted,
1007 * we are asking ldiskfs_ext_walk_space() to continue
1008 * scaning after that block
1010 cex->ec_len = le16_to_cpu(nex.ee_len);
1011 cex->ec_start = ldiskfs_ext_pblock(&nex);
1012 BUG_ON(le16_to_cpu(nex.ee_len) == 0);
1013 BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
1016 up_write((&LDISKFS_I(inode)->i_data_sem));
1017 ldiskfs_journal_stop(handle);
1022 CERROR("hmm. why do we find this extent?\n");
1023 CERROR("initial space: %lu:%u\n",
1024 bp->start, bp->init_num);
1025 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
1026 CERROR("current extent: %u/%u/%llu %d\n",
1027 cex->ec_block, cex->ec_len,
1028 (unsigned long long)cex->ec_start,
1031 CERROR("current extent: %u/%u/%llu\n",
1032 cex->ec_block, cex->ec_len,
1033 (unsigned long long)cex->ec_start);
1037 if (cex->ec_block < bp->start)
1038 i = bp->start - cex->ec_block;
1039 if (i >= cex->ec_len)
1040 CERROR("nothing to do?! i = %d, e_num = %u\n",
1042 for (; i < cex->ec_len && bp->num; i++) {
1043 *(bp->blocks) = cex->ec_start + i;
1044 /* unmap any underlying metadata from
1045 * the block device mapping. b=6998.
1048 clean_bdev_aliases(inode->i_sb->s_bdev,
1058 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
1059 int clen, sector_t *blocks, int create)
1061 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1062 struct bpointers bp;
1065 if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
1069 bp.start = index * blocks_per_page;
1070 bp.init_num = bp.num = clen * blocks_per_page;
1073 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
1074 bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
1076 err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
1077 ldiskfs_ext_new_extent_cb, &bp);
1078 ldiskfs_ext_invalidate_cache(inode);
1083 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
1084 struct page **page, int pages,
1085 sector_t *blocks, int create)
1087 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1088 pgoff_t bitmap_max_page_index;
1092 bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
1094 for (i = 0, b = blocks; i < pages; i++, page++) {
1095 if ((*page)->index + 1 >= bitmap_max_page_index) {
1099 rc = ldiskfs_map_inode_page(inode, *page, b, create);
1101 CERROR("ino %lu, blk %llu create %d: rc %d\n",
1103 (unsigned long long)*b, create, rc);
1106 b += blocks_per_page;
1111 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
1113 int pages, sector_t *blocks,
1116 int rc = 0, i = 0, clen = 0;
1117 struct page *fp = NULL;
1119 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1120 inode->i_ino, pages, (*page)->index);
1122 /* pages are sorted already. so, we just have to find
1123 * contig. space and process them properly */
1126 /* start new extent */
1131 } else if (fp->index + clen == (*page)->index) {
1132 /* continue the extent */
1139 /* process found extent */
1140 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1145 /* look for next extent */
1147 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1151 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1158 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1159 int pages, sector_t *blocks,
1164 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1165 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1169 rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1174 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1175 int pages, sector_t *blocks,
1178 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1180 struct page *fp = NULL;
1182 pgoff_t max_page_index;
1183 handle_t *handle = NULL;
1185 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1187 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1188 inode->i_ino, pages, (*page)->index);
1191 create = LDISKFS_GET_BLOCKS_CREATE;
1192 handle = ldiskfs_journal_current_handle();
1193 LASSERT(handle != NULL);
1194 rc = osd_attach_jinode(inode);
1198 /* pages are sorted already. so, we just have to find
1199 * contig. space and process them properly */
1201 long blen, total = 0;
1202 struct ldiskfs_map_blocks map = { 0 };
1204 if (fp == NULL) { /* start new extent */
1209 } else if (fp->index + clen == (*page)->index) {
1210 /* continue the extent */
1216 if (fp->index + clen >= max_page_index)
1217 GOTO(cleanup, rc = -EFBIG);
1218 /* process found extent */
1219 map.m_lblk = fp->index * blocks_per_page;
