Whamcloud - gitweb
LU-11673 tests: add space before ']' in test-framework
[fs/lustre-release.git] / lustre / osd-zfs / osd_io.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/osd-zfs/osd_io.c
33  *
34  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
35  * Author: Mike Pershin <tappro@whamcloud.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_OSD
39
40 #include <libcfs/libcfs.h>
41 #include <obd_support.h>
42 #include <lustre_net.h>
43 #include <obd.h>
44 #include <obd_class.h>
45 #include <lustre_disk.h>
46 #include <lustre_fid.h>
47 #include <lustre_quota.h>
48
49 #include "osd_internal.h"
50
51 #include <sys/dnode.h>
52 #include <sys/dbuf.h>
53 #include <sys/spa.h>
54 #include <sys/stat.h>
55 #include <sys/zap.h>
56 #include <sys/spa_impl.h>
57 #include <sys/zfs_znode.h>
58 #include <sys/dmu_tx.h>
59 #include <sys/dmu_objset.h>
60 #include <sys/dsl_prop.h>
61 #include <sys/sa_impl.h>
62 #include <sys/txg.h>
63
64 static char *osd_0copy_tag = "zerocopy";
65
66
67 static void dbuf_set_pending_evict(dmu_buf_t *db)
68 {
69         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
70         dbi->db_pending_evict = TRUE;
71 }
72
73 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
74 {
75         struct obd_histogram *h = osd->od_brw_stats.hist;
76
77         if (rw == READ) {
78                 atomic_inc(&osd->od_r_in_flight);
79                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
80                                  atomic_read(&osd->od_r_in_flight));
81                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
82
83         } else {
84                 atomic_inc(&osd->od_w_in_flight);
85                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
86                                  atomic_read(&osd->od_w_in_flight));
87                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
88
89         }
90 }
91
92 static void record_end_io(struct osd_device *osd, int rw,
93                           unsigned long elapsed, int disksize, int npages)
94 {
95         struct obd_histogram *h = osd->od_brw_stats.hist;
96
97         if (rw == READ)
98                 atomic_dec(&osd->od_r_in_flight);
99         else
100                 atomic_dec(&osd->od_w_in_flight);
101
102         lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
103         if (disksize > 0)
104                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
105         if (elapsed)
106                 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
107 }
108
109 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
110                           struct lu_buf *buf, loff_t *pos, size_t *size)
111 {
112         struct osd_object *obj = osd_dt_obj(dt);
113         uint64_t old_size;
114         int rc;
115
116         LASSERT(dt_object_exists(dt));
117         LASSERT(obj->oo_dn);
118
119         read_lock(&obj->oo_attr_lock);
120         old_size = obj->oo_attr.la_size;
121         read_unlock(&obj->oo_attr_lock);
122
123         if (*pos + *size > old_size) {
124                 if (old_size < *pos)
125                         return 0;
126
127                 *size = old_size - *pos;
128         }
129
130         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
131                           buf->lb_buf, DMU_READ_PREFETCH);
132         if (!rc) {
133                 rc = *size;
134                 *pos += *size;
135         }
136
137         return rc;
138 }
139
140 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
141                         struct lu_buf *buf, loff_t *pos)
142 {
143         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
144         size_t size = buf->lb_len;
145         ktime_t start;
146         s64 delta_ms;
147         int rc;
148
149         start = ktime_get();
150         record_start_io(osd, READ, 0);
151         rc = __osd_read(env, dt, buf, pos, &size);
152         delta_ms = ktime_ms_delta(ktime_get(), start);
153         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
154
155         return rc;
156 }
157
158 static inline ssize_t osd_read_no_record(const struct lu_env *env,
159                                          struct dt_object *dt,
160                                          struct lu_buf *buf, loff_t *pos)
161 {
162         size_t size = buf->lb_len;
163
164         return __osd_read(env, dt, buf, pos, &size);
165 }
166
167 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
168                                 const struct lu_buf *buf, loff_t pos,
169                                 struct thandle *th)
170 {
171         struct osd_object  *obj  = osd_dt_obj(dt);
172         struct osd_device  *osd = osd_obj2dev(obj);
173         struct osd_thandle *oh;
174         uint64_t            oid;
175         ENTRY;
176
177         oh = container_of0(th, struct osd_thandle, ot_super);
178
179         /* in some cases declare can race with creation (e.g. llog)
180          * and we need to wait till object is initialized. notice
181          * LOHA_EXISTs is supposed to be the last step in the
182          * initialization */
183
184         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
185         if (dt_object_exists(dt))
186                 oid = obj->oo_dn->dn_object;
187         else
188                 oid = DMU_NEW_OBJECT;
189
