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  *
  31  * lustre/osd-zfs/osd_io.c
  32  *
  33  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  34  * Author: Mike Pershin <tappro@whamcloud.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSD
  38
  39 #include <libcfs/libcfs.h>
  40 #include <obd_support.h>
  41 #include <lustre_net.h>
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <lustre_disk.h>
  45 #include <lustre_fid.h>
  46 #include <lustre_quota.h>
  47
  48 #include "osd_internal.h"
  49
  50 #include <sys/dnode.h>
  51 #include <sys/dbuf.h>
  52 #include <sys/spa.h>
  53 #include <sys/stat.h>
  54 #include <sys/zap.h>
  55 #include <sys/spa_impl.h>
  56 #include <sys/zfs_znode.h>
  57 #include <sys/dmu_tx.h>
  58 #include <sys/dmu_objset.h>
  59 #include <sys/dsl_prop.h>
  60 #include <sys/sa_impl.h>
  61 #include <sys/txg.h>
  62
  63 static char osd_0copy_tag[] = "zerocopy";
  64
  65 static void dbuf_set_pending_evict(dmu_buf_t *db)
  66 {
  67         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
  68         dbi->db_pending_evict = TRUE;
  69 }
  70
  71 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  72 {
  73         struct obd_histogram *h = osd->od_brw_stats.bs_hist;
  74
  75         if (rw == READ) {
  76                 atomic_inc(&osd->od_r_in_flight);
  77                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  78                                  atomic_read(&osd->od_r_in_flight));
  79                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  80         } else {
  81                 atomic_inc(&osd->od_w_in_flight);
  82                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  83                                  atomic_read(&osd->od_w_in_flight));
  84                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  85         }
  86 }
  87
  88 static void record_end_io(struct osd_device *osd, int rw,
  89                           unsigned long elapsed, int disksize, int npages)
  90 {
  91         struct obd_histogram *h = osd->od_brw_stats.bs_hist;
  92
  93         if (rw == READ)
  94                 atomic_dec(&osd->od_r_in_flight);
  95         else
  96                 atomic_dec(&osd->od_w_in_flight);
  97
  98         lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
  99         if (disksize > 0)
 100                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
 101         if (elapsed)
 102                 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
 103 }
 104
 105 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 106                           struct lu_buf *buf, loff_t *pos, size_t *size)
 107 {
 108         struct osd_object *obj = osd_dt_obj(dt);
 109         uint64_t old_size;
 110         int rc;
 111
 112         LASSERT(dt_object_exists(dt));
 113         LASSERT(obj->oo_dn);
 114
 115         read_lock(&obj->oo_attr_lock);
 116         old_size = obj->oo_attr.la_size;
 117         read_unlock(&obj->oo_attr_lock);
 118
 119         if (*pos + *size > old_size) {
 120                 if (old_size < *pos)
 121                         return 0;
 122
 123                 *size = old_size - *pos;
 124         }
 125
 126         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 127                           buf->lb_buf, DMU_READ_PREFETCH);
 128         if (!rc) {
 129                 rc = *size;
 130                 *pos += *size;
 131         }
 132
 133         return rc;
 134 }
 135
 136 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 137                         struct lu_buf *buf, loff_t *pos)
 138 {
 139         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 140         size_t size = buf->lb_len;
 141         hrtime_t start = gethrtime();
 142         s64 delta_ms;
 143         int rc;
 144
 145         record_start_io(osd, READ, 0);
 146         rc = __osd_read(env, dt, buf, pos, &size);
 147         delta_ms = gethrtime() - start;
 148         do_div(delta_ms, NSEC_PER_MSEC);
 149         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 150
 151         return rc;
 152 }
 153
 154 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 155                                          struct dt_object *dt,
 156                                          struct lu_buf *buf, loff_t *pos)
 157 {
 158         size_t size = buf->lb_len;
 159
 160         return __osd_read(env, dt, buf, pos, &size);
 161 }
 162
 163 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 164                                 const struct lu_buf *buf, loff_t pos,
 165                                 struct thandle *th)
 166 {
 167         struct osd_object  *obj  = osd_dt_obj(dt);
 168         struct osd_device  *osd = osd_obj2dev(obj);
 169         struct osd_thandle *oh;
 170         uint64_t            oid;
 171         ENTRY;
 172
 173         oh = container_of(th, struct osd_thandle, ot_super);
 174
 175         /* in some cases declare can race with creation (e.g. llog)
 176          * and we need to wait till object is initialized. notice
 177          * LOHA_EXISTs is supposed to be the last step in the
 178          * initialization */
 179
 180         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 181         if (dt_object_exists(dt))
 182                 oid = obj->oo_dn->dn_object;
 183         else
 184                 oid = DMU_NEW_OBJECT;
 185
 186         /* XXX: we still miss for append declaration support in ZFS
 187          *      -1 means append which is used by llog mostly, llog
 188          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 189         if (pos == -1)
 190                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 191                             obj->oo_attr.la_size + (2 << 20));
 192         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 193
 194         /* dt_declare_write() is usually called for system objects, such
 195          * as llog or last_rcvd files. We needn't enforce quota on those
 196          * objects, so always set the lqi_space as 0. */
 197         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 198                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 199                                  0, oh, NULL, OSD_QID_BLK));
 200 }
 201
 202 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 203                         const struct lu_buf *buf, loff_t *pos,
 204                          struct thandle *th)
 205 {
 206         struct osd_object  *obj  = osd_dt_obj(dt);
 207         struct osd_device  *osd = osd_obj2dev(obj);
 208         struct osd_thandle *oh;
 209         uint64_t            offset = *pos;
 210         int                 rc;
 211
 212         ENTRY;
 213
 214         LASSERT(dt_object_exists(dt));
 215         LASSERT(obj->oo_dn);
 216
 217         LASSERT(th != NULL);
 218         oh = container_of(th, struct osd_thandle, ot_super);
 219
 220         down_read(&obj->oo_guard);
 221         if (obj->oo_destroyed)
 222                 GOTO(out, rc = -ENOENT);
 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         up_read(&obj->oo_guard);
 246         RETURN(rc);
 247 }
 248
 249 /*
 250  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 251  *      purposes as it's not very well defined ...
