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 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 brw_stats *h = &osd->od_brw_stats;
  74
  75         if (rw == READ) {
  76                 atomic_inc(&osd->od_r_in_flight);
  77                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_RPC_HIST],
  78                                       atomic_read(&osd->od_r_in_flight));
  79                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_DISCONT_PAGES],
  80                                       discont_pages);
  81         } else {
  82                 atomic_inc(&osd->od_w_in_flight);
  83                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_W_RPC_HIST],
  84                                       atomic_read(&osd->od_w_in_flight));
  85                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_W_DISCONT_PAGES],
  86                                       discont_pages);
  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 brw_stats *h = &osd->od_brw_stats;
  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_pcpu(&h->bs_hist[BRW_R_PAGES + rw], npages);
 101         if (disksize > 0)
 102                 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_DISK_IOSIZE + rw],
 103                                            disksize);
 104         if (elapsed)
 105                 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_IO_TIME + rw],
 106                                             elapsed);
 107 }
 108
 109 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 110                           struct lu_buf *buf, loff_t *pos, size_t *size)
 111 {
 112         struct osd_object *obj = osd_dt_obj(dt);
 113         uint64_t old_size;
 114         int rc;
 115
 116         LASSERT(dt_object_exists(dt));
 117         LASSERT(obj->oo_dn);
 118
 119         read_lock(&obj->oo_attr_lock);
 120         old_size = obj->oo_attr.la_size;
 121         read_unlock(&obj->oo_attr_lock);
 122
 123         if (*pos + *size > old_size) {
 124                 if (old_size < *pos)
 125                         return 0;
 126
 127                 *size = old_size - *pos;
 128         }
 129
 130         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 131                           buf->lb_buf, DMU_READ_PREFETCH);
 132         if (!rc) {
 133                 rc = *size;
 134                 *pos += *size;
 135         }
 136
 137         return rc;
 138 }
 139
 140 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 141                         struct lu_buf *buf, loff_t *pos)
 142 {
 143         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 144         size_t size = buf->lb_len;
 145         hrtime_t start = gethrtime();
 146         s64 delta_ms;
 147         int rc;
 148
 149         record_start_io(osd, READ, 0);
 150         rc = __osd_read(env, dt, buf, pos, &size);
 151         delta_ms = gethrtime() - start;
 152         do_div(delta_ms, NSEC_PER_MSEC);
 153         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 154
 155         return rc;
 156 }
 157
 158 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 159                                          struct dt_object *dt,
 160                                          struct lu_buf *buf, loff_t *pos)
 161 {
 162         size_t size = buf->lb_len;
 163
 164         return __osd_read(env, dt, buf, pos, &size);
 165 }
 166
 167 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 168                                 const struct lu_buf *buf, loff_t pos,
 169                                 struct thandle *th)
 170 {
 171         struct osd_object  *obj  = osd_dt_obj(dt);
 172         struct osd_device  *osd = osd_obj2dev(obj);
 173         loff_t _pos = pos, max = 0;
 174         struct osd_thandle *oh;
 175         uint64_t            oid;
 176         ENTRY;
 177
 178         oh = container_of(th, struct osd_thandle, ot_super);
 179
 180         /* in some cases declare can race with creation (e.g. llog)
 181          * and we need to wait till object is initialized. notice
 182          * LOHA_EXISTs is supposed to be the last step in the
 183          * initialization */
 184
 185         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 186         if (dt_object_exists(dt))
 187                 oid = obj->oo_dn->dn_object;
 188         else
 189                 oid = DMU_NEW_OBJECT;
 190
 191         /* XXX: we still miss for append declaration support in ZFS
 192          *      -1 means append which is used by llog mostly, llog
 193          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 194         max = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 195                     obj->oo_attr.la_size + (2 << 20));
 196         if (pos == -1)
 197                 pos = max;
 198         if (obj->oo_dn) {
 199                 loff_t tstart, tend, end = pos + buf->lb_len;
 200                 dmu_tx_hold_t *txh;
 201
 202                 /* try to find a close declared window to fit/extend */
 203                 for (txh = list_head(&oh->ot_tx->tx_holds); txh != NULL;
 204                     txh = list_next(&oh->ot_tx->tx_holds, txh)) {
 205                         if (obj->oo_dn != txh->txh_dnode)
 206                                 continue;
 207                         if (txh->txh_type != THT_WRITE)
 208                                 continue;
 209
 210                         /* bytes already declared in this handle */
 211                         tstart = txh->txh_arg1;
 212                         tend = txh->txh_arg1 + txh->txh_arg2;
 213
 214                         if (pos < tstart)
 215                                 tstart = pos;
 216                         if (tend < end)
 217                                 tend = end;
 218                         /* if this is an append, then extend it */
 219                         if (_pos == -1 && txh->txh_arg1 == max)
