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 osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
223 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);
247 * XXX: for the moment I don't want to use lnb_flags for osd-internal
248 * purposes as it's not very well defined ...
249 * instead I use the lowest bit of the address so that:
250 * arc buffer: .lnb_data = abuf (arc we loan for write)
251 * dbuf buffer: .lnb_data = dbuf | 1 (dbuf we get for read)
252 * copy buffer: .lnb_page->mapping = obj (page we allocate for write)
256 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
257 struct niobuf_local *lnb, int npages)
259 struct osd_object *obj = osd_dt_obj(dt);
260 struct osd_device *osd = osd_obj2dev(obj);
264 LASSERT(dt_object_exists(dt));
267 for (i = 0; i < npages; i++) {
268 if (lnb[i].lnb_page == NULL)
270 if (lnb[i].lnb_page->mapping == (void *)obj) {
271 /* this is anonymous page allocated for copy-write */
272 lnb[i].lnb_page->mapping = NULL;
273 __free_page(lnb[i].lnb_page);
274 atomic_dec(&osd->od_zerocopy_alloc);
276 /* see comment in osd_bufs_get_read() */
277 ptr = (unsigned long)lnb[i].lnb_data;
280 dmu_buf_rele((void *)ptr, osd_0copy_tag);
281 atomic_dec(&osd->od_zerocopy_pin);
282 } else if (lnb[i].lnb_data != NULL) {
283 int j, apages, abufsz;
284 abufsz = arc_buf_size(lnb[i].lnb_data);
285 apages = abufsz >> PAGE_SHIFT;
286 /* these references to pages must be invalidated
287 * to prevent access in osd_bufs_put() */
288 for (j = 0; j < apages; j++)
289 lnb[i + j].lnb_page = NULL;
290 dmu_return_arcbuf(lnb[i].lnb_data);
291 atomic_dec(&osd->od_zerocopy_loan);
294 lnb[i].lnb_page = NULL;
295 lnb[i].lnb_data = NULL;
301 static inline struct page *kmem_to_page(void *addr)
303 LASSERT(!((unsigned long)addr & ~PAGE_MASK));
304 if (is_vmalloc_addr(addr))
305 return vmalloc_to_page(addr);
307 return virt_to_page(addr);
311 * Prepare buffers for read.
313 * The function maps the range described by \a off and \a len to \a lnb array.
314 * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
315 * we fill \a lnb array with the pages storing ARC buffers. Notice the current
316 * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
317 * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
318 * A better implementation would just return the buffers (potentially unfilled)
319 * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
321 * \param[in] env environment
322 * \param[in] obj object
323 * \param[in] off offset in bytes
324 * \param[in] len the number of bytes to access
325 * \param[out] lnb array of local niobufs pointing to the buffers with data
327 * \retval 0 for success
328 * \retval negative error number of failure
330 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
331 loff_t off, ssize_t len, struct niobuf_local *lnb,
334 struct osd_device *osd = osd_obj2dev(obj);
335 int rc, i, numbufs, npages = 0, drop_cache = 0;
336 hrtime_t start = gethrtime();
341 record_start_io(osd, READ, 0);
343 if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
346 /* grab buffers for read:
347 * OSD API let us to grab buffers first, then initiate IO(s)
348 * so that all required IOs will be done in parallel, but at the
349 * moment DMU doesn't provide us with a method to grab buffers.
350 * If we discover this is a vital for good performance we
351 * can get own replacement for dmu_buf_hold_array_by_bonus().
