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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osd-zfs/osd_io.c
38 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
39 * Author: Mike Pershin <tappro@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_OSD
44 #include <lustre_ver.h>
45 #include <libcfs/libcfs.h>
46 #include <obd_support.h>
47 #include <lustre_net.h>
49 #include <obd_class.h>
50 #include <lustre_disk.h>
51 #include <lustre_fid.h>
52 #include <lustre/lustre_idl.h> /* LLOG_MIN_CHUNK_SIZE definition */
54 #include "osd_internal.h"
56 #include <sys/dnode.h>
61 #include <sys/spa_impl.h>
62 #include <sys/zfs_znode.h>
63 #include <sys/dmu_tx.h>
64 #include <sys/dmu_objset.h>
65 #include <sys/dsl_prop.h>
66 #include <sys/sa_impl.h>
69 static char *osd_zerocopy_tag = "zerocopy";
72 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
74 struct obd_histogram *h = osd->od_brw_stats.hist;
77 atomic_inc(&osd->od_r_in_flight);
78 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
79 atomic_read(&osd->od_r_in_flight));
80 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
83 atomic_inc(&osd->od_w_in_flight);
84 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
85 atomic_read(&osd->od_w_in_flight));
86 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
91 static void record_end_io(struct osd_device *osd, int rw,
92 unsigned long elapsed, int disksize, int npages)
94 struct obd_histogram *h = osd->od_brw_stats.hist;
97 atomic_dec(&osd->od_r_in_flight);
98 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
100 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
102 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
105 atomic_dec(&osd->od_w_in_flight);
106 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
108 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
110 lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
114 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
115 struct lu_buf *buf, loff_t *pos)
117 struct osd_object *obj = osd_dt_obj(dt);
118 struct osd_device *osd = osd_obj2dev(obj);
120 int size = buf->lb_len;
124 LASSERT(dt_object_exists(dt));
127 start = cfs_time_current();
129 read_lock(&obj->oo_attr_lock);
130 old_size = obj->oo_attr.la_size;
131 read_unlock(&obj->oo_attr_lock);
133 if (*pos + size > old_size) {
137 size = old_size - *pos;
140 record_start_io(osd, READ, 0);
142 rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
143 buf->lb_buf, DMU_READ_PREFETCH);
145 record_end_io(osd, READ, cfs_time_current() - start, size,
154 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
155 const struct lu_buf *buf, loff_t pos,
158 struct osd_object *obj = osd_dt_obj(dt);
159 struct osd_device *osd = osd_obj2dev(obj);
160 struct osd_thandle *oh;
164 oh = container_of0(th, struct osd_thandle, ot_super);
166 /* in some cases declare can race with creation (e.g. llog)
167 * and we need to wait till object is initialized. notice
168 * LOHA_EXISTs is supposed to be the last step in the
171 /* size change (in dnode) will be declared by dmu_tx_hold_write() */
172 if (dt_object_exists(dt))
173 oid = obj->oo_db->db_object;
175 oid = DMU_NEW_OBJECT;
177 /* XXX: we still miss for append declaration support in ZFS
178 * -1 means append which is used by llog mostly, llog
179 * can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
181 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
182 obj->oo_attr.la_size + (2 << 20));
183 dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
185 /* dt_declare_write() is usually called for system objects, such
186 * as llog or last_rcvd files. We needn't enforce quota on those
187 * objects, so always set the lqi_space as 0. */
188 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
189 obj->oo_attr.la_gid, 0, oh, true, NULL,
193 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
194 const struct lu_buf *buf, loff_t *pos,
195 struct thandle *th, int ignore_quota)
197 struct osd_object *obj = osd_dt_obj(dt);
198 struct osd_device *osd = osd_obj2dev(obj);
199 struct osd_thandle *oh;
200 uint64_t offset = *pos;
205 LASSERT(dt_object_exists(dt));
209 oh = container_of0(th, struct osd_thandle, ot_super);
211 record_start_io(osd, WRITE, 0);
213 dmu_write(osd->od_os, obj->oo_db->db_object, offset,
214 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
215 write_lock(&obj->oo_attr_lock);
216 if (obj->oo_attr.la_size < offset + buf->lb_len) {
217 obj->oo_attr.la_size = offset + buf->lb_len;
218 write_unlock(&obj->oo_attr_lock);
219 /* osd_object_sa_update() will be copying directly from oo_attr
220 * into dbuf. any update within a single txg will copy the
222 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
223 &obj->oo_attr.la_size, 8, oh);
227 write_unlock(&obj->oo_attr_lock);
234 record_end_io(osd, WRITE, 0, buf->lb_len,
235 buf->lb_len >> PAGE_SHIFT);
241 * XXX: for the moment I don't want to use lnb_flags for osd-internal
242 * purposes as it's not very well defined ...