 220                                 tend += buf->lb_len;
 221                         /* don't let too big appends */
 222                         if (tend - tstart > 4*1024*1024)
 223                                 continue;
 224                         if (pos >= tend || end <= tstart)
 225                                 continue;
 226
 227                         txh->txh_arg1 = tstart;
 228                         txh->txh_arg2 = tend - tstart;
 229                         return 0;
 230                 }
 231         }
 232         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 233
 234         /* dt_declare_write() is usually called for system objects, such
 235          * as llog or last_rcvd files. We needn't enforce quota on those
 236          * objects, so always set the lqi_space as 0. */
 237         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 238                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 239                                  0, oh, NULL, OSD_QID_BLK));
 240 }
 241
 242 static dmu_buf_t *osd_get_dbuf(struct osd_object *obj, uint64_t offset)
 243 {
 244         dmu_buf_t **dbs = obj->oo_dbs;
 245         uint64_t blkid;
 246         int i;
 247
 248         blkid = dbuf_whichblock(obj->oo_dn, 0, offset);
 249         for (i = 0; i < OSD_MAX_DBUFS; i++) {
 250                 dmu_buf_impl_t *dbi = (void *)dbs[i];
 251                 if (!dbs[i])
 252                         continue;
 253                 if (dbi->db_blkid == blkid)
 254                         return dbs[i];
 255         }
 256         return (dmu_buf_t *)dbuf_hold(obj->oo_dn, blkid, osd_0copy_tag);
 257 }
 258
 259 static void osd_put_dbuf(struct osd_object *obj, dmu_buf_t *db)
 260 {
 261         dmu_buf_t **dbs = obj->oo_dbs;
 262         int i;
 263
 264         for (i = 0; i < OSD_MAX_DBUFS; i++) {
 265                 if (dbs[i] == db)
 266                         return;
 267         }
 268         /* get rid of dbuf with blkd > 0 */
 269         for (i = 0; i < OSD_MAX_DBUFS; i++) {
 270                 if (dbs[i] == NULL) {
 271                         dbs[i] = db;
 272                         return;
 273                 }
 274                 if (dbs[i]->db_offset > 0) {
 275                         /* replace this one */
 276                         dbuf_rele((dmu_buf_impl_t *)dbs[i], osd_0copy_tag);
 277                         dbs[i] = db;
 278                         return;
 279                 }
 280         }
 281         LBUG();
 282 }
 283
 284 static ssize_t osd_write_llog_header(struct osd_object *obj,
 285                                      const struct lu_buf *buf, loff_t *pos,
 286                                      struct osd_thandle *oh)
 287 {
 288         int bufoff, tocpy;
 289         int len = buf->lb_len;
 290         loff_t offset = *pos;
 291         char *data = buf->lb_buf;
 292
 293         while (len > 0) {
 294                 dmu_buf_t *db = osd_get_dbuf(obj, offset);
 295
 296                 bufoff = offset - db->db_offset;
 297                 tocpy = MIN(db->db_size - bufoff, len);
 298                 if (tocpy == db->db_size)
 299                         dmu_buf_will_fill(db, oh->ot_tx);
 300                 else
 301                         dmu_buf_will_dirty(db, oh->ot_tx);
 302                 LASSERT(offset >= db->db_offset);
 303                 LASSERT(offset + tocpy <= db->db_offset + db->db_size);
 304                 (void) memcpy((char *)db->db_data + bufoff, data, tocpy);
 305
 306                 if (tocpy == db->db_size)
 307                         dmu_buf_fill_done(db, oh->ot_tx);
 308
 309                 offset += tocpy;
 310                 data += tocpy;
 311                 len -= tocpy;
 312
 313                 osd_put_dbuf(obj, db);
 314         }
 315
 316         return 0;
 317 }
 318
 319 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 320                         const struct lu_buf *buf, loff_t *pos,
 321                         struct thandle *th)
 322 {
 323         struct osd_object  *obj  = osd_dt_obj(dt);
 324         struct osd_device  *osd = osd_obj2dev(obj);
 325         struct osd_thandle *oh;
 326         uint64_t            offset = *pos;
 327         int                 rc;
 328
 329         ENTRY;
 330
 331         LASSERT(dt_object_exists(dt));
 332         LASSERT(obj->oo_dn);
 333
 334         LASSERT(th != NULL);
 335         oh = container_of(th, struct osd_thandle, ot_super);
 336
 337         down_read(&obj->oo_guard);
 338         if (obj->oo_destroyed)
 339                 GOTO(out, rc = -ENOENT);
 340
 341         if (fid_is_llog(lu_object_fid(&dt->do_lu))) {
 342                 osd_write_llog_header(obj, buf, pos, oh);
 343         } else {
 344                 osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 345                               buf->lb_buf, oh->ot_tx);
 346         }
 347         write_lock(&obj->oo_attr_lock);
 348         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 349                 obj->oo_attr.la_size = offset + buf->lb_len;
 350                 write_unlock(&obj->oo_attr_lock);
 351                 /* osd_object_sa_update() will be copying directly from oo_attr
 352                  * into dbuf.  any update within a single txg will copy the
 353                  * most actual */
 354                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 355                                         &obj->oo_attr.la_size, 8, oh);
 356                 if (unlikely(rc))
 357                         GOTO(out, rc);
 358         } else {
 359                 write_unlock(&obj->oo_attr_lock);
 360         }
 361
 362         *pos += buf->lb_len;
 363         rc = buf->lb_len;
 364
 365 out:
 366         up_read(&obj->oo_guard);
 367         RETURN(rc);
 368 }
 369
 370 /*
 371  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 372  *      purposes as it's not very well defined ...
