4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 #define DEBUG_SUBSYSTEM S_OSD
43 /* prerequisite for linux/xattr.h */
44 #include <linux/types.h>
45 /* prerequisite for linux/xattr.h */
48 #include <linux/pagevec.h>
51 * struct OBD_{ALLOC,FREE}*()
54 #include <obd_support.h>
56 #include "osd_internal.h"
59 #include <ldiskfs/ldiskfs_extents.h>
61 static inline bool osd_use_page_cache(struct osd_device *d)
63 /* do not use pagecache if write and read caching are disabled */
64 if (d->od_writethrough_cache + d->od_read_cache == 0)
66 /* use pagecache by default */
70 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
71 int rw, int line, int pages)
75 LASSERTF(iobuf->dr_elapsed_valid == 0,
76 "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
77 atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
79 LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
81 init_waitqueue_head(&iobuf->dr_wait);
82 atomic_set(&iobuf->dr_numreqs, 0);
87 iobuf->dr_elapsed = ktime_set(0, 0);
88 /* must be counted before, so assert */
90 iobuf->dr_init_at = line;
92 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
93 if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
94 LASSERT(iobuf->dr_pg_buf.lb_len >=
95 pages * sizeof(iobuf->dr_pages[0]));
99 /* start with 1MB for 4K blocks */
101 while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
104 CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
105 (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
107 blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
108 iobuf->dr_max_pages = 0;
109 CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
110 (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
112 lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
113 iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
114 if (unlikely(iobuf->dr_blocks == NULL))
117 lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
118 iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
119 if (unlikely(iobuf->dr_pages == NULL))
122 lu_buf_realloc(&iobuf->dr_lnb_buf,
123 pages * sizeof(iobuf->dr_lnbs[0]));
124 iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
125 if (unlikely(iobuf->dr_lnbs == NULL))
128 iobuf->dr_max_pages = pages;
132 #define osd_init_iobuf(dev, iobuf, rw, pages) \
133 __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
135 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
136 struct niobuf_local *lnb)
138 LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
139 iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
140 iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
144 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
146 int rw = iobuf->dr_rw;
148 if (iobuf->dr_elapsed_valid) {
149 iobuf->dr_elapsed_valid = 0;
150 LASSERT(iobuf->dr_dev == d);
151 LASSERT(iobuf->dr_frags > 0);
152 lprocfs_oh_tally(&d->od_brw_stats.hist[BRW_R_DIO_FRAGS+rw],
154 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
155 ktime_to_ms(iobuf->dr_elapsed));
159 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
160 static void dio_complete_routine(struct bio *bio)
162 int error = bio->bi_status;
164 static void dio_complete_routine(struct bio *bio, int error)
167 struct osd_iobuf *iobuf = bio->bi_private;
170 /* CAVEAT EMPTOR: possibly in IRQ context
171 * DO NOT record procfs stats here!!!
174 if (unlikely(iobuf == NULL)) {
175 CERROR("***** bio->bi_private is NULL! This should never happen. Normally, I would crash here, but instead I will dump the bio contents to the console. Please report this to <https://jira.whamcloud.com/> , along with any interesting messages leading up to this point (like SCSI errors, perhaps). Because bi_private is NULL, I can't wake up the thread that initiated this IO - you will probably have to reboot this node.\n");
176 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
177 ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
178 bio->bi_next, (unsigned long)bio->bi_flags,
179 (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
180 bio_sectors(bio) << 9, bio->bi_end_io,
181 atomic_read(&bio->__bi_cnt),
186 /* the check is outside of the cycle for performance reason -bzzz */
187 if (!bio_data_dir(bio)) {
188 DECLARE_BVEC_ITER_ALL(iter_all);
190 bio_for_each_segment_all(bvl, bio, iter_all) {
191 if (likely(error == 0))
192 SetPageUptodate(bvl_to_page(bvl));
193 LASSERT(PageLocked(bvl_to_page(bvl)));
195 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
197 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
200 /* any real error is good enough -bzzz */
201 if (error != 0 && iobuf->dr_error == 0)
202 iobuf->dr_error = error;
205 * set dr_elapsed before dr_numreqs turns to 0, otherwise
206 * it's possible that service thread will see dr_numreqs
207 * is zero, but dr_elapsed is not set yet, leading to lost
208 * data in this processing and an assertion in a subsequent
211 if (atomic_read(&iobuf->dr_numreqs) == 1) {
212 ktime_t now = ktime_get();
214 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
215 iobuf->dr_elapsed_valid = 1;
217 if (atomic_dec_and_test(&iobuf->dr_numreqs))
218 wake_up(&iobuf->dr_wait);
220 /* Completed bios used to be chained off iobuf->dr_bios and freed in
221 * filter_clear_dreq(). It was then possible to exhaust the biovec-256
222 * mempool when serious on-disk fragmentation was encountered,
223 * deadlocking the OST. The bios are now released as soon as complete
224 * so the pool cannot be exhausted while IOs are competing. b=10076
229 static void record_start_io(struct osd_iobuf *iobuf, int size)
231 struct osd_device *osd = iobuf->dr_dev;
232 struct obd_histogram *h = osd->od_brw_stats.hist;
235 atomic_inc(&iobuf->dr_numreqs);
237 if (iobuf->dr_rw == 0) {
238 atomic_inc(&osd->od_r_in_flight);
239 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
240 atomic_read(&osd->od_r_in_flight));
241 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
242 } else if (iobuf->dr_rw == 1) {
243 atomic_inc(&osd->od_w_in_flight);
244 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
245 atomic_read(&osd->od_w_in_flight));
246 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
252 static void osd_submit_bio(int rw, struct bio *bio)
254 LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
255 #ifdef HAVE_SUBMIT_BIO_2ARGS
256 submit_bio(rw ? WRITE : READ, bio);
263 static int can_be_merged(struct bio *bio, sector_t sector)
268 return bio_end_sector(bio) == sector ? 1 : 0;
271 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
273 * This function will change the data written, thus it should only be
274 * used when checking data integrity feature
276 static void bio_integrity_fault_inject(struct bio *bio)
278 struct bio_vec *bvec;
279 DECLARE_BVEC_ITER_ALL(iter_all);
283 bio_for_each_segment_all(bvec, bio, iter_all) {
284 struct page *page = bvec->bv_page;
294 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
295 unsigned int sectors, int tuple_size)
297 __u16 *expected_guard;
301 expected_guard = expected_guard_buf;
302 for (i = 0; i < sectors; i++) {
303 bio_guard = (__u16 *)bio_prot_buf;
304 if (*bio_guard != *expected_guard) {
306 "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
307 i, *expected_guard, *bio_guard, sectors,
312 bio_prot_buf += tuple_size;
317 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
318 struct osd_iobuf *iobuf, int index)
320 struct blk_integrity *bi = bdev_get_integrity(bdev);
321 struct bio_integrity_payload *bip = bio->bi_integrity;