190         /* XXX: we still miss for append declaration support in ZFS
191          *      -1 means append which is used by llog mostly, llog
192          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
193         if (pos == -1)
194                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
195                             obj->oo_attr.la_size + (2 << 20));
196         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
197
198         /* dt_declare_write() is usually called for system objects, such
199          * as llog or last_rcvd files. We needn't enforce quota on those
200          * objects, so always set the lqi_space as 0. */
201         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
202                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
203                                  0, oh, NULL, OSD_QID_BLK));
204 }
205
206 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
207                         const struct lu_buf *buf, loff_t *pos,
208                          struct thandle *th)
209 {
210         struct osd_object  *obj  = osd_dt_obj(dt);
211         struct osd_device  *osd = osd_obj2dev(obj);
212         struct osd_thandle *oh;
213         uint64_t            offset = *pos;
214         int                 rc;
215
216         ENTRY;
217
218         LASSERT(dt_object_exists(dt));
219         LASSERT(obj->oo_dn);
220
221         LASSERT(th != NULL);
222         oh = container_of0(th, struct osd_thandle, ot_super);
223
224         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
225                       buf->lb_buf, oh->ot_tx);
226         write_lock(&obj->oo_attr_lock);
227         if (obj->oo_attr.la_size < offset + buf->lb_len) {
228                 obj->oo_attr.la_size = offset + buf->lb_len;
229                 write_unlock(&obj->oo_attr_lock);
230                 /* osd_object_sa_update() will be copying directly from oo_attr
231                  * into dbuf.  any update within a single txg will copy the
232                  * most actual */
233                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
234                                         &obj->oo_attr.la_size, 8, oh);
235                 if (unlikely(rc))
236                         GOTO(out, rc);
237         } else {
238                 write_unlock(&obj->oo_attr_lock);
239         }
240
241         *pos += buf->lb_len;
242         rc = buf->lb_len;
243
244 out:
245         RETURN(rc);
246 }
247
248 /*
249  * XXX: for the moment I don't want to use lnb_flags for osd-internal
250  *      purposes as it's not very well defined ...
251  *      instead I use the lowest bit of the address so that:
252  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
253  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
254  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
255  *
256  *      bzzz, to blame
257  */
258 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
259                         struct niobuf_local *lnb, int npages)
260 {
261         struct osd_object *obj  = osd_dt_obj(dt);
262         struct osd_device *osd = osd_obj2dev(obj);
263         unsigned long      ptr;
264         int                i;
265
266         LASSERT(dt_object_exists(dt));
267         LASSERT(obj->oo_dn);
268
269         for (i = 0; i < npages; i++) {
270                 if (lnb[i].lnb_page == NULL)
271                         continue;
272                 if (lnb[i].lnb_page->mapping == (void *)obj) {
273                         /* this is anonymous page allocated for copy-write */
274                         lnb[i].lnb_page->mapping = NULL;
275                         __free_page(lnb[i].lnb_page);
276                         atomic_dec(&osd->od_zerocopy_alloc);
277                 } else {
278                         /* see comment in osd_bufs_get_read() */
279                         ptr = (unsigned long)lnb[i].lnb_data;
280                         if (ptr & 1UL) {
281                                 ptr &= ~1UL;
282                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
283                                 atomic_dec(&osd->od_zerocopy_pin);
284                         } else if (lnb[i].lnb_data != NULL) {
285                                 int j, apages, abufsz;
286                                 abufsz = arc_buf_size(lnb[i].lnb_data);
287                                 apages = abufsz / PAGE_SIZE;
288                                 /* these references to pages must be invalidated
289                                  * to prevent access in osd_bufs_put() */
290                                 for (j = 0; j < apages; j++)
291                                         lnb[i + j].lnb_page = NULL;
292                                 dmu_return_arcbuf(lnb[i].lnb_data);
293                                 atomic_dec(&osd->od_zerocopy_loan);
294                         }
295                 }
296                 lnb[i].lnb_page = NULL;
297                 lnb[i].lnb_data = NULL;
298         }
299
300         return 0;
301 }
302
303 static inline struct page *kmem_to_page(void *addr)
304 {
305         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
306         if (is_vmalloc_addr(addr))
307                 return vmalloc_to_page(addr);
308         else
309                 return virt_to_page(addr);
310 }
311
312 /**
313  * Prepare buffers for read.