 373  *      instead I use the lowest bit of the address so that:
 374  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 375  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 376  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 377  *
 378  *      bzzz, to blame
 379  */
 380 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 381                         struct niobuf_local *lnb, int npages)
 382 {
 383         struct osd_object *obj  = osd_dt_obj(dt);
 384         struct osd_device *osd = osd_obj2dev(obj);
 385         unsigned long      ptr;
 386         int                i;
 387
 388         LASSERT(dt_object_exists(dt));
 389         LASSERT(obj->oo_dn);
 390
 391         for (i = 0; i < npages; i++) {
 392                 if (lnb[i].lnb_page == NULL)
 393                         continue;
 394                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 395                         /* this is anonymous page allocated for copy-write */
 396                         lnb[i].lnb_page->mapping = NULL;
 397                         __free_page(lnb[i].lnb_page);
 398                         atomic_dec(&osd->od_zerocopy_alloc);
 399                 } else {
 400                         /* see comment in osd_bufs_get_read() */
 401                         ptr = (unsigned long)lnb[i].lnb_data;
 402                         if (ptr & 1UL) {
 403                                 ptr &= ~1UL;
 404                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 405                                 atomic_dec(&osd->od_zerocopy_pin);
 406                         } else if (lnb[i].lnb_data != NULL) {
 407                                 int j, apages, abufsz;
 408                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 409                                 apages = abufsz >> PAGE_SHIFT;
 410                                 /* these references to pages must be invalidated
 411                                  * to prevent access in osd_bufs_put() */
 412                                 for (j = 0; j < apages; j++)
 413                                         lnb[i + j].lnb_page = NULL;
 414                                 dmu_return_arcbuf(lnb[i].lnb_data);
 415                                 atomic_dec(&osd->od_zerocopy_loan);
 416                         }
 417                 }
 418                 lnb[i].lnb_page = NULL;
 419                 lnb[i].lnb_data = NULL;
 420         }
 421
 422         return 0;
 423 }
 424
 425 static inline struct page *kmem_to_page(void *addr)
 426 {
 427         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 428         if (is_vmalloc_addr(addr))
 429                 return vmalloc_to_page(addr);
 430         else
 431                 return virt_to_page(addr);
 432 }
 433
 434 /**
 435  * Prepare buffers for read.
 436  *
 437  * The function maps the range described by \a off and \a len to \a lnb array.
 438  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 439  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 440  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 441  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 442  * A better implementation would just return the buffers (potentially unfilled)
 443  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 444  *
 445  * \param[in] env       environment
 446  * \param[in] obj       object
 447  * \param[in] off       offset in bytes
 448  * \param[in] len       the number of bytes to access
 449  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 450  *
 451  * \retval              0 for success
 452  * \retval              negative error number of failure
 453  */
 454 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 455                              loff_t off, ssize_t len, struct niobuf_local *lnb,
 456                              int maxlnb)
 457 {
 458         struct osd_device *osd = osd_obj2dev(obj);
 459         int rc, i, numbufs, npages = 0, drop_cache = 0;
 460         hrtime_t start = gethrtime();
 461         dmu_buf_t **dbp;
 462         s64 delta_ms;
 463
 464         ENTRY;
 465         record_start_io(osd, READ, 0);
 466
 467         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
 468                 drop_cache = 1;
 469
 470         /* grab buffers for read:
 471          * OSD API let us to grab buffers first, then initiate IO(s)
 472          * so that all required IOs will be done in parallel, but at the
 473          * moment DMU doesn't provide us with a method to grab buffers.
 474          * If we discover this is a vital for good performance we
 475          * can get own replacement for dmu_buf_hold_array_by_bonus().
 476          */
 477         while (len > 0 &&
 478                (obj->oo_dn->dn_datablkshift != 0 ||
 479                 off < obj->oo_dn->dn_datablksz)) {
 480                 if (obj->oo_dn->dn_datablkshift == 0 &&
 481                     off + len > obj->oo_dn->dn_datablksz)
 482                         len = obj->oo_dn->dn_datablksz - off;
 483
 484                 dbp = NULL;
 485                 if (unlikely(npages >= maxlnb))
 486                         GOTO(err, rc = -EOVERFLOW);
 487
 488                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 489                                                   off, len, TRUE, osd_0copy_tag,
 490                                                   &numbufs, &dbp);
 491                 if (unlikely(rc))
 492                         GOTO(err, rc);
 493
 494                 for (i = 0; i < numbufs; i++) {
 495                         int bufoff, tocpy, thispage;
 496                         void *dbf = dbp[i];
 497
 498                         LASSERT(len > 0);
 499
 500                         atomic_inc(&osd->od_zerocopy_pin);
 501
 502                         bufoff = off - dbp[i]->db_offset;
 503                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 504
 505                         /* kind of trick to differentiate dbuf vs. arcbuf */
 506                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 507                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 508
 509                         while (tocpy > 0) {
 510                                 if (unlikely(npages >= maxlnb))
 511                                         GOTO(err, rc = -EOVERFLOW);
 512
 513                                 thispage = PAGE_SIZE;
 514                                 thispage -= bufoff & (PAGE_SIZE - 1);
 515                                 thispage = min(tocpy, thispage);
 516
 517                                 lnb->lnb_rc = 0;
 518                                 lnb->lnb_file_offset = off;
 519                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 520                                 lnb->lnb_len = thispage;
 521                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 522                                                              bufoff);
 523                                 /* mark just a single slot: we need this
 524                                  * reference to dbuf to be released once */
 525                                 lnb->lnb_data = dbf;
 526                                 dbf = NULL;
 527
 528                                 tocpy -= thispage;
 529                                 len -= thispage;
 530                                 bufoff += thispage;
 531                                 off += thispage;
 532
 533                                 npages++;
 534                                 lnb++;
 535                         }
 536
 537                         if (drop_cache)
 538                                 dbuf_set_pending_evict(dbp[i]);
 539
 540                         /* steal dbuf so dmu_buf_rele_array() can't release
 541                          * it */
 542                         dbp[i] = NULL;
 543                 }
 544
 545                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 546         }
 547
 548         delta_ms = gethrtime() - start;
 549         do_div(delta_ms, NSEC_PER_MSEC);
 550         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 551
 552         RETURN(npages);
 553
 554 err:
 555         LASSERT(rc < 0);
 556         if (dbp)
 557                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 558         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 559         RETURN(rc);
 560 }
 561
 562 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 563 {
 564         arc_buf_t *abuf;
 565
 566         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 567         if (unlikely(!abuf))
 568                 return ERR_PTR(-ENOMEM);
 569
 570 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 571         /**
 572          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 573          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 574          * setting up page array for RDMA transfer. See LU-9305.