1220 map.m_len = blen = clen * blocks_per_page;
1222 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1225 for (; total < blen && c < map.m_len; c++, total++) {
1227 *(blocks + total) = 0;
1231 *(blocks + total) = map.m_pblk + c;
1232 /* unmap any possible underlying
1233 * metadata from the block device
1234 * mapping. bug 6998. */
1235 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1238 inode->i_sb->s_bdev,
1244 if (rc == 0 && total < blen) {
1245 map.m_lblk = fp->index * blocks_per_page + total;
1246 map.m_len = blen - total;
1252 /* look for next extent */
1254 blocks += blocks_per_page * clen;
1259 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1261 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1262 struct niobuf_local *lnb, int npages)
1264 struct osd_thread_info *oti = osd_oti_get(env);
1265 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1266 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1267 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1279 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1280 if (unlikely(rc != 0))
1283 isize = i_size_read(inode);
1284 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1286 if (osd->od_writethrough_cache)
1288 if (isize > osd->od_readcache_max_filesize)
1291 start = ktime_get();
1292 for (i = 0; i < npages; i++) {
1295 generic_error_remove_page(inode->i_mapping,
1299 * till commit the content of the page is undefined
1300 * we'll set it uptodate once bulk is done. otherwise
1301 * subsequent reads can access non-stable data
1303 ClearPageUptodate(lnb[i].lnb_page);
1305 if (lnb[i].lnb_len == PAGE_SIZE)
1308 if (maxidx >= lnb[i].lnb_page->index) {
1309 osd_iobuf_add_page(iobuf, &lnb[i]);
1312 char *p = kmap(lnb[i].lnb_page);
1314 off = lnb[i].lnb_page_offset;
1317 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1320 memset(p + off, 0, PAGE_SIZE - off);
1321 kunmap(lnb[i].lnb_page);
1325 timediff = ktime_us_delta(end, start);
1326 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1328 if (iobuf->dr_npages) {
1329 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1331 iobuf->dr_blocks, 0);
1332 if (likely(rc == 0)) {
1333 rc = osd_do_bio(osd, inode, iobuf);
1334 /* do IO stats for preparation reads */
1335 osd_fini_iobuf(osd, iobuf);
1341 struct osd_fextent {
1344 unsigned int mapped:1;
1347 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1348 struct osd_fextent *cached_extent)
1350 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1351 sector_t block = offset >> inode->i_blkbits;
1353 struct fiemap_extent_info fei = { 0 };
1354 struct fiemap_extent fe = { 0 };
1355 mm_segment_t saved_fs;
1358 if (block >= cached_extent->start && block < cached_extent->end)
1359 return cached_extent->mapped;
1361 if (i_size_read(inode) == 0)
1364 /* Beyond EOF, must not be mapped */
1365 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1368 fei.fi_extents_max = 1;
1369 fei.fi_extents_start = &fe;
1371 saved_fs = get_fs();
1373 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1378 start = fe.fe_logical >> inode->i_blkbits;
1380 if (start > block) {
1381 cached_extent->start = block;
1382 cached_extent->end = start;
1383 cached_extent->mapped = 0;
1385 cached_extent->start = start;
1386 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1388 cached_extent->mapped = 1;
1391 return cached_extent->mapped;
1394 static int osd_declare_write_commit(const struct lu_env *env,
1395 struct dt_object *dt,
1396 struct niobuf_local *lnb, int npages,
1397 struct thandle *handle)
1399 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1400 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1401 struct osd_thandle *oh;
1409 long long quota_space = 0;
1410 struct osd_fextent extent = { 0 };
1411 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1414 LASSERT(handle != NULL);
1415 oh = container_of0(handle, struct osd_thandle, ot_super);
1416 LASSERT(oh->ot_handle == NULL);
1420 /* calculate number of extents (probably better to pass nb) */
1421 for (i = 0; i < npages; i++) {
1422 if (i && lnb[i].lnb_file_offset !=
1423 lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1426 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1427 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1429 quota_space += PAGE_SIZE;
1431 /* ignore quota for the whole request if any page is from
1432 * client cache or written by root.
