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, 2014, 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_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,
116 struct lustre_capa *capa)
118 struct osd_object *obj = osd_dt_obj(dt);
119 struct osd_device *osd = osd_obj2dev(obj);
121 int size = buf->lb_len;
125 LASSERT(dt_object_exists(dt));
128 start = cfs_time_current();
130 read_lock(&obj->oo_attr_lock);
131 old_size = obj->oo_attr.la_size;
132 read_unlock(&obj->oo_attr_lock);
134 if (*pos + size > old_size) {
138 size = old_size - *pos;
141 record_start_io(osd, READ, 0);
143 rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
144 buf->lb_buf, DMU_READ_PREFETCH);
146 record_end_io(osd, READ, cfs_time_current() - start, size,
147 size >> PAGE_CACHE_SHIFT);
155 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
156 const struct lu_buf *buf, loff_t pos,
159 struct osd_object *obj = osd_dt_obj(dt);
160 struct osd_device *osd = osd_obj2dev(obj);
161 struct osd_thandle *oh;
165 oh = container_of0(th, struct osd_thandle, ot_super);
167 /* in some cases declare can race with creation (e.g. llog)
168 * and we need to wait till object is initialized. notice
169 * LOHA_EXISTs is supposed to be the last step in the
172 /* declare possible size change. notice we can't check
173 * current size here as another thread can change it */
175 if (dt_object_exists(dt)) {
177 oid = obj->oo_db->db_object;
179 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
181 oid = DMU_NEW_OBJECT;
182 dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
185 /* XXX: we still miss for append declaration support in ZFS
186 * -1 means append which is used by llog mostly, llog
187 * can grow upto LLOG_CHUNK_SIZE*8 records */
189 pos = max_t(loff_t, 256 * 8 * LLOG_CHUNK_SIZE,
190 obj->oo_attr.la_size + (2 << 20));
191 dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
193 /* dt_declare_write() is usually called for system objects, such
194 * as llog or last_rcvd files. We needn't enforce quota on those
195 * objects, so always set the lqi_space as 0. */
196 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
197 obj->oo_attr.la_gid, 0, oh, true, NULL,
201 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
202 const struct lu_buf *buf, loff_t *pos,
203 struct thandle *th, struct lustre_capa *capa,
206 struct osd_object *obj = osd_dt_obj(dt);
207 struct osd_device *osd = osd_obj2dev(obj);
208 struct osd_thandle *oh;
209 uint64_t offset = *pos;
214 LASSERT(dt_object_exists(dt));
218 oh = container_of0(th, struct osd_thandle, ot_super);
220 record_start_io(osd, WRITE, 0);
222 dmu_write(osd->od_os, obj->oo_db->db_object, offset,
223 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
224 write_lock(&obj->oo_attr_lock);
225 if (obj->oo_attr.la_size < offset + buf->lb_len) {
226 obj->oo_attr.la_size = offset + buf->lb_len;
227 write_unlock(&obj->oo_attr_lock);
228 /* osd_object_sa_update() will be copying directly from oo_attr
229 * into dbuf. any update within a single txg will copy the
231 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
232 &obj->oo_attr.la_size, 8, oh);
236 write_unlock(&obj->oo_attr_lock);
243 record_end_io(osd, WRITE, 0, buf->lb_len,
244 buf->lb_len >> PAGE_CACHE_SHIFT);
250 * XXX: for the moment I don't want to use lnb_flags for osd-internal
251 * purposes as it's not very well defined ...