 252  *      instead I use the lowest bit of the address so that:
 253  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 254  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 255  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 256  *
 257  *      bzzz, to blame
 258  */
 259 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 260                         struct niobuf_local *lnb, int npages)
 261 {
 262         struct osd_object *obj  = osd_dt_obj(dt);
 263         struct osd_device *osd = osd_obj2dev(obj);
 264         unsigned long      ptr;
 265         int                i;
 266
 267         LASSERT(dt_object_exists(dt));
 268         LASSERT(obj->oo_dn);
 269
 270         for (i = 0; i < npages; i++) {
 271                 if (lnb[i].lnb_page == NULL)
 272                         continue;
 273                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 274                         /* this is anonymous page allocated for copy-write */
 275                         lnb[i].lnb_page->mapping = NULL;
 276                         __free_page(lnb[i].lnb_page);
 277                         atomic_dec(&osd->od_zerocopy_alloc);
 278                 } else {
 279                         /* see comment in osd_bufs_get_read() */
 280                         ptr = (unsigned long)lnb[i].lnb_data;
 281                         if (ptr & 1UL) {
 282                                 ptr &= ~1UL;
 283                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 284                                 atomic_dec(&osd->od_zerocopy_pin);
 285                         } else if (lnb[i].lnb_data != NULL) {
 286                                 int j, apages, abufsz;
 287                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 288                                 apages = abufsz >> PAGE_SHIFT;
 289                                 /* these references to pages must be invalidated
 290                                  * to prevent access in osd_bufs_put() */
 291                                 for (j = 0; j < apages; j++)
 292                                         lnb[i + j].lnb_page = NULL;
 293                                 dmu_return_arcbuf(lnb[i].lnb_data);
 294                                 atomic_dec(&osd->od_zerocopy_loan);
 295                         }
 296                 }
 297                 lnb[i].lnb_page = NULL;
 298                 lnb[i].lnb_data = NULL;
 299         }
 300
 301         return 0;
 302 }
 303
 304 static inline struct page *kmem_to_page(void *addr)
 305 {
 306         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 307         if (is_vmalloc_addr(addr))
 308                 return vmalloc_to_page(addr);
 309         else
 310                 return virt_to_page(addr);
 311 }
 312
 313 /**
 314  * Prepare buffers for read.
 315  *
 316  * The function maps the range described by \a off and \a len to \a lnb array.
 317  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 318  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 319  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 320  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 321  * A better implementation would just return the buffers (potentially unfilled)
 322  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 323  *
 324  * \param[in] env       environment
 325  * \param[in] obj       object
 326  * \param[in] off       offset in bytes
 327  * \param[in] len       the number of bytes to access
 328  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 329  *
 330  * \retval              0 for success
 331  * \retval              negative error number of failure
 332  */
 333 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 334                              loff_t off, ssize_t len, struct niobuf_local *lnb,
 335                              int maxlnb)
 336 {
 337         struct osd_device *osd = osd_obj2dev(obj);
 338         int rc, i, numbufs, npages = 0, drop_cache = 0;
 339         hrtime_t start = gethrtime();
 340         dmu_buf_t **dbp;
 341         s64 delta_ms;
 342
 343         ENTRY;
 344         record_start_io(osd, READ, 0);
 345
 346         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
 347                 drop_cache = 1;
 348
 349         /* grab buffers for read:
 350          * OSD API let us to grab buffers first, then initiate IO(s)
 351          * so that all required IOs will be done in parallel, but at the
 352          * moment DMU doesn't provide us with a method to grab buffers.
 353          * If we discover this is a vital for good performance we
 354          * can get own replacement for dmu_buf_hold_array_by_bonus().
