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 static void dbuf_set_pending_evict(dmu_buf_t *db)
  67 {
  68         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
  69         dbi->db_pending_evict = TRUE;
  70 }
  71
  72 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  73 {
  74         struct obd_histogram *h = osd->od_brw_stats.hist;
  75
  76         if (rw == READ) {
  77                 atomic_inc(&osd->od_r_in_flight);
  78                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  79                                  atomic_read(&osd->od_r_in_flight));
  80                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  81
  82         } else {
  83                 atomic_inc(&osd->od_w_in_flight);
  84                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  85                                  atomic_read(&osd->od_w_in_flight));
  86                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  87
  88         }
  89 }
  90
  91 static void record_end_io(struct osd_device *osd, int rw,
  92                           unsigned long elapsed, int disksize, int npages)
  93 {
  94         struct obd_histogram *h = osd->od_brw_stats.hist;
  95
  96         if (rw == READ)
  97                 atomic_dec(&osd->od_r_in_flight);
  98         else
  99                 atomic_dec(&osd->od_w_in_flight);
 100
 101         lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
 102         if (disksize > 0)
 103                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
 104         if (elapsed)
 105                 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
 106 }
 107
 108 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 109                           struct lu_buf *buf, loff_t *pos, size_t *size)
 110 {
 111         struct osd_object *obj = osd_dt_obj(dt);
 112         uint64_t old_size;
 113         int rc;
 114
 115         LASSERT(dt_object_exists(dt));
 116         LASSERT(obj->oo_dn);
 117
 118         read_lock(&obj->oo_attr_lock);
 119         old_size = obj->oo_attr.la_size;
 120         read_unlock(&obj->oo_attr_lock);
 121
 122         if (*pos + *size > old_size) {
 123                 if (old_size < *pos)
 124                         return 0;
 125
 126                 *size = old_size - *pos;
 127         }
 128
 129         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 130                           buf->lb_buf, DMU_READ_PREFETCH);
 131         if (!rc) {
 132                 rc = *size;
 133                 *pos += *size;
 134         }
 135
 136         return rc;
 137 }
 138
 139 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 140                         struct lu_buf *buf, loff_t *pos)
 141 {
 142         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 143         size_t size = buf->lb_len;
 144         hrtime_t start = gethrtime();
 145         s64 delta_ms;
 146         int rc;
 147
 148         record_start_io(osd, READ, 0);
 149         rc = __osd_read(env, dt, buf, pos, &size);
 150         delta_ms = gethrtime() - start;
 151         do_div(delta_ms, NSEC_PER_MSEC);
 152         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 153
 154         return rc;
 155 }
 156
 157 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 158                                          struct dt_object *dt,
 159                                          struct lu_buf *buf, loff_t *pos)
 160 {
 161         size_t size = buf->lb_len;
 162
 163         return __osd_read(env, dt, buf, pos, &size);
 164 }
 165
 166 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 167                                 const struct lu_buf *buf, loff_t pos,
 168                                 struct thandle *th)
 169 {
 170         struct osd_object  *obj  = osd_dt_obj(dt);
 171         struct osd_device  *osd = osd_obj2dev(obj);
 172         struct osd_thandle *oh;
 173         uint64_t            oid;
 174         ENTRY;
 175
 176         oh = container_of(th, struct osd_thandle, ot_super);
 177
 178         /* in some cases declare can race with creation (e.g. llog)
 179          * and we need to wait till object is initialized. notice
 180          * LOHA_EXISTs is supposed to be the last step in the
 181          * initialization */
 182
 183         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 184         if (dt_object_exists(dt))
 185                 oid = obj->oo_dn->dn_object;
 186         else
 187                 oid = DMU_NEW_OBJECT;
 188
 189         /* XXX: we still miss for append declaration support in ZFS
 190          *      -1 means append which is used by llog mostly, llog
 191          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 192         if (pos == -1)
 193                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 194                             obj->oo_attr.la_size + (2 << 20));
 195         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 196
 197         /* dt_declare_write() is usually called for system objects, such
 198          * as llog or last_rcvd files. We needn't enforce quota on those
 199          * objects, so always set the lqi_space as 0. */
 200         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 201                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 202                                  0, oh, NULL, OSD_QID_BLK));
 203 }
 204
 205 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 206                         const struct lu_buf *buf, loff_t *pos,
 207                          struct thandle *th)
 208 {
 209         struct osd_object  *obj  = osd_dt_obj(dt);
 210         struct osd_device  *osd = osd_obj2dev(obj);
 211         struct osd_thandle *oh;
 212         uint64_t            offset = *pos;
 213         int                 rc;
 214
 215         ENTRY;
 216
 217         LASSERT(dt_object_exists(dt));
 218         LASSERT(obj->oo_dn);
 219
 220         LASSERT(th != NULL);
 221         oh = container_of(th, struct osd_thandle, ot_super);
 222
 223         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 224                       buf->lb_buf, oh->ot_tx);
 225         write_lock(&obj->oo_attr_lock);
 226         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 227                 obj->oo_attr.la_size = offset + buf->lb_len;
 228                 write_unlock(&obj->oo_attr_lock);
 229                 /* osd_object_sa_update() will be copying directly from oo_attr
 230                  * into dbuf.  any update within a single txg will copy the
 231                  * most actual */
 232                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 233                                         &obj->oo_attr.la_size, 8, oh);
 234                 if (unlikely(rc))
 235                         GOTO(out, rc);
 236         } else {
 237                 write_unlock(&obj->oo_attr_lock);
 238         }
 239
 240         *pos += buf->lb_len;
 241         rc = buf->lb_len;
 242
 243 out:
 244         RETURN(rc);
 245 }
 246
 247 /*
 248  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 249  *      purposes as it's not very well defined ...