322 struct niobuf_local *lnb;
323 unsigned short sector_size = blk_integrity_interval(bi);
324 void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
325 bip->bip_vec->bv_offset;
327 sector_t sector = bio_start_sector(bio);
328 unsigned int sectors, total;
329 DECLARE_BVEC_ITER_ALL(iter_all);
330 __u16 *expected_guard;
334 bio_for_each_segment_all(bv, bio, iter_all) {
335 lnb = iobuf->dr_lnbs[index];
336 expected_guard = lnb->lnb_guards;
337 sectors = bv->bv_len / sector_size;
338 if (lnb->lnb_guard_rpc) {
339 rc = bio_dif_compare(expected_guard, bio_prot_buf,
340 sectors, bi->tuple_size);
346 bio_prot_buf += sectors * bi->tuple_size;
347 total += sectors * bi->tuple_size;
348 LASSERT(total <= bip_size(bio->bi_integrity));
354 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
355 struct osd_iobuf *iobuf,
356 int start_page_idx, bool fault_inject,
357 bool integrity_enabled)
359 struct super_block *sb = osd_sb(osd);
360 integrity_gen_fn *generate_fn = NULL;
361 integrity_vrfy_fn *verify_fn = NULL;
366 if (!integrity_enabled)
369 rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
373 rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
377 /* Verify and inject fault only when writing */
378 if (iobuf->dr_rw == 1) {
379 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
380 rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
386 if (unlikely(fault_inject))
387 bio_integrity_fault_inject(bio);
393 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
394 # ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
395 static void dio_integrity_complete_routine(struct bio *bio)
397 static void dio_integrity_complete_routine(struct bio *bio, int error)
400 struct osd_bio_private *bio_private = bio->bi_private;
402 bio->bi_private = bio_private->obp_iobuf;
403 osd_dio_complete_routine(bio, error);
405 OBD_FREE_PTR(bio_private);
407 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
408 #else /* !CONFIG_BLK_DEV_INTEGRITY */
409 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
410 fault_inject, integrity_enabled) 0
411 #endif /* CONFIG_BLK_DEV_INTEGRITY */
413 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
414 bool integrity_enabled, int start_page_idx,
415 struct osd_bio_private **pprivate)
421 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
422 if (integrity_enabled) {
423 struct osd_bio_private *bio_private = NULL;
425 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
426 if (bio_private == NULL)
428 bio->bi_end_io = dio_integrity_complete_routine;
429 bio->bi_private = bio_private;
430 bio_private->obp_start_page_idx = start_page_idx;
431 bio_private->obp_iobuf = iobuf;
432 *pprivate = bio_private;
436 bio->bi_end_io = dio_complete_routine;
437 bio->bi_private = iobuf;
443 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
444 struct osd_iobuf *iobuf)
446 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
447 struct page **pages = iobuf->dr_pages;
448 int npages = iobuf->dr_npages;
449 sector_t *blocks = iobuf->dr_blocks;
450 int total_blocks = npages * blocks_per_page;
451 struct super_block *sb = inode->i_sb;
452 int sector_bits = sb->s_blocksize_bits - 9;
453 unsigned int blocksize = sb->s_blocksize;
454 struct block_device *bdev = sb->s_bdev;
455 struct osd_bio_private *bio_private = NULL;
456 struct bio *bio = NULL;
457 int bio_start_page_idx;
459 unsigned int page_offset;
467 bool integrity_enabled;
468 struct blk_plug plug;
471 fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
472 LASSERT(iobuf->dr_npages == npages);
474 integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
476 osd_brw_stats_update(osd, iobuf);
477 iobuf->dr_start_time = ktime_get();
479 blk_start_plug(&plug);
480 for (page_idx = 0, block_idx = 0;
482 page_idx++, block_idx += blocks_per_page) {
483 page = pages[page_idx];
484 LASSERT(block_idx + blocks_per_page <= total_blocks);
486 for (i = 0, page_offset = 0;
488 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] <<
509 if (bio && can_be_merged(bio, sector) &&
510 bio_add_page(bio, page, blocksize * nblocks,
512 continue; /* added this frag OK */
515 struct request_queue *q = bio_get_queue(bio);
516 unsigned int bi_size = bio_sectors(bio) << 9;
518 /* Dang! I have to fragment this I/O */
520 "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
521 bi_size, bio->bi_vcnt, bio->bi_max_vecs,
523 queue_max_sectors(q),
524 osd_bio_nr_segs(bio),
525 queue_max_segments(q));
526 rc = osd_bio_integrity_handle(osd, bio,
527 iobuf, bio_start_page_idx,
528 fault_inject, integrity_enabled);
534 record_start_io(iobuf, bi_size);
535 osd_submit_bio(iobuf->dr_rw, bio);
538 bio_start_page_idx = page_idx;
539 /* allocate new bio */
540 bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
541 (npages - page_idx) *
544 CERROR("Can't allocate bio %u*%u = %u pages\n",
545 (npages - page_idx), blocks_per_page,
546 (npages - page_idx) * blocks_per_page);
551 bio_set_dev(bio, bdev);
552 bio_set_sector(bio, sector);
553 bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
554 rc = osd_bio_init(bio, iobuf, integrity_enabled,
555 bio_start_page_idx, &bio_private);
561 rc = bio_add_page(bio, page,
562 blocksize * nblocks, page_offset);
568 rc = osd_bio_integrity_handle(osd, bio, iobuf,
577 record_start_io(iobuf, bio_sectors(bio) << 9);
578 osd_submit_bio(iobuf->dr_rw, bio);
583 blk_finish_plug(&plug);
585 /* in order to achieve better IO throughput, we don't wait for writes
586 * completion here. instead we proceed with transaction commit in
587 * parallel and wait for IO completion once transaction is stopped
588 * see osd_trans_stop() for more details -bzzz
590 if (iobuf->dr_rw == 0 || fault_inject) {
591 wait_event(iobuf->dr_wait,
592 atomic_read(&iobuf->dr_numreqs) == 0);
593 osd_fini_iobuf(osd, iobuf);
597 rc = iobuf->dr_error;
600 OBD_FREE_PTR(bio_private);
606 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
607 struct niobuf_local *lnb, int maxlnb)
615 int poff = offset & (PAGE_SIZE - 1);
616 int plen = PAGE_SIZE - poff;
618 if (*nrpages >= maxlnb) {
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;
636 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
647 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
648 loff_t offset, gfp_t gfp_mask, bool cache)
650 struct osd_thread_info *oti = osd_oti_get(env);
651 struct inode *inode = osd_dt_obj(dt)->oo_inode;
652 struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
659 page = find_or_create_page(inode->i_mapping,
660 offset >> PAGE_SHIFT, gfp_mask);
663 LASSERT(!PagePrivate2(page));
664 wait_on_page_writeback(page);
666 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
672 if (inode->i_mapping->nrpages) {
673 /* consult with pagecache, but do not create new pages */
674 /* this is normally used once */
675 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
677 wait_on_page_writeback(page);
682 LASSERT(oti->oti_dio_pages);
683 cur = oti->oti_dio_pages_used;
684 page = oti->oti_dio_pages[cur];
686 if (unlikely(!page)) {
687 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
688 page = alloc_page(gfp_mask);
691 oti->oti_dio_pages[cur] = page;
692 SetPagePrivate2(page);
696 ClearPageUptodate(page);
697 page->index = offset >> PAGE_SHIFT;
698 oti->oti_dio_pages_used++;
704 * there are following "locks":
715 * - lock pages, unlock
717 * - lock partial page
723 * Unlock and release pages loaded by osd_bufs_get()
725 * Unlock \a npages pages from \a lnb and drop the refcount on them.
