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) 2009, 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/
31 * lustre/osd-zfs/osd_io.c
33 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
34 * Author: Mike Pershin <tappro@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_OSD
39 #include <libcfs/libcfs.h>
40 #include <obd_support.h>
41 #include <lustre_net.h>
43 #include <obd_class.h>
44 #include <lustre_disk.h>
45 #include <lustre_fid.h>
46 #include <lustre_quota.h>
48 #include "osd_internal.h"
50 #include <sys/dnode.h>
55 #include <sys/spa_impl.h>
56 #include <sys/zfs_znode.h>
57 #include <sys/dmu_tx.h>
58 #include <sys/dmu_objset.h>
59 #include <sys/dsl_prop.h>
60 #include <sys/sa_impl.h>
63 static char osd_0copy_tag[] = "zerocopy";
65 static void dbuf_set_pending_evict(dmu_buf_t *db)
67 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
68 dbi->db_pending_evict = TRUE;
71 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
73 struct obd_histogram *h = osd->od_brw_stats.hist;
76 atomic_inc(&osd->od_r_in_flight);
77 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
78 atomic_read(&osd->od_r_in_flight));
79 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
82 atomic_inc(&osd->od_w_in_flight);
83 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
84 atomic_read(&osd->od_w_in_flight));
85 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
90 static void record_end_io(struct osd_device *osd, int rw,
91 unsigned long elapsed, int disksize, int npages)
93 struct obd_histogram *h = osd->od_brw_stats.hist;
96 atomic_dec(&osd->od_r_in_flight);
98 atomic_dec(&osd->od_w_in_flight);
100 lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
102 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
104 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
107 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
108 struct lu_buf *buf, loff_t *pos, size_t *size)
110 struct osd_object *obj = osd_dt_obj(dt);
114 LASSERT(dt_object_exists(dt));
117 read_lock(&obj->oo_attr_lock);
118 old_size = obj->oo_attr.la_size;
119 read_unlock(&obj->oo_attr_lock);
121 if (*pos + *size > old_size) {
125 *size = old_size - *pos;
128 rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
129 buf->lb_buf, DMU_READ_PREFETCH);
138 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
139 struct lu_buf *buf, loff_t *pos)
141 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
142 size_t size = buf->lb_len;
143 hrtime_t start = gethrtime();
147 record_start_io(osd, READ, 0);
148 rc = __osd_read(env, dt, buf, pos, &size);
149 delta_ms = gethrtime() - start;
150 do_div(delta_ms, NSEC_PER_MSEC);
151 record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
156 static inline ssize_t osd_read_no_record(const struct lu_env *env,
157 struct dt_object *dt,
158 struct lu_buf *buf, loff_t *pos)
160 size_t size = buf->lb_len;
162 return __osd_read(env, dt, buf, pos, &size);
165 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
166 const struct lu_buf *buf, loff_t pos,
169 struct osd_object *obj = osd_dt_obj(dt);
170 struct osd_device *osd = osd_obj2dev(obj);
171 struct osd_thandle *oh;
175 oh = container_of(th, struct osd_thandle, ot_super);
177 /* in some cases declare can race with creation (e.g. llog)
178 * and we need to wait till object is initialized. notice
179 * LOHA_EXISTs is supposed to be the last step in the
182 /* size change (in dnode) will be declared by dmu_tx_hold_write() */
183 if (dt_object_exists(dt))
184 oid = obj->oo_dn->dn_object;
186 oid = DMU_NEW_OBJECT;
188 /* XXX: we still miss for append declaration support in ZFS
189 * -1 means append which is used by llog mostly, llog
190 * can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
192 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
193 obj->oo_attr.la_size + (2 << 20));
194 osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
196 /* dt_declare_write() is usually called for system objects, such
197 * as llog or last_rcvd files. We needn't enforce quota on those
198 * objects, so always set the lqi_space as 0. */
199 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