 250  *      instead I use the lowest bit of the address so that:
 251  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 252  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 253  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 254  *
 255  *      bzzz, to blame
 256  */
 257 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 258                         struct niobuf_local *lnb, int npages)
 259 {
 260         struct osd_object *obj  = osd_dt_obj(dt);
 261         struct osd_device *osd = osd_obj2dev(obj);
 262         unsigned long      ptr;
 263         int                i;
 264
 265         LASSERT(dt_object_exists(dt));
 266         LASSERT(obj->oo_dn);
 267
 268         for (i = 0; i < npages; i++) {
 269                 if (lnb[i].lnb_page == NULL)
 270                         continue;
 271                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 272                         /* this is anonymous page allocated for copy-write */
 273                         lnb[i].lnb_page->mapping = NULL;
 274                         __free_page(lnb[i].lnb_page);
 275                         atomic_dec(&osd->od_zerocopy_alloc);
 276                 } else {
 277                         /* see comment in osd_bufs_get_read() */
 278                         ptr = (unsigned long)lnb[i].lnb_data;
 279                         if (ptr & 1UL) {
 280                                 ptr &= ~1UL;
 281                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 282                                 atomic_dec(&osd->od_zerocopy_pin);
 283                         } else if (lnb[i].lnb_data != NULL) {
 284                                 int j, apages, abufsz;
 285                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 286                                 apages = abufsz >> PAGE_SHIFT;
 287                                 /* these references to pages must be invalidated
 288                                  * to prevent access in osd_bufs_put() */
 289                                 for (j = 0; j < apages; j++)
 290                                         lnb[i + j].lnb_page = NULL;
 291                                 dmu_return_arcbuf(lnb[i].lnb_data);
 292                                 atomic_dec(&osd->od_zerocopy_loan);
 293                         }
 294                 }
 295                 lnb[i].lnb_page = NULL;
 296                 lnb[i].lnb_data = NULL;
 297         }
 298
 299         return 0;
 300 }
 301
 302 static inline struct page *kmem_to_page(void *addr)
 303 {
 304         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 305         if (is_vmalloc_addr(addr))
 306                 return vmalloc_to_page(addr);
 307         else
 308                 return virt_to_page(addr);
 309 }
 310
 311 /**
 312  * Prepare buffers for read.
 313  *
 314  * The function maps the range described by \a off and \a len to \a lnb array.
 315  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 316  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 317  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 318  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 319  * A better implementation would just return the buffers (potentially unfilled)
 320  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 321  *
 322  * \param[in] env       environment
 323  * \param[in] obj       object
 324  * \param[in] off       offset in bytes
 325  * \param[in] len       the number of bytes to access
 326  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 327  *
 328  * \retval              0 for success
 329  * \retval              negative error number of failure
 330  */
 331 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 332                              loff_t off, ssize_t len, struct niobuf_local *lnb,
 333                              int maxlnb)
 334 {
 335         struct osd_device *osd = osd_obj2dev(obj);
 336         int rc, i, numbufs, npages = 0, drop_cache = 0;
 337         hrtime_t start = gethrtime();
 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                 dbp = NULL;
 362                 if (unlikely(npages >= maxlnb))
 363                         GOTO(err, rc = -EOVERFLOW);
 364
 365                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 366                                                   off, len, TRUE, osd_0copy_tag,
 367                                                   &numbufs, &dbp);
 368                 if (unlikely(rc))
 369                         GOTO(err, rc);
 370
 371                 for (i = 0; i < numbufs; i++) {
 372                         int bufoff, tocpy, thispage;
 373                         void *dbf = dbp[i];
 374
 375                         LASSERT(len > 0);
 376
 377                         atomic_inc(&osd->od_zerocopy_pin);
 378
 379                         bufoff = off - dbp[i]->db_offset;
 380                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 381
 382                         /* kind of trick to differentiate dbuf vs. arcbuf */
 383                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 384                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 385
 386                         while (tocpy > 0) {
 387                                 if (unlikely(npages >= maxlnb))
 388                                         GOTO(err, rc = -EOVERFLOW);
 389
 390                                 thispage = PAGE_SIZE;
 391                                 thispage -= bufoff & (PAGE_SIZE - 1);
 392                                 thispage = min(tocpy, thispage);
 393
 394                                 lnb->lnb_rc = 0;