 575          */
 576         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 577                 dmu_return_arcbuf(abuf);
 578                 return NULL;
 579         }
 580 #endif
 581
 582         return abuf;
 583 }
 584
 585 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 586                               loff_t off, ssize_t len, struct niobuf_local *lnb,
 587                               int maxlnb)
 588 {
 589         struct osd_device *osd = osd_obj2dev(obj);
 590         int                poff, plen, off_in_block, sz_in_block;
 591         int                rc, i = 0, npages = 0;
 592         dnode_t *dn = obj->oo_dn;
 593         arc_buf_t *abuf;
 594         uint32_t bs = dn->dn_datablksz;
 595         ENTRY;
 596
 597         /*
 598          * currently only full blocks are subject to zerocopy approach:
 599          * so that we're sure nobody is trying to update the same block
 600          */
 601         while (len > 0) {
 602                 if (unlikely(npages >= maxlnb))
 603                         GOTO(out_err, rc = -EOVERFLOW);
 604
 605                 off_in_block = off & (bs - 1);
 606                 sz_in_block = min_t(int, bs - off_in_block, len);
 607
 608                 abuf = NULL;
 609                 if (sz_in_block == bs) {
 610                         /* full block, try to use zerocopy */
 611                         abuf = osd_request_arcbuf(dn, bs);
 612                         if (unlikely(IS_ERR(abuf)))
 613                                 GOTO(out_err, rc = PTR_ERR(abuf));
 614                 }
 615
 616                 if (abuf != NULL) {
 617                         atomic_inc(&osd->od_zerocopy_loan);
 618
 619                         /* go over pages arcbuf contains, put them as
 620                          * local niobufs for ptlrpc's bulks */
 621                         while (sz_in_block > 0) {
 622                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 623
 624                                 if (unlikely(npages >= maxlnb))
 625                                         GOTO(out_err, rc = -EOVERFLOW);
 626
 627                                 lnb[i].lnb_file_offset = off;
 628                                 lnb[i].lnb_page_offset = 0;
 629                                 lnb[i].lnb_len = plen;
 630                                 lnb[i].lnb_rc = 0;
 631                                 if (sz_in_block == bs)
 632                                         lnb[i].lnb_data = abuf;
 633                                 else
 634                                         lnb[i].lnb_data = NULL;
 635
 636                                 /* this one is not supposed to fail */
 637                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 638                                                         off_in_block);
 639                                 LASSERT(lnb[i].lnb_page);
 640
 641                                 lprocfs_counter_add(osd->od_stats,
 642                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 643
 644                                 sz_in_block -= plen;
 645                                 len -= plen;
 646                                 off += plen;
 647                                 off_in_block += plen;
 648                                 i++;
 649                                 npages++;
 650                         }
 651                 } else {
 652                         if (off_in_block == 0 && len < bs &&
 653                                         off + len >= obj->oo_attr.la_size)
 654                                 lprocfs_counter_add(osd->od_stats,
 655                                                 LPROC_OSD_TAIL_IO, 1);
 656
 657                         /* can't use zerocopy, allocate temp. buffers */
 658                         poff = off & (PAGE_SIZE - 1);
 659                         while (sz_in_block > 0) {
 660                                 plen = min_t(int, poff + sz_in_block,
 661                                              PAGE_SIZE);
 662                                 plen -= poff;
 663
 664                                 if (unlikely(npages >= maxlnb))
 665                                         GOTO(out_err, rc = -EOVERFLOW);
 666
 667                                 lnb[i].lnb_file_offset = off;
 668                                 lnb[i].lnb_page_offset = poff;
 669                                 poff = 0;
 670
 671                                 lnb[i].lnb_len = plen;
 672                                 lnb[i].lnb_rc = 0;
 673                                 lnb[i].lnb_data = NULL;
 674
 675                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 676                                 if (unlikely(lnb[i].lnb_page == NULL))
 677                                         GOTO(out_err, rc = -ENOMEM);
 678
 679                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 680                                 lnb[i].lnb_page->mapping = (void *)obj;
 681
 682                                 atomic_inc(&osd->od_zerocopy_alloc);
 683                                 lprocfs_counter_add(osd->od_stats,
 684                                                 LPROC_OSD_COPY_IO, 1);
 685
 686                                 sz_in_block -= plen;
 687                                 len -= plen;
 688                                 off += plen;
 689                                 i++;
 690                                 npages++;
 691                         }
 692                 }
 693         }
 694
 695         RETURN(npages);
 696
 697 out_err:
 698         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 699         RETURN(rc);
 700 }
 701
 702 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 703                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 704                         int maxlnb, enum dt_bufs_type rw)
 705 {
 706         struct osd_object *obj  = osd_dt_obj(dt);
 707         int                rc;
 708
 709         down_read(&obj->oo_guard);
 710
 711         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 712                 GOTO(out, rc = -ENOENT);
 713
 714         if (rw & DT_BUFS_TYPE_WRITE)
 715                 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
 716         else
 717                 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
 718
 719 out:
 720         up_read(&obj->oo_guard);
 721         return rc;
 722 }
 723
 724 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 725                         struct niobuf_local *lnb, int npages)
 726 {
 727         struct osd_object *obj = osd_dt_obj(dt);
 728
 729         LASSERT(dt_object_exists(dt));
 730         LASSERT(obj->oo_dn);
 731
 732         return 0;
 733 }
 734
 735 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 736                                            uint64_t offset,
 737                                            uint32_t blksz)
 738 {
 739         LASSERT(blksz > 0);
 740
 741         size += offset % blksz;
 742
 743         if (likely(is_power_of_2(blksz)))
 744                 return round_up(size, blksz);
 745         else
 746                 return DIV_ROUND_UP_ULL(size, blksz) * blksz;
 747 }
 748
 749 static int osd_declare_write_commit(const struct lu_env *env,
 750                                     struct dt_object *dt,
 751                                     struct niobuf_local *lnb, int npages,
 752                                     struct thandle *th)
 753 {
 754         struct osd_object  *obj = osd_dt_obj(dt);
 755         struct osd_device  *osd = osd_obj2dev(obj);
 756         struct osd_thandle *oh;
 757         uint64_t            offset = 0;
 758         uint32_t            size = 0;
 759         uint32_t blksz = obj->oo_dn->dn_datablksz;
 760         int                 i, rc;
 761         bool synced = false;
 762         long long           space = 0;
 763         struct page        *last_page = NULL;
 764         unsigned long       discont_pages = 0;
 765         enum osd_quota_local_flags local_flags = 0;
 766         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 767         ENTRY;
 768
 769         LASSERT(dt_object_exists(dt));
 770         LASSERT(obj->oo_dn);
 771
 772         LASSERT(lnb);
 773         LASSERT(npages > 0);
 774
 775         oh = container_of(th, struct osd_thandle, ot_super);
 776
 777         for (i = 0; i < npages; i++) {
 778                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 779                         ++discont_pages;
 780                 last_page = lnb[i].lnb_page;
 781                 if (lnb[i].lnb_rc)
 782                         /* ENOSPC, network RPC error, etc.