 355          */
 356         while (len > 0 &&
 357                (obj->oo_dn->dn_datablkshift != 0 ||
 358                 off < obj->oo_dn->dn_datablksz)) {
 359                 if (obj->oo_dn->dn_datablkshift == 0 &&
 360                     off + len > obj->oo_dn->dn_datablksz)
 361                         len = obj->oo_dn->dn_datablksz - off;
 362
 363                 dbp = NULL;
 364                 if (unlikely(npages >= maxlnb))
 365                         GOTO(err, rc = -EOVERFLOW);
 366
 367                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 368                                                   off, len, TRUE, osd_0copy_tag,
 369                                                   &numbufs, &dbp);
 370                 if (unlikely(rc))
 371                         GOTO(err, rc);
 372
 373                 for (i = 0; i < numbufs; i++) {
 374                         int bufoff, tocpy, thispage;
 375                         void *dbf = dbp[i];
 376
 377                         LASSERT(len > 0);
 378
 379                         atomic_inc(&osd->od_zerocopy_pin);
 380
 381                         bufoff = off - dbp[i]->db_offset;
 382                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 383
 384                         /* kind of trick to differentiate dbuf vs. arcbuf */
 385                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 386                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 387
 388                         while (tocpy > 0) {
 389                                 if (unlikely(npages >= maxlnb))
 390                                         GOTO(err, rc = -EOVERFLOW);
 391
 392                                 thispage = PAGE_SIZE;
 393                                 thispage -= bufoff & (PAGE_SIZE - 1);
 394                                 thispage = min(tocpy, thispage);
 395
 396                                 lnb->lnb_rc = 0;
 397                                 lnb->lnb_file_offset = off;
 398                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 399                                 lnb->lnb_len = thispage;
 400                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 401                                                              bufoff);
 402                                 /* mark just a single slot: we need this
 403                                  * reference to dbuf to be released once */
 404                                 lnb->lnb_data = dbf;
 405                                 dbf = NULL;
 406
 407                                 tocpy -= thispage;
 408                                 len -= thispage;
 409                                 bufoff += thispage;
 410                                 off += thispage;
 411
 412                                 npages++;
 413                                 lnb++;
 414                         }
 415
 416                         if (drop_cache)
 417                                 dbuf_set_pending_evict(dbp[i]);
 418
 419                         /* steal dbuf so dmu_buf_rele_array() can't release
 420                          * it */
 421                         dbp[i] = NULL;
 422                 }
 423
 424                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 425         }
 426
 427         delta_ms = gethrtime() - start;
 428         do_div(delta_ms, NSEC_PER_MSEC);
 429         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 430
 431         RETURN(npages);
 432
 433 err:
 434         LASSERT(rc < 0);
 435         if (dbp)
 436                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 437         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 438         RETURN(rc);
 439 }
 440
 441 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 442 {
 443         arc_buf_t *abuf;
 444
 445         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 446         if (unlikely(!abuf))
 447                 return ERR_PTR(-ENOMEM);
 448
 449 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 450         /**
 451          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 452          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 453          * setting up page array for RDMA transfer. See LU-9305.
 454          */
 455         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 456                 dmu_return_arcbuf(abuf);
 457                 return NULL;
 458         }
 459 #endif
 460
 461         return abuf;
 462 }
 463
 464 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 465                               loff_t off, ssize_t len, struct niobuf_local *lnb,
 466                               int maxlnb)
 467 {
 468         struct osd_device *osd = osd_obj2dev(obj);
 469         int                poff, plen, off_in_block, sz_in_block;
 470         int                rc, i = 0, npages = 0;
 471         dnode_t *dn = obj->oo_dn;
 472         arc_buf_t *abuf;
 473         uint32_t bs = dn->dn_datablksz;
 474         ENTRY;
 475
 476         /*
 477          * currently only full blocks are subject to zerocopy approach:
 478          * so that we're sure nobody is trying to update the same block
 479          */
 480         while (len > 0) {
 481                 if (unlikely(npages >= maxlnb))
 482                         GOTO(out_err, rc = -EOVERFLOW);
 483
 484                 off_in_block = off & (bs - 1);
 485                 sz_in_block = min_t(int, bs - off_in_block, len);
 486
 487                 abuf = NULL;
 488                 if (sz_in_block == bs) {
 489                         /* full block, try to use zerocopy */
 490                         abuf = osd_request_arcbuf(dn, bs);
 491                         if (unlikely(IS_ERR(abuf)))
 492                                 GOTO(out_err, rc = PTR_ERR(abuf));
 493                 }
 494
 495                 if (abuf != NULL) {
 496                         atomic_inc(&osd->od_zerocopy_loan);
 497
 498                         /* go over pages arcbuf contains, put them as
 499                          * local niobufs for ptlrpc's bulks */
 500                         while (sz_in_block > 0) {
 501                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 502
 503                                 if (unlikely(npages >= maxlnb))
 504                                         GOTO(out_err, rc = -EOVERFLOW);
 505
 506                                 lnb[i].lnb_file_offset = off;
 507                                 lnb[i].lnb_page_offset = 0;