252 * instead I use the lowest bit of the address so that:
253 * arc buffer: .lnb_data = abuf (arc we loan for write)
254 * dbuf buffer: .lnb_data = dbuf | 1 (dbuf we get for read)
255 * copy buffer: .lnb_page->mapping = obj (page we allocate for write)
259 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
260 struct niobuf_local *lnb, int npages)
262 struct osd_object *obj = osd_dt_obj(dt);
263 struct osd_device *osd = osd_obj2dev(obj);
267 LASSERT(dt_object_exists(dt));
270 for (i = 0; i < npages; i++) {
271 if (lnb[i].lnb_page == NULL)
273 if (lnb[i].lnb_page->mapping == (void *)obj) {
274 /* this is anonymous page allocated for copy-write */
275 lnb[i].lnb_page->mapping = NULL;
276 __free_page(lnb[i].lnb_page);
277 atomic_dec(&osd->od_zerocopy_alloc);
279 /* see comment in osd_bufs_get_read() */
280 ptr = (unsigned long)lnb[i].lnb_data;
283 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
284 atomic_dec(&osd->od_zerocopy_pin);
285 } else if (lnb[i].lnb_data != NULL) {
286 dmu_return_arcbuf(lnb[i].lnb_data);
287 atomic_dec(&osd->od_zerocopy_loan);
290 lnb[i].lnb_page = NULL;
291 lnb[i].lnb_data = NULL;
297 static inline struct page *kmem_to_page(void *addr)
299 if (is_vmalloc_addr(addr))
300 return vmalloc_to_page(addr);
302 return virt_to_page(addr);
306 * Prepare buffers for read.
308 * The function maps the range described by \a off and \a len to \a lnb array.
309 * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
310 * we fill \a lnb array with the pages storing ARC buffers. Notice the current
311 * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
312 * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
313 * A better implementation would just return the buffers (potentially unfilled)
314 * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
316 * \param[in] env environment
317 * \param[in] obj object
318 * \param[in] off offset in bytes
319 * \param[in] len the number of bytes to access
320 * \param[out] lnb array of local niobufs pointing to the buffers with data
322 * \retval 0 for success
323 * \retval negative error number of failure
325 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
326 loff_t off, ssize_t len, struct niobuf_local *lnb)
328 struct osd_device *osd = osd_obj2dev(obj);
329 unsigned long start = cfs_time_current();
330 int rc, i, numbufs, npages = 0;
334 record_start_io(osd, READ, 0);
336 /* grab buffers for read:
337 * OSD API let us to grab buffers first, then initiate IO(s)
338 * so that all required IOs will be done in parallel, but at the
339 * moment DMU doesn't provide us with a method to grab buffers.
340 * If we discover this is a vital for good performance we
341 * can get own replacement for dmu_buf_hold_array_by_bonus().
344 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
345 osd_zerocopy_tag, &numbufs,
350 for (i = 0; i < numbufs; i++) {
351 int bufoff, tocpy, thispage;
356 atomic_inc(&osd->od_zerocopy_pin);
358 bufoff = off - dbp[i]->db_offset;
359 tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
361 /* kind of trick to differentiate dbuf vs. arcbuf */
362 LASSERT(((unsigned long)dbp[i] & 1) == 0);
363 dbf = (void *) ((unsigned long)dbp[i] | 1);
366 thispage = PAGE_CACHE_SIZE;
367 thispage -= bufoff & (PAGE_CACHE_SIZE - 1);
368 thispage = min(tocpy, thispage);
371 lnb->lnb_file_offset = off;
372 lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK;
373 lnb->lnb_len = thispage;
374 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
376 /* mark just a single slot: we need this
377 * reference to dbuf to be release once */
390 /* steal dbuf so dmu_buf_rele_array() cant release it */
394 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
397 record_end_io(osd, READ, cfs_time_current() - start,
398 npages * PAGE_SIZE, npages);
404 osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
408 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
409 loff_t off, ssize_t len, struct niobuf_local *lnb)
411 struct osd_device *osd = osd_obj2dev(obj);
412 int plen, off_in_block, sz_in_block;
413 int rc, i = 0, npages = 0;
419 dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
422 * currently only full blocks are subject to zerocopy approach:
423 * so that we're sure nobody is trying to update the same block
426 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
428 off_in_block = off & (bs - 1);
