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.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
  81         } else {
  82                 atomic_inc(&osd->od_w_in_flight);
  83                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  84                                  atomic_read(&osd->od_w_in_flight));
  85                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  86
  87         }
  88 }
  89
  90 static void record_end_io(struct osd_device *osd, int rw,
  91                           unsigned long elapsed, int disksize, int npages)
  92 {
  93         struct obd_histogram *h = osd->od_brw_stats.hist;
  94
  95         if (rw == READ)
  96                 atomic_dec(&osd->od_r_in_flight);
  97         else
  98                 atomic_dec(&osd->od_w_in_flight);
  99
 100         lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
 101         if (disksize > 0)
 102                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
 103         if (elapsed)
 104                 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
 105 }
 106
 107 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 108                           struct lu_buf *buf, loff_t *pos, size_t *size)
 109 {
 110         struct osd_object *obj = osd_dt_obj(dt);
 111         uint64_t old_size;
 112         int rc;
 113
 114         LASSERT(dt_object_exists(dt));
 115         LASSERT(obj->oo_dn);
 116
 117         read_lock(&obj->oo_attr_lock);
 118         old_size = obj->oo_attr.la_size;
 119         read_unlock(&obj->oo_attr_lock);
 120
 121         if (*pos + *size > old_size) {
 122                 if (old_size < *pos)
 123                         return 0;
 124
 125                 *size = old_size - *pos;
 126         }
 127
 128         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 129                           buf->lb_buf, DMU_READ_PREFETCH);
 130         if (!rc) {
 131                 rc = *size;
 132                 *pos += *size;
 133         }
 134
 135         return rc;
 136 }
 137
 138 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 139                         struct lu_buf *buf, loff_t *pos)
 140 {
 141         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 142         size_t size = buf->lb_len;
 143         hrtime_t start = gethrtime();
 144         s64 delta_ms;
 145         int rc;
 146
 147         record_start_io(osd, READ, 0);
 148         rc = __osd_read(env, dt, buf, pos, &size);
 149         delta_ms = gethrtime() - start;
 150         do_div(delta_ms, NSEC_PER_MSEC);
 151         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 152
 153         return rc;
 154 }
 155
 156 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 157                                          struct dt_object *dt,
 158                                          struct lu_buf *buf, loff_t *pos)
 159 {
 160         size_t size = buf->lb_len;
 161
 162         return __osd_read(env, dt, buf, pos, &size);
 163 }
 164
 165 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 166                                 const struct lu_buf *buf, loff_t pos,
 167                                 struct thandle *th)
 168 {
 169         struct osd_object  *obj  = osd_dt_obj(dt);
 170         struct osd_device  *osd = osd_obj2dev(obj);
 171         struct osd_thandle *oh;
 172         uint64_t            oid;
 173         ENTRY;
 174
 175         oh = container_of(th, struct osd_thandle, ot_super);
 176
 177         /* in some cases declare can race with creation (e.g. llog)
 178          * and we need to wait till object is initialized. notice
 179          * LOHA_EXISTs is supposed to be the last step in the
 180          * initialization */
 181
 182         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 183         if (dt_object_exists(dt))
 184                 oid = obj->oo_dn->dn_object;
 185         else
 186                 oid = DMU_NEW_OBJECT;
 187
 188         /* XXX: we still miss for append declaration support in ZFS
 189          *      -1 means append which is used by llog mostly, llog
 190          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 191         if (pos == -1)
 192                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 193                             obj->oo_attr.la_size + (2 << 20));
 194         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 195
 196         /* dt_declare_write() is usually called for system objects, such
 197          * as llog or last_rcvd files. We needn't enforce quota on those
 198          * objects, so always set the lqi_space as 0. */
 199         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 200                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 201                                  0, oh, NULL, OSD_QID_BLK));
 202 }
 203
 204 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 205                         const struct lu_buf *buf, loff_t *pos,
 206                          struct thandle *th)
 207 {
 208         struct osd_object  *obj  = osd_dt_obj(dt);
 209         struct osd_device  *osd = osd_obj2dev(obj);
 210         struct osd_thandle *oh;
 211         uint64_t            offset = *pos;
 212         int                 rc;
 213
 214         ENTRY;
 215
 216         LASSERT(dt_object_exists(dt));
 217         LASSERT(obj->oo_dn);
 218
 219         LASSERT(th != NULL);
 220         oh = container_of(th, struct osd_thandle, ot_super);
 221
 222         down_read(&obj->oo_guard);
 223         if (obj->oo_destroyed)
 224                 GOTO(out, rc = -ENOENT);
 225
 226         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 227                       buf->lb_buf, oh->ot_tx);
 228         write_lock(&obj->oo_attr_lock);
 229         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 230                 obj->oo_attr.la_size = offset + buf->lb_len;
 231                 write_unlock(&obj->oo_attr_lock);
 232                 /* osd_object_sa_update() will be copying directly from oo_attr
 233                  * into dbuf.  any update within a single txg will copy the
 234                  * most actual */
 235                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 236                                         &obj->oo_attr.la_size, 8, oh);
 237                 if (unlikely(rc))
 238                         GOTO(out, rc);
 239         } else {
 240                 write_unlock(&obj->oo_attr_lock);
 241         }
 242
 243         *pos += buf->lb_len;
 244         rc = buf->lb_len;
 245
 246 out:
 247         up_read(&obj->oo_guard);
 248         RETURN(rc);
 249 }
 250
 251 /*
 252  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 253  *      purposes as it's not very well defined ...