1434 * XXX once we drop the 1.8 client support, the checking
1435 * for whether page is from cache can be simplified as:
1436 * !(lnb[i].flags & OBD_BRW_SYNC)
1438 * XXX we could handle this on per-lnb basis as done by
1440 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1441 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1443 declare_flags |= OSD_QID_FORCE;
1447 * each extent can go into new leaf causing a split
1448 * 5 is max tree depth: inode + 4 index blocks
1449 * with blockmaps, depth is 3 at most
1451 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1453 * many concurrent threads may grow tree by the time
1454 * our transaction starts. so, consider 2 is a min depth
1456 depth = ext_depth(inode);
1457 depth = max(depth, 1) + 1;
1459 credits++; /* inode */
1460 credits += depth * 2 * extents;
1464 credits++; /* inode */
1465 credits += depth * extents;
1468 /* quota space for metadata blocks */
1469 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1471 /* quota space should be reported in 1K blocks */
1472 quota_space = toqb(quota_space);
1474 /* each new block can go in different group (bitmap + gd) */
1476 /* we can't dirty more bitmap blocks than exist */
1477 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1478 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1480 credits += newblocks;
1482 /* we can't dirty more gd blocks than exist */
1483 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1484 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1486 credits += newblocks;
1488 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1490 /* make sure the over quota flags were not set */
1491 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1493 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1494 i_projid_read(inode), quota_space, oh,
1495 osd_dt_obj(dt), &flags, declare_flags);
1497 /* we need only to store the overquota flags in the first lnb for
1498 * now, once we support multiple objects BRW, this code needs be
1500 if (flags & QUOTA_FL_OVER_USRQUOTA)
1501 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1502 if (flags & QUOTA_FL_OVER_GRPQUOTA)
1503 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1504 if (flags & QUOTA_FL_OVER_PRJQUOTA)
1505 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1508 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1513 /* Check if a block is allocated or not */
1514 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1515 struct niobuf_local *lnb, int npages,
1516 struct thandle *thandle)
1518 struct osd_thread_info *oti = osd_oti_get(env);
1519 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1520 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1521 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1527 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1528 if (unlikely(rc != 0))
1531 isize = i_size_read(inode);
1532 ll_vfs_dq_init(inode);
1534 for (i = 0; i < npages; i++) {
1535 if (lnb[i].lnb_rc == -ENOSPC &&
1536 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1537 /* Allow the write to proceed if overwriting an
1542 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1543 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1545 LASSERT(lnb[i].lnb_page);
1546 generic_error_remove_page(inode->i_mapping,
1551 LASSERT(PageLocked(lnb[i].lnb_page));
1552 LASSERT(!PageWriteback(lnb[i].lnb_page));
1554 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1555 isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1558 * Since write and truncate are serialized by oo_sem, even
1559 * partial-page truncate should not leave dirty pages in the
1562 LASSERT(!PageDirty(lnb[i].lnb_page));
1564 SetPageUptodate(lnb[i].lnb_page);
1566 osd_iobuf_add_page(iobuf, &lnb[i]);
1569 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1571 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1573 } else if (iobuf->dr_npages > 0) {
1574 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1576 iobuf->dr_blocks, 1);
1578 /* no pages to write, no transno is needed */
1579 thandle->th_local = 1;
1582 if (likely(rc == 0)) {
1583 spin_lock(&inode->i_lock);
1584 if (isize > i_size_read(inode)) {
1585 i_size_write(inode, isize);
1586 LDISKFS_I(inode)->i_disksize = isize;
1587 spin_unlock(&inode->i_lock);
1588 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1590 spin_unlock(&inode->i_lock);
1593 rc = osd_do_bio(osd, inode, iobuf);
1594 /* we don't do stats here as in read path because
1595 * write is async: we'll do this in osd_put_bufs() */
1597 osd_fini_iobuf(osd, iobuf);
1600 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1602 if (unlikely(rc != 0)) {
1603 /* if write fails, we should drop pages from the cache */
1604 for (i = 0; i < npages; i++) {
1605 if (lnb[i].lnb_page == NULL)
1607 LASSERT(PageLocked(lnb[i].lnb_page));
1608 generic_error_remove_page(inode->i_mapping,
1616 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1617 struct niobuf_local *lnb, int npages)
1619 struct osd_thread_info *oti = osd_oti_get(env);
1620 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1621 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1622 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1623 int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1630 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1631 if (unlikely(rc != 0))
1634 isize = i_size_read(inode);
1636 if (osd->od_read_cache)
1638 if (isize > osd->od_readcache_max_filesize)
1641 start = ktime_get();
1642 for (i = 0; i < npages; i++) {
1644 if (isize <= lnb[i].lnb_file_offset)
1645 /* If there's no more data, abort early.