727 * \param env thread execution environment
728 * \param dt dt object undergoing IO (OSD object + methods)
729 * \param lnb array of pages undergoing IO
730 * \param npages number of pages in \a lnb
734 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
735 struct niobuf_local *lnb, int npages)
737 struct osd_thread_info *oti = osd_oti_get(env);
741 ll_pagevec_init(&pvec, 0);
743 for (i = 0; i < npages; i++) {
744 struct page *page = lnb[i].lnb_page;
749 /* if the page isn't cached, then reset uptodate
752 if (PagePrivate2(page)) {
753 oti->oti_dio_pages_used--;
755 if (lnb[i].lnb_locked)
757 if (pagevec_add(&pvec, page) == 0)
758 pagevec_release(&pvec);
761 lnb[i].lnb_page = NULL;
764 LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
766 /* Release any partial pagevec */
767 pagevec_release(&pvec);
773 * Load and lock pages undergoing IO
775 * Pages as described in the \a lnb array are fetched (from disk or cache)
776 * and locked for IO by the caller.
778 * DLM locking protects us from write and truncate competing for same region,
779 * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
780 * It's possible the writeout on a such a page is in progress when we access
781 * it. It's also possible that during this writeout we put new (partial) data
782 * into the page, but won't be able to proceed in filter_commitrw_write().
783 * Therefore, just wait for writeout completion as it should be rare enough.
785 * \param env thread execution environment
786 * \param dt dt object undergoing IO (OSD object + methods)
787 * \param pos byte offset of IO start
788 * \param len number of bytes of IO
789 * \param lnb array of extents undergoing IO
790 * \param rw read or write operation, and other flags
791 * \param capa capabilities
793 * \retval pages (zero or more) loaded successfully
794 * \retval -ENOMEM on memory/page allocation error
796 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
797 loff_t pos, ssize_t len, struct niobuf_local *lnb,
798 int maxlnb, enum dt_bufs_type rw)
800 struct osd_thread_info *oti = osd_oti_get(env);
801 struct osd_object *obj = osd_dt_obj(dt);
802 struct osd_device *osd = osd_obj2dev(obj);
803 int npages, i, iosize, rc = 0;
808 LASSERT(obj->oo_inode);
810 rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
814 write = rw & DT_BUFS_TYPE_WRITE;
816 fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
817 iosize = fsize - lnb[0].lnb_file_offset;
818 fsize = max(fsize, i_size_read(obj->oo_inode));
820 cache = rw & DT_BUFS_TYPE_READAHEAD;
824 cache = osd_use_page_cache(osd);
827 if (!osd->od_writethrough_cache) {
831 if (iosize > osd->od_writethrough_max_iosize) {
836 if (!osd->od_read_cache) {
840 if (iosize > osd->od_readcache_max_iosize) {
845 /* don't use cache on large files */
846 if (osd->od_readcache_max_filesize &&
847 fsize > osd->od_readcache_max_filesize)
853 if (!cache && unlikely(!oti->oti_dio_pages)) {
854 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
855 PTLRPC_MAX_BRW_PAGES);
856 if (!oti->oti_dio_pages)
860 /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
861 gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
863 for (i = 0; i < npages; i++, lnb++) {
864 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
866 if (lnb->lnb_page == NULL)
867 GOTO(cleanup, rc = -ENOMEM);
873 /* XXX: this version doesn't invalidate cached pages, but use them */
874 if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
875 /* do not allow data aliasing, invalidate pagecache */
876 /* XXX: can be quite expensive in mixed case */
877 invalidate_mapping_pages(obj->oo_inode->i_mapping,
878 lnb[0].lnb_file_offset >> PAGE_SHIFT,
879 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
887 osd_bufs_put(env, dt, lnb - i, i);
891 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
892 int pages, sector_t *blocks,
895 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
897 struct page *fp = NULL;
899 pgoff_t max_page_index;
900 handle_t *handle = NULL;
902 max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
904 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
905 inode->i_ino, pages, (*page)->index);
908 create = LDISKFS_GET_BLOCKS_CREATE;
909 handle = ldiskfs_journal_current_handle();
910 LASSERT(handle != NULL);
911 rc = osd_attach_jinode(inode);
915 /* pages are sorted already. so, we just have to find
916 * contig. space and process them properly
919 long blen, total = 0;
920 struct ldiskfs_map_blocks map = { 0 };
922 if (fp == NULL) { /* start new extent */
927 } else if (fp->index + clen == (*page)->index) {
928 /* continue the extent */
934 if (fp->index + clen >= max_page_index)
935 GOTO(cleanup, rc = -EFBIG);
936 /* process found extent */
937 map.m_lblk = fp->index * blocks_per_page;
938 map.m_len = blen = clen * blocks_per_page;
940 rc = ldiskfs_map_blocks(handle, inode, &map, create);
944 for (; total < blen && c < map.m_len; c++, total++) {
946 *(blocks + total) = 0;
950 *(blocks + total) = map.m_pblk + c;
951 /* unmap any possible underlying
952 * metadata from the block device
955 if ((map.m_flags & LDISKFS_MAP_NEW) &&
957 clean_bdev_aliases(inode->i_sb->s_bdev,
962 if (rc == 0 && total < blen) {
963 map.m_lblk = fp->index * blocks_per_page + total;
964 map.m_len = blen - total;
970 /* look for next extent */
972 blocks += blocks_per_page * clen;
978 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
979 struct niobuf_local *lnb, int npages)
981 struct osd_thread_info *oti = osd_oti_get(env);
982 struct osd_iobuf *iobuf = &oti->oti_iobuf;
983 struct inode *inode = osd_dt_obj(dt)->oo_inode;
984 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
993 rc = osd_init_iobuf(osd, iobuf, 0, npages);
994 if (unlikely(rc != 0))
997 isize = i_size_read(inode);
998 maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1000 start = ktime_get();
1001 for (i = 0; i < npages; i++) {
1004 * till commit the content of the page is undefined
1005 * we'll set it uptodate once bulk is done. otherwise
1006 * subsequent reads can access non-stable data
1008 ClearPageUptodate(lnb[i].lnb_page);
1010 if (lnb[i].lnb_len == PAGE_SIZE)
1013 if (maxidx >= lnb[i].lnb_page->index) {
1014 osd_iobuf_add_page(iobuf, &lnb[i]);
1017 char *p = kmap(lnb[i].lnb_page);
1019 off = lnb[i].lnb_page_offset;
1022 off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1025 memset(p + off, 0, PAGE_SIZE - off);
1026 kunmap(lnb[i].lnb_page);
1030 timediff = ktime_us_delta(end, start);
1031 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1033 if (iobuf->dr_npages) {
1034 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1036 iobuf->dr_blocks, 0);
1037 if (likely(rc == 0)) {
1038 rc = osd_do_bio(osd, inode, iobuf);
1039 /* do IO stats for preparation reads */
1040 osd_fini_iobuf(osd, iobuf);
1046 struct osd_fextent {
1049 unsigned int mapped:1;
1052 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1053 struct osd_fextent *cached_extent)
1055 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1056 sector_t block = offset >> inode->i_blkbits;
1058 struct fiemap_extent_info fei = { 0 };
1059 struct fiemap_extent fe = { 0 };
1060 mm_segment_t saved_fs;
1063 if (block >= cached_extent->start && block < cached_extent->end)
1064 return cached_extent->mapped;
1066 if (i_size_read(inode) == 0)
1069 /* Beyond EOF, must not be mapped */
1070 if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1073 fei.fi_extents_max = 1;
1074 fei.fi_extents_start = &fe;
1076 saved_fs = get_fs();