314  *
315  * The function maps the range described by \a off and \a len to \a lnb array.
316  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
317  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
318  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
319  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
320  * A better implementation would just return the buffers (potentially unfilled)
321  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
322  *
323  * \param[in] env       environment
324  * \param[in] obj       object
325  * \param[in] off       offset in bytes
326  * \param[in] len       the number of bytes to access
327  * \param[out] lnb      array of local niobufs pointing to the buffers with data
328  *
329  * \retval              0 for success
330  * \retval              negative error number of failure
331  */
332 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
333                              loff_t off, ssize_t len, struct niobuf_local *lnb)
334 {
335         struct osd_device *osd = osd_obj2dev(obj);
336         int rc, i, numbufs, npages = 0, drop_cache = 0;
337         ktime_t start = ktime_get();
338         dmu_buf_t **dbp;
339         s64 delta_ms;
340
341         ENTRY;
342         record_start_io(osd, READ, 0);
343
344         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
345                 drop_cache = 1;
346
347         /* grab buffers for read:
348          * OSD API let us to grab buffers first, then initiate IO(s)
349          * so that all required IOs will be done in parallel, but at the
350          * moment DMU doesn't provide us with a method to grab buffers.
351          * If we discover this is a vital for good performance we
352          * can get own replacement for dmu_buf_hold_array_by_bonus().
353          */
354         while (len > 0 &&
355                (obj->oo_dn->dn_datablkshift != 0 ||
356                 off < obj->oo_dn->dn_datablksz)) {
357                 if (obj->oo_dn->dn_datablkshift == 0 &&
358                     off + len > obj->oo_dn->dn_datablksz)
359                         len = obj->oo_dn->dn_datablksz - off;
360
361                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
362                                                   off, len, TRUE, osd_0copy_tag,
363                                                   &numbufs, &dbp);
364                 if (unlikely(rc))
365                         GOTO(err, rc);
366
367                 for (i = 0; i < numbufs; i++) {
368                         int bufoff, tocpy, thispage;
369                         void *dbf = dbp[i];
370
371                         LASSERT(len > 0);
372
373                         atomic_inc(&osd->od_zerocopy_pin);
374
375                         bufoff = off - dbp[i]->db_offset;
376                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
377
378                         /* kind of trick to differentiate dbuf vs. arcbuf */
379                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
380                         dbf = (void *) ((unsigned long)dbp[i] | 1);
381
382                         while (tocpy > 0) {
383                                 thispage = PAGE_SIZE;
384                                 thispage -= bufoff & (PAGE_SIZE - 1);
385                                 thispage = min(tocpy, thispage);
386
387                                 lnb->lnb_rc = 0;
388                                 lnb->lnb_file_offset = off;
389                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
390                                 lnb->lnb_len = thispage;
391                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
392                                                              bufoff);
393                                 /* mark just a single slot: we need this
394                                  * reference to dbuf to be released once */
395                                 lnb->lnb_data = dbf;
396                                 dbf = NULL;
397
398                                 tocpy -= thispage;
399                                 len -= thispage;
400                                 bufoff += thispage;
401                                 off += thispage;
402
403                                 npages++;
404                                 lnb++;
405                         }
406
407                         if (drop_cache)
408                                 dbuf_set_pending_evict(dbp[i]);
409
410                         /* steal dbuf so dmu_buf_rele_array() can't release
411                          * it */
412                         dbp[i] = NULL;
413                 }
414
415                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
416         }
417
418         delta_ms = ktime_ms_delta(ktime_get(), start);
419         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
420
421         RETURN(npages);
422
423 err:
424         LASSERT(rc < 0);
425         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
426         RETURN(rc);
427 }
428
429 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
430 {
431         arc_buf_t *abuf;
432
433         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
434         if (unlikely(!abuf))
435                 return ERR_PTR(-ENOMEM);
436
437 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
438         /**
439          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
440          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
441          * setting up page array for RDMA transfer. See LU-9305.