429 sz_in_block = min_t(int, bs - off_in_block, len);
431 if (sz_in_block == bs) {
432 /* full block, try to use zerocopy */
434 abuf = dmu_request_arcbuf(obj->oo_db, bs);
435 if (unlikely(abuf == NULL))
436 GOTO(out_err, rc = -ENOMEM);
438 atomic_inc(&osd->od_zerocopy_loan);
440 /* go over pages arcbuf contains, put them as
441 * local niobufs for ptlrpc's bulks */
442 while (sz_in_block > 0) {
443 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
445 lnb[i].lnb_file_offset = off;
446 lnb[i].lnb_page_offset = 0;
447 lnb[i].lnb_len = plen;
449 if (sz_in_block == bs)
450 lnb[i].lnb_data = abuf;
452 lnb[i].lnb_data = NULL;
454 /* this one is not supposed to fail */
455 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
457 LASSERT(lnb[i].lnb_page);
459 lprocfs_counter_add(osd->od_stats,
460 LPROC_OSD_ZEROCOPY_IO, 1);
465 off_in_block += plen;
470 if (off_in_block == 0 && len < bs &&
471 off + len >= obj->oo_attr.la_size)
472 lprocfs_counter_add(osd->od_stats,
473 LPROC_OSD_TAIL_IO, 1);
475 /* can't use zerocopy, allocate temp. buffers */
476 while (sz_in_block > 0) {
477 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
479 lnb[i].lnb_file_offset = off;
480 lnb[i].lnb_page_offset = 0;
481 lnb[i].lnb_len = plen;
483 lnb[i].lnb_data = NULL;
485 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
486 if (unlikely(lnb[i].lnb_page == NULL))
487 GOTO(out_err, rc = -ENOMEM);
489 LASSERT(lnb[i].lnb_page->mapping == NULL);
490 lnb[i].lnb_page->mapping = (void *)obj;
492 atomic_inc(&osd->od_zerocopy_alloc);
493 lprocfs_counter_add(osd->od_stats,
494 LPROC_OSD_COPY_IO, 1);
508 osd_bufs_put(env, &obj->oo_dt, lnb, npages);
512 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
513 loff_t offset, ssize_t len, struct niobuf_local *lnb,
514 int rw, struct lustre_capa *capa)
516 struct osd_object *obj = osd_dt_obj(dt);
519 LASSERT(dt_object_exists(dt));
523 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
525 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
530 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
531 struct niobuf_local *lnb, int npages)
533 struct osd_object *obj = osd_dt_obj(dt);
535 LASSERT(dt_object_exists(dt));
541 /* Return number of blocks that aren't mapped in the [start, start + size]
543 static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
546 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)obj->oo_db;
557 if (dn->dn_maxblkid == 0) {
558 if (start + size <= dn->dn_datablksz)
560 if (start < dn->dn_datablksz)
561 start = dn->dn_datablksz;
562 /* assume largest block size */
563 blkshift = SPA_MAXBLOCKSHIFT;
565 /* blocksize can't change */
566 blkshift = dn->dn_datablkshift;
569 /* compute address of last block */
570 end = (start + size - 1) >> blkshift;
571 /* align start on block boundaries */
574 /* size is null, can't be mapped */
575 if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
576 GOTO(out, size = (end - start + 1) << blkshift);
578 /* beyond EOF, can't be mapped */
579 if (start > dn->dn_maxblkid)
580 GOTO(out, size = (end - start + 1) << blkshift);
583 for (blkid = start; blkid <= end; blkid++) {
584 if (blkid == dn->dn_maxblkid)
585 /* this one is mapped for sure */
587 if (blkid > dn->dn_maxblkid) {
588 size += (end - blkid + 1) << blkshift;
592 rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
594 /* for ENOENT (block not mapped) and any other errors,
595 * assume the block isn't mapped */
596 size += 1 << blkshift;
608 static int osd_declare_write_commit(const struct lu_env *env,
609 struct dt_object *dt,
610 struct niobuf_local *lnb, int npages,
613 struct osd_object *obj = osd_dt_obj(dt);
614 struct osd_device *osd = osd_obj2dev(obj);
615 struct osd_thandle *oh;
618 int i, rc, flags = 0;
619 bool ignore_quota = false, synced = false;
621 struct page *last_page = NULL;
622 unsigned long discont_pages = 0;
625 LASSERT(dt_object_exists(dt));
631 oh = container_of0(th, struct osd_thandle, ot_super);
633 for (i = 0; i < npages; i++) {
634 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
636 last_page = lnb[i].lnb_page;
638 /* ENOSPC, network RPC error, etc.
639 * We don't want to book space for pages which will be
640 * skipped in osd_write_commit(). Hence we skip pages
641 * with lnb_rc != 0 here too */
643 /* ignore quota for the whole request if any page is from
644 * client cache or written by root.