1078 rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1083 start = fe.fe_logical >> inode->i_blkbits;
1085 if (start > block) {
1086 cached_extent->start = block;
1087 cached_extent->end = start;
1088 cached_extent->mapped = 0;
1090 cached_extent->start = start;
1091 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1093 cached_extent->mapped = 1;
1096 return cached_extent->mapped;
1099 static int osd_declare_write_commit(const struct lu_env *env,
1100 struct dt_object *dt,
1101 struct niobuf_local *lnb, int npages,
1102 struct thandle *handle)
1104 const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1105 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1106 struct osd_thandle *oh;
1113 long long quota_space = 0;
1114 struct osd_fextent mapped = { 0 }, extent = { 0 };
1115 enum osd_quota_local_flags local_flags = 0;
1116 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1119 LASSERT(handle != NULL);
1120 oh = container_of(handle, struct osd_thandle, ot_super);
1121 LASSERT(oh->ot_handle == NULL);
1123 /* calculate number of extents (probably better to pass nb) */
1124 for (i = 0; i < npages; i++) {
1125 /* ignore quota for the whole request if any page is from
1126 * client cache or written by root.
1128 * XXX once we drop the 1.8 client support, the checking
1129 * for whether page is from cache can be simplified as:
1130 * !(lnb[i].flags & OBD_BRW_SYNC)
1132 * XXX we could handle this on per-lnb basis as done by
1135 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1136 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1138 declare_flags |= OSD_QID_FORCE;
1140 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1141 lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1145 /* count only unmapped changes */
1147 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1149 extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1151 extent.end += lnb[i].lnb_len;
1154 quota_space += PAGE_SIZE;
1157 credits++; /* inode */
1159 * overwrite case, no need to modify tree and
1165 * each extent can go into new leaf causing a split
1166 * 5 is max tree depth: inode + 4 index blocks
1167 * with blockmaps, depth is 3 at most
1169 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1171 * many concurrent threads may grow tree by the time
1172 * our transaction starts. so, consider 2 is a min depth
1174 depth = ext_depth(inode);
1175 depth = max(depth, 1) + 1;
1177 credits += depth * 2 * extents;
1181 credits += depth * extents;
1184 /* quota space for metadata blocks */
1185 quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1187 /* quota space should be reported in 1K blocks */
1188 quota_space = toqb(quota_space);
1190 /* each new block can go in different group (bitmap + gd) */
1192 /* we can't dirty more bitmap blocks than exist */
1193 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1194 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1196 credits += newblocks;
1198 /* we can't dirty more gd blocks than exist */
1199 if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1200 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1202 credits += newblocks;
1205 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1207 /* make sure the over quota flags were not set */
1208 lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1210 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1211 i_projid_read(inode), quota_space, oh,
1212 osd_dt_obj(dt), &local_flags, declare_flags);
1214 /* we need only to store the overquota flags in the first lnb for
1215 * now, once we support multiple objects BRW, this code needs be
1218 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1219 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1220 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1221 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1222 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1223 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1226 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1231 /* Check if a block is allocated or not */
1232 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1233 struct niobuf_local *lnb, int npages,
1234 struct thandle *thandle, __u64 user_size)
1236 struct osd_thread_info *oti = osd_oti_get(env);
1237 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1238 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1239 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1245 rc = osd_init_iobuf(osd, iobuf, 1, npages);
1246 if (unlikely(rc != 0))
1249 disk_size = i_size_read(inode);
1250 /* if disk_size is already bigger than specified user_size,
1253 if (disk_size > user_size)
1255 dquot_initialize(inode);
1257 for (i = 0; i < npages; i++) {
1258 if (lnb[i].lnb_rc == -ENOSPC &&
1259 (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1260 /* Allow the write to proceed if overwriting an
1266 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1267 CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1269 LASSERT(lnb[i].lnb_page);
1270 generic_error_remove_page(inode->i_mapping,
1275 LASSERT(PageLocked(lnb[i].lnb_page));
1276 LASSERT(!PageWriteback(lnb[i].lnb_page));
1278 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size)
1279 disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1282 * Since write and truncate are serialized by oo_sem, even
1283 * partial-page truncate should not leave dirty pages in the
1286 LASSERT(!PageDirty(lnb[i].lnb_page));
1288 SetPageUptodate(lnb[i].lnb_page);
1290 osd_iobuf_add_page(iobuf, &lnb[i]);
1293 /* if file has grown, take user_size into account */
1294 if (user_size && disk_size > user_size)
1295 disk_size = user_size;
1297 osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1299 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1301 } else if (iobuf->dr_npages > 0) {
1302 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1304 iobuf->dr_blocks, 1);
1306 /* no pages to write, no transno is needed */
1307 thandle->th_local = 1;
1310 if (likely(rc == 0)) {
1311 spin_lock(&inode->i_lock);
1312 if (disk_size > i_size_read(inode)) {
1313 i_size_write(inode, disk_size);
1314 LDISKFS_I(inode)->i_disksize = disk_size;
1315 spin_unlock(&inode->i_lock);
1316 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1318 spin_unlock(&inode->i_lock);
1321 rc = osd_do_bio(osd, inode, iobuf);
1322 /* we don't do stats here as in read path because
1323 * write is async: we'll do this in osd_put_bufs()
1326 osd_fini_iobuf(osd, iobuf);
1329 osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1331 if (unlikely(rc != 0)) {
1332 /* if write fails, we should drop pages from the cache */
1333 for (i = 0; i < npages; i++) {
1334 if (lnb[i].lnb_page == NULL)
1336 if (!PagePrivate2(lnb[i].lnb_page)) {
1337 LASSERT(PageLocked(lnb[i].lnb_page));
1338 generic_error_remove_page(inode->i_mapping,
1347 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1348 struct niobuf_local *lnb, int npages)
1350 struct osd_thread_info *oti = osd_oti_get(env);
1351 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1352 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1353 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1354 int rc = 0, i, cache_hits = 0, cache_misses = 0;
1361 rc = osd_init_iobuf(osd, iobuf, 0, npages);
1362 if (unlikely(rc != 0))
1365 isize = i_size_read(inode);
1367 start = ktime_get();
1368 for (i = 0; i < npages; i++) {
1370 if (isize <= lnb[i].lnb_file_offset)
1371 /* If there's no more data, abort early.