 254  *      instead I use the lowest bit of the address so that:
 255  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 256  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 257  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 258  *
 259  *      bzzz, to blame
 260  */
 261 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 262                         struct niobuf_local *lnb, int npages)
 263 {
 264         struct osd_object *obj  = osd_dt_obj(dt);
 265         struct osd_device *osd = osd_obj2dev(obj);
 266         unsigned long      ptr;
 267         int                i;
 268
 269         LASSERT(dt_object_exists(dt));
 270         LASSERT(obj->oo_dn);
 271
 272         for (i = 0; i < npages; i++) {
 273                 if (lnb[i].lnb_page == NULL)
 274                         continue;
 275                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 276                         /* this is anonymous page allocated for copy-write */
 277                         lnb[i].lnb_page->mapping = NULL;
 278                         __free_page(lnb[i].lnb_page);
 279                         atomic_dec(&osd->od_zerocopy_alloc);
 280                 } else {
 281                         /* see comment in osd_bufs_get_read() */
 282                         ptr = (unsigned long)lnb[i].lnb_data;
 283                         if (ptr & 1UL) {
 284                                 ptr &= ~1UL;
 285                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 286                                 atomic_dec(&osd->od_zerocopy_pin);
 287                         } else if (lnb[i].lnb_data != NULL) {
 288                                 int j, apages, abufsz;
 289                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 290                                 apages = abufsz >> PAGE_SHIFT;
 291                                 /* these references to pages must be invalidated
 292                                  * to prevent access in osd_bufs_put() */
 293                                 for (j = 0; j < apages; j++)
 294                                         lnb[i + j].lnb_page = NULL;
 295                                 dmu_return_arcbuf(lnb[i].lnb_data);
 296                                 atomic_dec(&osd->od_zerocopy_loan);
 297                         }
 298                 }
 299                 lnb[i].lnb_page = NULL;
 300                 lnb[i].lnb_data = NULL;
 301         }
 302
 303         return 0;
 304 }
 305
 306 static inline struct page *kmem_to_page(void *addr)
 307 {
 308         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 309         if (is_vmalloc_addr(addr))
 310                 return vmalloc_to_page(addr);
 311         else
 312                 return virt_to_page(addr);
 313 }
 314
 315 /**
 316  * Prepare buffers for read.
 317  *
 318  * The function maps the range described by \a off and \a len to \a lnb array.
 319  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 320  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 321  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 322  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 323  * A better implementation would just return the buffers (potentially unfilled)
 324  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 325  *
 326  * \param[in] env       environment
 327  * \param[in] obj       object
 328  * \param[in] off       offset in bytes
 329  * \param[in] len       the number of bytes to access
 330  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 331  *
 332  * \retval              0 for success
 333  * \retval              negative error number of failure
 334  */
 335 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 336                              loff_t off, ssize_t len, struct niobuf_local *lnb,
 337                              int maxlnb)
 338 {
 339         struct osd_device *osd = osd_obj2dev(obj);
 340         int rc, i, numbufs, npages = 0, drop_cache = 0;
 341         hrtime_t start = gethrtime();
 342         dmu_buf_t **dbp;
 343         s64 delta_ms;
 344
 345         ENTRY;
 346         record_start_io(osd, READ, 0);
 347
 348         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
 349                 drop_cache = 1;
 350
 351         /* grab buffers for read:
 352          * OSD API let us to grab buffers first, then initiate IO(s)
 353          * so that all required IOs will be done in parallel, but at the
 354          * moment DMU doesn't provide us with a method to grab buffers.
 355          * If we discover this is a vital for good performance we
 356          * can get own replacement for dmu_buf_hold_array_by_bonus().