 783                          * We don't want to book space for pages which will be
 784                          * skipped in osd_write_commit(). Hence we skip pages
 785                          * with lnb_rc != 0 here too */
 786                         continue;
 787                 /* ignore quota for the whole request if any page is from
 788                  * client cache or written by root.
 789                  *
 790                  * XXX we could handle this on per-lnb basis as done by
 791                  * grant. */
 792                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 793                     (lnb[i].lnb_flags & OBD_BRW_SYS_RESOURCE) ||
 794                     !(lnb[i].lnb_flags & OBD_BRW_SYNC))
 795                         declare_flags |= OSD_QID_FORCE;
 796
 797                 if (size == 0) {
 798                         /* first valid lnb */
 799                         offset = lnb[i].lnb_file_offset;
 800                         size = lnb[i].lnb_len;
 801                         continue;
 802                 }
 803                 if (offset + size == lnb[i].lnb_file_offset) {
 804                         /* this lnb is contiguous to the previous one */
 805                         size += lnb[i].lnb_len;
 806                         continue;
 807                 }
 808
 809                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 810                                   obj->oo_dn, offset, size);
 811                 /* Estimating space to be consumed by a write is rather
 812                  * complicated with ZFS. As a consequence, we don't account for
 813                  * indirect blocks and just use as a rough estimate the worse
 814                  * case where the old space is being held by a snapshot. Quota
 815                  * overrun will be adjusted once the operation is committed, if
 816                  * required. */
 817                 space += osd_roundup2blocksz(size, offset, blksz);
 818
 819                 offset = lnb[i].lnb_file_offset;
 820                 size = lnb[i].lnb_len;
 821         }
 822
 823         if (size) {
 824                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 825                                   offset, size);
 826                 space += osd_roundup2blocksz(size, offset, blksz);
 827         }
 828
 829         /* backend zfs filesystem might be configured to store multiple data
 830          * copies */
 831         space  *= osd->od_os->os_copies;
 832         space   = toqb(space);
 833         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 834                npages, space);
 835
 836         record_start_io(osd, WRITE, discont_pages);
 837 retry:
 838         /* acquire quota space if needed */
 839         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 840                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 841                                space, oh, &local_flags, declare_flags);
 842
 843         if (!synced && rc == -EDQUOT &&
 844             (local_flags & QUOTA_FL_SYNC) != 0) {
 845                 dt_sync(env, th->th_dev);
 846                 synced = true;
 847                 CDEBUG(D_QUOTA, "retry after sync\n");
 848                 local_flags = 0;
 849                 goto retry;
 850         }
 851
 852         /* we need only to store the overquota flags in the first lnb for
 853          * now, once we support multiple objects BRW, this code needs be
 854          * revised. */
 855         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
 856                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 857         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
 858                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 859 #ifdef ZFS_PROJINHERIT
 860         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
 861                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 862         if (local_flags & QUOTA_FL_ROOT_PRJQUOTA)
 863                 lnb[0].lnb_flags |= OBD_BRW_ROOT_PRJQUOTA;
 864 #endif
 865
 866         RETURN(rc);
 867 }
 868
 869 /**
 870  * Policy to grow ZFS block size by write pattern.
 871  * For sequential write, it grows block size gradually until it reaches the
 872  * maximum blocksize the dataset can support. Otherwise, it will pick a
 873  * a block size by the writing region of this I/O.