200 obj->oo_attr.la_gid, obj->oo_attr.la_projid,
201 0, oh, NULL, OSD_QID_BLK));
204 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
205 const struct lu_buf *buf, loff_t *pos,
208 struct osd_object *obj = osd_dt_obj(dt);
209 struct osd_device *osd = osd_obj2dev(obj);
210 struct osd_thandle *oh;
211 uint64_t offset = *pos;
216 LASSERT(dt_object_exists(dt));
220 oh = container_of(th, struct osd_thandle, ot_super);
222 down_read(&obj->oo_guard);
223 if (obj->oo_destroyed)
224 GOTO(out, rc = -ENOENT);
226 osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
227 buf->lb_buf, oh->ot_tx);
228 write_lock(&obj->oo_attr_lock);
229 if (obj->oo_attr.la_size < offset + buf->lb_len) {
230 obj->oo_attr.la_size = offset + buf->lb_len;
231 write_unlock(&obj->oo_attr_lock);
232 /* osd_object_sa_update() will be copying directly from oo_attr
233 * into dbuf. any update within a single txg will copy the
235 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
236 &obj->oo_attr.la_size, 8, oh);
240 write_unlock(&obj->oo_attr_lock);
247 up_read(&obj->oo_guard);
252 * XXX: for the moment I don't want to use lnb_flags for osd-internal
253 * purposes as it's not very well defined ...
254 * instead I use the lowest bit of the address so that:
255 * arc buffer: .lnb_data = abuf (arc we loan for write)
256 * dbuf buffer: .lnb_data = dbuf | 1 (dbuf we get for read)
257 * copy buffer: .lnb_page->mapping = obj (page we allocate for write)
261 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
262 struct niobuf_local *lnb, int npages)
264 struct osd_object *obj = osd_dt_obj(dt);
265 struct osd_device *osd = osd_obj2dev(obj);
269 LASSERT(dt_object_exists(dt));
272 for (i = 0; i < npages; i++) {
273 if (lnb[i].lnb_page == NULL)
275 if (lnb[i].lnb_page->mapping == (void *)obj) {
276 /* this is anonymous page allocated for copy-write */
277 lnb[i].lnb_page->mapping = NULL;
278 __free_page(lnb[i].lnb_page);
279 atomic_dec(&osd->od_zerocopy_alloc);
281 /* see comment in osd_bufs_get_read() */
282 ptr = (unsigned long)lnb[i].lnb_data;
285 dmu_buf_rele((void *)ptr, osd_0copy_tag);
286 atomic_dec(&osd->od_zerocopy_pin);
287 } else if (lnb[i].lnb_data != NULL) {
288 int j, apages, abufsz;
289 abufsz = arc_buf_size(lnb[i].lnb_data);
290 apages = abufsz >> PAGE_SHIFT;
291 /* these references to pages must be invalidated
292 * to prevent access in osd_bufs_put() */
293 for (j = 0; j < apages; j++)
294 lnb[i + j].lnb_page = NULL;
295 dmu_return_arcbuf(lnb[i].lnb_data);
296 atomic_dec(&osd->od_zerocopy_loan);
299 lnb[i].lnb_page = NULL;
300 lnb[i].lnb_data = NULL;
306 static inline struct page *kmem_to_page(void *addr)
308 LASSERT(!((unsigned long)addr & ~PAGE_MASK));
309 if (is_vmalloc_addr(addr))
310 return vmalloc_to_page(addr);
312 return virt_to_page(addr);
316 * Prepare buffers for read.
318 * The function maps the range described by \a off and \a len to \a lnb array.
319 * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
320 * we fill \a lnb array with the pages storing ARC buffers. Notice the current
321 * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
322 * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
323 * A better implementation would just return the buffers (potentially unfilled)
324 * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
326 * \param[in] env environment
327 * \param[in] obj object
328 * \param[in] off offset in bytes
329 * \param[in] len the number of bytes to access
330 * \param[out] lnb array of local niobufs pointing to the buffers with data
332 * \retval 0 for success
333 * \retval negative error number of failure
335 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
336 loff_t off, ssize_t len, struct niobuf_local *lnb,
339 struct osd_device *osd = osd_obj2dev(obj);
340 int rc, i, numbufs, npages = 0, drop_cache = 0;
341 hrtime_t start = gethrtime();
346 record_start_io(osd, READ, 0);
348 if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
351 /* grab buffers for read:
352 * OSD API let us to grab buffers first, then initiate IO(s)
353 * so that all required IOs will be done in parallel, but at the
354 * moment DMU doesn't provide us with a method to grab buffers.