 357          */
 358         while (len > 0 &&
 359                (obj->oo_dn->dn_datablkshift != 0 ||
 360                 off < obj->oo_dn->dn_datablksz)) {
 361                 if (obj->oo_dn->dn_datablkshift == 0 &&
 362                     off + len > obj->oo_dn->dn_datablksz)
 363                         len = obj->oo_dn->dn_datablksz - off;
 364
 365                 dbp = NULL;
 366                 if (unlikely(npages >= maxlnb))
 367                         GOTO(err, rc = -EOVERFLOW);
 368
 369                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 370                                                   off, len, TRUE, osd_0copy_tag,
 371                                                   &numbufs, &dbp);
 372                 if (unlikely(rc))
 373                         GOTO(err, rc);
 374
 375                 for (i = 0; i < numbufs; i++) {
 376                         int bufoff, tocpy, thispage;
 377                         void *dbf = dbp[i];
 378
 379                         LASSERT(len > 0);
 380
 381                         atomic_inc(&osd->od_zerocopy_pin);
 382
 383                         bufoff = off - dbp[i]->db_offset;
 384                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 385
 386                         /* kind of trick to differentiate dbuf vs. arcbuf */
 387                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 388                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 389
 390                         while (tocpy > 0) {
 391                                 if (unlikely(npages >= maxlnb))
 392                                         GOTO(err, rc = -EOVERFLOW);
 393
 394                                 thispage = PAGE_SIZE;
 395                                 thispage -= bufoff & (PAGE_SIZE - 1);
 396                                 thispage = min(tocpy, thispage);
 397
 398                                 lnb->lnb_rc = 0;
 399                                 lnb->lnb_file_offset = off;
 400                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 401                                 lnb->lnb_len = thispage;
 402                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 403                                                              bufoff);
 404                                 /* mark just a single slot: we need this
 405                                  * reference to dbuf to be released once */
 406                                 lnb->lnb_data = dbf;
 407                                 dbf = NULL;
 408
 409                                 tocpy -= thispage;
 410                                 len -= thispage;
 411                                 bufoff += thispage;
 412                                 off += thispage;
 413
 414                                 npages++;
 415                                 lnb++;
 416                         }
 417
 418                         if (drop_cache)
 419                                 dbuf_set_pending_evict(dbp[i]);
 420
 421                         /* steal dbuf so dmu_buf_rele_array() can't release
 422                          * it */
 423                         dbp[i] = NULL;
 424                 }
 425
 426                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 427         }
 428
 429         delta_ms = gethrtime() - start;
 430         do_div(delta_ms, NSEC_PER_MSEC);
 431         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 432
 433         RETURN(npages);
 434
 435 err:
 436         LASSERT(rc < 0);
 437         if (dbp)
 438                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 439         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 440         RETURN(rc);
 441 }
 442
 443 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 444 {
 445         arc_buf_t *abuf;
 446
 447         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 448         if (unlikely(!abuf))
 449                 return ERR_PTR(-ENOMEM);
 450
 451 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 452         /**
 453          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 454          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 455          * setting up page array for RDMA transfer. See LU-9305.
 456          */
 457         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 458                 dmu_return_arcbuf(abuf);
 459                 return NULL;
 460         }
 461 #endif
 462
 463         return abuf;
 464 }
 465
 466 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 467                               loff_t off, ssize_t len, struct niobuf_local *lnb,
 468                               int maxlnb)
 469 {
 470         struct osd_device *osd = osd_obj2dev(obj);
 471         int                poff, plen, off_in_block, sz_in_block;
 472         int                rc, i = 0, npages = 0;
 473         dnode_t *dn = obj->oo_dn;
 474         arc_buf_t *abuf;
 475         uint32_t bs = dn->dn_datablksz;
 476         ENTRY;
 477
 478         /*
 479          * currently only full blocks are subject to zerocopy approach:
 480          * so that we're sure nobody is trying to update the same block
 481          */
 482         while (len > 0) {
 483                 if (unlikely(npages >= maxlnb))
 484                         GOTO(out_err, rc = -EOVERFLOW);
 485
 486                 off_in_block = off & (bs - 1);
 487                 sz_in_block = min_t(int, bs - off_in_block, len);
 488
 489                 abuf = NULL;
 490                 if (sz_in_block == bs) {
 491                         /* full block, try to use zerocopy */
 492                         abuf = osd_request_arcbuf(dn, bs);
 493                         if (unlikely(IS_ERR(abuf)))
 494                                 GOTO(out_err, rc = PTR_ERR(abuf));
 495                 }
 496
 497                 if (abuf != NULL) {
 498                         atomic_inc(&osd->od_zerocopy_loan);
 499
 500                         /* go over pages arcbuf contains, put them as
 501                          * local niobufs for ptlrpc's bulks */
 502                         while (sz_in_block > 0) {
 503                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 504
 505                                 if (unlikely(npages >= maxlnb))
 506                                         GOTO(out_err, rc = -EOVERFLOW);
 507
 508                                 lnb[i].lnb_file_offset = off;
 509                                 lnb[i].lnb_page_offset = 0;
 510                                 lnb[i].lnb_len = plen;
 511                                 lnb[i].lnb_rc = 0;