442          */
443         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
444                 dmu_return_arcbuf(abuf);
445                 return NULL;
446         }
447 #endif
448
449         return abuf;
450 }
451
452 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
453                               loff_t off, ssize_t len, struct niobuf_local *lnb)
454 {
455         struct osd_device *osd = osd_obj2dev(obj);
456         int                poff, plen, off_in_block, sz_in_block;
457         int                rc, i = 0, npages = 0;
458         dnode_t *dn = obj->oo_dn;
459         arc_buf_t *abuf;
460         uint32_t bs = dn->dn_datablksz;
461         ENTRY;
462
463         /*
464          * currently only full blocks are subject to zerocopy approach:
465          * so that we're sure nobody is trying to update the same block
466          */
467         while (len > 0) {
468                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
469
470                 off_in_block = off & (bs - 1);
471                 sz_in_block = min_t(int, bs - off_in_block, len);
472
473                 abuf = NULL;
474                 if (sz_in_block == bs) {
475                         /* full block, try to use zerocopy */
476                         abuf = osd_request_arcbuf(dn, bs);
477                         if (unlikely(IS_ERR(abuf)))
478                                 GOTO(out_err, rc = PTR_ERR(abuf));
479                 }
480
481                 if (abuf != NULL) {
482                         atomic_inc(&osd->od_zerocopy_loan);
483
484                         /* go over pages arcbuf contains, put them as
485                          * local niobufs for ptlrpc's bulks */
486                         while (sz_in_block > 0) {
487                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
488
489                                 lnb[i].lnb_file_offset = off;
490                                 lnb[i].lnb_page_offset = 0;
491                                 lnb[i].lnb_len = plen;
492                                 lnb[i].lnb_rc = 0;
493                                 if (sz_in_block == bs)
494                                         lnb[i].lnb_data = abuf;
495                                 else
496                                         lnb[i].lnb_data = NULL;
497
498                                 /* this one is not supposed to fail */
499                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
500                                                         off_in_block);
501                                 LASSERT(lnb[i].lnb_page);
502
503                                 lprocfs_counter_add(osd->od_stats,
504                                                 LPROC_OSD_ZEROCOPY_IO, 1);
505
506                                 sz_in_block -= plen;
507                                 len -= plen;
508                                 off += plen;
509                                 off_in_block += plen;
510                                 i++;
511                                 npages++;
512                         }
513                 } else {
514                         if (off_in_block == 0 && len < bs &&
515                                         off + len >= obj->oo_attr.la_size)
516                                 lprocfs_counter_add(osd->od_stats,
517                                                 LPROC_OSD_TAIL_IO, 1);
518
519                         /* can't use zerocopy, allocate temp. buffers */
520                         poff = off & (PAGE_SIZE - 1);
521                         while (sz_in_block > 0) {
522                                 plen = min_t(int, poff + sz_in_block,
523                                              PAGE_SIZE);
524                                 plen -= poff;
525
526                                 lnb[i].lnb_file_offset = off;
527                                 lnb[i].lnb_page_offset = poff;
528                                 poff = 0;
529
530                                 lnb[i].lnb_len = plen;
531                                 lnb[i].lnb_rc = 0;
532                                 lnb[i].lnb_data = NULL;
533
534                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
535                                 if (unlikely(lnb[i].lnb_page == NULL))
536                                         GOTO(out_err, rc = -ENOMEM);
537
538                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
539                                 lnb[i].lnb_page->mapping = (void *)obj;
540
541                                 atomic_inc(&osd->od_zerocopy_alloc);
542                                 lprocfs_counter_add(osd->od_stats,
543                                                 LPROC_OSD_COPY_IO, 1);
544
545                                 sz_in_block -= plen;
546                                 len -= plen;
547                                 off += plen;
548                                 i++;
549                                 npages++;
550                         }
551                 }
552         }
553
554         RETURN(npages);
555
556 out_err:
557         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
558         RETURN(rc);
559 }
560
561 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
562                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
563                         enum dt_bufs_type rw)
564 {
565         struct osd_object *obj  = osd_dt_obj(dt);
566         int                rc;
567
568         LASSERT(dt_object_exists(dt));
569         LASSERT(obj->oo_dn);
570
571         if (rw & DT_BUFS_TYPE_WRITE)
572                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
573         else
574                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
575
576         return rc;
577 }
578
579 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
580                         struct niobuf_local *lnb, int npages)
581 {
582         struct osd_object *obj = osd_dt_obj(dt);
