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.
31 * Copyright (c) 2011, 2012 Whamcloud, Inc.
32 * Use is subject to license terms.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/osd-zfs/osd_io.c
40 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
41 * Author: Mike Pershin <tappro@whamcloud.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_OSD
49 #include <lustre_ver.h>
50 #include <libcfs/libcfs.h>
51 #include <lustre_fsfilt.h>
52 #include <obd_support.h>
53 #include <lustre_net.h>
55 #include <obd_class.h>
56 #include <lustre_disk.h>
57 #include <lustre_fid.h>
59 #include "osd_internal.h"
61 #include <sys/dnode.h>
66 #include <sys/spa_impl.h>
67 #include <sys/zfs_znode.h>
68 #include <sys/dmu_tx.h>
69 #include <sys/dmu_objset.h>
70 #include <sys/dsl_prop.h>
71 #include <sys/sa_impl.h>
74 static char *osd_zerocopy_tag = "zerocopy";
76 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
77 struct lu_buf *buf, loff_t *pos,
78 struct lustre_capa *capa)
80 struct osd_object *obj = osd_dt_obj(dt);
81 struct osd_device *osd = osd_obj2dev(obj);
83 int size = buf->lb_len;
86 LASSERT(dt_object_exists(dt));
89 cfs_read_lock(&obj->oo_attr_lock);
90 old_size = obj->oo_attr.la_size;
91 cfs_read_unlock(&obj->oo_attr_lock);
93 if (*pos + size > old_size) {
97 size = old_size - *pos;
100 rc = -dmu_read(osd->od_objset.os, obj->oo_db->db_object, *pos, size,
101 buf->lb_buf, DMU_READ_PREFETCH);
106 /* XXX: workaround for bug in HEAD: fsfilt_ldiskfs_read() returns
107 * requested number of bytes, not actually read ones */
108 if (S_ISLNK(obj->oo_dt.do_lu.lo_header->loh_attr))
114 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
115 const loff_t size, loff_t pos,
118 struct osd_object *obj = osd_dt_obj(dt);
119 struct osd_thandle *oh;
123 oh = container_of0(th, struct osd_thandle, ot_super);
125 /* in some cases declare can race with creation (e.g. llog)
126 * and we need to wait till object is initialized. notice
127 * LOHA_EXISTs is supposed to be the last step in the
130 /* declare possible size change. notice we can't check
131 * current size here as another thread can change it */
133 if (dt_object_exists(dt)) {
135 oid = obj->oo_db->db_object;
137 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
139 oid = DMU_NEW_OBJECT;
140 dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
143 dmu_tx_hold_write(oh->ot_tx, oid, pos, size);
148 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
149 const struct lu_buf *buf, loff_t *pos,
150 struct thandle *th, struct lustre_capa *capa,
153 struct osd_object *obj = osd_dt_obj(dt);
154 struct osd_device *osd = osd_obj2dev(obj);
155 udmu_objset_t *uos = &osd->od_objset;
156 struct osd_thandle *oh;
157 uint64_t offset = *pos;
161 LASSERT(dt_object_exists(dt));
165 oh = container_of0(th, struct osd_thandle, ot_super);
167 dmu_write(osd->od_objset.os, obj->oo_db->db_object, offset,
168 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
169 cfs_write_lock(&obj->oo_attr_lock);
170 if (obj->oo_attr.la_size < offset + buf->lb_len) {
171 obj->oo_attr.la_size = offset + buf->lb_len;
172 cfs_write_unlock(&obj->oo_attr_lock);
173 /* osd_object_sa_update() will be copying directly from oo_attr
174 * into dbuf. any update within a single txg will copy the
176 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
177 &obj->oo_attr.la_size, 8, oh);
181 cfs_write_unlock(&obj->oo_attr_lock);
192 * XXX: for the moment I don't want to use lnb_flags for osd-internal
193 * purposes as it's not very well defined ...