 512                                 if (sz_in_block == bs)
 513                                         lnb[i].lnb_data = abuf;
 514                                 else
 515                                         lnb[i].lnb_data = NULL;
 516
 517                                 /* this one is not supposed to fail */
 518                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 519                                                         off_in_block);
 520                                 LASSERT(lnb[i].lnb_page);
 521
 522                                 lprocfs_counter_add(osd->od_stats,
 523                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 524
 525                                 sz_in_block -= plen;
 526                                 len -= plen;
 527                                 off += plen;
 528                                 off_in_block += plen;
 529                                 i++;
 530                                 npages++;
 531                         }
 532                 } else {
 533                         if (off_in_block == 0 && len < bs &&
 534                                         off + len >= obj->oo_attr.la_size)
 535                                 lprocfs_counter_add(osd->od_stats,
 536                                                 LPROC_OSD_TAIL_IO, 1);
 537
 538                         /* can't use zerocopy, allocate temp. buffers */
 539                         poff = off & (PAGE_SIZE - 1);
 540                         while (sz_in_block > 0) {
 541                                 plen = min_t(int, poff + sz_in_block,
 542                                              PAGE_SIZE);
 543                                 plen -= poff;
 544
 545                                 if (unlikely(npages >= maxlnb))
 546                                         GOTO(out_err, rc = -EOVERFLOW);
 547
 548                                 lnb[i].lnb_file_offset = off;
 549                                 lnb[i].lnb_page_offset = poff;
 550                                 poff = 0;
 551
 552                                 lnb[i].lnb_len = plen;
 553                                 lnb[i].lnb_rc = 0;
 554                                 lnb[i].lnb_data = NULL;
 555
 556                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 557                                 if (unlikely(lnb[i].lnb_page == NULL))
 558                                         GOTO(out_err, rc = -ENOMEM);
 559
 560                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 561                                 lnb[i].lnb_page->mapping = (void *)obj;
 562
 563                                 atomic_inc(&osd->od_zerocopy_alloc);
 564                                 lprocfs_counter_add(osd->od_stats,
 565                                                 LPROC_OSD_COPY_IO, 1);
 566
 567                                 sz_in_block -= plen;
 568                                 len -= plen;
 569                                 off += plen;
 570                                 i++;
 571                                 npages++;
 572                         }
 573                 }
 574         }
 575
 576         RETURN(npages);
 577
 578 out_err:
 579         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 580         RETURN(rc);
 581 }
 582
 583 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 584                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 585                         int maxlnb, enum dt_bufs_type rw)
 586 {
 587         struct osd_object *obj  = osd_dt_obj(dt);
 588         int                rc;
 589
 590         LASSERT(dt_object_exists(dt));
 591         LASSERT(obj->oo_dn);
 592
 593         if (rw & DT_BUFS_TYPE_WRITE)
 594                 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
 595         else
 596                 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
 597
 598         return rc;
 599 }
 600
 601 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 602                         struct niobuf_local *lnb, int npages)
 603 {
 604         struct osd_object *obj = osd_dt_obj(dt);
 605
 606         LASSERT(dt_object_exists(dt));
 607         LASSERT(obj->oo_dn);
 608
 609         return 0;
 610 }
 611
 612 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 613                                            uint64_t offset,
 614                                            uint32_t blksz)
 615 {
 616         LASSERT(blksz > 0);
 617
 618         size += offset % blksz;
 619
 620         if (likely(is_power_of_2(blksz)))
 621                 return round_up(size, blksz);
 622         else
 623                 return DIV_ROUND_UP_ULL(size, blksz) * blksz;
 624 }
 625
 626 static int osd_declare_write_commit(const struct lu_env *env,
 627                                     struct dt_object *dt,
 628                                     struct niobuf_local *lnb, int npages,
 629                                     struct thandle *th)
 630 {
 631         struct osd_object  *obj = osd_dt_obj(dt);
 632         struct osd_device  *osd = osd_obj2dev(obj);
 633         struct osd_thandle *oh;
 634         uint64_t            offset = 0;
 635         uint32_t            size = 0;
 636         uint32_t blksz = obj->oo_dn->dn_datablksz;
 637         int                 i, rc;
 638         bool synced = false;
 639         long long           space = 0;
 640         struct page        *last_page = NULL;
 641         unsigned long       discont_pages = 0;
 642         enum osd_quota_local_flags local_flags = 0;
 643         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 644         ENTRY;
 645
 646         LASSERT(dt_object_exists(dt));
 647         LASSERT(obj->oo_dn);
 648
 649         LASSERT(lnb);
 650         LASSERT(npages > 0);
 651
 652         oh = container_of(th, struct osd_thandle, ot_super);
 653
 654         for (i = 0; i < npages; i++) {
 655                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 656                         ++discont_pages;
 657                 last_page = lnb[i].lnb_page;
 658                 if (lnb[i].lnb_rc)
 659                         /* ENOSPC, network RPC error, etc.
 660                          * We don't want to book space for pages which will be
 661                          * skipped in osd_write_commit(). Hence we skip pages
 662                          * with lnb_rc != 0 here too */
 663                         continue;
 664                 /* ignore quota for the whole request if any page is from
 665                  * client cache or written by root.