583
584         LASSERT(dt_object_exists(dt));
585         LASSERT(obj->oo_dn);
586
587         return 0;
588 }
589
590 static inline uint64_t osd_roundup2blocksz(uint64_t size,
591                                            uint64_t offset,
592                                            uint32_t blksz)
593 {
594         LASSERT(blksz > 0);
595
596         size += offset % blksz;
597
598         if (likely(is_power_of_2(blksz)))
599                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
600
601         size += blksz - 1;
602         do_div(size, blksz);
603         return size * blksz;
604 }
605
606 static int osd_declare_write_commit(const struct lu_env *env,
607                                     struct dt_object *dt,
608                                     struct niobuf_local *lnb, int npages,
609                                     struct thandle *th)
610 {
611         struct osd_object  *obj = osd_dt_obj(dt);
612         struct osd_device  *osd = osd_obj2dev(obj);
613         struct osd_thandle *oh;
614         uint64_t            offset = 0;
615         uint32_t            size = 0;
616         uint32_t blksz = obj->oo_dn->dn_datablksz;
617         int                 i, rc, flags = 0;
618         bool synced = false;
619         long long           space = 0;
620         struct page        *last_page = NULL;
621         unsigned long       discont_pages = 0;
622         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
623         ENTRY;
624
625         LASSERT(dt_object_exists(dt));
626         LASSERT(obj->oo_dn);
627
628         LASSERT(lnb);
629         LASSERT(npages > 0);
630
631         oh = container_of0(th, struct osd_thandle, ot_super);
632
633         for (i = 0; i < npages; i++) {
634                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
635                         ++discont_pages;
636                 last_page = lnb[i].lnb_page;
637                 if (lnb[i].lnb_rc)
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 */
642                         continue;
643                 /* ignore quota for the whole request if any page is from
644                  * client cache or written by root.
645                  *
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)
649                  *
650                  * XXX we could handle this on per-lnb basis as done by
651                  * grant. */
652                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
653                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
654                     OBD_BRW_FROM_GRANT)
655                         declare_flags |= OSD_QID_FORCE;
656
657                 if (size == 0) {
658                         /* first valid lnb */
659                         offset = lnb[i].lnb_file_offset;
660                         size = lnb[i].lnb_len;
661                         continue;
662                 }
663                 if (offset + size == lnb[i].lnb_file_offset) {
664                         /* this lnb is contiguous to the previous one */
665                         size += lnb[i].lnb_len;
666                         continue;
667                 }
668
669                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
670                                   obj->oo_dn, offset, size);
671                 /* Estimating space to be consumed by a write is rather
672                  * complicated with ZFS. As a consequence, we don't account for
673                  * indirect blocks and just use as a rough estimate the worse
674                  * case where the old space is being held by a snapshot. Quota
675                  * overrun will be adjusted once the operation is committed, if
676                  * required. */
677                 space += osd_roundup2blocksz(size, offset, blksz);
678
679                 offset = lnb[i].lnb_file_offset;
680                 size = lnb[i].lnb_len;
681         }
682
683         if (size) {
684                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
685                                   offset, size);
686                 space += osd_roundup2blocksz(size, offset, blksz);
687         }
688
689         /* backend zfs filesystem might be configured to store multiple data
690          * copies */
691         space  *= osd->od_os->os_copies;
692         space   = toqb(space);
693         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
694                npages, space);
695
696         record_start_io(osd, WRITE, discont_pages);
697 retry:
698         /* acquire quota space if needed */
699         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
700                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
701                                space, oh, &flags, declare_flags);
702
703         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
704                 dt_sync(env, th->th_dev);
705                 synced = true;
706                 CDEBUG(D_QUOTA, "retry after sync\n");
707                 flags = 0;
708                 goto retry;
709         }
710
711         /* we need only to store the overquota flags in the first lnb for
712          * now, once we support multiple objects BRW, this code needs be
713          * revised. */
714         if (flags & QUOTA_FL_OVER_USRQUOTA)
715                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
716         if (flags & QUOTA_FL_OVER_GRPQUOTA)
717                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
718 #ifdef ZFS_PROJINHERIT
719         if (flags & QUOTA_FL_OVER_PRJQUOTA)
720                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
721 #endif
722
723         RETURN(rc);
724 }
725
726 /**
727  * Policy to grow ZFS block size by write pattern.
728  * For sequential write, it grows block size gradually until it reaches the
729  * maximum blocksize the dataset can support. Otherwise, it will pick a
730  * a block size by the writing region of this I/O.