 395                                 lnb->lnb_file_offset = off;
 396                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 397                                 lnb->lnb_len = thispage;
 398                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 399                                                              bufoff);
 400                                 /* mark just a single slot: we need this
 401                                  * reference to dbuf to be released once */
 402                                 lnb->lnb_data = dbf;
 403                                 dbf = NULL;
 404
 405                                 tocpy -= thispage;
 406                                 len -= thispage;
 407                                 bufoff += thispage;
 408                                 off += thispage;
 409
 410                                 npages++;
 411                                 lnb++;
 412                         }
 413
 414                         if (drop_cache)
 415                                 dbuf_set_pending_evict(dbp[i]);
 416
 417                         /* steal dbuf so dmu_buf_rele_array() can't release
 418                          * it */
 419                         dbp[i] = NULL;
 420                 }
 421
 422                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 423         }
 424
 425         delta_ms = gethrtime() - start;
 426         do_div(delta_ms, NSEC_PER_MSEC);
 427         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 428
 429         RETURN(npages);
 430
 431 err:
 432         LASSERT(rc < 0);
 433         if (dbp)
 434                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 435         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 436         RETURN(rc);
 437 }
 438
 439 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 440 {
 441         arc_buf_t *abuf;
 442
 443         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 444         if (unlikely(!abuf))
 445                 return ERR_PTR(-ENOMEM);
 446
 447 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 448         /**
 449          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 450          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 451          * setting up page array for RDMA transfer. See LU-9305.
 452          */
 453         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 454                 dmu_return_arcbuf(abuf);
 455                 return NULL;
 456         }
 457 #endif
 458
 459         return abuf;
 460 }
 461
 462 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 463                               loff_t off, ssize_t len, struct niobuf_local *lnb,
 464                               int maxlnb)
 465 {
 466         struct osd_device *osd = osd_obj2dev(obj);
 467         int                poff, plen, off_in_block, sz_in_block;
 468         int                rc, i = 0, npages = 0;
 469         dnode_t *dn = obj->oo_dn;
 470         arc_buf_t *abuf;
 471         uint32_t bs = dn->dn_datablksz;
 472         ENTRY;
 473
 474         /*
 475          * currently only full blocks are subject to zerocopy approach:
 476          * so that we're sure nobody is trying to update the same block
 477          */
 478         while (len > 0) {
 479                 if (unlikely(npages >= maxlnb))
 480                         GOTO(out_err, rc = -EOVERFLOW);
 481
 482                 off_in_block = off & (bs - 1);
 483                 sz_in_block = min_t(int, bs - off_in_block, len);
 484
 485                 abuf = NULL;
 486                 if (sz_in_block == bs) {
 487                         /* full block, try to use zerocopy */
 488                         abuf = osd_request_arcbuf(dn, bs);
 489                         if (unlikely(IS_ERR(abuf)))
 490                                 GOTO(out_err, rc = PTR_ERR(abuf));
 491                 }
 492
 493                 if (abuf != NULL) {
 494                         atomic_inc(&osd->od_zerocopy_loan);
 495
 496                         /* go over pages arcbuf contains, put them as
 497                          * local niobufs for ptlrpc's bulks */
 498                         while (sz_in_block > 0) {
 499                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 500
 501                                 if (unlikely(npages >= maxlnb))
 502                                         GOTO(out_err, rc = -EOVERFLOW);
 503
 504                                 lnb[i].lnb_file_offset = off;
 505                                 lnb[i].lnb_page_offset = 0;
 506                                 lnb[i].lnb_len = plen;
 507                                 lnb[i].lnb_rc = 0;
 508                                 if (sz_in_block == bs)
 509                                         lnb[i].lnb_data = abuf;
 510                                 else
 511                                         lnb[i].lnb_data = NULL;
 512
 513                                 /* this one is not supposed to fail */
 514                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 515                                                         off_in_block);
 516                                 LASSERT(lnb[i].lnb_page);
 517
 518                                 lprocfs_counter_add(osd->od_stats,
 519                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 520
 521                                 sz_in_block -= plen;
 522                                 len -= plen;
 523                                 off += plen;
 524                                 off_in_block += plen;
 525                                 i++;
 526                                 npages++;