 666                  *
 667                  * XXX once we drop the 1.8 client support, the checking
 668                  * for whether page is from cache can be simplified as:
 669                  * !(lnb[i].flags & OBD_BRW_SYNC)
 670                  *
 671                  * XXX we could handle this on per-lnb basis as done by
 672                  * grant. */
 673                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 674                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 675                     OBD_BRW_FROM_GRANT)
 676                         declare_flags |= OSD_QID_FORCE;
 677
 678                 if (size == 0) {
 679                         /* first valid lnb */
 680                         offset = lnb[i].lnb_file_offset;
 681                         size = lnb[i].lnb_len;
 682                         continue;
 683                 }
 684                 if (offset + size == lnb[i].lnb_file_offset) {
 685                         /* this lnb is contiguous to the previous one */
 686                         size += lnb[i].lnb_len;
 687                         continue;
 688                 }
 689
 690                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 691                                   obj->oo_dn, offset, size);
 692                 /* Estimating space to be consumed by a write is rather
 693                  * complicated with ZFS. As a consequence, we don't account for
 694                  * indirect blocks and just use as a rough estimate the worse
 695                  * case where the old space is being held by a snapshot. Quota
 696                  * overrun will be adjusted once the operation is committed, if
 697                  * required. */
 698                 space += osd_roundup2blocksz(size, offset, blksz);
 699
 700                 offset = lnb[i].lnb_file_offset;
 701                 size = lnb[i].lnb_len;
 702         }
 703
 704         if (size) {
 705                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 706                                   offset, size);
 707                 space += osd_roundup2blocksz(size, offset, blksz);
 708         }
 709
 710         /* backend zfs filesystem might be configured to store multiple data
 711          * copies */
 712         space  *= osd->od_os->os_copies;
 713         space   = toqb(space);
 714         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 715                npages, space);
 716
 717         record_start_io(osd, WRITE, discont_pages);
 718 retry:
 719         /* acquire quota space if needed */
 720         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 721                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 722                                space, oh, &local_flags, declare_flags);
 723
 724         if (!synced && rc == -EDQUOT &&
 725             (local_flags & QUOTA_FL_SYNC) != 0) {
 726                 dt_sync(env, th->th_dev);
 727                 synced = true;
 728                 CDEBUG(D_QUOTA, "retry after sync\n");
 729                 local_flags = 0;
 730                 goto retry;
 731         }
 732
 733         /* we need only to store the overquota flags in the first lnb for
 734          * now, once we support multiple objects BRW, this code needs be
 735          * revised. */
 736         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
 737                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 738         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
 739                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 740 #ifdef ZFS_PROJINHERIT
 741         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
 742                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 743 #endif
 744
 745         RETURN(rc);
 746 }
 747
 748 /**
 749  * Policy to grow ZFS block size by write pattern.
 750  * For sequential write, it grows block size gradually until it reaches the
 751  * maximum blocksize the dataset can support. Otherwise, it will pick a
 752  * a block size by the writing region of this I/O.
 753  */
 754 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 755                               uint64_t start, uint64_t end)
 756 {
 757         struct osd_device       *osd = osd_obj2dev(obj);
 758         dnode_t *dn = obj->oo_dn;
 759         uint32_t                 blksz;
 760         int                      rc = 0;
 761
 762         ENTRY;
 763
 764         if (dn->dn_maxblkid > 0) /* can't change block size */
 765                 GOTO(out, rc);
 766
 767         if (dn->dn_datablksz >= osd->od_max_blksz)
 768                 GOTO(out, rc);
 769
 770         down_write(&obj->oo_guard);
 771
 772         blksz = dn->dn_datablksz;
 773         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 774                 GOTO(out_unlock, rc);
 775
 776         /* now ZFS can support up to 16MB block size, and if the write
 777          * is sequential, it just increases the block size gradually */
 778         if (start <= blksz) { /* sequential */
 779                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 780         } else { /* sparse, pick a block size by write region */
 781                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 782                                         end - start);
 783         }
 784
 785         if (!is_power_of_2(blksz))
 786                 blksz = size_roundup_power2(blksz);
 787
 788         if (blksz > dn->dn_datablksz) {
 789                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 790                                                blksz, 0, oh->ot_tx);
 791                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 792                 if (rc < 0)
 793                         CDEBUG(D_INODE, "object "DFID": change block size"
 794                                "%u -> %u error rc = %d\n",
 795                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 796                                dn->dn_datablksz, blksz, rc);
 797         }
 798         EXIT;
 799 out_unlock:
 800         up_write(&obj->oo_guard);
 801 out:
 802         return rc;
 803 }
 804
 805 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 806                                         loff_t off, ssize_t len)
 807 {
 808         dmu_buf_t **dbp;
 809         int i, rc, numbufs;
 810
 811         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 812                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 813         if (unlikely(rc))
 814                 return;
 815
 816         for (i = 0; i < numbufs; i++)
 817                 dbuf_set_pending_evict(dbp[i]);
 818
 819         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 820 }
 821
 822 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 823                         struct niobuf_local *lnb, int npages,
 824                         struct thandle *th, __u64 user_size)
 825 {
 826         struct osd_object  *obj  = osd_dt_obj(dt);
 827         struct osd_device  *osd = osd_obj2dev(obj);
 828         struct osd_thandle *oh;
 829         uint64_t            new_size = 0;
 830         int                 i, abufsz, rc = 0, drop_cache = 0;
 831         unsigned long      iosize = 0;
 832         ENTRY;
 833
 834         LASSERT(dt_object_exists(dt));
 835         LASSERT(obj->oo_dn);
 836
 837         LASSERT(th != NULL);
 838         oh = container_of(th, struct osd_thandle, ot_super);
 839
 840         /* adjust block size. Assume the buffers are sorted. */
 841         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 842                                  lnb[npages - 1].lnb_file_offset +
 843                                  lnb[npages - 1].lnb_len);
 844
 845         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 846             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 847             osd->od_readcache_max_filesize)
 848                 drop_cache = 1;
 849
 850         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 851                 RETURN(-ENOSPC);
 852
 853         /* if la_size is already bigger than specified user_size,
 854          * ignore user_size
 855          */
 856         if (obj->oo_attr.la_size > user_size)
 857                 user_size = 0;
 858
 859         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 860          * size and assigning arcbuf take place at the same time.