 508                                 lnb[i].lnb_len = plen;
 509                                 lnb[i].lnb_rc = 0;
 510                                 if (sz_in_block == bs)
 511                                         lnb[i].lnb_data = abuf;
 512                                 else
 513                                         lnb[i].lnb_data = NULL;
 514
 515                                 /* this one is not supposed to fail */
 516                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 517                                                         off_in_block);
 518                                 LASSERT(lnb[i].lnb_page);
 519
 520                                 lprocfs_counter_add(osd->od_stats,
 521                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 522
 523                                 sz_in_block -= plen;
 524                                 len -= plen;
 525                                 off += plen;
 526                                 off_in_block += plen;
 527                                 i++;
 528                                 npages++;
 529                         }
 530                 } else {
 531                         if (off_in_block == 0 && len < bs &&
 532                                         off + len >= obj->oo_attr.la_size)
 533                                 lprocfs_counter_add(osd->od_stats,
 534                                                 LPROC_OSD_TAIL_IO, 1);
 535
 536                         /* can't use zerocopy, allocate temp. buffers */
 537                         poff = off & (PAGE_SIZE - 1);
 538                         while (sz_in_block > 0) {
 539                                 plen = min_t(int, poff + sz_in_block,
 540                                              PAGE_SIZE);
 541                                 plen -= poff;
 542
 543                                 if (unlikely(npages >= maxlnb))
 544                                         GOTO(out_err, rc = -EOVERFLOW);
 545
 546                                 lnb[i].lnb_file_offset = off;
 547                                 lnb[i].lnb_page_offset = poff;
 548                                 poff = 0;
 549
 550                                 lnb[i].lnb_len = plen;
 551                                 lnb[i].lnb_rc = 0;
 552                                 lnb[i].lnb_data = NULL;
 553
 554                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 555                                 if (unlikely(lnb[i].lnb_page == NULL))
 556                                         GOTO(out_err, rc = -ENOMEM);
 557
 558                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 559                                 lnb[i].lnb_page->mapping = (void *)obj;
 560
 561                                 atomic_inc(&osd->od_zerocopy_alloc);
 562                                 lprocfs_counter_add(osd->od_stats,
 563                                                 LPROC_OSD_COPY_IO, 1);
 564
 565                                 sz_in_block -= plen;
 566                                 len -= plen;
 567                                 off += plen;
 568                                 i++;
 569                                 npages++;
 570                         }
 571                 }
 572         }
 573
 574         RETURN(npages);
 575
 576 out_err:
 577         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 578         RETURN(rc);
 579 }
 580
 581 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 582                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 583                         int maxlnb, enum dt_bufs_type rw)
 584 {
 585         struct osd_object *obj  = osd_dt_obj(dt);
 586         int                rc;
 587
 588         LASSERT(dt_object_exists(dt));
 589         LASSERT(obj->oo_dn);
 590
 591         if (rw & DT_BUFS_TYPE_WRITE)
 592                 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
 593         else
 594                 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
 595
 596         return rc;
 597 }
 598
 599 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 600                         struct niobuf_local *lnb, int npages)
 601 {
 602         struct osd_object *obj = osd_dt_obj(dt);
 603
 604         LASSERT(dt_object_exists(dt));
 605         LASSERT(obj->oo_dn);
 606
 607         return 0;
 608 }
 609
 610 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 611                                            uint64_t offset,
 612                                            uint32_t blksz)
 613 {
 614         LASSERT(blksz > 0);
 615
 616         size += offset % blksz;
 617
 618         if (likely(is_power_of_2(blksz)))
 619                 return round_up(size, blksz);
 620         else
 621                 return DIV_ROUND_UP_ULL(size, blksz) * 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 we could handle this on per-lnb basis as done by
 666                  * grant. */
 667                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 668                     (lnb[i].lnb_flags & OBD_BRW_SYS_RESOURCE) ||
 669                     !(lnb[i].lnb_flags & OBD_BRW_SYNC))
 670                         declare_flags |= OSD_QID_FORCE;
 671
 672                 if (size == 0) {
 673                         /* first valid lnb */
 674                         offset = lnb[i].lnb_file_offset;
 675                         size = lnb[i].lnb_len;
 676                         continue;
 677                 }
 678                 if (offset + size == lnb[i].lnb_file_offset) {
 679                         /* this lnb is contiguous to the previous one */
 680                         size += lnb[i].lnb_len;
 681                         continue;
 682                 }
 683
 684                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 685                                   obj->oo_dn, offset, size);
 686                 /* Estimating space to be consumed by a write is rather
 687                  * complicated with ZFS. As a consequence, we don't account for
 688                  * indirect blocks and just use as a rough estimate the worse
 689                  * case where the old space is being held by a snapshot. Quota
 690                  * overrun will be adjusted once the operation is committed, if
 691                  * required. */
 692                 space += osd_roundup2blocksz(size, offset, blksz);
 693
 694                 offset = lnb[i].lnb_file_offset;
 695                 size = lnb[i].lnb_len;