1372 * lnb->lnb_rc == 0, so it's easy to detect later.
1376 /* instead of looking if we go beyong isize, send complete
1377 * pages all the time
1379 lnb[i].lnb_rc = lnb[i].lnb_len;
1381 /* Bypass disk read if fail_loc is set properly */
1382 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1383 SetPageUptodate(lnb[i].lnb_page);
1385 if (PageUptodate(lnb[i].lnb_page)) {
1387 unlock_page(lnb[i].lnb_page);
1390 osd_iobuf_add_page(iobuf, &lnb[i]);
1392 /* no need to unlock in osd_bufs_put(), the sooner page is
1393 * unlocked, the earlier another client can access it.
1394 * notice real unlock_page() can be called few lines
1395 * below after osd_do_bio(). lnb is a per-thread, so it's
1396 * fine to have PG_locked and lnb_locked inconsistent here
1398 lnb[i].lnb_locked = 0;
1401 timediff = ktime_us_delta(end, start);
1402 lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1404 if (cache_hits != 0)
1405 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1407 if (cache_misses != 0)
1408 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1410 if (cache_hits + cache_misses != 0)
1411 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1412 cache_hits + cache_misses);
1414 if (iobuf->dr_npages) {
1415 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1417 iobuf->dr_blocks, 0);
1418 rc = osd_do_bio(osd, inode, iobuf);
1420 /* IO stats will be done in osd_bufs_put() */
1422 /* early release to let others read data during the bulk */
1423 for (i = 0; i < iobuf->dr_npages; i++) {
1424 LASSERT(PageLocked(iobuf->dr_pages[i]));
1425 if (!PagePrivate2(iobuf->dr_pages[i]))
1426 unlock_page(iobuf->dr_pages[i]);
1434 * XXX: Another layering violation for now.
1436 * We don't want to use ->f_op->read methods, because generic file write
1438 * - serializes on ->i_sem, and
1440 * - does a lot of extra work like balance_dirty_pages(),
1442 * which doesn't work for globally shared files like /last_rcvd.
1444 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1446 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1448 memcpy(buffer, (char *)ei->i_data, buflen);
1453 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1455 struct buffer_head *bh;
1456 unsigned long block;
1462 /* prevent reading after eof */
1463 spin_lock(&inode->i_lock);
1464 if (i_size_read(inode) < *offs + size) {
1465 loff_t diff = i_size_read(inode) - *offs;
1467 spin_unlock(&inode->i_lock);
1470 "size %llu is too short to read @%llu\n",
1471 i_size_read(inode), *offs);
1473 } else if (diff == 0) {
1479 spin_unlock(&inode->i_lock);
1482 blocksize = 1 << inode->i_blkbits;
1485 block = *offs >> inode->i_blkbits;
1486 boffs = *offs & (blocksize - 1);
1487 csize = min(blocksize - boffs, size);
1488 bh = __ldiskfs_bread(NULL, inode, block, 0);
1490 CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1491 osd_ino2name(inode), csize, *offs, inode->i_ino,
1497 memcpy(buf, bh->b_data + boffs, csize);
1500 memset(buf, 0, csize);
1510 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1511 struct lu_buf *buf, loff_t *pos)
1513 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1516 /* Read small symlink from inode body as we need to maintain correct
1517 * on-disk symlinks for ldiskfs.
1519 if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1520 loff_t size = i_size_read(inode);
1522 if (buf->lb_len < size)
1525 if (size < sizeof(LDISKFS_I(inode)->i_data))
1526 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1528 rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1530 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1536 static inline int osd_extents_enabled(struct super_block *sb,
1537 struct inode *inode)
1539 if (inode != NULL) {
1540 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1542 } else if (ldiskfs_has_feature_extents(sb)) {
1548 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1549 const loff_t size, const loff_t pos,
1552 int credits, bits, bs, i;
1554 bits = sb->s_blocksize_bits;
1557 /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1558 * we do not expect blockmaps on the large files,
1559 * so let's shrink it to 2 levels (4GB files)
1562 /* this is default reservation: 2 levels */
1563 credits = (blocks + 2) * 3;
1565 /* actual offset is unknown, hard to optimize */
1569 /* now check for few specific cases to optimize */
1570 if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1573 /* allocate if not allocated */
1574 if (inode == NULL) {
1575 credits += blocks * 2;
1578 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1579 LASSERT(i < LDISKFS_NDIR_BLOCKS);
1580 if (LDISKFS_I(inode)->i_data[i] == 0)
1583 } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1584 /* single indirect */
1585 credits = blocks * 3;
1586 if (inode == NULL ||
1587 LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1590 /* The indirect block may be modified. */
1597 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1598 const struct lu_buf *buf, loff_t _pos,
1599 struct thandle *handle)
1601 struct osd_object *obj = osd_dt_obj(dt);
1602 struct inode *inode = obj->oo_inode;
1603 struct super_block *sb = osd_sb(osd_obj2dev(obj));
1604 struct osd_thandle *oh;
1605 int rc = 0, est = 0, credits, blocks, allocated = 0;
1611 LASSERT(buf != NULL);
1612 LASSERT(handle != NULL);
1614 oh = container_of(handle, struct osd_thandle, ot_super);
1615 LASSERT(oh->ot_handle == NULL);
1618 bits = sb->s_blocksize_bits;
1622 /* if this is an append, then we
1623 * should expect cross-block record
1630 /* blocks to modify */
1631 blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1632 LASSERT(blocks > 0);
1634 if (inode != NULL && _pos != -1) {
1635 /* object size in blocks */
1636 est = (i_size_read(inode) + bs - 1) >> bits;
1637 allocated = inode->i_blocks >> (bits - 9);
1638 if (pos + size <= i_size_read(inode) && est <= allocated) {
1639 /* looks like an overwrite, no need to modify tree */
1641 /* no need to modify i_size */
1646 if (osd_extents_enabled(sb, inode)) {
1648 * many concurrent threads may grow tree by the time
1649 * our transaction starts. so, consider 2 is a min depth
1650 * for every level we may need to allocate a new block
1651 * and take some entries from the old one. so, 3 blocks
1652 * to allocate (bitmap, gd, itself) + old block - 4 per
1655 depth = inode != NULL ? ext_depth(inode) : 0;
1656 depth = max(depth, 1) + 1;
1658 /* if not append, then split may need to modify
1659 * existing blocks moving entries into the new ones
1663 /* blocks to store data: bitmap,gd,itself */
1664 credits += blocks * 3;
1666 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1668 /* if inode is created as part of the transaction,
1669 * then it's counted already by the creation method
1676 osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1678 /* dt_declare_write() is usually called for system objects, such
1679 * as llog or last_rcvd files. We needn't enforce quota on those
1680 * objects, so always set the lqi_space as 0.