 861          *
 862          * Thread 1:
 863          * osd_write_commit()
 864          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 865          *   -> dmu_object_set_blocksize()
 866          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 867          *       write lock held
 868          *     -> dbuf_new_size()
 869          *      -> dmu_buf_will_dirty()
 870          *       -> dbuf_read()
 871          *        -> wait for the dbuf state to change
 872          * Thread 2:
 873          * osd_write_commit()
 874          *  -> dmu_assign_arcbuf()
 875          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 876          *    -> dbuf_dirty()
 877          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 878          *
 879          * By taking the read lock, it can avoid thread 2 to enter into the
 880          * critical section of assigning the arcbuf, while thread 1 is
 881          * changing the block size.
 882          */
 883         down_read(&obj->oo_guard);
 884         if (obj->oo_destroyed) {
 885                 up_read(&obj->oo_guard);
 886                 RETURN(-ENOENT);
 887         }
 888
 889         for (i = 0; i < npages; i++) {
 890                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 891                         (unsigned) lnb[i].lnb_len,
 892                         (unsigned) lnb[i].lnb_file_offset);
 893
 894                 if (lnb[i].lnb_rc) {
 895                         /* ENOSPC, network RPC error, etc.
 896                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 897                          * so we skip this page if lnb_rc is set to -ENOSPC */
 898                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 899                                 PFID(lu_object_fid(&dt->do_lu)), i,
 900                                 lnb[i].lnb_rc);
 901                         continue;
 902                 }
 903
 904                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 905                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 906                 if (lnb[i].lnb_page == NULL)
 907                         continue;
 908
 909                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 910                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 911                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
 912                                       lnb[i].lnb_page_offset, oh->ot_tx);
 913                         kunmap(lnb[i].lnb_page);
 914                         iosize += lnb[i].lnb_len;
 915                         abufsz = lnb[i].lnb_len; /* to drop cache below */
 916                 } else if (lnb[i].lnb_data) {
 917                         int j, apages;
 918                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 919                         /* buffer loaned for zerocopy, try to use it.
 920                          * notice that dmu_assign_arcbuf() is smart
 921                          * enough to recognize changed blocksize
 922                          * in this case it fallbacks to dmu_write() */
 923                         abufsz = arc_buf_size(lnb[i].lnb_data);
 924                         LASSERT(abufsz & PAGE_MASK);
 925                         apages = abufsz >> PAGE_SHIFT;
 926                         LASSERT(i + apages <= npages);
 927                         /* these references to pages must be invalidated
 928                          * to prevent access in osd_bufs_put() */
 929                         for (j = 0; j < apages; j++)
 930                                 lnb[i + j].lnb_page = NULL;
 931                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 932                                           lnb[i].lnb_file_offset,
 933                                           lnb[i].lnb_data, oh->ot_tx);
 934                         /* drop the reference, otherwise osd_put_bufs()
 935                          * will be releasing it - bad! */
 936                         lnb[i].lnb_data = NULL;
 937                         atomic_dec(&osd->od_zerocopy_loan);
 938                         iosize += abufsz;
 939                 } else {
 940                         /* we don't want to deal with cache if nothing
 941                          * has been send to ZFS at this step */
 942                         continue;
 943                 }
 944
 945                 if (!drop_cache)
 946                         continue;
 947
 948                 /* we have to mark dbufs for eviction here because
 949                  * dmu_assign_arcbuf() may create a new dbuf for
 950                  * loaned abuf */
 951                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
 952                                             abufsz);
 953         }
 954
 955         if (unlikely(new_size == 0)) {
 956                 /* no pages to write, no transno is needed */
 957                 th->th_local = 1;
 958                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 959                  * since ofd_commitrw_write() retries several times on ENOSPC */
 960                 up_read(&obj->oo_guard);
 961                 record_end_io(osd, WRITE, 0, 0, 0);
 962                 RETURN(0);
 963         }
 964
 965         /* if file has grown, take user_size into account */
 966         if (user_size && new_size > user_size)
 967                 new_size = user_size;
 968         write_lock(&obj->oo_attr_lock);
 969         if (obj->oo_attr.la_size < new_size) {
 970                 obj->oo_attr.la_size = new_size;
 971                 write_unlock(&obj->oo_attr_lock);
 972                 /* osd_object_sa_update() will be copying directly from
 973                  * oo_attr into dbuf. any update within a single txg will copy
 974                  * the most actual */
 975                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 976                                           &obj->oo_attr.la_size, 8, oh);
 977         } else {
 978                 write_unlock(&obj->oo_attr_lock);
 979         }
 980
 981         up_read(&obj->oo_guard);
 982
 983         record_end_io(osd, WRITE, 0, iosize, npages);
 984
 985         RETURN(rc);
 986 }
 987
 988 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 989                         struct niobuf_local *lnb, int npages)
 990 {
 991         struct osd_object *obj  = osd_dt_obj(dt);
 992         int                i;
 993         loff_t             eof;
 994
 995         LASSERT(dt_object_exists(dt));
 996         LASSERT(obj->oo_dn);
 997
 998         read_lock(&obj->oo_attr_lock);
 999         eof = obj->oo_attr.la_size;
1000         read_unlock(&obj->oo_attr_lock);
1001
1002         for (i = 0; i < npages; i++) {
1003                 if (unlikely(lnb[i].lnb_rc < 0))
1004                         continue;
1005
1006                 lnb[i].lnb_rc = lnb[i].lnb_len;
1007
1008                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
1009                         /* send complete pages all the time */
1010                         if (eof <= lnb[i].lnb_file_offset)
1011                                 lnb[i].lnb_rc = 0;
1012
1013                         /* all subsequent rc should be 0 */
1014                         while (++i < npages)
1015                                 lnb[i].lnb_rc = 0;
1016                         break;
1017                 }
1018         }
1019
1020         return 0;
1021 }
1022
1023 /*
1024  * Punch/truncate an object
1025  *
1026  *      IN:     db  - dmu_buf of the object to free data in.