731  */
732 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
733                               uint64_t start, uint64_t end)
734 {
735         struct osd_device       *osd = osd_obj2dev(obj);
736         dnode_t *dn = obj->oo_dn;
737         uint32_t                 blksz;
738         int                      rc = 0;
739
740         ENTRY;
741
742         if (dn->dn_maxblkid > 0) /* can't change block size */
743                 GOTO(out, rc);
744
745         if (dn->dn_datablksz >= osd->od_max_blksz)
746                 GOTO(out, rc);
747
748         down_write(&obj->oo_guard);
749
750         blksz = dn->dn_datablksz;
751         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
752                 GOTO(out_unlock, rc);
753
754         /* now ZFS can support up to 16MB block size, and if the write
755          * is sequential, it just increases the block size gradually */
756         if (start <= blksz) { /* sequential */
757                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
758         } else { /* sparse, pick a block size by write region */
759                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
760                                         end - start);
761         }
762
763         if (!is_power_of_2(blksz))
764                 blksz = size_roundup_power2(blksz);
765
766         if (blksz > dn->dn_datablksz) {
767                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
768                                                blksz, 0, oh->ot_tx);
769                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
770                 if (rc < 0)
771                         CDEBUG(D_INODE, "object "DFID": change block size"
772                                "%u -> %u error rc = %d\n",
773                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
774                                dn->dn_datablksz, blksz, rc);
775         }
776         EXIT;
777 out_unlock:
778         up_write(&obj->oo_guard);
779 out:
780         return rc;
781 }
782
783 static void osd_evict_dbufs_after_write(struct osd_object *obj,
784                                         loff_t off, ssize_t len)
785 {
786         dmu_buf_t **dbp;
787         int i, rc, numbufs;
788
789         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
790                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
791         if (unlikely(rc))
792                 return;
793
794         for (i = 0; i < numbufs; i++)
795                 dbuf_set_pending_evict(dbp[i]);
796
797         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
798 }
799
800 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
801                         struct niobuf_local *lnb, int npages,
802                         struct thandle *th)
803 {
804         struct osd_object  *obj  = osd_dt_obj(dt);
805         struct osd_device  *osd = osd_obj2dev(obj);
806         struct osd_thandle *oh;
807         uint64_t            new_size = 0;
808         int                 i, abufsz, rc = 0, drop_cache = 0;
809         unsigned long      iosize = 0;
810         ENTRY;
811
812         LASSERT(dt_object_exists(dt));
813         LASSERT(obj->oo_dn);
814
815         LASSERT(th != NULL);
816         oh = container_of0(th, struct osd_thandle, ot_super);
817
818         /* adjust block size. Assume the buffers are sorted. */
819         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
820                                  lnb[npages - 1].lnb_file_offset +
821                                  lnb[npages - 1].lnb_len);
822
823         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
824             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
825             osd->od_readcache_max_filesize)
826                 drop_cache = 1;
827
828         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
829                 RETURN(-ENOSPC);
830
831         /* LU-8791: take oo_guard to avoid the deadlock that changing block
832          * size and assigning arcbuf take place at the same time.
833          *
834          * Thread 1:
835          * osd_write_commit()
836          *  -> osd_grow_blocksize() with osd_object::oo_guard held
837          *   -> dmu_object_set_blocksize()
838          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
839          *       write lock held
840          *     -> dbuf_new_size()
841          *      -> dmu_buf_will_dirty()
842          *       -> dbuf_read()
843          *        -> wait for the dbuf state to change
844          * Thread 2:
845          * osd_write_commit()
846          *  -> dmu_assign_arcbuf()
847          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
848          *    -> dbuf_dirty()
849          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
850          *
851          * By taking the read lock, it can avoid thread 2 to enter into the
852          * critical section of assigning the arcbuf, while thread 1 is
853          * changing the block size.
854          */
855         down_read(&obj->oo_guard);
856         for (i = 0; i < npages; i++) {
857                 CDEBUG(D_INODE, "write %u bytes at %u\n",
858                         (unsigned) lnb[i].lnb_len,
859                         (unsigned) lnb[i].lnb_file_offset);
860
861                 if (lnb[i].lnb_rc) {
862                         /* ENOSPC, network RPC error, etc.