243 * instead I use the lowest bit of the address so that:
244 * arc buffer: .lnb_data = abuf (arc we loan for write)
245 * dbuf buffer: .lnb_data = dbuf | 1 (dbuf we get for read)
246 * copy buffer: .lnb_page->mapping = obj (page we allocate for write)
250 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
251 struct niobuf_local *lnb, int npages)
253 struct osd_object *obj = osd_dt_obj(dt);
254 struct osd_device *osd = osd_obj2dev(obj);
258 LASSERT(dt_object_exists(dt));
261 for (i = 0; i < npages; i++) {
262 if (lnb[i].lnb_page == NULL)
264 if (lnb[i].lnb_page->mapping == (void *)obj) {
265 /* this is anonymous page allocated for copy-write */
266 lnb[i].lnb_page->mapping = NULL;
267 __free_page(lnb[i].lnb_page);
268 atomic_dec(&osd->od_zerocopy_alloc);
270 /* see comment in osd_bufs_get_read() */
271 ptr = (unsigned long)lnb[i].lnb_data;
274 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
275 atomic_dec(&osd->od_zerocopy_pin);
276 } else if (lnb[i].lnb_data != NULL) {
277 dmu_return_arcbuf(lnb[i].lnb_data);
278 atomic_dec(&osd->od_zerocopy_loan);
281 lnb[i].lnb_page = NULL;
282 lnb[i].lnb_data = NULL;
288 static inline struct page *kmem_to_page(void *addr)
290 if (is_vmalloc_addr(addr))
291 return vmalloc_to_page(addr);
293 return virt_to_page(addr);
297 * Prepare buffers for read.
299 * The function maps the range described by \a off and \a len to \a lnb array.
300 * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
301 * we fill \a lnb array with the pages storing ARC buffers. Notice the current
302 * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
303 * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
304 * A better implementation would just return the buffers (potentially unfilled)
305 * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
307 * \param[in] env environment
308 * \param[in] obj object
309 * \param[in] off offset in bytes
310 * \param[in] len the number of bytes to access
311 * \param[out] lnb array of local niobufs pointing to the buffers with data
313 * \retval 0 for success
314 * \retval negative error number of failure
316 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
317 loff_t off, ssize_t len, struct niobuf_local *lnb)
319 struct osd_device *osd = osd_obj2dev(obj);
320 unsigned long start = cfs_time_current();
321 int rc, i, numbufs, npages = 0;
325 record_start_io(osd, READ, 0);
327 /* grab buffers for read:
328 * OSD API let us to grab buffers first, then initiate IO(s)
329 * so that all required IOs will be done in parallel, but at the
330 * moment DMU doesn't provide us with a method to grab buffers.
331 * If we discover this is a vital for good performance we
332 * can get own replacement for dmu_buf_hold_array_by_bonus().
335 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
336 osd_zerocopy_tag, &numbufs,
341 for (i = 0; i < numbufs; i++) {
342 int bufoff, tocpy, thispage;
347 atomic_inc(&osd->od_zerocopy_pin);
349 bufoff = off - dbp[i]->db_offset;
350 tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
352 /* kind of trick to differentiate dbuf vs. arcbuf */
353 LASSERT(((unsigned long)dbp[i] & 1) == 0);
354 dbf = (void *) ((unsigned long)dbp[i] | 1);
357 thispage = PAGE_SIZE;
358 thispage -= bufoff & (PAGE_SIZE - 1);
359 thispage = min(tocpy, thispage);
362 lnb->lnb_file_offset = off;
363 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
364 lnb->lnb_len = thispage;
365 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
367 /* mark just a single slot: we need this
368 * reference to dbuf to be released once */
381 /* steal dbuf so dmu_buf_rele_array() can't release
386 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
389 record_end_io(osd, READ, cfs_time_current() - start,
390 npages * PAGE_SIZE, npages);
396 osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
400 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