355 * If we discover this is a vital for good performance we
356 * can get own replacement for dmu_buf_hold_array_by_bonus().
359 (obj->oo_dn->dn_datablkshift != 0 ||
360 off < obj->oo_dn->dn_datablksz)) {
361 if (obj->oo_dn->dn_datablkshift == 0 &&
362 off + len > obj->oo_dn->dn_datablksz)
363 len = obj->oo_dn->dn_datablksz - off;
366 if (unlikely(npages >= maxlnb))
367 GOTO(err, rc = -EOVERFLOW);
369 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
370 off, len, TRUE, osd_0copy_tag,
375 for (i = 0; i < numbufs; i++) {
376 int bufoff, tocpy, thispage;
381 atomic_inc(&osd->od_zerocopy_pin);
383 bufoff = off - dbp[i]->db_offset;
384 tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
386 /* kind of trick to differentiate dbuf vs. arcbuf */
387 LASSERT(((unsigned long)dbp[i] & 1) == 0);
388 dbf = (void *) ((unsigned long)dbp[i] | 1);
391 if (unlikely(npages >= maxlnb))
392 GOTO(err, rc = -EOVERFLOW);
394 thispage = PAGE_SIZE;
395 thispage -= bufoff & (PAGE_SIZE - 1);
396 thispage = min(tocpy, thispage);
399 lnb->lnb_file_offset = off;
400 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
401 lnb->lnb_len = thispage;
402 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
404 /* mark just a single slot: we need this
405 * reference to dbuf to be released once */
419 dbuf_set_pending_evict(dbp[i]);
421 /* steal dbuf so dmu_buf_rele_array() can't release
426 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
429 delta_ms = gethrtime() - start;
430 do_div(delta_ms, NSEC_PER_MSEC);
431 record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
438 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
439 osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
443 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
447 abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
449 return ERR_PTR(-ENOMEM);
451 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
453 * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
454 * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
455 * setting up page array for RDMA transfer. See LU-9305.
457 if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
458 dmu_return_arcbuf(abuf);
466 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
467 loff_t off, ssize_t len, struct niobuf_local *lnb,
470 struct osd_device *osd = osd_obj2dev(obj);
471 int poff, plen, off_in_block, sz_in_block;
472 int rc, i = 0, npages = 0;
473 dnode_t *dn = obj->oo_dn;
475 uint32_t bs = dn->dn_datablksz;
479 * currently only full blocks are subject to zerocopy approach:
480 * so that we're sure nobody is trying to update the same block
483 if (unlikely(npages >= maxlnb))
484 GOTO(out_err, rc = -EOVERFLOW);
486 off_in_block = off & (bs - 1);
487 sz_in_block = min_t(int, bs - off_in_block, len);
490 if (sz_in_block == bs) {
491 /* full block, try to use zerocopy */
492 abuf = osd_request_arcbuf(dn, bs);
493 if (unlikely(IS_ERR(abuf)))
494 GOTO(out_err, rc = PTR_ERR(abuf));
498 atomic_inc(&osd->od_zerocopy_loan);
500 /* go over pages arcbuf contains, put them as
501 * local niobufs for ptlrpc's bulks */
502 while (sz_in_block > 0) {
503 plen = min_t(int, sz_in_block, PAGE_SIZE);