863                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
864                          * so we skip this page if lnb_rc is set to -ENOSPC */
865                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
866                                 PFID(lu_object_fid(&dt->do_lu)), i,
867                                 lnb[i].lnb_rc);
868                         continue;
869                 }
870
871                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
872                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
873                 if (lnb[i].lnb_page == NULL)
874                         continue;
875
876                 if (lnb[i].lnb_page->mapping == (void *)obj) {
877                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
878                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
879                                       lnb[i].lnb_page_offset, oh->ot_tx);
880                         kunmap(lnb[i].lnb_page);
881                         iosize += lnb[i].lnb_len;
882                         abufsz = lnb[i].lnb_len; /* to drop cache below */
883                 } else if (lnb[i].lnb_data) {
884                         int j, apages;
885                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
886                         /* buffer loaned for zerocopy, try to use it.
887                          * notice that dmu_assign_arcbuf() is smart
888                          * enough to recognize changed blocksize
889                          * in this case it fallbacks to dmu_write() */
890                         abufsz = arc_buf_size(lnb[i].lnb_data);
891                         LASSERT(abufsz & PAGE_MASK);
892                         apages = abufsz / PAGE_SIZE;
893                         LASSERT(i + apages <= npages);
894                         /* these references to pages must be invalidated
895                          * to prevent access in osd_bufs_put() */
896                         for (j = 0; j < apages; j++)
897                                 lnb[i + j].lnb_page = NULL;
898                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
899                                           lnb[i].lnb_file_offset,
900                                           lnb[i].lnb_data, oh->ot_tx);
901                         /* drop the reference, otherwise osd_put_bufs()
902                          * will be releasing it - bad! */
903                         lnb[i].lnb_data = NULL;
904                         atomic_dec(&osd->od_zerocopy_loan);
905                         iosize += abufsz;
906                 } else {
907                         /* we don't want to deal with cache if nothing
908                          * has been send to ZFS at this step */
909                         continue;
910                 }
911
912                 if (!drop_cache)
913                         continue;
914
915                 /* we have to mark dbufs for eviction here because
916                  * dmu_assign_arcbuf() may create a new dbuf for
917                  * loaned abuf */
918                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
919                                             abufsz);
920         }
921         up_read(&obj->oo_guard);
922
923         if (unlikely(new_size == 0)) {
924                 /* no pages to write, no transno is needed */
925                 th->th_local = 1;
926                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
927                  * since ofd_commitrw_write() retries several times on ENOSPC */
928                 record_end_io(osd, WRITE, 0, 0, 0);
929                 RETURN(0);
930         }
931
932         write_lock(&obj->oo_attr_lock);
933         if (obj->oo_attr.la_size < new_size) {
934                 obj->oo_attr.la_size = new_size;
935                 write_unlock(&obj->oo_attr_lock);
936                 /* osd_object_sa_update() will be copying directly from
937                  * oo_attr into dbuf. any update within a single txg will copy
938                  * the most actual */
939                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
940                                           &obj->oo_attr.la_size, 8, oh);
941         } else {
942                 write_unlock(&obj->oo_attr_lock);
943         }
944
945         record_end_io(osd, WRITE, 0, iosize, npages);
946
947         RETURN(rc);
948 }
949
950 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
951                         struct niobuf_local *lnb, int npages)
952 {
953         struct osd_object *obj  = osd_dt_obj(dt);
954         int                i;
955         loff_t             eof;
956
957         LASSERT(dt_object_exists(dt));
958         LASSERT(obj->oo_dn);
959
960         read_lock(&obj->oo_attr_lock);
961         eof = obj->oo_attr.la_size;
962         read_unlock(&obj->oo_attr_lock);
963
964         for (i = 0; i < npages; i++) {
965                 if (unlikely(lnb[i].lnb_rc < 0))
966                         continue;
967
968                 lnb[i].lnb_rc = lnb[i].lnb_len;
969
970                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
971                         if (eof <= lnb[i].lnb_file_offset)
972                                 lnb[i].lnb_rc = 0;
973                         else
974                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
975
976                         /* all subsequent rc should be 0 */
977                         while (++i < npages)
978                                 lnb[i].lnb_rc = 0;
979                         break;
980                 }
981         }
982
983         return 0;
984 }
985
986 /*
987  * Punch/truncate an object
988  *
989  *      IN:     db  - dmu_buf of the object to free data in.
990  *              off - start of section to free.
991  *              len - length of section to free (DMU_OBJECT_END => to EOF).