401 loff_t off, ssize_t len, struct niobuf_local *lnb)
403 struct osd_device *osd = osd_obj2dev(obj);
404 int plen, off_in_block, sz_in_block;
405 int rc, i = 0, npages = 0;
411 dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
414 * currently only full blocks are subject to zerocopy approach:
415 * so that we're sure nobody is trying to update the same block
418 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
420 off_in_block = off & (bs - 1);
421 sz_in_block = min_t(int, bs - off_in_block, len);
423 if (sz_in_block == bs) {
424 /* full block, try to use zerocopy */
426 abuf = dmu_request_arcbuf(obj->oo_db, bs);
427 if (unlikely(abuf == NULL))
428 GOTO(out_err, rc = -ENOMEM);
430 atomic_inc(&osd->od_zerocopy_loan);
432 /* go over pages arcbuf contains, put them as
433 * local niobufs for ptlrpc's bulks */
434 while (sz_in_block > 0) {
435 plen = min_t(int, sz_in_block, PAGE_SIZE);
437 lnb[i].lnb_file_offset = off;
438 lnb[i].lnb_page_offset = 0;
439 lnb[i].lnb_len = plen;
441 if (sz_in_block == bs)
442 lnb[i].lnb_data = abuf;
444 lnb[i].lnb_data = NULL;
446 /* this one is not supposed to fail */
447 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
449 LASSERT(lnb[i].lnb_page);
451 lprocfs_counter_add(osd->od_stats,
452 LPROC_OSD_ZEROCOPY_IO, 1);
457 off_in_block += plen;
462 if (off_in_block == 0 && len < bs &&
463 off + len >= obj->oo_attr.la_size)
464 lprocfs_counter_add(osd->od_stats,
465 LPROC_OSD_TAIL_IO, 1);
467 /* can't use zerocopy, allocate temp. buffers */
468 while (sz_in_block > 0) {
469 plen = min_t(int, sz_in_block, PAGE_SIZE);
471 lnb[i].lnb_file_offset = off;
472 lnb[i].lnb_page_offset = 0;
473 lnb[i].lnb_len = plen;
475 lnb[i].lnb_data = NULL;
477 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
478 if (unlikely(lnb[i].lnb_page == NULL))
479 GOTO(out_err, rc = -ENOMEM);
481 LASSERT(lnb[i].lnb_page->mapping == NULL);
482 lnb[i].lnb_page->mapping = (void *)obj;
484 atomic_inc(&osd->od_zerocopy_alloc);
485 lprocfs_counter_add(osd->od_stats,
486 LPROC_OSD_COPY_IO, 1);
500 osd_bufs_put(env, &obj->oo_dt, lnb, npages);
504 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
505 loff_t offset, ssize_t len, struct niobuf_local *lnb,
508 struct osd_object *obj = osd_dt_obj(dt);
511 LASSERT(dt_object_exists(dt));
515 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
517 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
522 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
523 struct niobuf_local *lnb, int npages)
525 struct osd_object *obj = osd_dt_obj(dt);
527 LASSERT(dt_object_exists(dt));
533 static inline uint32_t osd_get_blocksz(struct osd_object *obj)
540 dmu_object_size_from_db(obj->oo_db, &blksz, &unused);
544 static inline uint64_t osd_roundup2blocksz(uint64_t size,
550 size += offset % blksz;
552 if (likely(IS_PO2(blksz)))
553 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
560 static int osd_declare_write_commit(const struct lu_env *env,
561 struct dt_object *dt,
562 struct niobuf_local *lnb, int npages,
565 struct osd_object *obj = osd_dt_obj(dt);
566 struct osd_device *osd = osd_obj2dev(obj);
567 struct osd_thandle *oh;
570 uint32_t blksz = osd_get_blocksz(obj);
571 int i, rc, flags = 0;
572 bool ignore_quota = false, synced = false;
574 struct page *last_page = NULL;
575 unsigned long discont_pages = 0;
578 LASSERT(dt_object_exists(dt));
584 oh = container_of0(th, struct osd_thandle, ot_super);
586 for (i = 0; i < npages; i++) {
587 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
589 last_page = lnb[i].lnb_page;
591 /* ENOSPC, network RPC error, etc.
592 * We don't want to book space for pages which will be
593 * skipped in osd_write_commit(). Hence we skip pages
594 * with lnb_rc != 0 here too */
596 /* ignore quota for the whole request if any page is from
597 * client cache or written by root.