 696         }
 697
 698         if (size) {
 699                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 700                                   offset, size);
 701                 space += osd_roundup2blocksz(size, offset, blksz);
 702         }
 703
 704         /* backend zfs filesystem might be configured to store multiple data
 705          * copies */
 706         space  *= osd->od_os->os_copies;
 707         space   = toqb(space);
 708         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 709                npages, space);
 710
 711         record_start_io(osd, WRITE, discont_pages);
 712 retry:
 713         /* acquire quota space if needed */
 714         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 715                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 716                                space, oh, &local_flags, declare_flags);
 717
 718         if (!synced && rc == -EDQUOT &&
 719             (local_flags & QUOTA_FL_SYNC) != 0) {
 720                 dt_sync(env, th->th_dev);
 721                 synced = true;
 722                 CDEBUG(D_QUOTA, "retry after sync\n");
 723                 local_flags = 0;
 724                 goto retry;
 725         }
 726
 727         /* we need only to store the overquota flags in the first lnb for
 728          * now, once we support multiple objects BRW, this code needs be
 729          * revised. */
 730         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
 731                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 732         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
 733                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 734 #ifdef ZFS_PROJINHERIT
 735         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
 736                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 737 #endif
 738
 739         RETURN(rc);
 740 }
 741
 742 /**
 743  * Policy to grow ZFS block size by write pattern.
 744  * For sequential write, it grows block size gradually until it reaches the
 745  * maximum blocksize the dataset can support. Otherwise, it will pick a
 746  * a block size by the writing region of this I/O.
 747  */
 748 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 749                               uint64_t start, uint64_t end)
 750 {
 751         struct osd_device       *osd = osd_obj2dev(obj);
 752         dnode_t *dn = obj->oo_dn;
 753         uint32_t                 blksz;
 754         int                      rc = 0;
 755
 756         ENTRY;
 757
 758         if (dn->dn_maxblkid > 0) /* can't change block size */
 759                 GOTO(out, rc);
 760
 761         if (dn->dn_datablksz >= osd->od_max_blksz)
 762                 GOTO(out, rc);
 763
 764         down_write(&obj->oo_guard);
 765
 766         blksz = dn->dn_datablksz;
 767         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 768                 GOTO(out_unlock, rc);
 769
 770         /* now ZFS can support up to 16MB block size, and if the write
 771          * is sequential, it just increases the block size gradually */
 772         if (start <= blksz) { /* sequential */
 773                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 774         } else { /* sparse, pick a block size by write region */
 775                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 776                                         end - start);
 777         }
 778
 779         if (!is_power_of_2(blksz))
 780                 blksz = size_roundup_power2(blksz);
 781
 782         if (blksz > dn->dn_datablksz) {
 783                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 784                                                blksz, 0, oh->ot_tx);
 785                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 786                 if (rc < 0)
 787                         CDEBUG(D_INODE, "object "DFID": change block size"
 788                                "%u -> %u error rc = %d\n",
 789                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 790                                dn->dn_datablksz, blksz, rc);
 791         }
 792         EXIT;
 793 out_unlock:
 794         up_write(&obj->oo_guard);
 795 out:
 796         return rc;
 797 }
 798
 799 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 800                                         loff_t off, ssize_t len)
 801 {
 802         dmu_buf_t **dbp;
 803         int i, rc, numbufs;
 804
 805         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 806                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 807         if (unlikely(rc))
 808                 return;
 809
 810         for (i = 0; i < numbufs; i++)
 811                 dbuf_set_pending_evict(dbp[i]);
 812
 813         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 814 }
 815
 816 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 817                         struct niobuf_local *lnb, int npages,
 818                         struct thandle *th, __u64 user_size)
 819 {
 820         struct osd_object  *obj  = osd_dt_obj(dt);
 821         struct osd_device  *osd = osd_obj2dev(obj);
 822         struct osd_thandle *oh;
 823         uint64_t            new_size = 0;
 824         int                 i, abufsz, rc = 0, drop_cache = 0;
 825         unsigned long      iosize = 0;
 826         ENTRY;
 827
 828         LASSERT(dt_object_exists(dt));
 829         LASSERT(obj->oo_dn);
 830
 831         LASSERT(th != NULL);
 832         oh = container_of(th, struct osd_thandle, ot_super);
 833
 834         /* adjust block size. Assume the buffers are sorted. */
 835         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 836                                  lnb[npages - 1].lnb_file_offset +
 837                                  lnb[npages - 1].lnb_len);
 838
 839         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 840             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 841             osd->od_readcache_max_filesize)
 842                 drop_cache = 1;
 843
 844         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 845                 RETURN(-ENOSPC);
 846
 847         /* if la_size is already bigger than specified user_size,
 848          * ignore user_size
 849          */
 850         if (obj->oo_attr.la_size > user_size)
 851                 user_size = 0;
 852
 853         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 854          * size and assigning arcbuf take place at the same time.