1646 * lnb->lnb_rc == 0, so it's easy to detect later. */
1649 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1650 lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1652 lnb[i].lnb_rc = lnb[i].lnb_len;
1654 /* Bypass disk read if fail_loc is set properly */
1655 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1656 SetPageUptodate(lnb[i].lnb_page);
1658 if (PageUptodate(lnb[i].lnb_page)) {
1662 osd_iobuf_add_page(iobuf, &lnb[i]);
1666 generic_error_remove_page(inode->i_mapping,
1670 timediff = ktime_us_delta(end, start);
1671 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1673 if (cache_hits != 0)
1674 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1676 if (cache_misses != 0)
1677 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1679 if (cache_hits + cache_misses != 0)
1680 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1681 cache_hits + cache_misses);
1683 if (iobuf->dr_npages) {
1684 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1686 iobuf->dr_blocks, 0);
1687 rc = osd_do_bio(osd, inode, iobuf);
1689 /* IO stats will be done in osd_bufs_put() */
1696 * XXX: Another layering violation for now.
1698 * We don't want to use ->f_op->read methods, because generic file write
1700 * - serializes on ->i_sem, and
1702 * - does a lot of extra work like balance_dirty_pages(),
1704 * which doesn't work for globally shared files like /last_rcvd.
1706 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1708 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1710 memcpy(buffer, (char *)ei->i_data, buflen);
1715 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1717 struct buffer_head *bh;
1718 unsigned long block;
1724 /* prevent reading after eof */
1725 spin_lock(&inode->i_lock);
1726 if (i_size_read(inode) < *offs + size) {
1727 loff_t diff = i_size_read(inode) - *offs;
1728 spin_unlock(&inode->i_lock);
1730 CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1731 i_size_read(inode), *offs);
1733 } else if (diff == 0) {
1739 spin_unlock(&inode->i_lock);
1742 blocksize = 1 << inode->i_blkbits;
1745 block = *offs >> inode->i_blkbits;
1746 boffs = *offs & (blocksize - 1);
1747 csize = min(blocksize - boffs, size);
1748 bh = __ldiskfs_bread(NULL, inode, block, 0);
1750 CERROR("%s: can't read %u@%llu on ino %lu: "
1751 "rc = %ld\n", osd_ino2name(inode),
1752 csize, *offs, inode->i_ino,
1758 memcpy(buf, bh->b_data + boffs, csize);
1761 memset(buf, 0, csize);
1771 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1772 struct lu_buf *buf, loff_t *pos)
1774 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1777 /* Read small symlink from inode body as we need to maintain correct
1778 * on-disk symlinks for ldiskfs.
1780 if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1781 (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1782 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1784 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1789 static inline int osd_extents_enabled(struct super_block *sb,
1790 struct inode *inode)
1792 if (inode != NULL) {
1793 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1795 } else if (ldiskfs_has_feature_extents(sb)) {
1801 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1802 const loff_t size, const loff_t pos,
1805 int credits, bits, bs, i;
1807 bits = sb->s_blocksize_bits;
1810 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1811 * we do not expect blockmaps on the large files,
1812 * so let's shrink it to 2 levels (4GB files) */
1814 /* this is default reservation: 2 levels */
1815 credits = (blocks + 2) * 3;
1817 /* actual offset is unknown, hard to optimize */
1821 /* now check for few specific cases to optimize */
1822 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1825 /* allocate if not allocated */
1826 if (inode == NULL) {
1827 credits += blocks * 2;
1830 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1831 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1832 if (LDISKFS_I(inode)->i_data[i] == 0)
1835 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1836 /* single indirect */
1837 credits = blocks * 3;
1838 if (inode == NULL ||
1839 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1842 /* The indirect block may be modified. */
1849 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1850 const struct lu_buf *buf, loff_t _pos,
1851 struct thandle *handle)
1853 struct osd_object *obj = osd_dt_obj(dt);
1854 struct inode *inode = obj->oo_inode;
1855 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1856 struct osd_thandle *oh;
1857 int rc = 0, est = 0, credits, blocks, allocated = 0;
1863 LASSERT(buf != NULL);
1864 LASSERT(handle != NULL);
1866 oh = container_of0(handle, struct osd_thandle, ot_super);
1867 LASSERT(oh->ot_handle == NULL);