1683 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1685 i_projid_read(inode), 0,
1686 oh, obj, NULL, OSD_QID_BLK);
1689 rc = osd_trunc_lock(obj, oh, true);
1694 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1696 /* LU-2634: clear the extent format for fast symlink */
1697 ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1699 memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1700 spin_lock(&inode->i_lock);
1701 LDISKFS_I(inode)->i_disksize = buflen;
1702 i_size_write(inode, buflen);
1703 spin_unlock(&inode->i_lock);
1704 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1709 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1710 int bufsize, int write_NUL, loff_t *offs,
1713 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1714 struct buffer_head *bh = NULL;
1715 loff_t offset = *offs;
1716 loff_t new_size = i_size_read(inode);
1717 unsigned long block;
1718 int blocksize = 1 << inode->i_blkbits;
1719 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1723 int dirty_inode = 0;
1724 bool create, sparse, sync = false;
1728 * long symlink write does not count the NUL terminator in
1729 * bufsize, we write it, and the inode's file size does not
1730 * count the NUL terminator as well.
1732 ((char *)buf)[bufsize] = '\0';
1736 dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1739 /* sparse checking is racy, but sparse is very rare case, leave as is */
1740 sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1741 ((new_size - 1) >> inode->i_blkbits) + 1);
1743 while (bufsize > 0) {
1744 int credits = handle->h_buffer_credits;
1745 unsigned long last_block = (new_size == 0) ? 0 :
1746 (new_size - 1) >> inode->i_blkbits;
1751 block = offset >> inode->i_blkbits;
1752 boffs = offset & (blocksize - 1);
1753 size = min(blocksize - boffs, bufsize);
1754 sync = (block > last_block || new_size == 0 || sparse);
1757 down(&ei->i_append_sem);
1759 bh = __ldiskfs_bread(handle, inode, block, 0);
1761 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1763 "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1765 offset, block, bufsize, *offs);
1767 if (IS_ERR_OR_NULL(bh)) {
1768 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1769 int flags = LDISKFS_GET_BLOCKS_CREATE;
1771 /* while the file system is being mounted, avoid
1772 * preallocation otherwise mount can take a long
1773 * time as mballoc cache is cold.
1774 * XXX: this is a workaround until we have a proper
1776 * XXX: works with extent-based files only */
1777 if (!osd->od_cl_seq)
1778 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1779 bh = __ldiskfs_bread(handle, inode, block, flags);
1783 up(&ei->i_append_sem);
1788 if (IS_ERR_OR_NULL(bh)) {
1797 "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1798 inode->i_sb->s_id, offset, block, bufsize, *offs,
1799 credits, handle->h_buffer_credits, err);
1803 err = ldiskfs_journal_get_write_access(handle, bh);
1805 CERROR("journal_get_write_access() returned error %d\n",
1809 LASSERTF(boffs + size <= bh->b_size,
1810 "boffs %d size %d bh->b_size %lu\n",
1811 boffs, size, (unsigned long)bh->b_size);
1813 memset(bh->b_data, 0, bh->b_size);
1815 up(&ei->i_append_sem);
1819 memcpy(bh->b_data + boffs, buf, size);
1820 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1824 if (offset + size > new_size)
1825 new_size = offset + size;
1831 up(&ei->i_append_sem);
1838 /* correct in-core and on-disk sizes */
1839 if (new_size > i_size_read(inode)) {
1840 spin_lock(&inode->i_lock);
1841 if (new_size > i_size_read(inode))
1842 i_size_write(inode, new_size);
1843 if (i_size_read(inode) > ei->i_disksize) {
1844 ei->i_disksize = i_size_read(inode);
1847 spin_unlock(&inode->i_lock);
1850 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1857 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1858 const struct lu_buf *buf, loff_t *pos,
1859 struct thandle *handle)
1861 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1862 struct osd_thandle *oh;
1866 LASSERT(dt_object_exists(dt));
1868 LASSERT(handle != NULL);
1869 LASSERT(inode != NULL);
1870 dquot_initialize(inode);
1872 /* XXX: don't check: one declared chunk can be used many times */
1873 /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1875 oh = container_of(handle, struct osd_thandle, ot_super);
1876 LASSERT(oh->ot_handle->h_transaction != NULL);
1877 osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1879 /* Write small symlink to inode body as we need to maintain correct
1880 * on-disk symlinks for ldiskfs.
1881 * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1882 * does not count it in.
1884 is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1885 if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1886 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1888 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1889 is_link, pos, oh->ot_handle);
1891 result = buf->lb_len;
1893 osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1898 static int osd_declare_fallocate(const struct lu_env *env,
1899 struct dt_object *dt, __u64 start, __u64 end,
1900 int mode, struct thandle *th)
1902 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
1903 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1904 struct inode *inode = osd_dt_obj(dt)->oo_inode;
1905 long long quota_space = 0;
1906 /* 5 is max tree depth. (inode + 4 index blocks) */
1913 * Only mode == 0 (which is standard prealloc) is supported now.
1914 * Rest of mode options is not supported yet.
1916 if (mode & ~FALLOC_FL_KEEP_SIZE)
1917 RETURN(-EOPNOTSUPP);
1922 /* quota space for metadata blocks
1923 * approximate metadata estimate should be good enough.
1925 quota_space += PAGE_SIZE;
1926 quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1928 /* quota space should be reported in 1K blocks */
1929 quota_space = toqb(quota_space) + toqb(end - start) +
1930 LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
1932 /* We don't need to reserve credits for whole fallocate here.