 855          *
 856          * Thread 1:
 857          * osd_write_commit()
 858          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 859          *   -> dmu_object_set_blocksize()
 860          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 861          *       write lock held
 862          *     -> dbuf_new_size()
 863          *      -> dmu_buf_will_dirty()
 864          *       -> dbuf_read()
 865          *        -> wait for the dbuf state to change
 866          * Thread 2:
 867          * osd_write_commit()
 868          *  -> dmu_assign_arcbuf()
 869          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 870          *    -> dbuf_dirty()
 871          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 872          *
 873          * By taking the read lock, it can avoid thread 2 to enter into the
 874          * critical section of assigning the arcbuf, while thread 1 is
 875          * changing the block size.
 876          */
 877         down_read(&obj->oo_guard);
 878         if (obj->oo_destroyed) {
 879                 up_read(&obj->oo_guard);
 880                 RETURN(-ENOENT);
 881         }
 882
 883         for (i = 0; i < npages; i++) {
 884                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 885                         (unsigned) lnb[i].lnb_len,
 886                         (unsigned) lnb[i].lnb_file_offset);
 887
 888                 if (lnb[i].lnb_rc) {
 889                         /* ENOSPC, network RPC error, etc.
 890                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 891                          * so we skip this page if lnb_rc is set to -ENOSPC */
 892                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 893                                 PFID(lu_object_fid(&dt->do_lu)), i,
 894                                 lnb[i].lnb_rc);
 895                         continue;
 896                 }
 897
 898                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 899                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 900                 if (lnb[i].lnb_page == NULL)
 901                         continue;
 902
 903                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 904                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 905                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
 906                                       lnb[i].lnb_page_offset, oh->ot_tx);
 907                         kunmap(lnb[i].lnb_page);
 908                         iosize += lnb[i].lnb_len;
 909                         abufsz = lnb[i].lnb_len; /* to drop cache below */
 910                 } else if (lnb[i].lnb_data) {
 911                         int j, apages;
 912                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 913                         /* buffer loaned for zerocopy, try to use it.
 914                          * notice that dmu_assign_arcbuf() is smart
 915                          * enough to recognize changed blocksize
 916                          * in this case it fallbacks to dmu_write() */
 917                         abufsz = arc_buf_size(lnb[i].lnb_data);
 918                         LASSERT(abufsz & PAGE_MASK);
 919                         apages = abufsz >> PAGE_SHIFT;
 920                         LASSERT(i + apages <= npages);
 921                         /* these references to pages must be invalidated
 922                          * to prevent access in osd_bufs_put() */
 923                         for (j = 0; j < apages; j++)
 924                                 lnb[i + j].lnb_page = NULL;
 925                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 926                                           lnb[i].lnb_file_offset,
 927                                           lnb[i].lnb_data, oh->ot_tx);
 928                         /* drop the reference, otherwise osd_put_bufs()
 929                          * will be releasing it - bad! */
 930                         lnb[i].lnb_data = NULL;
 931                         atomic_dec(&osd->od_zerocopy_loan);
 932                         iosize += abufsz;
 933                 } else {
 934                         /* we don't want to deal with cache if nothing
 935                          * has been send to ZFS at this step */
 936                         continue;
 937                 }
 938
 939                 if (!drop_cache)
 940                         continue;
 941
 942                 /* we have to mark dbufs for eviction here because
 943                  * dmu_assign_arcbuf() may create a new dbuf for
 944                  * loaned abuf */
 945                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
 946                                             abufsz);
 947         }
 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                 up_read(&obj->oo_guard);
 955                 record_end_io(osd, WRITE, 0, 0, 0);
 956                 RETURN(0);
 957         }
 958
 959         /* if file has grown, take user_size into account */
 960         if (user_size && new_size > user_size)
 961                 new_size = user_size;
 962         write_lock(&obj->oo_attr_lock);
 963         if (obj->oo_attr.la_size < new_size) {
 964                 obj->oo_attr.la_size = new_size;
 965                 write_unlock(&obj->oo_attr_lock);
 966                 /* osd_object_sa_update() will be copying directly from
 967                  * oo_attr into dbuf. any update within a single txg will copy
 968                  * the most actual */
 969                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 970                                           &obj->oo_attr.la_size, 8, oh);
 971         } else {
 972                 write_unlock(&obj->oo_attr_lock);
 973         }
 974
 975         up_read(&obj->oo_guard);
 976
 977         record_end_io(osd, WRITE, 0, iosize, npages);
 978
 979         RETURN(rc);
 980 }
 981
 982 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 983                         struct niobuf_local *lnb, int npages)
 984 {
 985         struct osd_object *obj  = osd_dt_obj(dt);
 986         int                i;
 987         loff_t             eof;
 988
 989         LASSERT(dt_object_exists(dt));
 990         LASSERT(obj->oo_dn);
 991
 992         read_lock(&obj->oo_attr_lock);
 993         eof = obj->oo_attr.la_size;
 994         read_unlock(&obj->oo_attr_lock);
 995
 996         for (i = 0; i < npages; i++) {
 997                 if (unlikely(lnb[i].lnb_rc < 0))
 998                         continue;
 999
1000                 lnb[i].lnb_rc = lnb[i].lnb_len;
1001
1002                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
1003                         /* send complete pages all the time */
1004                         if (eof <= lnb[i].lnb_file_offset)
1005                                 lnb[i].lnb_rc = 0;
1006
1007                         /* all subsequent rc should be 0 */
1008                         while (++i < npages)
1009                                 lnb[i].lnb_rc = 0;
1010                         break;
1011                 }
1012         }
1013
1014         return 0;
1015 }
1016
1017 /*
1018  * Punch/truncate an object
1019  *
1020  *      IN:     db  - dmu_buf of the object to free data in.