 527                         }
 528                 } else {
 529                         if (off_in_block == 0 && len < bs &&
 530                                         off + len >= obj->oo_attr.la_size)
 531                                 lprocfs_counter_add(osd->od_stats,
 532                                                 LPROC_OSD_TAIL_IO, 1);
 533
 534                         /* can't use zerocopy, allocate temp. buffers */
 535                         poff = off & (PAGE_SIZE - 1);
 536                         while (sz_in_block > 0) {
 537                                 plen = min_t(int, poff + sz_in_block,
 538                                              PAGE_SIZE);
 539                                 plen -= poff;
 540
 541                                 if (unlikely(npages >= maxlnb))
 542                                         GOTO(out_err, rc = -EOVERFLOW);
 543
 544                                 lnb[i].lnb_file_offset = off;
 545                                 lnb[i].lnb_page_offset = poff;
 546                                 poff = 0;
 547
 548                                 lnb[i].lnb_len = plen;
 549                                 lnb[i].lnb_rc = 0;
 550                                 lnb[i].lnb_data = NULL;
 551
 552                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 553                                 if (unlikely(lnb[i].lnb_page == NULL))
 554                                         GOTO(out_err, rc = -ENOMEM);
 555
 556                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 557                                 lnb[i].lnb_page->mapping = (void *)obj;
 558
 559                                 atomic_inc(&osd->od_zerocopy_alloc);
 560                                 lprocfs_counter_add(osd->od_stats,
 561                                                 LPROC_OSD_COPY_IO, 1);
 562
 563                                 sz_in_block -= plen;
 564                                 len -= plen;
 565                                 off += plen;
 566                                 i++;
 567                                 npages++;
 568                         }
 569                 }
 570         }
 571
 572         RETURN(npages);
 573
 574 out_err:
 575         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 576         RETURN(rc);
 577 }
 578
 579 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 580                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 581                         int maxlnb, enum dt_bufs_type rw)
 582 {
 583         struct osd_object *obj  = osd_dt_obj(dt);
 584         int                rc;
 585
 586         LASSERT(dt_object_exists(dt));
 587         LASSERT(obj->oo_dn);
 588
 589         if (rw & DT_BUFS_TYPE_WRITE)
 590                 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
 591         else
 592                 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
 593
 594         return rc;
 595 }
 596
 597 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 598                         struct niobuf_local *lnb, int npages)
 599 {
 600         struct osd_object *obj = osd_dt_obj(dt);
 601
 602         LASSERT(dt_object_exists(dt));
 603         LASSERT(obj->oo_dn);
 604
 605         return 0;
 606 }
 607
 608 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 609                                            uint64_t offset,
 610                                            uint32_t blksz)
 611 {
 612         LASSERT(blksz > 0);
 613
 614         size += offset % blksz;
 615
 616         if (likely(is_power_of_2(blksz)))
 617                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 618
 619         size += blksz - 1;
 620         do_div(size, blksz);
 621         return size * blksz;
 622 }
 623
 624 static int osd_declare_write_commit(const struct lu_env *env,
 625                                     struct dt_object *dt,
 626                                     struct niobuf_local *lnb, int npages,
 627                                     struct thandle *th)
 628 {
 629         struct osd_object  *obj = osd_dt_obj(dt);
 630         struct osd_device  *osd = osd_obj2dev(obj);
 631         struct osd_thandle *oh;
 632         uint64_t            offset = 0;
 633         uint32_t            size = 0;
 634         uint32_t blksz = obj->oo_dn->dn_datablksz;
 635         int                 i, rc;
 636         bool synced = false;
 637         long long           space = 0;
 638         struct page        *last_page = NULL;
 639         unsigned long       discont_pages = 0;
 640         enum osd_quota_local_flags local_flags = 0;
 641         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 642         ENTRY;
 643
 644         LASSERT(dt_object_exists(dt));
 645         LASSERT(obj->oo_dn);
 646
 647         LASSERT(lnb);
 648         LASSERT(npages > 0);
 649
 650         oh = container_of(th, struct osd_thandle, ot_super);
 651
 652         for (i = 0; i < npages; i++) {
 653                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 654                         ++discont_pages;
 655                 last_page = lnb[i].lnb_page;
 656                 if (lnb[i].lnb_rc)
 657                         /* ENOSPC, network RPC error, etc.
 658                          * We don't want to book space for pages which will be
 659                          * skipped in osd_write_commit(). Hence we skip pages
 660                          * with lnb_rc != 0 here too */
 661                         continue;
 662                 /* ignore quota for the whole request if any page is from
 663                  * client cache or written by root.