505 if (unlikely(npages >= maxlnb))
506 GOTO(out_err, rc = -EOVERFLOW);
508 lnb[i].lnb_file_offset = off;
509 lnb[i].lnb_page_offset = 0;
510 lnb[i].lnb_len = plen;
512 if (sz_in_block == bs)
513 lnb[i].lnb_data = abuf;
515 lnb[i].lnb_data = NULL;
517 /* this one is not supposed to fail */
518 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
520 LASSERT(lnb[i].lnb_page);
522 lprocfs_counter_add(osd->od_stats,
523 LPROC_OSD_ZEROCOPY_IO, 1);
528 off_in_block += plen;
533 if (off_in_block == 0 && len < bs &&
534 off + len >= obj->oo_attr.la_size)
535 lprocfs_counter_add(osd->od_stats,
536 LPROC_OSD_TAIL_IO, 1);
538 /* can't use zerocopy, allocate temp. buffers */
539 poff = off & (PAGE_SIZE - 1);
540 while (sz_in_block > 0) {
541 plen = min_t(int, poff + sz_in_block,
545 if (unlikely(npages >= maxlnb))
546 GOTO(out_err, rc = -EOVERFLOW);
548 lnb[i].lnb_file_offset = off;
549 lnb[i].lnb_page_offset = poff;
552 lnb[i].lnb_len = plen;
554 lnb[i].lnb_data = NULL;
556 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
557 if (unlikely(lnb[i].lnb_page == NULL))
558 GOTO(out_err, rc = -ENOMEM);
560 LASSERT(lnb[i].lnb_page->mapping == NULL);
561 lnb[i].lnb_page->mapping = (void *)obj;
563 atomic_inc(&osd->od_zerocopy_alloc);
564 lprocfs_counter_add(osd->od_stats,
565 LPROC_OSD_COPY_IO, 1);
579 osd_bufs_put(env, &obj->oo_dt, lnb, npages);
583 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
584 loff_t offset, ssize_t len, struct niobuf_local *lnb,
585 int maxlnb, enum dt_bufs_type rw)
587 struct osd_object *obj = osd_dt_obj(dt);
590 LASSERT(dt_object_exists(dt));
593 if (rw & DT_BUFS_TYPE_WRITE)
594 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
596 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
601 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
602 struct niobuf_local *lnb, int npages)
604 struct osd_object *obj = osd_dt_obj(dt);
606 LASSERT(dt_object_exists(dt));
612 static inline uint64_t osd_roundup2blocksz(uint64_t size,
618 size += offset % blksz;
620 if (likely(is_power_of_2(blksz)))
621 return round_up(size, blksz);
623 return DIV_ROUND_UP_ULL(size, blksz) * blksz;
626 static int osd_declare_write_commit(const struct lu_env *env,
627 struct dt_object *dt,
628 struct niobuf_local *lnb, int npages,
631 struct osd_object *obj = osd_dt_obj(dt);
632 struct osd_device *osd = osd_obj2dev(obj);
633 struct osd_thandle *oh;
636 uint32_t blksz = obj->oo_dn->dn_datablksz;
640 struct page *last_page = NULL;
641 unsigned long discont_pages = 0;
642 enum osd_quota_local_flags local_flags = 0;
643 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
646 LASSERT(dt_object_exists(dt));
652 oh = container_of(th, struct osd_thandle, ot_super);
654 for (i = 0; i < npages; i++) {
655 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
657 last_page = lnb[i].lnb_page;
659 /* ENOSPC, network RPC error, etc.
660 * We don't want to book space for pages which will be
661 * skipped in osd_write_commit(). Hence we skip pages
662 * with lnb_rc != 0 here too */
664 /* ignore quota for the whole request if any page is from
665 * client cache or written by root.