1870 bits = sb->s_blocksize_bits;
1874 /* if this is an append, then we
1875 * should expect cross-block record */
1881 /* blocks to modify */
1882 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1883 LASSERT(blocks > 0);
1885 if (inode != NULL && _pos != -1) {
1886 /* object size in blocks */
1887 est = (i_size_read(inode) + bs - 1) >> bits;
1888 allocated = inode->i_blocks >> (bits - 9);
1889 if (pos + size <= i_size_read(inode) && est <= allocated) {
1890 /* looks like an overwrite, no need to modify tree */
1892 /* no need to modify i_size */
1897 if (osd_extents_enabled(sb, inode)) {
1899 * many concurrent threads may grow tree by the time
1900 * our transaction starts. so, consider 2 is a min depth
1901 * for every level we may need to allocate a new block
1902 * and take some entries from the old one. so, 3 blocks
1903 * to allocate (bitmap, gd, itself) + old block - 4 per
1906 depth = inode != NULL ? ext_depth(inode) : 0;
1907 depth = max(depth, 1) + 1;
1909 /* if not append, then split may need to modify
1910 * existing blocks moving entries into the new ones */
1913 /* blocks to store data: bitmap,gd,itself */
1914 credits += blocks * 3;
1916 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1918 /* if inode is created as part of the transaction,
1919 * then it's counted already by the creation method */
1925 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1927 /* dt_declare_write() is usually called for system objects, such
1928 * as llog or last_rcvd files. We needn't enforce quota on those
1929 * objects, so always set the lqi_space as 0. */
1931 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1933 i_projid_read(inode), 0,
1934 oh, obj, NULL, OSD_QID_BLK);
1937 rc = osd_trunc_lock(obj, oh, true);
1942 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1944 /* LU-2634: clear the extent format for fast symlink */
1945 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1947 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1948 spin_lock(&inode->i_lock);
1949 LDISKFS_I(inode)->i_disksize = buflen;
1950 i_size_write(inode, buflen);
1951 spin_unlock(&inode->i_lock);
1952 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1957 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1958 int write_NUL, loff_t *offs, handle_t *handle)
1960 struct buffer_head *bh = NULL;
1961 loff_t offset = *offs;
1962 loff_t new_size = i_size_read(inode);
1963 unsigned long block;
1964 int blocksize = 1 << inode->i_blkbits;
1968 int dirty_inode = 0;
1972 * long symlink write does not count the NUL terminator in
1973 * bufsize, we write it, and the inode's file size does not
1974 * count the NUL terminator as well.
1976 ((char *)buf)[bufsize] = '\0';
1980 while (bufsize > 0) {
1981 int credits = handle->h_buffer_credits;
1986 block = offset >> inode->i_blkbits;
1987 boffs = offset & (blocksize - 1);
1988 size = min(blocksize - boffs, bufsize);
1989 bh = __ldiskfs_bread(handle, inode, block, 1);
1990 if (IS_ERR_OR_NULL(bh)) {
1998 CERROR("%s: error reading offset %llu (block %lu, "
1999 "size %d, offs %llu), credits %d/%d: rc = %d\n",
2000 inode->i_sb->s_id, offset, block, bufsize, *offs,
2001 credits, handle->h_buffer_credits, err);
2005 err = ldiskfs_journal_get_write_access(handle, bh);
2007 CERROR("journal_get_write_access() returned error %d\n",
2011 LASSERTF(boffs + size <= bh->b_size,
2012 "boffs %d size %d bh->b_size %lu\n",
2013 boffs, size, (unsigned long)bh->b_size);
2014 memcpy(bh->b_data + boffs, buf, size);
2015 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2019 if (offset + size > new_size)
2020 new_size = offset + size;
2030 /* correct in-core and on-disk sizes */
2031 if (new_size > i_size_read(inode)) {
2032 spin_lock(&inode->i_lock);
2033 if (new_size > i_size_read(inode))
2034 i_size_write(inode, new_size);
2035 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
2036 LDISKFS_I(inode)->i_disksize = i_size_read(inode);
2039 spin_unlock(&inode->i_lock);
2041 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2049 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2050 const struct lu_buf *buf, loff_t *pos,
2051 struct thandle *handle)
2053 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2054 struct osd_thandle *oh;
2058 LASSERT(dt_object_exists(dt));
2060 LASSERT(handle != NULL);
2061 LASSERT(inode != NULL);
2062 ll_vfs_dq_init(inode);
2064 /* XXX: don't check: one declared chunk can be used many times */
2065 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2067 oh = container_of(handle, struct osd_thandle, ot_super);
2068 LASSERT(oh->ot_handle->h_transaction != NULL);
2069 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2071 /* Write small symlink to inode body as we need to maintain correct
2072 * on-disk symlinks for ldiskfs.