 664                  *
 665                  * XXX once we drop the 1.8 client support, the checking
 666                  * for whether page is from cache can be simplified as:
 667                  * !(lnb[i].flags & OBD_BRW_SYNC)
 668                  *
 669                  * XXX we could handle this on per-lnb basis as done by
 670                  * grant. */
 671                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 672                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 673                     OBD_BRW_FROM_GRANT)
 674                         declare_flags |= OSD_QID_FORCE;
 675
 676                 if (size == 0) {
 677                         /* first valid lnb */
 678                         offset = lnb[i].lnb_file_offset;
 679                         size = lnb[i].lnb_len;
 680                         continue;
 681                 }
 682                 if (offset + size == lnb[i].lnb_file_offset) {
 683                         /* this lnb is contiguous to the previous one */
 684                         size += lnb[i].lnb_len;
 685                         continue;
 686                 }
 687
 688                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 689                                   obj->oo_dn, offset, size);
 690                 /* Estimating space to be consumed by a write is rather
 691                  * complicated with ZFS. As a consequence, we don't account for
 692                  * indirect blocks and just use as a rough estimate the worse
 693                  * case where the old space is being held by a snapshot. Quota
 694                  * overrun will be adjusted once the operation is committed, if
 695                  * required. */
 696                 space += osd_roundup2blocksz(size, offset, blksz);
 697
 698                 offset = lnb[i].lnb_file_offset;
 699                 size = lnb[i].lnb_len;
 700         }
 701
 702         if (size) {
 703                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 704                                   offset, size);
 705                 space += osd_roundup2blocksz(size, offset, blksz);
 706         }
 707
 708         /* backend zfs filesystem might be configured to store multiple data
 709          * copies */
 710         space  *= osd->od_os->os_copies;
 711         space   = toqb(space);
 712         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 713                npages, space);
 714
 715         record_start_io(osd, WRITE, discont_pages);
 716 retry:
 717         /* acquire quota space if needed */
 718         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 719                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 720                                space, oh, &local_flags, declare_flags);
 721
 722         if (!synced && rc == -EDQUOT &&
 723             (local_flags & QUOTA_FL_SYNC) != 0) {
 724                 dt_sync(env, th->th_dev);
 725                 synced = true;
 726                 CDEBUG(D_QUOTA, "retry after sync\n");
 727                 local_flags = 0;
 728                 goto retry;
 729         }
 730
 731         /* we need only to store the overquota flags in the first lnb for
 732          * now, once we support multiple objects BRW, this code needs be
 733          * revised. */
 734         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
 735                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 736         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
 737                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 738 #ifdef ZFS_PROJINHERIT
 739         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
 740                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 741 #endif
 742
 743         RETURN(rc);
 744 }
 745
 746 /**
 747  * Policy to grow ZFS block size by write pattern.
 748  * For sequential write, it grows block size gradually until it reaches the
 749  * maximum blocksize the dataset can support. Otherwise, it will pick a
 750  * a block size by the writing region of this I/O.
 751  */
 752 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 753                               uint64_t start, uint64_t end)
 754 {
 755         struct osd_device       *osd = osd_obj2dev(obj);
 756         dnode_t *dn = obj->oo_dn;
 757         uint32_t                 blksz;
 758         int                      rc = 0;
 759
 760         ENTRY;
 761
 762         if (dn->dn_maxblkid > 0) /* can't change block size */
 763                 GOTO(out, rc);
 764
 765         if (dn->dn_datablksz >= osd->od_max_blksz)
 766                 GOTO(out, rc);
 767
 768         down_write(&obj->oo_guard);
 769
 770         blksz = dn->dn_datablksz;
 771         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 772                 GOTO(out_unlock, rc);
 773
 774         /* now ZFS can support up to 16MB block size, and if the write
 775          * is sequential, it just increases the block size gradually */
 776         if (start <= blksz) { /* sequential */
 777                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 778         } else { /* sparse, pick a block size by write region */
 779                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 780                                         end - start);
 781         }
 782
 783         if (!is_power_of_2(blksz))
 784                 blksz = size_roundup_power2(blksz);
 785
 786         if (blksz > dn->dn_datablksz) {
 787                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 788                                                blksz, 0, oh->ot_tx);
 789                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 790                 if (rc < 0)
 791                         CDEBUG(D_INODE, "object "DFID": change block size"
 792                                "%u -> %u error rc = %d\n",
 793                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 794                                dn->dn_datablksz, blksz, rc);
 795         }
 796         EXIT;
 797 out_unlock:
 798         up_write(&obj->oo_guard);
 799 out:
 800         return rc;
 801 }
 802
 803 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 804                                         loff_t off, ssize_t len)
 805 {
 806         dmu_buf_t **dbp;
 807         int i, rc, numbufs;
 808
 809         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 810                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 811         if (unlikely(rc))
 812                 return;
 813
 814         for (i = 0; i < numbufs; i++)
 815                 dbuf_set_pending_evict(dbp[i]);
 816
 817         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 818 }
 819
 820 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 821                         struct niobuf_local *lnb, int npages,
 822                         struct thandle *th, __u64 user_size)
 823 {
 824         struct osd_object  *obj  = osd_dt_obj(dt);
 825         struct osd_device  *osd = osd_obj2dev(obj);
 826         struct osd_thandle *oh;
 827         uint64_t            new_size = 0;
 828         int                 i, abufsz, rc = 0, drop_cache = 0;
 829         unsigned long      iosize = 0;
 830         ENTRY;
 831
 832         LASSERT(dt_object_exists(dt));
 833         LASSERT(obj->oo_dn);
 834
 835         LASSERT(th != NULL);
 836         oh = container_of(th, struct osd_thandle, ot_super);
 837
 838         /* adjust block size. Assume the buffers are sorted. */
 839         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 840                                  lnb[npages - 1].lnb_file_offset +
 841                                  lnb[npages - 1].lnb_len);
 842
 843         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 844             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 845             osd->od_readcache_max_filesize)
 846                 drop_cache = 1;
 847
 848         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 849                 RETURN(-ENOSPC);
 850
 851         /* if la_size is already bigger than specified user_size,
 852          * ignore user_size
 853          */
 854         if (obj->oo_attr.la_size > user_size)
 855                 user_size = 0;
 856
 857         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 858          * size and assigning arcbuf take place at the same time.