667 * XXX we could handle this on per-lnb basis as done by
669 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
670 (lnb[i].lnb_flags & OBD_BRW_SYS_RESOURCE) ||
671 !(lnb[i].lnb_flags & OBD_BRW_SYNC))
672 declare_flags |= OSD_QID_FORCE;
675 /* first valid lnb */
676 offset = lnb[i].lnb_file_offset;
677 size = lnb[i].lnb_len;
680 if (offset + size == lnb[i].lnb_file_offset) {
681 /* this lnb is contiguous to the previous one */
682 size += lnb[i].lnb_len;
686 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
687 obj->oo_dn, offset, size);
688 /* Estimating space to be consumed by a write is rather
689 * complicated with ZFS. As a consequence, we don't account for
690 * indirect blocks and just use as a rough estimate the worse
691 * case where the old space is being held by a snapshot. Quota
692 * overrun will be adjusted once the operation is committed, if
694 space += osd_roundup2blocksz(size, offset, blksz);
696 offset = lnb[i].lnb_file_offset;
697 size = lnb[i].lnb_len;
701 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
703 space += osd_roundup2blocksz(size, offset, blksz);
706 /* backend zfs filesystem might be configured to store multiple data
708 space *= osd->od_os->os_copies;
710 CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
713 record_start_io(osd, WRITE, discont_pages);
715 /* acquire quota space if needed */
716 rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
717 obj->oo_attr.la_gid, obj->oo_attr.la_projid,
718 space, oh, &local_flags, declare_flags);
720 if (!synced && rc == -EDQUOT &&
721 (local_flags & QUOTA_FL_SYNC) != 0) {
722 dt_sync(env, th->th_dev);
724 CDEBUG(D_QUOTA, "retry after sync\n");
729 /* we need only to store the overquota flags in the first lnb for
730 * now, once we support multiple objects BRW, this code needs be
732 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
733 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
734 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
735 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
736 #ifdef ZFS_PROJINHERIT
737 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
738 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
745 * Policy to grow ZFS block size by write pattern.
746 * For sequential write, it grows block size gradually until it reaches the
747 * maximum blocksize the dataset can support. Otherwise, it will pick a
748 * a block size by the writing region of this I/O.
750 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
751 uint64_t start, uint64_t end)
753 struct osd_device *osd = osd_obj2dev(obj);
754 dnode_t *dn = obj->oo_dn;
760 if (dn->dn_maxblkid > 0) /* can't change block size */
763 if (dn->dn_datablksz >= osd->od_max_blksz)
766 down_write(&obj->oo_guard);
768 blksz = dn->dn_datablksz;
769 if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
770 GOTO(out_unlock, rc);
772 /* now ZFS can support up to 16MB block size, and if the write
773 * is sequential, it just increases the block size gradually */
774 if (start <= blksz) { /* sequential */
775 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
776 } else { /* sparse, pick a block size by write region */
777 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
781 if (!is_power_of_2(blksz))
782 blksz = size_roundup_power2(blksz);
784 if (blksz > dn->dn_datablksz) {
785 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
786 blksz, 0, oh->ot_tx);
787 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
789 CDEBUG(D_INODE, "object "DFID": change block size"
790 "%u -> %u error rc = %d\n",
791 PFID(lu_object_fid(&obj->oo_dt.do_lu)),
792 dn->dn_datablksz, blksz, rc);
796 up_write(&obj->oo_guard);
801 static void osd_evict_dbufs_after_write(struct osd_object *obj,
802 loff_t off, ssize_t len)
807 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
808 TRUE, osd_0copy_tag, &numbufs, &dbp);
812 for (i = 0; i < numbufs; i++)
813 dbuf_set_pending_evict(dbp[i]);
815 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
818 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
819 struct niobuf_local *lnb, int npages,
820 struct thandle *th, __u64 user_size)
822 struct osd_object *obj = osd_dt_obj(dt);
823 struct osd_device *osd = osd_obj2dev(obj);
824 struct osd_thandle *oh;
825 uint64_t new_size = 0;
826 int i, abufsz, rc = 0, drop_cache = 0;
827 unsigned long iosize = 0;
830 LASSERT(dt_object_exists(dt));
834 oh = container_of(th, struct osd_thandle, ot_super);
836 /* adjust block size. Assume the buffers are sorted. */
837 (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
838 lnb[npages - 1].lnb_file_offset +
839 lnb[npages - 1].lnb_len);
841 if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
842 lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
843 osd->od_readcache_max_filesize)
846 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
849 /* if la_size is already bigger than specified user_size,
852 if (obj->oo_attr.la_size > user_size)
855 /* LU-8791: take oo_guard to avoid the deadlock that changing block
856 * size and assigning arcbuf take place at the same time.