354 (obj->oo_dn->dn_datablkshift != 0 ||
355 off < obj->oo_dn->dn_datablksz)) {
356 if (obj->oo_dn->dn_datablkshift == 0 &&
357 off + len > obj->oo_dn->dn_datablksz)
358 len = obj->oo_dn->dn_datablksz - off;
361 if (unlikely(npages >= maxlnb))
362 GOTO(err, rc = -EOVERFLOW);
364 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
365 off, len, TRUE, osd_0copy_tag,
370 for (i = 0; i < numbufs; i++) {
371 int bufoff, tocpy, thispage;
376 atomic_inc(&osd->od_zerocopy_pin);
378 bufoff = off - dbp[i]->db_offset;
379 tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
381 /* kind of trick to differentiate dbuf vs. arcbuf */
382 LASSERT(((unsigned long)dbp[i] & 1) == 0);
383 dbf = (void *) ((unsigned long)dbp[i] | 1);
386 if (unlikely(npages >= maxlnb))
387 GOTO(err, rc = -EOVERFLOW);
389 thispage = PAGE_SIZE;
390 thispage -= bufoff & (PAGE_SIZE - 1);
391 thispage = min(tocpy, thispage);
394 lnb->lnb_file_offset = off;
395 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
396 lnb->lnb_len = thispage;
397 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
399 /* mark just a single slot: we need this
400 * reference to dbuf to be released once */
414 dbuf_set_pending_evict(dbp[i]);
416 /* steal dbuf so dmu_buf_rele_array() can't release
421 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
424 delta_ms = gethrtime() - start;
425 do_div(delta_ms, NSEC_PER_MSEC);
426 record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
433 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
434 osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
438 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
442 abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
444 return ERR_PTR(-ENOMEM);
446 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
448 * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
449 * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
450 * setting up page array for RDMA transfer. See LU-9305.
452 if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
453 dmu_return_arcbuf(abuf);
461 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
462 loff_t off, ssize_t len, struct niobuf_local *lnb,
465 struct osd_device *osd = osd_obj2dev(obj);
466 int poff, plen, off_in_block, sz_in_block;
467 int rc, i = 0, npages = 0;
468 dnode_t *dn = obj->oo_dn;
470 uint32_t bs = dn->dn_datablksz;
474 * currently only full blocks are subject to zerocopy approach:
475 * so that we're sure nobody is trying to update the same block
478 if (unlikely(npages >= maxlnb))
479 GOTO(out_err, rc = -EOVERFLOW);
481 off_in_block = off & (bs - 1);
482 sz_in_block = min_t(int, bs - off_in_block, len);
485 if (sz_in_block == bs) {
486 /* full block, try to use zerocopy */
487 abuf = osd_request_arcbuf(dn, bs);
488 if (unlikely(IS_ERR(abuf)))
489 GOTO(out_err, rc = PTR_ERR(abuf));
493 atomic_inc(&osd->od_zerocopy_loan);
495 /* go over pages arcbuf contains, put them as
496 * local niobufs for ptlrpc's bulks */
497 while (sz_in_block > 0) {
498 plen = min_t(int, sz_in_block, PAGE_SIZE);
500 if (unlikely(npages >= maxlnb))
501 GOTO(out_err, rc = -EOVERFLOW);
503 lnb[i].lnb_file_offset = off;
504 lnb[i].lnb_page_offset = 0;
505 lnb[i].lnb_len = plen;
507 if (sz_in_block == bs)
508 lnb[i].lnb_data = abuf;
510 lnb[i].lnb_data = NULL;
512 /* this one is not supposed to fail */
513 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
515 LASSERT(lnb[i].lnb_page);
517 lprocfs_counter_add(osd->od_stats,
518 LPROC_OSD_ZEROCOPY_IO, 1);
523 off_in_block += plen;
528 if (off_in_block == 0 && len < bs &&
529 off + len >= obj->oo_attr.la_size)
530 lprocfs_counter_add(osd->od_stats,
531 LPROC_OSD_TAIL_IO, 1);
533 /* can't use zerocopy, allocate temp. buffers */
534 poff = off & (PAGE_SIZE - 1);
535 while (sz_in_block > 0) {
536 plen = min_t(int, poff + sz_in_block,
540 if (unlikely(npages >= maxlnb))
541 GOTO(out_err, rc = -EOVERFLOW);
543 lnb[i].lnb_file_offset = off;
544 lnb[i].lnb_page_offset = poff;
547 lnb[i].lnb_len = plen;
549 lnb[i].lnb_data = NULL;
551 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
552 if (unlikely(lnb[i].lnb_page == NULL))
553 GOTO(out_err, rc = -ENOMEM);
555 LASSERT(lnb[i].lnb_page->mapping == NULL);
556 lnb[i].lnb_page->mapping = (void *)obj;
558 atomic_inc(&osd->od_zerocopy_alloc);
559 lprocfs_counter_add(osd->od_stats,
560 LPROC_OSD_COPY_IO, 1);