1021  *              off - start of section to free.
1022  *              len - length of section to free (DMU_OBJECT_END => to EOF).
1023  *
1024  *      RETURN: 0 if success
1025  *              error code if failure
1026  *
1027  * The transaction passed to this routine must have
1028  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1029  * called and then assigned to a transaction group.
1030  */
1031 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1032                               dmu_tx_t *tx, uint64_t off, uint64_t len)
1033 {
1034         dnode_t *dn = obj->oo_dn;
1035         uint64_t size = obj->oo_attr.la_size;
1036         int rc = 0;
1037
1038         /* Assert that the transaction has been assigned to a
1039            transaction group. */
1040         LASSERT(tx->tx_txg != 0);
1041         /*
1042          * Nothing to do if file already at desired length.
1043          */
1044         if (len == DMU_OBJECT_END && size == off)
1045                 return 0;
1046
1047         /* if object holds encrypted content, we need to make sure we truncate
1048          * on an encryption unit boundary, or subsequent reads will get
1049          * corrupted content
1050          */
1051         if (len != DMU_OBJECT_END)
1052                 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1053                         (off & ~LUSTRE_ENCRYPTION_MASK);
1054         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1055             off & ~LUSTRE_ENCRYPTION_MASK)
1056                 off = (off & LUSTRE_ENCRYPTION_MASK) +
1057                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1058
1059
1060         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1061         if (off < size)
1062                 dmu_free_range(os, dn->dn_object, off, len, tx);
1063
1064         return rc;
1065 }
1066
1067 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1068                         __u64 start, __u64 end, struct thandle *th)
1069 {
1070         struct osd_object  *obj = osd_dt_obj(dt);
1071         struct osd_device  *osd = osd_obj2dev(obj);
1072         struct osd_thandle *oh;
1073         __u64               len;
1074         int                 rc = 0;
1075         ENTRY;
1076
1077         LASSERT(dt_object_exists(dt));
1078         LASSERT(osd_invariant(obj));
1079
1080         LASSERT(th != NULL);
1081         oh = container_of(th, struct osd_thandle, ot_super);
1082
1083         write_lock(&obj->oo_attr_lock);
1084         /* truncate */
1085         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1086                 len = DMU_OBJECT_END;
1087         else
1088                 len = end - start;
1089         write_unlock(&obj->oo_attr_lock);
1090
1091         down_read(&obj->oo_guard);
1092         if (obj->oo_destroyed)
1093                 GOTO(out, rc = -ENOENT);
1094
1095         rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1096
1097         /* set new size */
1098         if (len == DMU_OBJECT_END) {
1099                 write_lock(&obj->oo_attr_lock);
1100                 obj->oo_attr.la_size = start;
1101                 write_unlock(&obj->oo_attr_lock);
1102                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1103                                           &obj->oo_attr.la_size, 8, oh);
1104         }
1105 out:
1106         up_read(&obj->oo_guard);
1107         RETURN(rc);
1108 }
1109
1110 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1111                         __u64 start, __u64 end, struct thandle *handle)
1112 {
1113         struct osd_object  *obj = osd_dt_obj(dt);
1114         struct osd_device  *osd = osd_obj2dev(obj);
1115         struct osd_thandle *oh;
1116         __u64               len;
1117         ENTRY;
1118
1119         oh = container_of(handle, struct osd_thandle, ot_super);
1120
1121         read_lock(&obj->oo_attr_lock);
1122         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1123                 len = DMU_OBJECT_END;
1124         else
1125                 len = end - start;
1126
1127         /* declare we'll free some blocks ... */
1128         /* if object holds encrypted content, we need to make sure we truncate
1129          * on an encryption unit boundary, or subsequent reads will get
1130          * corrupted content
1131          */
1132         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1133             start & ~LUSTRE_ENCRYPTION_MASK)
1134                 start = (start & LUSTRE_ENCRYPTION_MASK) +
1135                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1136         if (start < obj->oo_attr.la_size) {
1137                 read_unlock(&obj->oo_attr_lock);
1138                 dmu_tx_mark_netfree(oh->ot_tx);
1139                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1140         } else {
1141                 read_unlock(&obj->oo_attr_lock);
1142         }
1143
1144         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1145                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1146                                  0, oh, NULL, OSD_QID_BLK));
1147 }
1148