1933 * We reserve space only for metadata. Fallocate credits are
1934 * extended as required
1936 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1937 i_projid_read(inode), quota_space, oh,
1938 osd_dt_obj(dt), NULL, OSD_QID_BLK);
1942 /* Borrow @ext4_chunk_trans_blocks */
1943 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
1945 ldiskfs_group_t groups;
1951 depth = ext_depth(inode);
1952 idxblocks = depth * 2;
1955 * Now let's see how many group bitmaps and group descriptors need
1958 groups = idxblocks + 1;
1960 if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
1961 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
1962 if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
1963 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
1965 /* bitmaps and block group descriptor blocks */
1966 ret = idxblocks + groups + gdpblocks;
1968 /* Blocks for super block, inode, quota and xattr blocks */
1969 ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
1974 static int osd_extend_restart_trans(handle_t *handle, int needed)
1978 if (ldiskfs_handle_has_enough_credits(handle, needed))
1981 rc = ldiskfs_journal_extend(handle, needed - handle->h_buffer_credits);
1985 rc = ldiskfs_journal_restart(handle, needed);
1990 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1991 __u64 start, __u64 end, int mode, struct thandle *th)
1993 struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
1994 handle_t *handle = ldiskfs_journal_current_handle();
1995 unsigned int save_credits = oh->ot_credits;
1996 struct osd_object *obj = osd_dt_obj(dt);
1997 struct inode *inode = obj->oo_inode;
1998 struct ldiskfs_map_blocks map;
1999 unsigned int credits;
2000 ldiskfs_lblk_t blen;
2001 ldiskfs_lblk_t boff;
2002 loff_t new_size = 0;
2009 LASSERT(dt_object_exists(dt));
2010 LASSERT(osd_invariant(obj));
2011 LASSERT(inode != NULL);
2013 CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2014 inode->i_ino, start, end, mode);
2016 dquot_initialize(inode);
2020 boff = start >> inode->i_blkbits;
2021 blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2023 flags = LDISKFS_GET_BLOCKS_CREATE;
2024 if (mode & FALLOC_FL_KEEP_SIZE)
2025 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2030 * We only support preallocation for extent-based file only.
2032 if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)))
2033 GOTO(out, rc = -EOPNOTSUPP);
2035 if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2036 end > LDISKFS_I(inode)->i_disksize)) {
2038 rc = inode_newsize_ok(inode, new_size);
2043 inode_dio_wait(inode);
2048 /* Don't normalize the request if it can fit in one extent so
2049 * that it doesn't get unnecessarily split into multiple extents.
2051 if (blen <= EXT_UNWRITTEN_MAX_LEN)
2052 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2055 * credits to insert 1 extent into extent tree.
2057 credits = osd_chunk_trans_blocks(inode, blen);
2058 depth = ext_depth(inode);
2060 while (rc >= 0 && blen) {
2064 * Recalculate credits when extent tree depth changes.
2066 if (depth != ext_depth(inode)) {
2067 credits = osd_chunk_trans_blocks(inode, blen);
2068 depth = ext_depth(inode);
2071 /* TODO: quota check */
2072 rc = osd_extend_restart_trans(handle, credits);
2076 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2079 "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2080 inode->i_ino, map.m_lblk, map.m_len, rc);
2081 ldiskfs_mark_inode_dirty(handle, inode);
2086 map.m_len = blen = blen - rc;
2087 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2088 inode->i_ctime = current_time(inode);
2092 if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2093 inode->i_mtime = inode->i_ctime;
2095 if (epos > inode->i_size)
2096 ldiskfs_set_inode_flag(inode,
2097 LDISKFS_INODE_EOFBLOCKS);
2100 ldiskfs_mark_inode_dirty(handle, inode);
2104 inode_unlock(inode);
2106 /* extand credits if needed for operations such as attribute set */
2108 rc = osd_extend_restart_trans(handle, save_credits);
2113 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2114 __u64 start, __u64 end, struct thandle *th)
2116 struct osd_thandle *oh;
2117 struct inode *inode;
2122 oh = container_of(th, struct osd_thandle, ot_super);
2125 * we don't need to reserve credits for whole truncate
2126 * it's not possible as truncate may need to free too many
2127 * blocks and that won't fit a single transaction. instead
2128 * we reserve credits to change i_size and put inode onto
2129 * orphan list. if needed truncate will extend or restart
2132 osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2133 osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2135 inode = osd_dt_obj(dt)->oo_inode;
2138 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2139 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2143 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2148 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2149 __u64 start, __u64 end, struct thandle *th)
2151 struct osd_object *obj = osd_dt_obj(dt);
2152 struct osd_device *osd = osd_obj2dev(obj);
2153 struct inode *inode = obj->oo_inode;
2154 struct osd_access_lock *al;
2155 struct osd_thandle *oh;
2156 int rc = 0, found = 0;
2160 LASSERT(dt_object_exists(dt));
2161 LASSERT(osd_invariant(obj));
2162 LASSERT(inode != NULL);
2163 dquot_initialize(inode);
2166 oh = container_of(th, struct osd_thandle, ot_super);
2167 LASSERT(oh->ot_handle->h_transaction != NULL);
2169 /* we used to skip truncate to current size to
2170 * optimize truncates on OST. with DoM we can
2171 * get attr_set to set specific size (MDS_REINT)
2172 * and then get truncate RPC which essentially
2173 * would be skipped. this is bad.. so, disable
2174 * this optimization on MDS till the client stop
2175 * to sent MDS_REINT (LU-11033) -bzzz
2177 if (osd->od_is_ost && i_size_read(inode) == start)
2180 osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2182 spin_lock(&inode->i_lock);
2183 if (i_size_read(inode) < start)
2185 i_size_write(inode, start);
2186 spin_unlock(&inode->i_lock);
2187 /* if object holds encrypted content, we need to make sure we truncate
2188 * on an encryption unit boundary, or subsequent reads will get
2191 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2192 start & ~LUSTRE_ENCRYPTION_MASK)
2193 start = (start & LUSTRE_ENCRYPTION_MASK) +
2194 LUSTRE_ENCRYPTION_UNIT_SIZE;
2195 ll_truncate_pagecache(inode, start);
2197 /* optimize grow case */
2199 osd_execute_truncate(obj);
2204 /* add to orphan list to ensure truncate completion
2205 * if this transaction succeed. ldiskfs_truncate()
2206 * will take the inode out of the list
2208 rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2209 inode_unlock(inode);
2213 list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2214 if (obj != al->tl_obj)
2216 LASSERT(al->tl_shared == 0);
2218 /* do actual truncate in osd_trans_stop() */
2219 al->tl_truncate = 1;
2228 static int fiemap_check_ranges(struct inode *inode,
2229 u64 start, u64 len, u64 *new_len)
2238 if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2239 maxbytes = inode->i_sb->s_maxbytes;
2241 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2243 if (start > maxbytes)
2247 * Shrink request scope to what the fs can actually handle.