 859          *
 860          * Thread 1:
 861          * osd_write_commit()
 862          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 863          *   -> dmu_object_set_blocksize()
 864          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 865          *       write lock held
 866          *     -> dbuf_new_size()
 867          *      -> dmu_buf_will_dirty()
 868          *       -> dbuf_read()
 869          *        -> wait for the dbuf state to change
 870          * Thread 2:
 871          * osd_write_commit()
 872          *  -> dmu_assign_arcbuf()
 873          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 874          *    -> dbuf_dirty()
 875          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 876          *
 877          * By taking the read lock, it can avoid thread 2 to enter into the
 878          * critical section of assigning the arcbuf, while thread 1 is
 879          * changing the block size.
 880          */
 881         down_read(&obj->oo_guard);
 882         for (i = 0; i < npages; i++) {
 883                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 884                         (unsigned) lnb[i].lnb_len,
 885                         (unsigned) lnb[i].lnb_file_offset);
 886
 887                 if (lnb[i].lnb_rc) {
 888                         /* ENOSPC, network RPC error, etc.
 889                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 890                          * so we skip this page if lnb_rc is set to -ENOSPC */
 891                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 892                                 PFID(lu_object_fid(&dt->do_lu)), i,
 893                                 lnb[i].lnb_rc);
 894                         continue;
 895                 }
 896
 897                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 898                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 899                 if (lnb[i].lnb_page == NULL)
 900                         continue;
 901
 902                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 903                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 904                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
 905                                       lnb[i].lnb_page_offset, oh->ot_tx);
 906                         kunmap(lnb[i].lnb_page);
 907                         iosize += lnb[i].lnb_len;
 908                         abufsz = lnb[i].lnb_len; /* to drop cache below */
 909                 } else if (lnb[i].lnb_data) {
 910                         int j, apages;
 911                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 912                         /* buffer loaned for zerocopy, try to use it.
 913                          * notice that dmu_assign_arcbuf() is smart
 914                          * enough to recognize changed blocksize
 915                          * in this case it fallbacks to dmu_write() */
 916                         abufsz = arc_buf_size(lnb[i].lnb_data);
 917                         LASSERT(abufsz & PAGE_MASK);
 918                         apages = abufsz >> PAGE_SHIFT;
 919                         LASSERT(i + apages <= npages);
 920                         /* these references to pages must be invalidated
 921                          * to prevent access in osd_bufs_put() */
 922                         for (j = 0; j < apages; j++)
 923                                 lnb[i + j].lnb_page = NULL;
 924                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 925                                           lnb[i].lnb_file_offset,
 926                                           lnb[i].lnb_data, oh->ot_tx);
 927                         /* drop the reference, otherwise osd_put_bufs()
 928                          * will be releasing it - bad! */
 929                         lnb[i].lnb_data = NULL;
 930                         atomic_dec(&osd->od_zerocopy_loan);
 931                         iosize += abufsz;
 932                 } else {
 933                         /* we don't want to deal with cache if nothing
 934                          * has been send to ZFS at this step */
 935                         continue;
 936                 }
 937
 938                 if (!drop_cache)
 939                         continue;
 940
 941                 /* we have to mark dbufs for eviction here because
 942                  * dmu_assign_arcbuf() may create a new dbuf for
 943                  * loaned abuf */
 944                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
 945                                             abufsz);
 946         }
 947         up_read(&obj->oo_guard);
 948
 949         if (unlikely(new_size == 0)) {
 950                 /* no pages to write, no transno is needed */
 951                 th->th_local = 1;
 952                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 953                  * since ofd_commitrw_write() retries several times on ENOSPC */
 954                 record_end_io(osd, WRITE, 0, 0, 0);
 955                 RETURN(0);
 956         }
 957
 958         /* if file has grown, take user_size into account */
 959         if (user_size && new_size > user_size)
 960                 new_size = user_size;
 961         write_lock(&obj->oo_attr_lock);
 962         if (obj->oo_attr.la_size < new_size) {
 963                 obj->oo_attr.la_size = new_size;