574 osd_bufs_put(env, &obj->oo_dt, lnb, npages);
578 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
579 loff_t offset, ssize_t len, struct niobuf_local *lnb,
580 int maxlnb, enum dt_bufs_type rw)
582 struct osd_object *obj = osd_dt_obj(dt);
585 LASSERT(dt_object_exists(dt));
588 if (rw & DT_BUFS_TYPE_WRITE)
589 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
591 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
596 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
597 struct niobuf_local *lnb, int npages)
599 struct osd_object *obj = osd_dt_obj(dt);
601 LASSERT(dt_object_exists(dt));
607 static inline uint64_t osd_roundup2blocksz(uint64_t size,
613 size += offset % blksz;
615 if (likely(is_power_of_2(blksz)))
616 return round_up(size, blksz);
618 return DIV_ROUND_UP_ULL(size, blksz) * blksz;
621 static int osd_declare_write_commit(const struct lu_env *env,
622 struct dt_object *dt,
623 struct niobuf_local *lnb, int npages,
626 struct osd_object *obj = osd_dt_obj(dt);
627 struct osd_device *osd = osd_obj2dev(obj);
628 struct osd_thandle *oh;
631 uint32_t blksz = obj->oo_dn->dn_datablksz;
635 struct page *last_page = NULL;
636 unsigned long discont_pages = 0;
637 enum osd_quota_local_flags local_flags = 0;
638 enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
641 LASSERT(dt_object_exists(dt));
647 oh = container_of(th, struct osd_thandle, ot_super);
649 for (i = 0; i < npages; i++) {
650 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
652 last_page = lnb[i].lnb_page;
654 /* ENOSPC, network RPC error, etc.
655 * We don't want to book space for pages which will be
656 * skipped in osd_write_commit(). Hence we skip pages
657 * with lnb_rc != 0 here too */
659 /* ignore quota for the whole request if any page is from
660 * client cache or written by root.
662 * XXX once we drop the 1.8 client support, the checking
663 * for whether page is from cache can be simplified as:
664 * !(lnb[i].flags & OBD_BRW_SYNC)
666 * XXX we could handle this on per-lnb basis as done by
668 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
669 (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
671 declare_flags |= OSD_QID_FORCE;
674 /* first valid lnb */
675 offset = lnb[i].lnb_file_offset;
676 size = lnb[i].lnb_len;
679 if (offset + size == lnb[i].lnb_file_offset) {
680 /* this lnb is contiguous to the previous one */
681 size += lnb[i].lnb_len;
685 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
686 obj->oo_dn, offset, size);
687 /* Estimating space to be consumed by a write is rather
688 * complicated with ZFS. As a consequence, we don't account for
689 * indirect blocks and just use as a rough estimate the worse
690 * case where the old space is being held by a snapshot. Quota
691 * overrun will be adjusted once the operation is committed, if
693 space += osd_roundup2blocksz(size, offset, blksz);
695 offset = lnb[i].lnb_file_offset;
696 size = lnb[i].lnb_len;
700 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
702 space += osd_roundup2blocksz(size, offset, blksz);
705 /* backend zfs filesystem might be configured to store multiple data
707 space *= osd->od_os->os_copies;
709 CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
712 record_start_io(osd, WRITE, discont_pages);
714 /* acquire quota space if needed */
715 rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
716 obj->oo_attr.la_gid, obj->oo_attr.la_projid,
717 space, oh, &local_flags, declare_flags);
719 if (!synced && rc == -EDQUOT &&
720 (local_flags & QUOTA_FL_SYNC) != 0) {
721 dt_sync(env, th->th_dev);
723 CDEBUG(D_QUOTA, "retry after sync\n");
728 /* we need only to store the overquota flags in the first lnb for
729 * now, once we support multiple objects BRW, this code needs be
731 if (local_flags & QUOTA_FL_OVER_USRQUOTA)
732 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
733 if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
734 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
735 #ifdef ZFS_PROJINHERIT
736 if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
737 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
744 * Policy to grow ZFS block size by write pattern.
745 * For sequential write, it grows block size gradually until it reaches the
746 * maximum blocksize the dataset can support. Otherwise, it will pick a
747 * a block size by the writing region of this I/O.