2073 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2074 * does not count it in.
2076 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2077 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2078 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2080 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
2081 buf->lb_len, is_link, pos,
2084 result = buf->lb_len;
2086 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2091 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2092 __u64 start, __u64 end, struct thandle *th)
2094 struct osd_thandle *oh;
2095 struct inode *inode;
2100 oh = container_of(th, struct osd_thandle, ot_super);
2103 * we don't need to reserve credits for whole truncate
2104 * it's not possible as truncate may need to free too many
2105 * blocks and that won't fit a single transaction. instead
2106 * we reserve credits to change i_size and put inode onto
2107 * orphan list. if needed truncate will extend or restart
2110 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2111 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2113 inode = osd_dt_obj(dt)->oo_inode;
2116 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2117 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2121 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2126 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2127 __u64 start, __u64 end, struct thandle *th)
2129 struct osd_object *obj = osd_dt_obj(dt);
2130 struct osd_device *osd = osd_obj2dev(obj);
2131 struct inode *inode = obj->oo_inode;
2132 struct osd_access_lock *al;
2133 struct osd_thandle *oh;
2134 int rc = 0, found = 0;
2138 LASSERT(end == OBD_OBJECT_EOF);
2139 LASSERT(dt_object_exists(dt));
2140 LASSERT(osd_invariant(obj));
2141 LASSERT(inode != NULL);
2142 ll_vfs_dq_init(inode);
2145 oh = container_of(th, struct osd_thandle, ot_super);
2146 LASSERT(oh->ot_handle->h_transaction != NULL);
2148 /* we used to skip truncate to current size to
2149 * optimize truncates on OST. with DoM we can
2150 * get attr_set to set specific size (MDS_REINT)
2151 * and then get truncate RPC which essentially
2152 * would be skipped. this is bad.. so, disable
2153 * this optimization on MDS till the client stop
2154 * to sent MDS_REINT (LU-11033) -bzzz */
2155 if (osd->od_is_ost && i_size_read(inode) == start)
2158 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2160 spin_lock(&inode->i_lock);
2161 if (i_size_read(inode) < start)
2163 i_size_write(inode, start);
2164 spin_unlock(&inode->i_lock);
2165 ll_truncate_pagecache(inode, start);
2167 /* optimize grow case */
2169 osd_execute_truncate(obj);
2173 /* add to orphan list to ensure truncate completion
2174 * if this transaction succeed. ldiskfs_truncate()
2175 * will take the inode out of the list */
2176 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2180 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2181 if (obj != al->tl_obj)
2183 LASSERT(al->tl_shared == 0);
2185 /* do actual truncate in osd_trans_stop() */
2186 al->tl_truncate = 1;
2195 static int fiemap_check_ranges(struct inode *inode,
2196 u64 start, u64 len, u64 *new_len)
2205 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2206 maxbytes = inode->i_sb->s_maxbytes;
2208 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2210 if (start > maxbytes)
2214 * Shrink request scope to what the fs can actually handle.