 964                 write_unlock(&obj->oo_attr_lock);
 965                 /* osd_object_sa_update() will be copying directly from
 966                  * oo_attr into dbuf. any update within a single txg will copy
 967                  * the most actual */
 968                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 969                                           &obj->oo_attr.la_size, 8, oh);
 970         } else {
 971                 write_unlock(&obj->oo_attr_lock);
 972         }
 973
 974         record_end_io(osd, WRITE, 0, iosize, npages);
 975
 976         RETURN(rc);
 977 }
 978
 979 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 980                         struct niobuf_local *lnb, int npages)
 981 {
 982         struct osd_object *obj  = osd_dt_obj(dt);
 983         int                i;
 984         loff_t             eof;
 985
 986         LASSERT(dt_object_exists(dt));
 987         LASSERT(obj->oo_dn);
 988
 989         read_lock(&obj->oo_attr_lock);
 990         eof = obj->oo_attr.la_size;
 991         read_unlock(&obj->oo_attr_lock);
 992
 993         for (i = 0; i < npages; i++) {
 994                 if (unlikely(lnb[i].lnb_rc < 0))
 995                         continue;
 996
 997                 lnb[i].lnb_rc = lnb[i].lnb_len;
 998
 999                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
1000                         /* send complete pages all the time */
1001                         if (eof <= lnb[i].lnb_file_offset)
1002                                 lnb[i].lnb_rc = 0;
1003
1004                         /* all subsequent rc should be 0 */
1005                         while (++i < npages)
1006                                 lnb[i].lnb_rc = 0;
1007                         break;
1008                 }
1009         }
1010
1011         return 0;
1012 }
1013
1014 /*
1015  * Punch/truncate an object
1016  *
1017  *      IN:     db  - dmu_buf of the object to free data in.
1018  *              off - start of section to free.
1019  *              len - length of section to free (DMU_OBJECT_END => to EOF).
1020  *
1021  *      RETURN: 0 if success
1022  *              error code if failure
1023  *
1024  * The transaction passed to this routine must have
1025  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1026  * called and then assigned to a transaction group.
1027  */
1028 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1029                               dmu_tx_t *tx, uint64_t off, uint64_t len)
1030 {
1031         dnode_t *dn = obj->oo_dn;
1032         uint64_t size = obj->oo_attr.la_size;
1033         int rc = 0;
1034
1035         /* Assert that the transaction has been assigned to a
1036            transaction group. */
1037         LASSERT(tx->tx_txg != 0);
1038         /*
1039          * Nothing to do if file already at desired length.
1040          */
1041         if (len == DMU_OBJECT_END && size == off)
1042                 return 0;
1043
1044         /* if object holds encrypted content, we need to make sure we truncate
1045          * on an encryption unit boundary, or subsequent reads will get
1046          * corrupted content
1047          */
1048         if (len != DMU_OBJECT_END)
1049                 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1050                         (off & ~LUSTRE_ENCRYPTION_MASK);
1051         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1052             off & ~LUSTRE_ENCRYPTION_MASK)
1053                 off = (off & LUSTRE_ENCRYPTION_MASK) +
1054                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1055
1056
1057         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1058         if (off < size)
1059                 dmu_free_range(os, dn->dn_object, off, len, tx);
1060
1061         return rc;
1062 }
1063
1064 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1065                         __u64 start, __u64 end, struct thandle *th)
1066 {
1067         struct osd_object  *obj = osd_dt_obj(dt);
1068         struct osd_device  *osd = osd_obj2dev(obj);
1069         struct osd_thandle *oh;
1070         __u64               len;
1071         int                 rc = 0;
1072         ENTRY;
1073
1074         LASSERT(dt_object_exists(dt));
1075         LASSERT(osd_invariant(obj));
1076
1077         LASSERT(th != NULL);
1078         oh = container_of(th, struct osd_thandle, ot_super);
1079
1080         write_lock(&obj->oo_attr_lock);
1081         /* truncate */
1082         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1083                 len = DMU_OBJECT_END;
1084         else
1085                 len = end - start;
1086         write_unlock(&obj->oo_attr_lock);
1087
1088         rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1089
1090         /* set new size */
1091         if (len == DMU_OBJECT_END) {
1092                 write_lock(&obj->oo_attr_lock);
1093                 obj->oo_attr.la_size = start;
1094                 write_unlock(&obj->oo_attr_lock);
1095                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1096                                           &obj->oo_attr.la_size, 8, oh);
1097         }
1098         RETURN(rc);
1099 }
1100
1101 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1102                         __u64 start, __u64 end, struct thandle *handle)
1103 {
1104         struct osd_object  *obj = osd_dt_obj(dt);
1105         struct osd_device  *osd = osd_obj2dev(obj);
1106         struct osd_thandle *oh;