749 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
750 uint64_t start, uint64_t end)
752 struct osd_device *osd = osd_obj2dev(obj);
753 dnode_t *dn = obj->oo_dn;
759 if (dn->dn_maxblkid > 0) /* can't change block size */
762 if (dn->dn_datablksz >= osd->od_max_blksz)
765 down_write(&obj->oo_guard);
767 blksz = dn->dn_datablksz;
768 if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
769 GOTO(out_unlock, rc);
771 /* now ZFS can support up to 16MB block size, and if the write
772 * is sequential, it just increases the block size gradually */
773 if (start <= blksz) { /* sequential */
774 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
775 } else { /* sparse, pick a block size by write region */
776 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
780 if (!is_power_of_2(blksz))
781 blksz = size_roundup_power2(blksz);
783 if (blksz > dn->dn_datablksz) {
784 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
785 blksz, 0, oh->ot_tx);
786 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
788 CDEBUG(D_INODE, "object "DFID": change block size"
789 "%u -> %u error rc = %d\n",
790 PFID(lu_object_fid(&obj->oo_dt.do_lu)),
791 dn->dn_datablksz, blksz, rc);
795 up_write(&obj->oo_guard);
800 static void osd_evict_dbufs_after_write(struct osd_object *obj,
801 loff_t off, ssize_t len)
806 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
807 TRUE, osd_0copy_tag, &numbufs, &dbp);
811 for (i = 0; i < numbufs; i++)
812 dbuf_set_pending_evict(dbp[i]);
814 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
817 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
818 struct niobuf_local *lnb, int npages,
819 struct thandle *th, __u64 user_size)
821 struct osd_object *obj = osd_dt_obj(dt);
822 struct osd_device *osd = osd_obj2dev(obj);
823 struct osd_thandle *oh;
824 uint64_t new_size = 0;
825 int i, abufsz, rc = 0, drop_cache = 0;
826 unsigned long iosize = 0;
829 LASSERT(dt_object_exists(dt));
833 oh = container_of(th, struct osd_thandle, ot_super);
835 /* adjust block size. Assume the buffers are sorted. */
836 (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
837 lnb[npages - 1].lnb_file_offset +
838 lnb[npages - 1].lnb_len);
840 if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
841 lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
842 osd->od_readcache_max_filesize)
845 if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
848 /* if la_size is already bigger than specified user_size,
851 if (obj->oo_attr.la_size > user_size)
854 /* LU-8791: take oo_guard to avoid the deadlock that changing block
855 * size and assigning arcbuf take place at the same time.
859 * -> osd_grow_blocksize() with osd_object::oo_guard held
860 * -> dmu_object_set_blocksize()
861 * -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
864 * -> dmu_buf_will_dirty()
866 * -> wait for the dbuf state to change
869 * -> dmu_assign_arcbuf()
870 * -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
872 * -> try to hold the read lock of dnode_t::dn_struct_rwlock
874 * By taking the read lock, it can avoid thread 2 to enter into the
875 * critical section of assigning the arcbuf, while thread 1 is
876 * changing the block size.
878 down_read(&obj->oo_guard);
879 for (i = 0; i < npages; i++) {
880 CDEBUG(D_INODE, "write %u bytes at %u\n",
881 (unsigned) lnb[i].lnb_len,
882 (unsigned) lnb[i].lnb_file_offset);
885 /* ENOSPC, network RPC error, etc.
886 * Unlike ldiskfs, zfs allocates new blocks on rewrite,
887 * so we skip this page if lnb_rc is set to -ENOSPC */
888 CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
889 PFID(lu_object_fid(&dt->do_lu)), i,
894 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
895 new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
896 if (lnb[i].lnb_page == NULL)
899 if (lnb[i].lnb_page->mapping == (void *)obj) {
900 osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
901 lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
902 lnb[i].lnb_page_offset, oh->ot_tx);
903 kunmap(lnb[i].lnb_page);
904 iosize += lnb[i].lnb_len;
905 abufsz = lnb[i].lnb_len; /* to drop cache below */
906 } else if (lnb[i].lnb_data) {
908 LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
909 /* buffer loaned for zerocopy, try to use it.