 874  */
 875 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 876                               uint64_t start, uint64_t end)
 877 {
 878         struct osd_device       *osd = osd_obj2dev(obj);
 879         dnode_t *dn = obj->oo_dn;
 880         uint32_t                 blksz;
 881         int                      rc = 0;
 882
 883         ENTRY;
 884
 885         if (dn->dn_maxblkid > 0) /* can't change block size */
 886                 GOTO(out, rc);
 887
 888         if (dn->dn_datablksz >= osd->od_max_blksz)
 889                 GOTO(out, rc);
 890
 891         down_write(&obj->oo_guard);
 892
 893         blksz = dn->dn_datablksz;
 894         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 895                 GOTO(out_unlock, rc);
 896
 897         /* now ZFS can support up to 16MB block size, and if the write
 898          * is sequential, it just increases the block size gradually */
 899         if (start <= blksz) { /* sequential */
 900                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 901         } else { /* sparse, pick a block size by write region */
 902                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 903                                         end - start);
 904         }
 905
 906         if (!is_power_of_2(blksz))
 907                 blksz = size_roundup_power2(blksz);
 908
 909         if (blksz > dn->dn_datablksz) {
 910                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 911                                                blksz, 0, oh->ot_tx);
 912                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 913                 if (rc < 0)
 914                         CDEBUG(D_INODE, "object "DFID": change block size"
 915                                "%u -> %u error rc = %d\n",
 916                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 917                                dn->dn_datablksz, blksz, rc);
 918         }
 919         EXIT;
 920 out_unlock:
 921         up_write(&obj->oo_guard);
 922 out:
 923         return rc;
 924 }
 925
 926 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 927                                         loff_t off, ssize_t len)
 928 {
 929         dmu_buf_t **dbp;
 930         int i, rc, numbufs;
 931
 932         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 933                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 934         if (unlikely(rc))
 935                 return;
 936
 937         for (i = 0; i < numbufs; i++)
 938                 dbuf_set_pending_evict(dbp[i]);
 939
 940         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 941 }
 942
 943 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 944                         struct niobuf_local *lnb, int npages,
 945                         struct thandle *th, __u64 user_size)
 946 {
 947         struct osd_object  *obj  = osd_dt_obj(dt);
 948         struct osd_device  *osd = osd_obj2dev(obj);
 949         struct osd_thandle *oh;
 950         uint64_t            new_size = 0;
 951         int                 i, abufsz, rc = 0, drop_cache = 0;
 952         unsigned long      iosize = 0;
 953         ENTRY;
 954
 955         LASSERT(dt_object_exists(dt));
 956         LASSERT(obj->oo_dn);
 957
 958         LASSERT(th != NULL);
 959         oh = container_of(th, struct osd_thandle, ot_super);
 960
 961         /* adjust block size. Assume the buffers are sorted. */
 962         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 963                                  lnb[npages - 1].lnb_file_offset +
 964                                  lnb[npages - 1].lnb_len);
 965
 966         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 967             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 968             osd->od_readcache_max_filesize)
 969                 drop_cache = 1;
 970
 971         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 972                 RETURN(-ENOSPC);
 973
 974         /* if la_size is already bigger than specified user_size,
 975          * ignore user_size
 976          */
 977         if (obj->oo_attr.la_size > user_size)
 978                 user_size = 0;
 979
 980         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 981          * size and assigning arcbuf take place at the same time.
 982          *
 983          * Thread 1:
 984          * osd_write_commit()
 985          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 986          *   -> dmu_object_set_blocksize()
 987          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 988          *       write lock held
 989          *     -> dbuf_new_size()
 990          *      -> dmu_buf_will_dirty()
 991          *       -> dbuf_read()
 992          *        -> wait for the dbuf state to change
 993          * Thread 2:
 994          * osd_write_commit()
 995          *  -> dmu_assign_arcbuf()
 996          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 997          *    -> dbuf_dirty()
 998          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 999          *
1000          * By taking the read lock, it can avoid thread 2 to enter into the
1001          * critical section of assigning the arcbuf, while thread 1 is
1002          * changing the block size.
1003          */
1004         down_read(&obj->oo_guard);
1005         if (obj->oo_destroyed) {
1006                 up_read(&obj->oo_guard);
1007                 RETURN(-ENOENT);
1008         }
1009
1010         for (i = 0; i < npages; i++) {
1011                 CDEBUG(D_INODE, "write %u bytes at %u\n",
1012                         (unsigned) lnb[i].lnb_len,
1013                         (unsigned) lnb[i].lnb_file_offset);
1014
1015                 if (lnb[i].lnb_rc) {
1016                         /* ENOSPC, network RPC error, etc.
1017                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
1018                          * so we skip this page if lnb_rc is set to -ENOSPC */
1019                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
1020                                 PFID(lu_object_fid(&dt->do_lu)), i,
1021                                 lnb[i].lnb_rc);
1022                         continue;
1023                 }
1024
1025                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1026                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1027                 if (lnb[i].lnb_page == NULL)
1028                         continue;
1029
1030                 if (lnb[i].lnb_page->mapping == (void *)obj) {
1031                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
1032                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
1033                                       lnb[i].lnb_page_offset, oh->ot_tx);
1034                         kunmap(lnb[i].lnb_page);
1035                         iosize += lnb[i].lnb_len;
1036                         abufsz = lnb[i].lnb_len; /* to drop cache below */
1037                 } else if (lnb[i].lnb_data) {
1038                         int j, apages;
1039                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
1040                         /* buffer loaned for zerocopy, try to use it.