599 * XXX once we drop the 1.8 client support, the checking
600 * for whether page is from cache can be simplified as:
601 * !(lnb[i].flags & OBD_BRW_SYNC)
603 * XXX we could handle this on per-lnb basis as done by
605 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
606 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
610 /* first valid lnb */
611 offset = lnb[i].lnb_file_offset;
612 size = lnb[i].lnb_len;
615 if (offset + size == lnb[i].lnb_file_offset) {
616 /* this lnb is contiguous to the previous one */
617 size += lnb[i].lnb_len;
621 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
623 /* Estimating space to be consumed by a write is rather
624 * complicated with ZFS. As a consequence, we don't account for
625 * indirect blocks and just use as a rough estimate the worse
626 * case where the old space is being held by a snapshot. Quota
627 * overrun will be adjusted once the operation is committed, if
629 space += osd_roundup2blocksz(size, offset, blksz);
631 offset = lnb[i].lnb_file_offset;
632 size = lnb[i].lnb_len;
636 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
638 space += osd_roundup2blocksz(size, offset, blksz);
641 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
643 oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
645 /* backend zfs filesystem might be configured to store multiple data
647 space *= osd->od_os->os_copies;
649 CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
652 record_start_io(osd, WRITE, discont_pages);
654 /* acquire quota space if needed */
655 rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
656 obj->oo_attr.la_gid, space, oh, true, &flags,
659 if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
660 dt_sync(env, th->th_dev);
662 CDEBUG(D_QUOTA, "retry after sync\n");
667 /* we need only to store the overquota flags in the first lnb for
668 * now, once we support multiple objects BRW, this code needs be
670 if (flags & QUOTA_FL_OVER_USRQUOTA)
671 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
672 if (flags & QUOTA_FL_OVER_GRPQUOTA)
673 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
679 * Policy to grow ZFS block size by write pattern.
680 * For sequential write, it grows block size gradually until it reaches the
681 * maximum blocksize the dataset can support. Otherwise, it will pick a
682 * a block size by the writing region of this I/O.
684 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
685 uint64_t start, uint64_t end)
687 struct osd_device *osd = osd_obj2dev(obj);
688 dmu_buf_impl_t *db = (dmu_buf_impl_t *)obj->oo_db;
698 if (dn->dn_maxblkid > 0) /* can't change block size */
701 if (dn->dn_datablksz >= osd->od_max_blksz)
704 down_write(&obj->oo_guard);
706 blksz = dn->dn_datablksz;
707 if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
708 GOTO(out_unlock, rc);
710 /* now ZFS can support up to 16MB block size, and if the write
711 * is sequential, it just increases the block size gradually */
712 if (start <= blksz) { /* sequential */
713 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
714 } else { /* sparse, pick a block size by write region */
715 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
719 if (!is_power_of_2(blksz))
720 blksz = size_roundup_power2(blksz);
722 if (blksz > dn->dn_datablksz) {
723 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
724 blksz, 0, oh->ot_tx);
725 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
727 CDEBUG(D_INODE, "object "DFID": change block size"
728 "%u -> %u error rc = %d\n",
729 PFID(lu_object_fid(&obj->oo_dt.do_lu)),
730 dn->dn_datablksz, blksz, rc);
734 up_write(&obj->oo_guard);
740 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
741 struct niobuf_local *lnb, int npages,
744 struct osd_object *obj = osd_dt_obj(dt);
745 struct osd_device *osd = osd_obj2dev(obj);
746 struct osd_thandle *oh;
747 uint64_t new_size = 0;
749 unsigned long iosize = 0;
752 LASSERT(dt_object_exists(dt));
756 oh = container_of0(th, struct osd_thandle, ot_super);
758 /* adjust block size. Assume the buffers are sorted. */
759 (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
760 lnb[npages - 1].lnb_file_offset +
761 lnb[npages - 1].lnb_len);
762 for (i = 0; i < npages; i++) {
763 CDEBUG(D_INODE, "write %u bytes at %u\n",
764 (unsigned) lnb[i].lnb_len,
765 (unsigned) lnb[i].lnb_file_offset);
768 /* ENOSPC, network RPC error, etc.
769 * Unlike ldiskfs, zfs allocates new blocks on rewrite,
770 * so we skip this page if lnb_rc is set to -ENOSPC */
771 CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
772 PFID(lu_object_fid(&dt->do_lu)), i,
777 if (lnb[i].lnb_page->mapping == (void *)obj) {
778 dmu_write(osd->od_os, obj->oo_db->db_object,
779 lnb[i].lnb_file_offset, lnb[i].lnb_len,
780 kmap(lnb[i].lnb_page), oh->ot_tx);
781 kunmap(lnb[i].lnb_page);
782 } else if (lnb[i].lnb_data) {
783 LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
784 /* buffer loaned for zerocopy, try to use it.