910 * notice that dmu_assign_arcbuf() is smart
911 * enough to recognize changed blocksize
912 * in this case it fallbacks to dmu_write() */
913 abufsz = arc_buf_size(lnb[i].lnb_data);
914 LASSERT(abufsz & PAGE_MASK);
915 apages = abufsz >> PAGE_SHIFT;
916 LASSERT(i + apages <= npages);
917 /* these references to pages must be invalidated
918 * to prevent access in osd_bufs_put() */
919 for (j = 0; j < apages; j++)
920 lnb[i + j].lnb_page = NULL;
921 dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
922 lnb[i].lnb_file_offset,
923 lnb[i].lnb_data, oh->ot_tx);
924 /* drop the reference, otherwise osd_put_bufs()
925 * will be releasing it - bad! */
926 lnb[i].lnb_data = NULL;
927 atomic_dec(&osd->od_zerocopy_loan);
930 /* we don't want to deal with cache if nothing
931 * has been send to ZFS at this step */
938 /* we have to mark dbufs for eviction here because
939 * dmu_assign_arcbuf() may create a new dbuf for
941 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
944 up_read(&obj->oo_guard);
946 if (unlikely(new_size == 0)) {
947 /* no pages to write, no transno is needed */
949 /* it is important to return 0 even when all lnb_rc == -ENOSPC
950 * since ofd_commitrw_write() retries several times on ENOSPC */
951 record_end_io(osd, WRITE, 0, 0, 0);
955 /* if file has grown, take user_size into account */
956 if (user_size && new_size > user_size)
957 new_size = user_size;
958 write_lock(&obj->oo_attr_lock);
959 if (obj->oo_attr.la_size < new_size) {
960 obj->oo_attr.la_size = new_size;
961 write_unlock(&obj->oo_attr_lock);
962 /* osd_object_sa_update() will be copying directly from
963 * oo_attr into dbuf. any update within a single txg will copy
965 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
966 &obj->oo_attr.la_size, 8, oh);
968 write_unlock(&obj->oo_attr_lock);
971 record_end_io(osd, WRITE, 0, iosize, npages);
976 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
977 struct niobuf_local *lnb, int npages)
979 struct osd_object *obj = osd_dt_obj(dt);
983 LASSERT(dt_object_exists(dt));
986 read_lock(&obj->oo_attr_lock);
987 eof = obj->oo_attr.la_size;
988 read_unlock(&obj->oo_attr_lock);
990 for (i = 0; i < npages; i++) {
991 if (unlikely(lnb[i].lnb_rc < 0))
994 lnb[i].lnb_rc = lnb[i].lnb_len;
996 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
997 /* send complete pages all the time */
998 if (eof <= lnb[i].lnb_file_offset)
1001 /* all subsequent rc should be 0 */
1002 while (++i < npages)
1012 * Punch/truncate an object
1014 * IN: db - dmu_buf of the object to free data in.
1015 * off - start of section to free.
1016 * len - length of section to free (DMU_OBJECT_END => to EOF).
1018 * RETURN: 0 if success
1019 * error code if failure
1021 * The transaction passed to this routine must have
1022 * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1023 * called and then assigned to a transaction group.
1025 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1026 dmu_tx_t *tx, uint64_t off, uint64_t len)
1028 dnode_t *dn = obj->oo_dn;
1029 uint64_t size = obj->oo_attr.la_size;
1032 /* Assert that the transaction has been assigned to a
1033 transaction group. */
1034 LASSERT(tx->tx_txg != 0);
1036 * Nothing to do if file already at desired length.