1027  *              off - start of section to free.
1028  *              len - length of section to free (DMU_OBJECT_END => to EOF).
1029  *
1030  *      RETURN: 0 if success
1031  *              error code if failure
1032  *
1033  * The transaction passed to this routine must have
1034  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1035  * called and then assigned to a transaction group.
1036  */
1037 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1038                               dmu_tx_t *tx, uint64_t off, uint64_t len)
1039 {
1040         dnode_t *dn = obj->oo_dn;
1041         uint64_t size = obj->oo_attr.la_size;
1042         int rc = 0;
1043
1044         /* Assert that the transaction has been assigned to a
1045            transaction group. */
1046         LASSERT(tx->tx_txg != 0);
1047         /*
1048          * Nothing to do if file already at desired length.
1049          */
1050         if (len == DMU_OBJECT_END && size == off)
1051                 return 0;
1052
1053         /* if object holds encrypted content, we need to make sure we truncate
1054          * on an encryption unit boundary, or subsequent reads will get
1055          * corrupted content
1056          */
1057         if (len != DMU_OBJECT_END)
1058                 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1059                         (off & ~LUSTRE_ENCRYPTION_MASK);
1060         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1061             off & ~LUSTRE_ENCRYPTION_MASK)
1062                 off = (off & LUSTRE_ENCRYPTION_MASK) +
1063                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1064
1065
1066         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1067         if (off < size)
1068                 dmu_free_range(os, dn->dn_object, off, len, tx);
1069
1070         return rc;
1071 }
1072
1073 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1074                         __u64 start, __u64 end, struct thandle *th)
1075 {
1076         struct osd_object  *obj = osd_dt_obj(dt);
1077         struct osd_device  *osd = osd_obj2dev(obj);
1078         struct osd_thandle *oh;
1079         __u64               len;
1080         int                 rc = 0;
1081         ENTRY;
1082
1083         LASSERT(dt_object_exists(dt));
1084         LASSERT(osd_invariant(obj));
1085
1086         LASSERT(th != NULL);
1087         oh = container_of(th, struct osd_thandle, ot_super);
1088
1089         write_lock(&obj->oo_attr_lock);
1090         /* truncate */
1091         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1092                 len = DMU_OBJECT_END;
1093         else
1094                 len = end - start;
1095         write_unlock(&obj->oo_attr_lock);
1096
1097         down_read(&obj->oo_guard);
1098         if (obj->oo_destroyed)
1099                 GOTO(out, rc = -ENOENT);
1100
1101         rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1102
1103         /* set new size */
1104         if (len == DMU_OBJECT_END) {
1105                 write_lock(&obj->oo_attr_lock);
1106                 obj->oo_attr.la_size = start;
1107                 write_unlock(&obj->oo_attr_lock);
1108                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1109                                           &obj->oo_attr.la_size, 8, oh);
1110         }
1111 out:
1112         up_read(&obj->oo_guard);
1113         RETURN(rc);
1114 }
1115
1116 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1117                         __u64 start, __u64 end, struct thandle *handle)
1118 {
1119         struct osd_object  *obj = osd_dt_obj(dt);
1120         struct osd_device  *osd = osd_obj2dev(obj);
1121         struct osd_thandle *oh;
1122         __u64               len;
1123         ENTRY;
1124
1125         oh = container_of(handle, struct osd_thandle, ot_super);
1126
1127         read_lock(&obj->oo_attr_lock);
1128         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1129                 len = DMU_OBJECT_END;
1130         else
1131                 len = end - start;
1132
1133         /* declare we'll free some blocks ... */
1134         /* if object holds encrypted content, we need to make sure we truncate
1135          * on an encryption unit boundary, or subsequent reads will get
1136          * corrupted content
1137          */
1138         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1139             start & ~LUSTRE_ENCRYPTION_MASK)
1140                 start = (start & LUSTRE_ENCRYPTION_MASK) +
1141                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1142         if (start < obj->oo_attr.la_size) {
1143                 read_unlock(&obj->oo_attr_lock);
1144                 dmu_tx_mark_netfree(oh->ot_tx);
1145                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1146         } else {
1147                 read_unlock(&obj->oo_attr_lock);
1148         }
1149
1150         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1151                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1152                                  0, oh, NULL, OSD_QID_BLK));
1153 }
1154