194 * instead I use the lowest bit of the address so that:
195 * arc buffer: .lnb_obj = abuf (arc we loan for write)
196 * dbuf buffer: .lnb_obj = dbuf | 1 (dbuf we get for read)
197 * copy buffer: .lnb_page->mapping = obj (page we allocate for write)
201 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
202 struct niobuf_local *lnb, int npages)
204 struct osd_object *obj = osd_dt_obj(dt);
205 struct osd_device *osd = osd_obj2dev(obj);
209 LASSERT(dt_object_exists(dt));
212 for (i = 0; i < npages; i++) {
213 if (lnb[i].page == NULL)
215 if (lnb[i].page->mapping == (void *)obj) {
216 /* this is anonymous page allocated for copy-write */
217 lnb[i].page->mapping = NULL;
218 __free_page(lnb[i].page);
219 cfs_atomic_dec(&osd->od_zerocopy_alloc);
221 /* see comment in osd_bufs_get_read() */
222 ptr = (unsigned long)lnb[i].dentry;
225 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
226 cfs_atomic_dec(&osd->od_zerocopy_pin);
227 } else if (lnb[i].dentry != NULL) {
228 dmu_return_arcbuf((void *)lnb[i].dentry);
229 cfs_atomic_dec(&osd->od_zerocopy_loan);
233 lnb[i].dentry = NULL;
239 static struct page *kmem_to_page(void *addr)
244 page = vmalloc_to_page(addr);
246 page = virt_to_page(addr);
251 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
252 loff_t off, ssize_t len, struct niobuf_local *lnb)
254 struct osd_device *osd = osd_obj2dev(obj);
256 int rc, i, numbufs, npages = 0;
259 /* grab buffers for read:
260 * OSD API let us to grab buffers first, then initiate IO(s)
261 * so that all required IOs will be done in parallel, but at the
262 * moment DMU doesn't provide us with a method to grab buffers.
263 * If we discover this is a vital for good performance we
264 * can get own replacement for dmu_buf_hold_array_by_bonus().
267 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
268 osd_zerocopy_tag, &numbufs,
272 for (i = 0; i < numbufs; i++) {
273 int bufoff, tocpy, thispage;
278 cfs_atomic_inc(&osd->od_zerocopy_pin);
280 bufoff = off - dbp[i]->db_offset;
281 tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
283 /* kind of trick to differentiate dbuf vs. arcbuf */
284 LASSERT(((unsigned long)dbp[i] & 1) == 0);
285 dbf = (void *) ((unsigned long)dbp[i] | 1);
288 thispage = CFS_PAGE_SIZE;
289 thispage -= bufoff & (CFS_PAGE_SIZE - 1);
290 thispage = min(tocpy, thispage);
293 lnb->lnb_file_offset = off;
294 lnb->offset = bufoff & ~CFS_PAGE_MASK;
296 lnb->page = kmem_to_page(dbp[i]->db_data +
298 /* mark just a single slot: we need this
299 * reference to dbuf to be release once */
312 /* steal dbuf so dmu_buf_rele_array() cant release it */
316 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
322 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
323 loff_t off, ssize_t len, struct niobuf_local *lnb)
325 struct osd_device *osd = osd_obj2dev(obj);
326 int plen, off_in_block, sz_in_block;
327 int i = 0, npages = 0;
333 dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
336 * currently only full blocks are subject to zerocopy approach:
337 * so that we're sure nobody is trying to update the same block
340 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
342 off_in_block = off & (bs - 1);
343 sz_in_block = min_t(int, bs - off_in_block, len);
345 if (sz_in_block == bs) {
346 /* full block, try to use zerocopy */
348 abuf = dmu_request_arcbuf(obj->oo_db, bs);
349 if (unlikely(abuf == NULL))
350 GOTO(out_err, -ENOMEM);
352 cfs_atomic_inc(&osd->od_zerocopy_loan);
354 /* go over pages arcbuf contains, put them as
355 * local niobufs for ptlrpc's bulks */
356 while (sz_in_block > 0) {
357 plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