2216 if (len > maxbytes || (maxbytes - len) < start)
2217 *new_len = maxbytes - start;
2222 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2223 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2225 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2228 struct fiemap_extent_info fieinfo = {0, };
2229 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2235 if (inode->i_op->fiemap == NULL)
2238 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2241 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2245 fieinfo.fi_flags = fm->fm_flags;
2246 fieinfo.fi_extents_max = fm->fm_extent_count;
2247 fieinfo.fi_extents_start = fm->fm_extents;
2249 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2250 filemap_write_and_wait(inode->i_mapping);
2252 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2253 fm->fm_flags = fieinfo.fi_flags;
2254 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2259 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2260 __u64 start, __u64 end, enum lu_ladvise_type advice)
2263 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2267 case LU_LADVISE_DONTNEED:
2270 invalidate_mapping_pages(inode->i_mapping,
2271 start >> PAGE_SHIFT,
2272 (end - 1) >> PAGE_SHIFT);
2283 * in some cases we may need declare methods for objects being created
2284 * e.g., when we create symlink
2286 const struct dt_body_operations osd_body_ops_new = {
2287 .dbo_declare_write = osd_declare_write,
2290 const struct dt_body_operations osd_body_ops = {
2291 .dbo_read = osd_read,
2292 .dbo_declare_write = osd_declare_write,
2293 .dbo_write = osd_write,
2294 .dbo_bufs_get = osd_bufs_get,
2295 .dbo_bufs_put = osd_bufs_put,
2296 .dbo_write_prep = osd_write_prep,
2297 .dbo_declare_write_commit = osd_declare_write_commit,
2298 .dbo_write_commit = osd_write_commit,
2299 .dbo_read_prep = osd_read_prep,
2300 .dbo_declare_punch = osd_declare_punch,
2301 .dbo_punch = osd_punch,
2302 .dbo_fiemap_get = osd_fiemap_get,
2303 .dbo_ladvise = osd_ladvise,
2307 * Get a truncate lock
2309 * In order to take multi-transaction truncate out of main transaction we let
2310 * the caller grab a lock on the object passed. the lock can be shared (for
2311 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2312 * and write in the same transaction handle (do not confuse with big ldiskfs
2313 * transaction containing lots of handles).
2314 * The lock must be taken at declaration.
2316 * \param obj object to lock
2318 * \shared shared or exclusive
2320 * \retval 0 lock is granted
2321 * \retval -NOMEM no memory to allocate lock
2323 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2325 struct osd_access_lock *al, *tmp;
2330 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2331 if (tmp->tl_obj != obj)
2333 LASSERT(tmp->tl_shared == shared);
2334 /* found same lock */
2339 if (unlikely(al == NULL))
2342 al->tl_truncate = false;
2344 down_read(&obj->oo_ext_idx_sem);
2346 down_write(&obj->oo_ext_idx_sem);
2347 al->tl_shared = shared;
2349 list_add(&al->tl_list, &oh->ot_trunc_locks);
2354 void osd_trunc_unlock_all(struct list_head *list)
2356 struct osd_access_lock *al, *tmp;
2357 list_for_each_entry_safe(al, tmp, list, tl_list) {
2359 up_read(&al->tl_obj->oo_ext_idx_sem);
2361 up_write(&al->tl_obj->oo_ext_idx_sem);
2362 list_del(&al->tl_list);
2367 void osd_execute_truncate(struct osd_object *obj)
2369 struct osd_device *d = osd_obj2dev(obj);
2370 struct inode *inode = obj->oo_inode;
2373 /* simulate crash before (in the middle) of delayed truncate */
2374 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2375 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2376 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2378 mutex_lock(&sbi->s_orphan_lock);
2379 list_del_init(&ei->i_orphan);
2380 mutex_unlock(&sbi->s_orphan_lock);
2384 ldiskfs_truncate(inode);
2387 * For a partial-page truncate, flush the page to disk immediately to
2388 * avoid data corruption during direct disk write. b=17397
2390 size = i_size_read(inode);
2391 if ((size & ~PAGE_MASK) == 0)
2393 if (osd_use_page_cache(d)) {
2394 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2396 /* Notice we use "wait" version to ensure I/O is complete */
2397 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2398 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2399 size >> PAGE_SHIFT);
2403 void osd_process_truncates(struct list_head *list)
2405 struct osd_access_lock *al;
2407 LASSERT(journal_current_handle() == NULL);
2409 list_for_each_entry(al, list, tl_list) {
2412 if (!al->tl_truncate)
2414 osd_execute_truncate(al->tl_obj);