1041                          * notice that dmu_assign_arcbuf() is smart
1042                          * enough to recognize changed blocksize
1043                          * in this case it fallbacks to dmu_write() */
1044                         abufsz = arc_buf_size(lnb[i].lnb_data);
1045                         LASSERT(abufsz & PAGE_MASK);
1046                         apages = abufsz >> PAGE_SHIFT;
1047                         LASSERT(i + apages <= npages);
1048                         /* these references to pages must be invalidated
1049                          * to prevent access in osd_bufs_put() */
1050                         for (j = 0; j < apages; j++)
1051                                 lnb[i + j].lnb_page = NULL;
1052                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
1053                                           lnb[i].lnb_file_offset,
1054                                           lnb[i].lnb_data, oh->ot_tx);
1055                         /* drop the reference, otherwise osd_put_bufs()
1056                          * will be releasing it - bad! */
1057                         lnb[i].lnb_data = NULL;
1058                         atomic_dec(&osd->od_zerocopy_loan);
1059                         iosize += abufsz;
1060                 } else {
1061                         /* we don't want to deal with cache if nothing
1062                          * has been send to ZFS at this step */
1063                         continue;
1064                 }
1065
1066                 if (!drop_cache)
1067                         continue;
1068
1069                 /* we have to mark dbufs for eviction here because
1070                  * dmu_assign_arcbuf() may create a new dbuf for
1071                  * loaned abuf */
1072                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
1073                                             abufsz);
1074         }
1075
1076         if (unlikely(new_size == 0)) {
1077                 /* no pages to write, no transno is needed */
1078                 th->th_local = 1;
1079                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
1080                  * since ofd_commitrw_write() retries several times on ENOSPC */
1081                 up_read(&obj->oo_guard);
1082                 record_end_io(osd, WRITE, 0, 0, 0);
1083                 RETURN(0);
1084         }
1085
1086         /* if file has grown, take user_size into account */
1087         if (user_size && new_size > user_size)
1088                 new_size = user_size;
1089         write_lock(&obj->oo_attr_lock);
1090         if (obj->oo_attr.la_size < new_size) {
1091                 obj->oo_attr.la_size = new_size;
1092                 write_unlock(&obj->oo_attr_lock);
1093                 /* osd_object_sa_update() will be copying directly from
1094                  * oo_attr into dbuf. any update within a single txg will copy
1095                  * the most actual */
1096                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1097                                           &obj->oo_attr.la_size, 8, oh);
1098         } else {
1099                 write_unlock(&obj->oo_attr_lock);
1100         }
1101
1102         up_read(&obj->oo_guard);
1103
1104         record_end_io(osd, WRITE, 0, iosize, npages);
1105
1106         RETURN(rc);
1107 }
1108
1109 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1110                         struct niobuf_local *lnb, int npages)
1111 {
1112         struct osd_object *obj  = osd_dt_obj(dt);
1113         int                i;
1114         loff_t             eof;
1115
1116         LASSERT(dt_object_exists(dt));
1117         LASSERT(obj->oo_dn);
1118
1119         read_lock(&obj->oo_attr_lock);
1120         eof = obj->oo_attr.la_size;
1121         read_unlock(&obj->oo_attr_lock);
1122
1123         for (i = 0; i < npages; i++) {
1124                 if (unlikely(lnb[i].lnb_rc < 0))
1125                         continue;
1126
1127                 lnb[i].lnb_rc = lnb[i].lnb_len;
1128
1129                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
1130                         /* send complete pages all the time */
1131                         if (eof <= lnb[i].lnb_file_offset)
1132                                 lnb[i].lnb_rc = 0;
1133
1134                         /* all subsequent rc should be 0 */
1135                         while (++i < npages)
1136                                 lnb[i].lnb_rc = 0;
1137                         break;
1138                 }
1139         }
1140
1141         return 0;
1142 }
1143
1144 /*
1145  * Punch/truncate an object
1146  *
1147  *      IN:     db  - dmu_buf of the object to free data in.
1148  *              off - start of section to free.
1149  *              len - length of section to free (DMU_OBJECT_END => to EOF).
1150  *
1151  *      RETURN: 0 if success
1152  *              error code if failure
1153  *
1154  * The transaction passed to this routine must have
1155  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1156  * called and then assigned to a transaction group.
1157  */
1158 static int __osd_object_punch(struct osd_object *obj, objset_t *os,
1159                               dmu_tx_t *tx, uint64_t off, uint64_t len)
1160 {
1161         dnode_t *dn = obj->oo_dn;
1162         uint64_t size = obj->oo_attr.la_size;
1163         int rc = 0;
1164
1165         /* Assert that the transaction has been assigned to a
1166            transaction group. */
1167         LASSERT(tx->tx_txg != 0);
1168         /*
1169          * Nothing to do if file already at desired length.
1170          */
1171         if (len == DMU_OBJECT_END && size == off)
1172                 return 0;
1173
1174         /* if object holds encrypted content, we need to make sure we truncate
1175          * on an encryption unit boundary, or subsequent reads will get
1176          * corrupted content
1177          */
1178         if (len != DMU_OBJECT_END)
1179                 len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
1180                         (off & ~LUSTRE_ENCRYPTION_MASK);
1181         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1182             off & ~LUSTRE_ENCRYPTION_MASK)
1183                 off = (off & LUSTRE_ENCRYPTION_MASK) +
1184                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1185
1186
1187         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1188         if (off < size)
1189                 dmu_free_range(os, dn->dn_object, off, len, tx);
1190
1191         return rc;
1192 }
1193
1194 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1195                         __u64 start, __u64 end, struct thandle *th)
1196 {
1197         struct osd_object  *obj = osd_dt_obj(dt);
1198         struct osd_device  *osd = osd_obj2dev(obj);
1199         struct osd_thandle *oh;
1200         __u64               len;
1201         int                 rc = 0;
1202         ENTRY;
1203
1204         LASSERT(dt_object_exists(dt));
1205         LASSERT(osd_invariant(obj));
1206
1207         LASSERT(th != NULL);
1208         oh = container_of(th, struct osd_thandle, ot_super);
1209
1210         write_lock(&obj->oo_attr_lock);
1211         /* truncate */
1212         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1213                 len = DMU_OBJECT_END;
1214         else
1215                 len = end - start;
1216         write_unlock(&obj->oo_attr_lock);
1217
1218         down_read(&obj->oo_guard);
1219         if (obj->oo_destroyed)