2249 if (len > maxbytes || (maxbytes - len) < start)
2250 *new_len = maxbytes - start;
2255 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2256 #define FIEMAP_MAX_EXTENTS (UINT_MAX / sizeof(struct fiemap_extent))
2258 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2261 struct fiemap_extent_info fieinfo = {0, };
2262 struct inode *inode = osd_dt_obj(dt)->oo_inode;
2265 mm_segment_t cur_fs;
2268 if (inode->i_op->fiemap == NULL)
2271 if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2274 rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2278 fieinfo.fi_flags = fm->fm_flags;
2279 fieinfo.fi_extents_max = fm->fm_extent_count;
2280 fieinfo.fi_extents_start = fm->fm_extents;
2282 if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2283 filemap_write_and_wait(inode->i_mapping);
2285 /* Save previous value address limit */
2287 /* Set the address limit of the kernel */
2290 rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2291 fm->fm_flags = fieinfo.fi_flags;
2292 fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2294 /* Restore the previous address limt */
2300 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2301 __u64 start, __u64 end, enum lu_ladvise_type advice)
2303 struct osd_object *obj = osd_dt_obj(dt);
2308 case LU_LADVISE_DONTNEED:
2310 invalidate_mapping_pages(obj->oo_inode->i_mapping,
2311 start >> PAGE_SHIFT,
2312 (end - 1) >> PAGE_SHIFT);
2322 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2323 loff_t offset, int whence)
2325 struct osd_object *obj = osd_dt_obj(dt);
2326 struct inode *inode = obj->oo_inode;
2332 LASSERT(dt_object_exists(dt));
2333 LASSERT(osd_invariant(obj));
2335 LASSERT(offset >= 0);
2337 file = osd_quasi_file(env, inode);
2338 result = file->f_op->llseek(file, offset, whence);
2341 * If 'offset' is beyond end of object file then treat it as not error
2342 * but valid case for SEEK_HOLE and return 'offset' as result.
2343 * LOV will decide if it is beyond real end of file or not.
2345 if (whence == SEEK_HOLE && result == -ENXIO)
2348 CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2349 "hole" : "data", offset, result);
2354 * in some cases we may need declare methods for objects being created
2355 * e.g., when we create symlink
2357 const struct dt_body_operations osd_body_ops_new = {
2358 .dbo_declare_write = osd_declare_write,
2361 const struct dt_body_operations osd_body_ops = {
2362 .dbo_read = osd_read,
2363 .dbo_declare_write = osd_declare_write,
2364 .dbo_write = osd_write,
2365 .dbo_bufs_get = osd_bufs_get,
2366 .dbo_bufs_put = osd_bufs_put,
2367 .dbo_write_prep = osd_write_prep,
2368 .dbo_declare_write_commit = osd_declare_write_commit,
2369 .dbo_write_commit = osd_write_commit,
2370 .dbo_read_prep = osd_read_prep,
2371 .dbo_declare_punch = osd_declare_punch,
2372 .dbo_punch = osd_punch,
2373 .dbo_fiemap_get = osd_fiemap_get,
2374 .dbo_ladvise = osd_ladvise,
2375 .dbo_declare_fallocate = osd_declare_fallocate,
2376 .dbo_fallocate = osd_fallocate,
2377 .dbo_lseek = osd_lseek,
2381 * Get a truncate lock
2383 * In order to take multi-transaction truncate out of main transaction we let
2384 * the caller grab a lock on the object passed. the lock can be shared (for
2385 * writes) and exclusive (for truncate). It's not allowed to mix truncate
2386 * and write in the same transaction handle (do not confuse with big ldiskfs
2387 * transaction containing lots of handles).
2388 * The lock must be taken at declaration.
2390 * \param obj object to lock
2392 * \shared shared or exclusive
2394 * \retval 0 lock is granted
2395 * \retval -NOMEM no memory to allocate lock
2397 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2399 struct osd_access_lock *al, *tmp;
2404 list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2405 if (tmp->tl_obj != obj)
2407 LASSERT(tmp->tl_shared == shared);
2408 /* found same lock */
2413 if (unlikely(al == NULL))
2416 al->tl_truncate = false;
2418 down_read(&obj->oo_ext_idx_sem);
2420 down_write(&obj->oo_ext_idx_sem);
2421 al->tl_shared = shared;
2422 lu_object_get(&obj->oo_dt.do_lu);
2424 list_add(&al->tl_list, &oh->ot_trunc_locks);
2429 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2431 struct osd_access_lock *al, *tmp;
2433 list_for_each_entry_safe(al, tmp, list, tl_list) {
2435 up_read(&al->tl_obj->oo_ext_idx_sem);
2437 up_write(&al->tl_obj->oo_ext_idx_sem);
2438 osd_object_put(env, al->tl_obj);
2439 list_del(&al->tl_list);
2444 void osd_execute_truncate(struct osd_object *obj)
2446 struct osd_device *d = osd_obj2dev(obj);
2447 struct inode *inode = obj->oo_inode;
2450 /* simulate crash before (in the middle) of delayed truncate */
2451 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2452 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2453 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2455 mutex_lock(&sbi->s_orphan_lock);
2456 list_del_init(&ei->i_orphan);
2457 mutex_unlock(&sbi->s_orphan_lock);
2461 size = i_size_read(inode);
2463 /* if object holds encrypted content, we need to make sure we truncate
2464 * on an encryption unit boundary, or block content will get corrupted
2466 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2467 size & ~LUSTRE_ENCRYPTION_MASK)
2468 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2469 LUSTRE_ENCRYPTION_UNIT_SIZE;
2470 ldiskfs_truncate(inode);
2471 inode_unlock(inode);
2472 if (inode->i_size != size) {
2473 spin_lock(&inode->i_lock);
2474 i_size_write(inode, size);
2475 LDISKFS_I(inode)->i_disksize = size;
2476 spin_unlock(&inode->i_lock);
2477 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2481 * For a partial-page truncate, flush the page to disk immediately to
2482 * avoid data corruption during direct disk write. b=17397
2484 if ((size & ~PAGE_MASK) == 0)
2486 if (osd_use_page_cache(d)) {
2487 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2489 /* Notice we use "wait" version to ensure I/O is complete */
2490 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2491 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2492 size >> PAGE_SHIFT);
2496 void osd_process_truncates(struct list_head *list)
2498 struct osd_access_lock *al;
2500 LASSERT(journal_current_handle() == NULL);
2502 list_for_each_entry(al, list, tl_list) {
2505 if (!al->tl_truncate)
2507 osd_execute_truncate(al->tl_obj);