646 * XXX once we drop the 1.8 client support, the checking
647 * for whether page is from cache can be simplified as:
648 * !(lnb[i].flags & OBD_BRW_SYNC)
650 * XXX we could handle this on per-lnb basis as done by
652 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
653 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
657 /* first valid lnb */
658 offset = lnb[i].lnb_file_offset;
659 size = lnb[i].lnb_len;
662 if (offset + size == lnb[i].lnb_file_offset) {
663 /* this lnb is contiguous to the previous one */
664 size += lnb[i].lnb_len;
668 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
670 /* estimating space that will be consumed by a write is rather
671 * complicated with ZFS. As a consequence, we don't account for
672 * indirect blocks and quota overrun will be adjusted once the
673 * operation is committed, if required. */
674 space += osd_count_not_mapped(obj, offset, size);
676 offset = lnb[i].lnb_file_offset;
677 size = lnb[i].lnb_len;
681 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
683 space += osd_count_not_mapped(obj, offset, size);
686 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
688 oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
690 /* backend zfs filesystem might be configured to store multiple data
692 space *= osd->od_os->os_copies;
694 CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
695 "space\n", npages, space);
697 record_start_io(osd, WRITE, discont_pages);
699 /* acquire quota space if needed */
700 rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
701 obj->oo_attr.la_gid, space, oh, true, &flags,
704 if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
705 dt_sync(env, th->th_dev);
707 CDEBUG(D_QUOTA, "retry after sync\n");
712 /* we need only to store the overquota flags in the first lnb for
713 * now, once we support multiple objects BRW, this code needs be
715 if (flags & QUOTA_FL_OVER_USRQUOTA)
716 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
717 if (flags & QUOTA_FL_OVER_GRPQUOTA)
718 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
723 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
724 struct niobuf_local *lnb, int npages,
727 struct osd_object *obj = osd_dt_obj(dt);
728 struct osd_device *osd = osd_obj2dev(obj);
729 struct osd_thandle *oh;
730 uint64_t new_size = 0;
732 unsigned long iosize = 0;
735 LASSERT(dt_object_exists(dt));
739 oh = container_of0(th, struct osd_thandle, ot_super);
741 for (i = 0; i < npages; i++) {
742 CDEBUG(D_INODE, "write %u bytes at %u\n",
743 (unsigned) lnb[i].lnb_len,
744 (unsigned) lnb[i].lnb_file_offset);
747 /* ENOSPC, network RPC error, etc.
748 * Unlike ldiskfs, zfs allocates new blocks on rewrite,
749 * so we skip this page if lnb_rc is set to -ENOSPC */
750 CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
751 PFID(lu_object_fid(&dt->do_lu)), i,
756 if (lnb[i].lnb_page->mapping == (void *)obj) {
757 dmu_write(osd->od_os, obj->oo_db->db_object,
758 lnb[i].lnb_file_offset, lnb[i].lnb_len,
759 kmap(lnb[i].lnb_page), oh->ot_tx);
760 kunmap(lnb[i].lnb_page);
761 } else if (lnb[i].lnb_data) {
762 LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
763 /* buffer loaned for zerocopy, try to use it.