1220                 GOTO(out, rc = -ENOENT);
1221
1222         rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
1223
1224         /* set new size */
1225         if (len == DMU_OBJECT_END) {
1226                 write_lock(&obj->oo_attr_lock);
1227                 obj->oo_attr.la_size = start;
1228                 write_unlock(&obj->oo_attr_lock);
1229                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1230                                           &obj->oo_attr.la_size, 8, oh);
1231         }
1232 out:
1233         up_read(&obj->oo_guard);
1234         RETURN(rc);
1235 }
1236
1237 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1238                         __u64 start, __u64 end, struct thandle *handle)
1239 {
1240         struct osd_object  *obj = osd_dt_obj(dt);
1241         struct osd_device  *osd = osd_obj2dev(obj);
1242         struct osd_thandle *oh;
1243         __u64               len;
1244         ENTRY;
1245
1246         oh = container_of(handle, struct osd_thandle, ot_super);
1247
1248         read_lock(&obj->oo_attr_lock);
1249         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1250                 len = DMU_OBJECT_END;
1251         else
1252                 len = end - start;
1253
1254         /* declare we'll free some blocks ... */
1255         /* if object holds encrypted content, we need to make sure we truncate
1256          * on an encryption unit boundary, or subsequent reads will get
1257          * corrupted content
1258          */
1259         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
1260             start & ~LUSTRE_ENCRYPTION_MASK)
1261                 start = (start & LUSTRE_ENCRYPTION_MASK) +
1262                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1263         if (start < obj->oo_attr.la_size) {
1264                 read_unlock(&obj->oo_attr_lock);
1265                 dmu_tx_mark_netfree(oh->ot_tx);
1266                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1267         } else {
1268                 read_unlock(&obj->oo_attr_lock);
1269         }
1270
1271         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1272                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1273                                  0, oh, NULL, OSD_QID_BLK));
1274 }
1275
1276 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1277                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1278 {
1279         int     rc;
1280         ENTRY;
1281
1282         switch (advice) {
1283         default:
1284                 rc = -ENOTSUPP;
1285                 break;
1286         }
1287
1288         RETURN(rc);
1289 }
1290
1291 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1292                          __u64 start, __u64 end, int mode, struct thandle *th)
1293 {
1294         int rc = -EOPNOTSUPP;
1295         ENTRY;
1296
1297          /*
1298           * space preallocation is not supported for ZFS
1299           * Returns -EOPNOTSUPP for now
1300           */
1301         RETURN(rc);
1302 }
1303
1304 static int osd_declare_fallocate(const struct lu_env *env,
1305                                  struct dt_object *dt, __u64 start, __u64 end,
1306                                  int mode, struct thandle *th)
1307 {
1308         int rc = -EOPNOTSUPP;
1309         ENTRY;
1310
1311          /*
1312           * space preallocation is not supported for ZFS
1313           * Returns -EOPNOTSUPP for now
1314           */
1315         RETURN(rc);
1316 }
1317
1318 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
1319                         loff_t offset, int whence)
1320 {
1321         struct osd_object *obj = osd_dt_obj(dt);
1322         struct osd_device *osd = osd_obj2dev(obj);
1323         uint64_t size = obj->oo_attr.la_size;
1324         uint64_t result = offset;
1325         int rc;
1326         boolean_t hole = whence == SEEK_HOLE;
1327
1328         ENTRY;
1329
1330         LASSERT(dt_object_exists(dt));
1331         LASSERT(osd_invariant(obj));
1332         LASSERT(offset >= 0);
1333
1334         /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
1335          * hole. LOV to decide after all if that real hole or not.
1336          */
1337         if (offset >= size)
1338                 RETURN(hole ? offset : -ENXIO);
1339
1340         /* Currently ZFS reports no valid DATA offset if object has dirty data
1341          * and we cannot just switch to generic way with reporting DATA on all
1342          * file offsets and HOLE beyond end of file, because we may get HOLE
1343          * reported correctly at some offset inside file then DATA will find
1344          * dirty state and be reported also at that offset by generic approach.
1345          * This is because for HOLE report ZFS doesn't check dirty state but
1346          * does for DATA.
1347          * The only way to get reliable results is to call txg_wait_synced()
1348          * when ZFS reports EBUSY result and repeat lseek call and that is
1349          * controlled via od_sync_on_lseek option.
1350          */
1351         if (!osd->od_sync_on_lseek)
1352                 result = hole ? size : offset;
1353
1354 again:
1355         rc = osd_dmu_offset_next(osd->od_os, obj->oo_dn->dn_object, hole,
1356                                  &result);
1357         /* dirty inode, lseek result is unreliable without sync */
1358         if (rc == EBUSY) {
1359                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1360                 goto again;
1361         }
1362
1363         if (rc == ESRCH)
1364                 RETURN(-ENXIO);
1365
1366         /* ZFS is not exported all needed function, so fall back to the
1367          * generic logic: for HOLE return file size, for DATA return
1368          * the current offset
1369          */
1370         if (rc == EOPNOTSUPP)
1371                 result = hole ? size : offset;
1372         else if (rc)
1373                 return -rc;
1374
1375         /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
1376          * result as end of the last block instead of logical EOF which we need
1377          */
1378         if (result > size)
1379                 result = size;
1380
1381         RETURN(result);
1382 }
1383
1384 const struct dt_body_operations osd_body_ops = {
1385         .dbo_read                       = osd_read,
1386         .dbo_declare_write              = osd_declare_write,
1387         .dbo_write                      = osd_write,
1388         .dbo_bufs_get                   = osd_bufs_get,
1389         .dbo_bufs_put                   = osd_bufs_put,
1390         .dbo_write_prep                 = osd_write_prep,
1391         .dbo_declare_write_commit       = osd_declare_write_commit,
1392         .dbo_write_commit               = osd_write_commit,
1393         .dbo_read_prep                  = osd_read_prep,
1394         .dbo_declare_punch              = osd_declare_punch,
1395         .dbo_punch                      = osd_punch,
1396         .dbo_ladvise                    = osd_ladvise,
1397         .dbo_declare_fallocate          = osd_declare_fallocate,
1398         .dbo_fallocate                  = osd_fallocate,
1399         .dbo_lseek                      = osd_lseek,
1400 };
1401
1402 const struct dt_body_operations osd_body_scrub_ops = {
1403         .dbo_read                       = osd_read_no_record,
1404         .dbo_declare_write              = osd_declare_write,
1405         .dbo_write                      = osd_write,
1406 };