1038 if (len == DMU_OBJECT_END && size == off)
1041 /* if object holds encrypted content, we need to make sure we truncate
1042 * on an encryption unit boundary, or subsequent reads will get
1045 if (len != DMU_OBJECT_END)
1046 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1047 (off & ~LUSTRE_ENCRYPTION_MASK);
1048 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1049 off & ~LUSTRE_ENCRYPTION_MASK)
1050 off = (off & LUSTRE_ENCRYPTION_MASK) +
1051 LUSTRE_ENCRYPTION_UNIT_SIZE;
1054 /* XXX: dnode_free_range() can be used to save on dnode lookup */
1056 dmu_free_range(os, dn->dn_object, off, len, tx);
1061 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1062 __u64 start, __u64 end, struct thandle *th)
1064 struct osd_object *obj = osd_dt_obj(dt);
1065 struct osd_device *osd = osd_obj2dev(obj);
1066 struct osd_thandle *oh;
1071 LASSERT(dt_object_exists(dt));
1072 LASSERT(osd_invariant(obj));
1074 LASSERT(th != NULL);
1075 oh = container_of(th, struct osd_thandle, ot_super);
1077 write_lock(&obj->oo_attr_lock);
1079 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1080 len = DMU_OBJECT_END;
1083 write_unlock(&obj->oo_attr_lock);
1085 rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1088 if (len == DMU_OBJECT_END) {
1089 write_lock(&obj->oo_attr_lock);
1090 obj->oo_attr.la_size = start;
1091 write_unlock(&obj->oo_attr_lock);
1092 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1093 &obj->oo_attr.la_size, 8, oh);
1098 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1099 __u64 start, __u64 end, struct thandle *handle)
1101 struct osd_object *obj = osd_dt_obj(dt);
1102 struct osd_device *osd = osd_obj2dev(obj);
1103 struct osd_thandle *oh;
1107 oh = container_of(handle, struct osd_thandle, ot_super);
1109 read_lock(&obj->oo_attr_lock);
1110 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1111 len = DMU_OBJECT_END;
1115 /* declare we'll free some blocks ... */
1116 /* if object holds encrypted content, we need to make sure we truncate
1117 * on an encryption unit boundary, or subsequent reads will get
1120 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1121 start & ~LUSTRE_ENCRYPTION_MASK)
1122 start = (start & LUSTRE_ENCRYPTION_MASK) +
1123 LUSTRE_ENCRYPTION_UNIT_SIZE;
1124 if (start < obj->oo_attr.la_size) {
1125 read_unlock(&obj->oo_attr_lock);
1126 dmu_tx_mark_netfree(oh->ot_tx);
1127 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1129 read_unlock(&obj->oo_attr_lock);
1132 RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1133 obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1134 0, oh, NULL, OSD_QID_BLK));
1137 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1138 __u64 start, __u64 end, enum lu_ladvise_type advice)
1152 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1153 __u64 start, __u64 end, int mode, struct thandle *th)
1155 int rc = -EOPNOTSUPP;
1159 * space preallocation is not supported for ZFS
1160 * Returns -EOPNOTSUPP for now
1165 static int osd_declare_fallocate(const struct lu_env *env,
1166 struct dt_object *dt, __u64 start, __u64 end,
1167 int mode, struct thandle *th)
1169 int rc = -EOPNOTSUPP;
1173 * space preallocation is not supported for ZFS
1174 * Returns -EOPNOTSUPP for now
1179 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1180 loff_t offset, int whence)
1182 struct osd_object *obj = osd_dt_obj(dt);
1183 uint64_t size = obj->oo_attr.la_size;
1184 uint64_t result = offset;
1186 boolean_t hole = whence == SEEK_HOLE;
1190 LASSERT(dt_object_exists(dt));
1191 LASSERT(osd_invariant(obj));
1192 LASSERT(offset >= 0);
1194 /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
1195 * hole. LOV to decide after all if that real hole or not.
1198 RETURN(hole ? offset : -ENXIO);
1200 rc = osd_dmu_offset_next(osd_obj2dev(obj)->od_os,
1201 obj->oo_dn->dn_object, hole, &result);
1205 /* file was dirty, so fall back to using generic logic:
1206 * For HOLE return file size, for DATA the result is set
1207 * already to the 'offset' parameter value.
1209 if (rc == EBUSY && hole)
1212 /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
1213 * result as end of the last block instead of logical EOF which we need
1221 const struct dt_body_operations osd_body_ops = {
1222 .dbo_read = osd_read,
1223 .dbo_declare_write = osd_declare_write,
1224 .dbo_write = osd_write,
1225 .dbo_bufs_get = osd_bufs_get,
1226 .dbo_bufs_put = osd_bufs_put,
1227 .dbo_write_prep = osd_write_prep,
1228 .dbo_declare_write_commit = osd_declare_write_commit,
1229 .dbo_write_commit = osd_write_commit,
1230 .dbo_read_prep = osd_read_prep,
1231 .dbo_declare_punch = osd_declare_punch,
1232 .dbo_punch = osd_punch,
1233 .dbo_ladvise = osd_ladvise,
1234 .dbo_declare_fallocate = osd_declare_fallocate,
1235 .dbo_fallocate = osd_fallocate,
1236 .dbo_lseek = osd_lseek,
1239 const struct dt_body_operations osd_body_scrub_ops = {
1240 .dbo_read = osd_read_no_record,
1241 .dbo_declare_write = osd_declare_write,
1242 .dbo_write = osd_write,