359 lnb[i].lnb_file_offset = off;
363 if (sz_in_block == bs)
364 lnb[i].dentry = (void *)abuf;
366 lnb[i].dentry = NULL;
368 /* this one is not supposed to fail */
369 lnb[i].page = kmem_to_page(abuf->b_data +
371 LASSERT(lnb[i].page);
373 lprocfs_counter_add(osd->od_stats,
374 LPROC_OSD_ZEROCOPY_IO, 1);
379 off_in_block += plen;
384 if (off_in_block == 0 && len < bs &&
385 off + len >= obj->oo_attr.la_size)
386 lprocfs_counter_add(osd->od_stats,
387 LPROC_OSD_TAIL_IO, 1);
389 /* can't use zerocopy, allocate temp. buffers */
390 while (sz_in_block > 0) {
391 plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
393 lnb[i].lnb_file_offset = off;
397 lnb[i].dentry = NULL;
399 lnb[i].page = alloc_page(OSD_GFP_IO);
400 if (unlikely(lnb[i].page == NULL))
401 GOTO(out_err, -ENOMEM);
403 LASSERT(lnb[i].page->mapping == NULL);
404 lnb[i].page->mapping = (void *)obj;
406 cfs_atomic_inc(&osd->od_zerocopy_alloc);
407 lprocfs_counter_add(osd->od_stats,
408 LPROC_OSD_COPY_IO, 1);
422 osd_bufs_put(env, &obj->oo_dt, lnb, npages);
426 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
427 loff_t offset, ssize_t len, struct niobuf_local *lnb,
428 int rw, struct lustre_capa *capa)
430 struct osd_object *obj = osd_dt_obj(dt);
433 LASSERT(dt_object_exists(dt));
437 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
439 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
444 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
445 struct niobuf_local *lnb, int npages)
447 struct osd_object *obj = osd_dt_obj(dt);
449 LASSERT(dt_object_exists(dt));
455 static int osd_declare_write_commit(const struct lu_env *env,
456 struct dt_object *dt,
457 struct niobuf_local *lnb, int npages,
460 struct osd_object *obj = osd_dt_obj(dt);
461 struct osd_thandle *oh;
467 LASSERT(dt_object_exists(dt));
473 oh = container_of0(th, struct osd_thandle, ot_super);
475 for (i = 0; i < npages; i++) {
477 /* ENOSPC, network RPC error, etc.
478 * We don't want to book space for pages which will be
479 * skipped in osd_write_commit(). Hence we skip pages
480 * with lnb_rc != 0 here too */
483 /* first valid lnb */
484 offset = lnb[i].lnb_file_offset;
488 if (offset + size == lnb[i].lnb_file_offset) {
489 /* this lnb is contiguous to the previous one */
494 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
496 offset = lnb->lnb_file_offset;
501 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
503 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
505 oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
510 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
511 struct niobuf_local *lnb, int npages,
514 struct osd_object *obj = osd_dt_obj(dt);
515 struct osd_device *osd = osd_obj2dev(obj);
516 udmu_objset_t *uos = &osd->od_objset;
517 struct osd_thandle *oh;
518 uint64_t new_size = 0;
522 LASSERT(dt_object_exists(dt));
526 oh = container_of0(th, struct osd_thandle, ot_super);
528 for (i = 0; i < npages; i++) {
529 CDEBUG(D_INODE, "write %u bytes at %u\n",
530 (unsigned) lnb[i].len,
531 (unsigned) lnb[i].lnb_file_offset);
534 /* ENOSPC, network RPC error, etc.
535 * Unlike ldiskfs, zfs allocates new blocks on rewrite,
536 * so we skip this page if lnb_rc is set to -ENOSPC */
537 CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
538 PFID(lu_object_fid(&dt->do_lu)), i,
543 if (lnb[i].page->mapping == (void *)obj) {
544 dmu_write(osd->od_objset.os, obj->oo_db->db_object,
545 lnb[i].lnb_file_offset, lnb[i].len,
546 kmap(lnb[i].page), oh->ot_tx);
548 } else if (lnb[i].dentry) {
549 LASSERT(((unsigned long)lnb[i].dentry & 1) == 0);
550 /* buffer loaned for zerocopy, try to use it.