764 * notice that dmu_assign_arcbuf() is smart
765 * enough to recognize changed blocksize
766 * in this case it fallbacks to dmu_write() */
767 dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
768 lnb[i].lnb_data, oh->ot_tx);
769 /* drop the reference, otherwise osd_put_bufs()
770 * will be releasing it - bad! */
771 lnb[i].lnb_data = NULL;
772 atomic_dec(&osd->od_zerocopy_loan);
775 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
776 new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
777 iosize += lnb[i].lnb_len;
780 if (unlikely(new_size == 0)) {
781 /* no pages to write, no transno is needed */
783 /* it is important to return 0 even when all lnb_rc == -ENOSPC
784 * since ofd_commitrw_write() retries several times on ENOSPC */
785 record_end_io(osd, WRITE, 0, 0, 0);
789 write_lock(&obj->oo_attr_lock);
790 if (obj->oo_attr.la_size < new_size) {
791 obj->oo_attr.la_size = new_size;
792 write_unlock(&obj->oo_attr_lock);
793 /* osd_object_sa_update() will be copying directly from
794 * oo_attr into dbuf. any update within a single txg will copy
796 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
797 &obj->oo_attr.la_size, 8, oh);
799 write_unlock(&obj->oo_attr_lock);
802 record_end_io(osd, WRITE, 0, iosize, npages);
807 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
808 struct niobuf_local *lnb, int npages)
810 struct osd_object *obj = osd_dt_obj(dt);
812 unsigned long size = 0;
815 LASSERT(dt_object_exists(dt));
818 read_lock(&obj->oo_attr_lock);
819 eof = obj->oo_attr.la_size;
820 read_unlock(&obj->oo_attr_lock);
822 for (i = 0; i < npages; i++) {
823 if (unlikely(lnb[i].lnb_rc < 0))
826 lnb[i].lnb_rc = lnb[i].lnb_len;
827 size += lnb[i].lnb_rc;
829 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > eof) {
830 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
831 if (lnb[i].lnb_rc < 0)
834 /* all subsequent rc should be 0 */
845 * Punch/truncate an object
847 * IN: db - dmu_buf of the object to free data in.
848 * off - start of section to free.
849 * len - length of section to free (DMU_OBJECT_END => to EOF).
851 * RETURN: 0 if success
852 * error code if failure
854 * The transaction passed to this routine must have
855 * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
856 * called and then assigned to a transaction group.
858 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
859 uint64_t size, uint64_t off, uint64_t len)
863 /* Assert that the transaction has been assigned to a
864 transaction group. */
865 LASSERT(tx->tx_txg != 0);
867 * Nothing to do if file already at desired length.
869 if (len == DMU_OBJECT_END && size == off)
873 rc = -dmu_free_range(os, db->db_object, off, len, tx);
878 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
879 __u64 start, __u64 end, struct thandle *th,
880 struct lustre_capa *capa)
882 struct osd_object *obj = osd_dt_obj(dt);
883 struct osd_device *osd = osd_obj2dev(obj);
884 struct osd_thandle *oh;
889 LASSERT(dt_object_exists(dt));
890 LASSERT(osd_invariant(obj));
893 oh = container_of0(th, struct osd_thandle, ot_super);
895 write_lock(&obj->oo_attr_lock);
897 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
898 len = DMU_OBJECT_END;
901 write_unlock(&obj->oo_attr_lock);
903 rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
904 obj->oo_attr.la_size, start, len);
906 if (len == DMU_OBJECT_END) {
907 write_lock(&obj->oo_attr_lock);
908 obj->oo_attr.la_size = start;
909 write_unlock(&obj->oo_attr_lock);
910 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
911 &obj->oo_attr.la_size, 8, oh);
916 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
917 __u64 start, __u64 end, struct thandle *handle)
919 struct osd_object *obj = osd_dt_obj(dt);
920 struct osd_device *osd = osd_obj2dev(obj);
921 struct osd_thandle *oh;
925 oh = container_of0(handle, struct osd_thandle, ot_super);
927 read_lock(&obj->oo_attr_lock);
928 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
929 len = DMU_OBJECT_END;
933 /* declare we'll free some blocks ... */
934 if (start < obj->oo_attr.la_size) {
935 read_unlock(&obj->oo_attr_lock);
936 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
938 read_unlock(&obj->oo_attr_lock);
941 /* ... and we'll modify size attribute */
942 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
944 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
945 obj->oo_attr.la_gid, 0, oh, true, NULL,
950 struct dt_body_operations osd_body_ops = {
951 .dbo_read = osd_read,
952 .dbo_declare_write = osd_declare_write,
953 .dbo_write = osd_write,
954 .dbo_bufs_get = osd_bufs_get,
955 .dbo_bufs_put = osd_bufs_put,
956 .dbo_write_prep = osd_write_prep,
957 .dbo_declare_write_commit = osd_declare_write_commit,
958 .dbo_write_commit = osd_write_commit,
959 .dbo_read_prep = osd_read_prep,
960 .dbo_declare_punch = osd_declare_punch,
961 .dbo_punch = osd_punch,