1149 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1150                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1151 {
1152         int     rc;
1153         ENTRY;
1154
1155         switch (advice) {
1156         default:
1157                 rc = -ENOTSUPP;
1158                 break;
1159         }
1160
1161         RETURN(rc);
1162 }
1163
1164 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1165                          __u64 start, __u64 end, int mode, struct thandle *th)
1166 {
1167         int rc = -EOPNOTSUPP;
1168         ENTRY;
1169
1170          /*
1171           * space preallocation is not supported for ZFS
1172           * Returns -EOPNOTSUPP for now
1173           */
1174         RETURN(rc);
1175 }
1176
1177 static int osd_declare_fallocate(const struct lu_env *env,
1178                                  struct dt_object *dt, __u64 start, __u64 end,
1179                                  int mode, struct thandle *th)
1180 {
1181         int rc = -EOPNOTSUPP;
1182         ENTRY;
1183
1184          /*
1185           * space preallocation is not supported for ZFS
1186           * Returns -EOPNOTSUPP for now
1187           */
1188         RETURN(rc);
1189 }
1190
1191 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1192                         loff_t offset, int whence)
1193 {
1194         struct osd_object *obj = osd_dt_obj(dt);
1195         struct osd_device *osd = osd_obj2dev(obj);
1196         uint64_t size = obj->oo_attr.la_size;
1197         uint64_t result = offset;
1198         int rc;
1199         boolean_t hole = whence == SEEK_HOLE;
1200
1201         ENTRY;
1202
1203         LASSERT(dt_object_exists(dt));
1204         LASSERT(osd_invariant(obj));
1205         LASSERT(offset >= 0);
1206
1207         /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
1208          * hole. LOV to decide after all if that real hole or not.
1209          */
1210         if (offset >= size)
1211                 RETURN(hole ? offset : -ENXIO);
1212
1213         /* Currently ZFS reports no valid DATA offset if object has dirty data
1214          * and we cannot just switch to generic way with reporting DATA on all
1215          * file offsets and HOLE beyond end of file, because we may get HOLE
1216          * reported correctly at some offset inside file then DATA will find
1217          * dirty state and be reported also at that offset by generic approach.
1218          * This is because for HOLE report ZFS doesn't check dirty state but
1219          * does for DATA.
1220          * The only way to get reliable results is to call txg_wait_synced()
1221          * when ZFS reports EBUSY result and repeat lseek call and that is
1222          * controlled via od_sync_on_lseek option.
1223          */
1224         if (!osd->od_sync_on_lseek)
1225                 result = hole ? size : offset;
1226
1227 again:
1228         rc = osd_dmu_offset_next(osd->od_os, obj->oo_dn->dn_object, hole,
1229                                  &result);
1230         /* dirty inode, lseek result is unreliable without sync */
1231         if (rc == EBUSY) {
1232                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1233                 goto again;
1234         }
1235
1236         if (rc == ESRCH)
1237                 RETURN(-ENXIO);
1238
1239         /* ZFS is not exported all needed function, so fall back to the
1240          * generic logic: for HOLE return file size, for DATA return
1241          * the current offset
1242          */
1243         if (rc == EOPNOTSUPP)
1244                 result = hole ? size : offset;
1245         else if (rc)
1246                 return -rc;
1247
1248         /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
1249          * result as end of the last block instead of logical EOF which we need
1250          */
1251         if (result > size)
1252                 result = size;
1253
1254         RETURN(result);
1255 }
1256
1257 const struct dt_body_operations osd_body_ops = {
1258         .dbo_read                       = osd_read,
1259         .dbo_declare_write              = osd_declare_write,
1260         .dbo_write                      = osd_write,
1261         .dbo_bufs_get                   = osd_bufs_get,
1262         .dbo_bufs_put                   = osd_bufs_put,
1263         .dbo_write_prep                 = osd_write_prep,
1264         .dbo_declare_write_commit       = osd_declare_write_commit,
1265         .dbo_write_commit               = osd_write_commit,
1266         .dbo_read_prep                  = osd_read_prep,
1267         .dbo_declare_punch              = osd_declare_punch,
1268         .dbo_punch                      = osd_punch,
1269         .dbo_ladvise                    = osd_ladvise,
1270         .dbo_declare_fallocate          = osd_declare_fallocate,
1271         .dbo_fallocate                  = osd_fallocate,
1272         .dbo_lseek                      = osd_lseek,
1273 };
1274
1275 const struct dt_body_operations osd_body_scrub_ops = {
1276         .dbo_read                       = osd_read_no_record,
1277         .dbo_declare_write              = osd_declare_write,
1278         .dbo_write                      = osd_write,
1279 };