551 * notice that dmu_assign_arcbuf() is smart
552 * enough to recognize changed blocksize
553 * in this case it fallbacks to dmu_write() */
554 dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
555 (void *)lnb[i].dentry, oh->ot_tx);
556 /* drop the reference, otherwise osd_put_bufs()
557 * will be releasing it - bad! */
558 lnb[i].dentry = NULL;
559 cfs_atomic_dec(&osd->od_zerocopy_loan);
562 if (new_size < lnb[i].lnb_file_offset + lnb[i].len)
563 new_size = lnb[i].lnb_file_offset + lnb[i].len;
566 if (unlikely(new_size == 0)) {
567 /* no pages to write, no transno is needed */
569 /* it is important to return 0 even when all lnb_rc == -ENOSPC
570 * since ofd_commitrw_write() retries several times on ENOSPC */
574 cfs_write_lock(&obj->oo_attr_lock);
575 if (obj->oo_attr.la_size < new_size) {
576 obj->oo_attr.la_size = new_size;
577 cfs_write_unlock(&obj->oo_attr_lock);
578 /* osd_object_sa_update() will be copying directly from
579 * oo_attr into dbuf. any update within a single txg will copy
581 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
582 &obj->oo_attr.la_size, 8, oh);
584 cfs_write_unlock(&obj->oo_attr_lock);
590 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
591 struct niobuf_local *lnb, int npages)
593 struct osd_object *obj = osd_dt_obj(dt);
598 LASSERT(dt_object_exists(dt));
601 for (i = 0; i < npages; i++) {
602 buf.lb_buf = kmap(lnb[i].page);
603 buf.lb_len = lnb[i].len;
604 offset = lnb[i].lnb_file_offset;
606 CDEBUG(D_OTHER, "read %u bytes at %u\n",
607 (unsigned) lnb[i].len,
608 (unsigned) lnb[i].lnb_file_offset);
609 lnb[i].rc = osd_read(env, dt, &buf, &offset, NULL);
612 if (lnb[i].rc < buf.lb_len) {
613 /* all subsequent rc should be 0 */
624 * Punch/truncate an object
626 * IN: db - dmu_buf of the object to free data in.
627 * off - start of section to free.
628 * len - length of section to free (DMU_OBJECT_END => to EOF).
630 * RETURN: 0 if success
631 * error code if failure
633 * The transaction passed to this routine must have
634 * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
635 * called and then assigned to a transaction group.
637 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
638 uint64_t size, uint64_t off, uint64_t len)
642 /* Assert that the transaction has been assigned to a
643 transaction group. */
644 LASSERT(tx->tx_txg != 0);
646 * Nothing to do if file already at desired length.
648 if (len == DMU_OBJECT_END && size == off)
652 rc = -dmu_free_range(os, db->db_object, off, len, tx);
657 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
658 __u64 start, __u64 end, struct thandle *th,
659 struct lustre_capa *capa)
661 struct osd_object *obj = osd_dt_obj(dt);
662 struct osd_device *osd = osd_obj2dev(obj);
663 udmu_objset_t *uos = &osd->od_objset;
664 struct osd_thandle *oh;
669 LASSERT(dt_object_exists(dt));
670 LASSERT(osd_invariant(obj));
673 oh = container_of0(th, struct osd_thandle, ot_super);
675 cfs_write_lock(&obj->oo_attr_lock);
677 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
678 len = DMU_OBJECT_END;
681 cfs_write_unlock(&obj->oo_attr_lock);
683 rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx,
684 obj->oo_attr.la_size, start, len);
686 if (len == DMU_OBJECT_END) {
687 cfs_write_lock(&obj->oo_attr_lock);
688 obj->oo_attr.la_size = start;
689 cfs_write_unlock(&obj->oo_attr_lock);
690 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
691 &obj->oo_attr.la_size, 8, oh);
696 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
697 __u64 start, __u64 end, struct thandle *handle)
699 struct osd_object *obj = osd_dt_obj(dt);
700 struct osd_thandle *oh;
704 oh = container_of0(handle, struct osd_thandle, ot_super);
706 cfs_read_lock(&obj->oo_attr_lock);
707 if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
708 len = DMU_OBJECT_END;
712 /* declare we'll free some blocks ... */
713 if (start < obj->oo_attr.la_size) {
714 cfs_read_unlock(&obj->oo_attr_lock);
715 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
717 cfs_read_unlock(&obj->oo_attr_lock);
720 /* ... and we'll modify size attribute */
721 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
727 struct dt_body_operations osd_body_ops = {
728 .dbo_read = osd_read,
729 .dbo_declare_write = osd_declare_write,
730 .dbo_write = osd_write,
731 .dbo_bufs_get = osd_bufs_get,
732 .dbo_bufs_put = osd_bufs_put,
733 .dbo_write_prep = osd_write_prep,
734 .dbo_declare_write_commit = osd_declare_write_commit,
735 .dbo_write_commit = osd_write_commit,
736 .dbo_read_prep = osd_read_prep,
737 .do_declare_punch = osd_declare_punch,
738 .do_punch = osd_punch,