860 * -> osd_grow_blocksize() with osd_object::oo_guard held
861 * -> dmu_object_set_blocksize()
862 * -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
865 * -> dmu_buf_will_dirty()
867 * -> wait for the dbuf state to change
870 * -> dmu_assign_arcbuf()
871 * -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
873 * -> try to hold the read lock of dnode_t::dn_struct_rwlock
875 * By taking the read lock, it can avoid thread 2 to enter into the
876 * critical section of assigning the arcbuf, while thread 1 is
877 * changing the block size.
879 down_read(&obj->oo_guard);
880 if (obj->oo_destroyed) {
881 up_read(&obj->oo_guard);
885 for (i = 0; i < npages; i++) {
886 CDEBUG(D_INODE, "write %u bytes at %u\n",
887 (unsigned) lnb[i].lnb_len,
888 (unsigned) lnb[i].lnb_file_offset);
891 /* ENOSPC, network RPC error, etc.
892 * Unlike ldiskfs, zfs allocates new blocks on rewrite,
893 * so we skip this page if lnb_rc is set to -ENOSPC */
894 CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
895 PFID(lu_object_fid(&dt->do_lu)), i,
900 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
901 new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
902 if (lnb[i].lnb_page == NULL)
905 if (lnb[i].lnb_page->mapping == (void *)obj) {
906 osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
907 lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
908 lnb[i].lnb_page_offset, oh->ot_tx);
909 kunmap(lnb[i].lnb_page);
910 iosize += lnb[i].lnb_len;
911 abufsz = lnb[i].lnb_len; /* to drop cache below */
912 } else if (lnb[i].lnb_data) {
914 LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
915 /* buffer loaned for zerocopy, try to use it.
916 * notice that dmu_assign_arcbuf() is smart
917 * enough to recognize changed blocksize
918 * in this case it fallbacks to dmu_write() */
919 abufsz = arc_buf_size(lnb[i].lnb_data);
920 LASSERT(abufsz & PAGE_MASK);
921 apages = abufsz >> PAGE_SHIFT;
922 LASSERT(i + apages <= npages);
923 /* these references to pages must be invalidated
924 * to prevent access in osd_bufs_put() */
925 for (j = 0; j < apages; j++)
926 lnb[i + j].lnb_page = NULL;
927 dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
928 lnb[i].lnb_file_offset,
929 lnb[i].lnb_data, oh->ot_tx);
930 /* drop the reference, otherwise osd_put_bufs()
931 * will be releasing it - bad! */
932 lnb[i].lnb_data = NULL;
933 atomic_dec(&osd->od_zerocopy_loan);
936 /* we don't want to deal with cache if nothing
937 * has been send to ZFS at this step */
944 /* we have to mark dbufs for eviction here because
945 * dmu_assign_arcbuf() may create a new dbuf for
947 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
951 if (unlikely(new_size == 0)) {
952 /* no pages to write, no transno is needed */
954 /* it is important to return 0 even when all lnb_rc == -ENOSPC
955 * since ofd_commitrw_write() retries several times on ENOSPC */
956 up_read(&obj->oo_guard);
957 record_end_io(osd, WRITE, 0, 0, 0);
961 /* if file has grown, take user_size into account */
962 if (user_size && new_size > user_size)
963 new_size = user_size;
964 write_lock(&obj->oo_attr_lock);
965 if (obj->oo_attr.la_size < new_size) {
966 obj->oo_attr.la_size = new_size;
967 write_unlock(&obj->oo_attr_lock);
968 /* osd_object_sa_update() will be copying directly from
969 * oo_attr into dbuf. any update within a single txg will copy
971 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
972 &obj->oo_attr.la_size, 8, oh);
974 write_unlock(&obj->oo_attr_lock);
977 up_read(&obj->oo_guard);
979 record_end_io(osd, WRITE, 0, iosize, npages);
984 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
985 struct niobuf_local *lnb, int npages)
987 struct osd_object *obj = osd_dt_obj(dt);
991 LASSERT(dt_object_exists(dt));
994 read_lock(&obj->oo_attr_lock);
995 eof = obj->oo_attr.la_size;
996 read_unlock(&obj->oo_attr_lock);
998 for (i = 0; i < npages; i++) {
999 if (unlikely(lnb[i].lnb_rc < 0))
1002 lnb[i].lnb_rc = lnb[i].lnb_len;
1004 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
1005 /* send complete pages all the time */
1006 if (eof <= lnb[i].lnb_file_offset)
1009 /* all subsequent rc should be 0 */
1010 while (++i < npages)
1020 * Punch/truncate an object
1022 * IN: db - dmu_buf of the object to free data in.