785 * notice that dmu_assign_arcbuf() is smart
786 * enough to recognize changed blocksize
787 * in this case it fallbacks to dmu_write() */
788 dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
789 lnb[i].lnb_data, oh->ot_tx);
790 /* drop the reference, otherwise osd_put_bufs()
791 * will be releasing it - bad! */
792 lnb[i].lnb_data = NULL;
793 atomic_dec(&osd->od_zerocopy_loan);
796 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
797 new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
798 iosize += lnb[i].lnb_len;
801 if (unlikely(new_size == 0)) {
802 /* no pages to write, no transno is needed */
804 /* it is important to return 0 even when all lnb_rc == -ENOSPC
805 * since ofd_commitrw_write() retries several times on ENOSPC */
806 record_end_io(osd, WRITE, 0, 0, 0);
810 write_lock(&obj->oo_attr_lock);
811 if (obj->oo_attr.la_size < new_size) {
812 obj->oo_attr.la_size = new_size;
813 write_unlock(&obj->oo_attr_lock);
814 /* osd_object_sa_update() will be copying directly from
815 * oo_attr into dbuf. any update within a single txg will copy
817 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
818 &obj->oo_attr.la_size, 8, oh);
820 write_unlock(&obj->oo_attr_lock);
823 record_end_io(osd, WRITE, 0, iosize, npages);
828 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
829 struct niobuf_local *lnb, int npages)
831 struct osd_object *obj = osd_dt_obj(dt);
835 LASSERT(dt_object_exists(dt));
838 read_lock(&obj->oo_attr_lock);
839 eof = obj->oo_attr.la_size;
840 read_unlock(&obj->oo_attr_lock);
842 for (i = 0; i < npages; i++) {
843 if (unlikely(lnb[i].lnb_rc < 0))
846 lnb[i].lnb_rc = lnb[i].lnb_len;
848 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
849 if (eof <= lnb[i].lnb_file_offset)
852 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
854 /* all subsequent rc should be 0 */
865 * Punch/truncate an object
867 * IN: db - dmu_buf of the object to free data in.
868 * off - start of section to free.
869 * len - length of section to free (DMU_OBJECT_END => to EOF).
871 * RETURN: 0 if success
872 * error code if failure
874 * The transaction passed to this routine must have
875 * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
876 * called and then assigned to a transaction group.
878 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
879 uint64_t size, uint64_t off, uint64_t len)
883 /* Assert that the transaction has been assigned to a
884 transaction group. */
885 LASSERT(tx->tx_txg != 0);
887 * Nothing to do if file already at desired length.
889 if (len == DMU_OBJECT_END && size == off)
893 rc = -dmu_free_range(os, db->db_object, off, len, tx);
898 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
899 __u64 start, __u64 end, struct thandle *th)
901 struct osd_object *obj = osd_dt_obj(dt);
902 struct osd_device *osd = osd_obj2dev(obj);
903 struct osd_thandle *oh;
908 LASSERT(dt_object_exists(dt));
909 LASSERT(osd_invariant(obj));
912 oh = container_of0(th, struct osd_thandle, ot_super);
914 write_lock(&obj->oo_attr_lock);
916 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
917 len = DMU_OBJECT_END;
920 write_unlock(&obj->oo_attr_lock);
922 rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
923 obj->oo_attr.la_size, start, len);
925 if (len == DMU_OBJECT_END) {
926 write_lock(&obj->oo_attr_lock);
927 obj->oo_attr.la_size = start;
928 write_unlock(&obj->oo_attr_lock);
929 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
930 &obj->oo_attr.la_size, 8, oh);
935 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
936 __u64 start, __u64 end, struct thandle *handle)
938 struct osd_object *obj = osd_dt_obj(dt);
939 struct osd_device *osd = osd_obj2dev(obj);
940 struct osd_thandle *oh;
944 oh = container_of0(handle, struct osd_thandle, ot_super);
946 read_lock(&obj->oo_attr_lock);
947 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
948 len = DMU_OBJECT_END;
952 /* declare we'll free some blocks ... */
953 if (start < obj->oo_attr.la_size) {
954 read_unlock(&obj->oo_attr_lock);
955 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
957 read_unlock(&obj->oo_attr_lock);
960 /* ... and we'll modify size attribute */
961 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
963 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
964 obj->oo_attr.la_gid, 0, oh, true, NULL,
968 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
969 __u64 start, __u64 end, enum lu_ladvise_type advice)
983 struct dt_body_operations osd_body_ops = {
984 .dbo_read = osd_read,
985 .dbo_declare_write = osd_declare_write,
986 .dbo_write = osd_write,
987 .dbo_bufs_get = osd_bufs_get,
988 .dbo_bufs_put = osd_bufs_put,
989 .dbo_write_prep = osd_write_prep,
990 .dbo_declare_write_commit = osd_declare_write_commit,
991 .dbo_write_commit = osd_write_commit,
992 .dbo_read_prep = osd_read_prep,
993 .dbo_declare_punch = osd_declare_punch,
994 .dbo_punch = osd_punch,
995 .dbo_ladvise = osd_ladvise,