992  *
993  *      RETURN: 0 if success
994  *              error code if failure
995  *
996  * The transaction passed to this routine must have
997  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
998  * called and then assigned to a transaction group.
999  */
1000 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
1001                                 uint64_t size, uint64_t off, uint64_t len)
1002 {
1003         int rc = 0;
1004
1005         /* Assert that the transaction has been assigned to a
1006            transaction group. */
1007         LASSERT(tx->tx_txg != 0);
1008         /*
1009          * Nothing to do if file already at desired length.
1010          */
1011         if (len == DMU_OBJECT_END && size == off)
1012                 return 0;
1013
1014         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1015         if (off < size)
1016                 dmu_free_range(os, dn->dn_object, off, len, tx);
1017
1018         return rc;
1019 }
1020
1021 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1022                         __u64 start, __u64 end, struct thandle *th)
1023 {
1024         struct osd_object  *obj = osd_dt_obj(dt);
1025         struct osd_device  *osd = osd_obj2dev(obj);
1026         struct osd_thandle *oh;
1027         __u64               len;
1028         int                 rc = 0;
1029         ENTRY;
1030
1031         LASSERT(dt_object_exists(dt));
1032         LASSERT(osd_invariant(obj));
1033
1034         LASSERT(th != NULL);
1035         oh = container_of0(th, struct osd_thandle, ot_super);
1036
1037         write_lock(&obj->oo_attr_lock);
1038         /* truncate */
1039         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1040                 len = DMU_OBJECT_END;
1041         else
1042                 len = end - start;
1043         write_unlock(&obj->oo_attr_lock);
1044
1045         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
1046                                 obj->oo_attr.la_size, start, len);
1047         /* set new size */
1048         if (len == DMU_OBJECT_END) {
1049                 write_lock(&obj->oo_attr_lock);
1050                 obj->oo_attr.la_size = start;
1051                 write_unlock(&obj->oo_attr_lock);
1052                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1053                                           &obj->oo_attr.la_size, 8, oh);
1054         }
1055         RETURN(rc);
1056 }
1057
1058 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1059                         __u64 start, __u64 end, struct thandle *handle)
1060 {
1061         struct osd_object  *obj = osd_dt_obj(dt);
1062         struct osd_device  *osd = osd_obj2dev(obj);
1063         struct osd_thandle *oh;
1064         __u64               len;
1065         ENTRY;
1066
1067         oh = container_of0(handle, struct osd_thandle, ot_super);
1068
1069         read_lock(&obj->oo_attr_lock);
1070         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1071                 len = DMU_OBJECT_END;
1072         else
1073                 len = end - start;
1074
1075         /* declare we'll free some blocks ... */
1076         if (start < obj->oo_attr.la_size) {
1077                 read_unlock(&obj->oo_attr_lock);
1078                 dmu_tx_mark_netfree(oh->ot_tx);
1079                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1080         } else {
1081                 read_unlock(&obj->oo_attr_lock);
1082         }
1083
1084         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1085                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1086                                  0, oh, NULL, OSD_QID_BLK));
1087 }
1088
1089 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1090                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1091 {
1092         int     rc;
1093         ENTRY;
1094
1095         switch (advice) {
1096         default:
1097                 rc = -ENOTSUPP;
1098                 break;
1099         }
1100
1101         RETURN(rc);
1102 }
1103
1104 struct dt_body_operations osd_body_ops = {
1105         .dbo_read                       = osd_read,
1106         .dbo_declare_write              = osd_declare_write,
1107         .dbo_write                      = osd_write,
1108         .dbo_bufs_get                   = osd_bufs_get,
1109         .dbo_bufs_put                   = osd_bufs_put,
1110         .dbo_write_prep                 = osd_write_prep,
1111         .dbo_declare_write_commit       = osd_declare_write_commit,
1112         .dbo_write_commit               = osd_write_commit,
1113         .dbo_read_prep                  = osd_read_prep,
1114         .dbo_declare_punch              = osd_declare_punch,
1115         .dbo_punch                      = osd_punch,
1116         .dbo_ladvise                    = osd_ladvise,
1117 };
1118
1119 struct dt_body_operations osd_body_scrub_ops = {
1120         .dbo_read                       = osd_read_no_record,
1121         .dbo_declare_write              = osd_declare_write,
1122         .dbo_write                      = osd_write,
1123 };