1023 * off - start of section to free.
1024 * len - length of section to free (DMU_OBJECT_END => to EOF).
1026 * RETURN: 0 if success
1027 * error code if failure
1029 * The transaction passed to this routine must have
1030 * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1031 * called and then assigned to a transaction group.
1033 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1034 dmu_tx_t *tx, uint64_t off, uint64_t len)
1036 dnode_t *dn = obj->oo_dn;
1037 uint64_t size = obj->oo_attr.la_size;
1040 /* Assert that the transaction has been assigned to a
1041 transaction group. */
1042 LASSERT(tx->tx_txg != 0);
1044 * Nothing to do if file already at desired length.
1046 if (len == DMU_OBJECT_END && size == off)
1049 /* if object holds encrypted content, we need to make sure we truncate
1050 * on an encryption unit boundary, or subsequent reads will get
1053 if (len != DMU_OBJECT_END)
1054 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1055 (off & ~LUSTRE_ENCRYPTION_MASK);
1056 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1057 off & ~LUSTRE_ENCRYPTION_MASK)
1058 off = (off & LUSTRE_ENCRYPTION_MASK) +
1059 LUSTRE_ENCRYPTION_UNIT_SIZE;
1062 /* XXX: dnode_free_range() can be used to save on dnode lookup */
1064 dmu_free_range(os, dn->dn_object, off, len, tx);
1069 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1070 __u64 start, __u64 end, struct thandle *th)
1072 struct osd_object *obj = osd_dt_obj(dt);
1073 struct osd_device *osd = osd_obj2dev(obj);
1074 struct osd_thandle *oh;
1079 LASSERT(dt_object_exists(dt));
1080 LASSERT(osd_invariant(obj));
1082 LASSERT(th != NULL);
1083 oh = container_of(th, struct osd_thandle, ot_super);
1085 write_lock(&obj->oo_attr_lock);
1087 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1088 len = DMU_OBJECT_END;
1091 write_unlock(&obj->oo_attr_lock);
1093 down_read(&obj->oo_guard);
1094 if (obj->oo_destroyed)
1095 GOTO(out, rc = -ENOENT);
1097 rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1100 if (len == DMU_OBJECT_END) {
1101 write_lock(&obj->oo_attr_lock);
1102 obj->oo_attr.la_size = start;
1103 write_unlock(&obj->oo_attr_lock);
1104 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1105 &obj->oo_attr.la_size, 8, oh);
1108 up_read(&obj->oo_guard);
1112 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1113 __u64 start, __u64 end, struct thandle *handle)
1115 struct osd_object *obj = osd_dt_obj(dt);
1116 struct osd_device *osd = osd_obj2dev(obj);
1117 struct osd_thandle *oh;
1121 oh = container_of(handle, struct osd_thandle, ot_super);
1123 read_lock(&obj->oo_attr_lock);
1124 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1125 len = DMU_OBJECT_END;
1129 /* declare we'll free some blocks ... */
1130 /* if object holds encrypted content, we need to make sure we truncate
1131 * on an encryption unit boundary, or subsequent reads will get
1134 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1135 start & ~LUSTRE_ENCRYPTION_MASK)
1136 start = (start & LUSTRE_ENCRYPTION_MASK) +
1137 LUSTRE_ENCRYPTION_UNIT_SIZE;
1138 if (start < obj->oo_attr.la_size) {
1139 read_unlock(&obj->oo_attr_lock);
1140 dmu_tx_mark_netfree(oh->ot_tx);
1141 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1143 read_unlock(&obj->oo_attr_lock);