1155 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1156                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1157 {
1158         int     rc;
1159         ENTRY;
1160
1161         switch (advice) {
1162         default:
1163                 rc = -ENOTSUPP;
1164                 break;
1165         }
1166
1167         RETURN(rc);
1168 }
1169
1170 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1171                          __u64 start, __u64 end, int mode, struct thandle *th)
1172 {
1173         int rc = -EOPNOTSUPP;
1174         ENTRY;
1175
1176          /*
1177           * space preallocation is not supported for ZFS
1178           * Returns -EOPNOTSUPP for now
1179           */
1180         RETURN(rc);
1181 }
1182
1183 static int osd_declare_fallocate(const struct lu_env *env,
1184                                  struct dt_object *dt, __u64 start, __u64 end,
1185                                  int mode, struct thandle *th)
1186 {
1187         int rc = -EOPNOTSUPP;
1188         ENTRY;
1189
1190          /*
1191           * space preallocation is not supported for ZFS
1192           * Returns -EOPNOTSUPP for now
1193           */
1194         RETURN(rc);
1195 }
1196
1197 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1198                         loff_t offset, int whence)
1199 {
1200         struct osd_object *obj = osd_dt_obj(dt);
1201         struct osd_device *osd = osd_obj2dev(obj);
1202         uint64_t size = obj->oo_attr.la_size;
1203         uint64_t result = offset;
1204         int rc;
1205         boolean_t hole = whence == SEEK_HOLE;
1206
1207         ENTRY;
1208
1209         LASSERT(dt_object_exists(dt));
1210         LASSERT(osd_invariant(obj));
1211         LASSERT(offset >= 0);
1212
1213         /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
1214          * hole. LOV to decide after all if that real hole or not.
1215          */
1216         if (offset >= size)
1217                 RETURN(hole ? offset : -ENXIO);
1218
1219         /* Currently ZFS reports no valid DATA offset if object has dirty data
1220          * and we cannot just switch to generic way with reporting DATA on all
1221          * file offsets and HOLE beyond end of file, because we may get HOLE
1222          * reported correctly at some offset inside file then DATA will find
1223          * dirty state and be reported also at that offset by generic approach.
1224          * This is because for HOLE report ZFS doesn't check dirty state but
1225          * does for DATA.
1226          * The only way to get reliable results is to call txg_wait_synced()
1227          * when ZFS reports EBUSY result and repeat lseek call and that is
1228          * controlled via od_sync_on_lseek option.
1229          */
1230         if (!osd->od_sync_on_lseek)
1231                 result = hole ? size : offset;
1232
1233 again:
1234         rc = osd_dmu_offset_next(osd->od_os, obj->oo_dn->dn_object, hole,
1235                                  &result);
1236         /* dirty inode, lseek result is unreliable without sync */
1237         if (rc == EBUSY) {
1238                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1239                 goto again;
1240         }
1241
1242         if (rc == ESRCH)
1243                 RETURN(-ENXIO);
1244
1245         /* ZFS is not exported all needed function, so fall back to the
1246          * generic logic: for HOLE return file size, for DATA return
1247          * the current offset
1248          */
1249         if (rc == EOPNOTSUPP)
1250                 result = hole ? size : offset;
1251         else if (rc)
1252                 return -rc;
1253
1254         /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
1255          * result as end of the last block instead of logical EOF which we need
1256          */
1257         if (result > size)
1258                 result = size;
1259
1260         RETURN(result);
1261 }
1262
1263 const struct dt_body_operations osd_body_ops = {
1264         .dbo_read                       = osd_read,
1265         .dbo_declare_write              = osd_declare_write,
1266         .dbo_write                      = osd_write,
1267         .dbo_bufs_get                   = osd_bufs_get,
1268         .dbo_bufs_put                   = osd_bufs_put,
1269         .dbo_write_prep                 = osd_write_prep,
1270         .dbo_declare_write_commit       = osd_declare_write_commit,
1271         .dbo_write_commit               = osd_write_commit,
1272         .dbo_read_prep                  = osd_read_prep,
1273         .dbo_declare_punch              = osd_declare_punch,
1274         .dbo_punch                      = osd_punch,
1275         .dbo_ladvise                    = osd_ladvise,
1276         .dbo_declare_fallocate          = osd_declare_fallocate,
1277         .dbo_fallocate                  = osd_fallocate,
1278         .dbo_lseek                      = osd_lseek,
1279 };
1280
1281 const struct dt_body_operations osd_body_scrub_ops = {
1282         .dbo_read                       = osd_read_no_record,
1283         .dbo_declare_write              = osd_declare_write,
1284         .dbo_write                      = osd_write,
1285 };