1107         __u64               len;
1108         ENTRY;
1109
1110         oh = container_of(handle, struct osd_thandle, ot_super);
1111
1112         read_lock(&obj->oo_attr_lock);
1113         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1114                 len = DMU_OBJECT_END;
1115         else
1116                 len = end - start;
1117
1118         /* declare we'll free some blocks ... */
1119         /* if object holds encrypted content, we need to make sure we truncate
1120          * on an encryption unit boundary, or subsequent reads will get
1121          * corrupted content
1122          */
1123         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1124             start & ~LUSTRE_ENCRYPTION_MASK)
1125                 start = (start & LUSTRE_ENCRYPTION_MASK) +
1126                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1127         if (start < obj->oo_attr.la_size) {
1128                 read_unlock(&obj->oo_attr_lock);
1129                 dmu_tx_mark_netfree(oh->ot_tx);
1130                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1131         } else {
1132                 read_unlock(&obj->oo_attr_lock);
1133         }
1134
1135         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1136                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1137                                  0, oh, NULL, OSD_QID_BLK));
1138 }
1139
1140 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1141                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1142 {
1143         int     rc;
1144         ENTRY;
1145
1146         switch (advice) {
1147         default:
1148                 rc = -ENOTSUPP;
1149                 break;
1150         }
1151
1152         RETURN(rc);
1153 }
1154
1155 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1156                          __u64 start, __u64 end, int mode, struct thandle *th)
1157 {
1158         int rc = -EOPNOTSUPP;
1159         ENTRY;
1160
1161          /*
1162           * space preallocation is not supported for ZFS
1163           * Returns -EOPNOTSUPP for now
1164           */
1165         RETURN(rc);
1166 }
1167
1168 static int osd_declare_fallocate(const struct lu_env *env,
1169                                  struct dt_object *dt, __u64 start, __u64 end,
1170                                  int mode, struct thandle *th)
1171 {
1172         int rc = -EOPNOTSUPP;
1173         ENTRY;
1174
1175          /*
1176           * space preallocation is not supported for ZFS
1177           * Returns -EOPNOTSUPP for now
1178           */
1179         RETURN(rc);
1180 }
1181
1182 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1183                         loff_t offset, int whence)
1184 {
1185         struct osd_object *obj = osd_dt_obj(dt);
1186         uint64_t size = obj->oo_attr.la_size;
1187         uint64_t result = offset;
1188         int rc;
1189         boolean_t hole = whence == SEEK_HOLE;
1190
1191         ENTRY;
1192
1193         LASSERT(dt_object_exists(dt));
1194         LASSERT(osd_invariant(obj));
1195
1196         if (offset < 0 || offset >= size)
1197                 RETURN(-ENXIO);
1198
1199 #ifdef HAVE_DMU_OFFSET_NEXT
1200         rc = dmu_offset_next(osd_obj2dev(obj)->od_os, obj->oo_dn->dn_object,
1201                              hole, &result);
1202         if (rc == ESRCH)
1203                 RETURN(-ENXIO);
1204 #else
1205         /*
1206          * In absence of dmu_offset_next() just do nothing but
1207          * return EBUSY as does dmu_offset_next() and that means
1208          * generic approach should be used.
1209          */
1210         rc = EBUSY;
1211 #endif
1212         /* file was dirty, so fall back to using generic logic */
1213         if (rc == EBUSY && hole)
1214                 RETURN(-ENXIO); /* see comment below */
1215
1216         /* when result is out of file range then it must be virtual hole
1217          * at the end of file, but this is not real file end, so return
1218          * just -ENXIO and LOV will translate it properly.
1219          */
1220         if (result >= size)
1221                 RETURN(-ENXIO);
1222
1223         RETURN(result);
1224 }
1225
1226 struct dt_body_operations osd_body_ops = {
1227         .dbo_read                       = osd_read,
1228         .dbo_declare_write              = osd_declare_write,
1229         .dbo_write                      = osd_write,
1230         .dbo_bufs_get                   = osd_bufs_get,
1231         .dbo_bufs_put                   = osd_bufs_put,
1232         .dbo_write_prep                 = osd_write_prep,
1233         .dbo_declare_write_commit       = osd_declare_write_commit,
1234         .dbo_write_commit               = osd_write_commit,
1235         .dbo_read_prep                  = osd_read_prep,
1236         .dbo_declare_punch              = osd_declare_punch,
1237         .dbo_punch                      = osd_punch,
1238         .dbo_ladvise                    = osd_ladvise,
1239         .dbo_declare_fallocate          = osd_declare_fallocate,
1240         .dbo_fallocate                  = osd_fallocate,
1241         .dbo_lseek                      = osd_lseek,
1242 };
1243
1244 struct dt_body_operations osd_body_scrub_ops = {
1245         .dbo_read                       = osd_read_no_record,
1246         .dbo_declare_write              = osd_declare_write,
1247         .dbo_write                      = osd_write,
1248 };