1146 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1147 obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1148 0, oh, NULL, OSD_QID_BLK));
1151 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1152 __u64 start, __u64 end, enum lu_ladvise_type advice)
1166 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1167 __u64 start, __u64 end, int mode, struct thandle *th)
1169 int rc = -EOPNOTSUPP;
1173 * space preallocation is not supported for ZFS
1174 * Returns -EOPNOTSUPP for now
1179 static int osd_declare_fallocate(const struct lu_env *env,
1180 struct dt_object *dt, __u64 start, __u64 end,
1181 int mode, struct thandle *th)
1183 int rc = -EOPNOTSUPP;
1187 * space preallocation is not supported for ZFS
1188 * Returns -EOPNOTSUPP for now
1193 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1194 loff_t offset, int whence)
1196 struct osd_object *obj = osd_dt_obj(dt);
1197 struct osd_device *osd = osd_obj2dev(obj);
1198 uint64_t size = obj->oo_attr.la_size;
1199 uint64_t result = offset;
1201 boolean_t hole = whence == SEEK_HOLE;
1205 LASSERT(dt_object_exists(dt));
1206 LASSERT(osd_invariant(obj));
1207 LASSERT(offset >= 0);
1209 /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
1210 * hole. LOV to decide after all if that real hole or not.
1213 RETURN(hole ? offset : -ENXIO);
1215 /* Currently ZFS reports no valid DATA offset if object has dirty data
1216 * and we cannot just switch to generic way with reporting DATA on all
1217 * file offsets and HOLE beyond end of file, because we may get HOLE
1218 * reported correctly at some offset inside file then DATA will find
1219 * dirty state and be reported also at that offset by generic approach.
1220 * This is because for HOLE report ZFS doesn't check dirty state but
1222 * The only way to get reliable results is to call txg_wait_synced()
1223 * when ZFS reports EBUSY result and repeat lseek call and that is
1224 * controlled via od_sync_on_lseek option.
1226 if (!osd->od_sync_on_lseek)
1227 result = hole ? size : offset;
1230 rc = osd_dmu_offset_next(osd->od_os, obj->oo_dn->dn_object, hole,
1232 /* dirty inode, lseek result is unreliable without sync */
1234 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1241 /* ZFS is not exported all needed function, so fall back to the
1242 * generic logic: for HOLE return file size, for DATA return
1243 * the current offset
1245 if (rc == EOPNOTSUPP)
1246 result = hole ? size : offset;
1250 /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
1251 * result as end of the last block instead of logical EOF which we need
1259 const struct dt_body_operations osd_body_ops = {
1260 .dbo_read = osd_read,
1261 .dbo_declare_write = osd_declare_write,
1262 .dbo_write = osd_write,
1263 .dbo_bufs_get = osd_bufs_get,
1264 .dbo_bufs_put = osd_bufs_put,
1265 .dbo_write_prep = osd_write_prep,
1266 .dbo_declare_write_commit = osd_declare_write_commit,
1267 .dbo_write_commit = osd_write_commit,
1268 .dbo_read_prep = osd_read_prep,
1269 .dbo_declare_punch = osd_declare_punch,
1270 .dbo_punch = osd_punch,
1271 .dbo_ladvise = osd_ladvise,
1272 .dbo_declare_fallocate = osd_declare_fallocate,
1273 .dbo_fallocate = osd_fallocate,
1274 .dbo_lseek = osd_lseek,
1277 const struct dt_body_operations osd_body_scrub_ops = {
1278 .dbo_read = osd_read_no_record,
1279 .dbo_declare_write = osd_declare_write,
1280 .dbo_write = osd_write,