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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * Copyright (c) 2012, Intel Corporation.
  32  * Use is subject to license terms.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/osd-zfs/osd_io.c
  39  *
  40  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  41  * Author: Mike Pershin <tappro@whamcloud.com>
  42  */
  43
  44 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_OSD
  48
  49 #include <lustre_ver.h>
  50 #include <libcfs/libcfs.h>
  51 #include <lustre_fsfilt.h>
  52 #include <obd_support.h>
  53 #include <lustre_net.h>
  54 #include <obd.h>
  55 #include <obd_class.h>
  56 #include <lustre_disk.h>
  57 #include <lustre_fid.h>
  58
  59 #include "osd_internal.h"
  60
  61 #include <sys/dnode.h>
  62 #include <sys/dbuf.h>
  63 #include <sys/spa.h>
  64 #include <sys/stat.h>
  65 #include <sys/zap.h>
  66 #include <sys/spa_impl.h>
  67 #include <sys/zfs_znode.h>
  68 #include <sys/dmu_tx.h>
  69 #include <sys/dmu_objset.h>
  70 #include <sys/dsl_prop.h>
  71 #include <sys/sa_impl.h>
  72 #include <sys/txg.h>
  73
  74 static char *osd_zerocopy_tag = "zerocopy";
  75
  76 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
  77                         struct lu_buf *buf, loff_t *pos,
  78                         struct lustre_capa *capa)
  79 {
  80         struct osd_object *obj  = osd_dt_obj(dt);
  81         struct osd_device *osd = osd_obj2dev(obj);
  82         uint64_t           old_size;
  83         int                size = buf->lb_len;
  84         int                rc;
  85
  86         LASSERT(dt_object_exists(dt));
  87         LASSERT(obj->oo_db);
  88
  89         read_lock(&obj->oo_attr_lock);
  90         old_size = obj->oo_attr.la_size;
  91         read_unlock(&obj->oo_attr_lock);
  92
  93         if (*pos + size > old_size) {
  94                 if (old_size < *pos)
  95                         return 0;
  96                 else
  97                         size = old_size - *pos;
  98         }
  99
 100         rc = -dmu_read(osd->od_objset.os, obj->oo_db->db_object, *pos, size,
 101                         buf->lb_buf, DMU_READ_PREFETCH);
 102         if (rc == 0) {
 103                 rc = size;
 104                 *pos += size;
 105
 106                 /* XXX: workaround for bug in HEAD: fsfilt_ldiskfs_read() returns
 107                  * requested number of bytes, not actually read ones */
 108                 if (S_ISLNK(obj->oo_dt.do_lu.lo_header->loh_attr))
 109                         rc = buf->lb_len;
 110         }
 111         return rc;
 112 }
 113
 114 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 115                                 const loff_t size, loff_t pos,
 116                                 struct thandle *th)
 117 {
 118         struct osd_object  *obj  = osd_dt_obj(dt);
 119         struct osd_device  *osd = osd_obj2dev(obj);
 120         struct osd_thandle *oh;
 121         uint64_t            oid;
 122         ENTRY;
 123
 124         oh = container_of0(th, struct osd_thandle, ot_super);
 125
 126         /* in some cases declare can race with creation (e.g. llog)
 127          * and we need to wait till object is initialized. notice
 128          * LOHA_EXISTs is supposed to be the last step in the
 129          * initialization */
 130
 131         /* declare possible size change. notice we can't check
 132          * current size here as another thread can change it */
 133
 134         if (dt_object_exists(dt)) {
 135                 LASSERT(obj->oo_db);
 136                 oid = obj->oo_db->db_object;
 137
 138                 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 139         } else {
 140                 oid = DMU_NEW_OBJECT;
 141                 dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
 142         }
 143
 144         dmu_tx_hold_write(oh->ot_tx, oid, pos, size);
 145
 146         /* dt_declare_write() is usually called for system objects, such
 147          * as llog or last_rcvd files. We needn't enforce quota on those
 148          * objects, so always set the lqi_space as 0. */
 149         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 150                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 151                                  false));
 152 }
 153
 154 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 155                         const struct lu_buf *buf, loff_t *pos,
 156                         struct thandle *th, struct lustre_capa *capa,
 157                         int ignore_quota)
 158 {
 159         struct osd_object  *obj  = osd_dt_obj(dt);
 160         struct osd_device  *osd = osd_obj2dev(obj);
 161         udmu_objset_t      *uos = &osd->od_objset;
 162         struct osd_thandle *oh;
 163         uint64_t            offset = *pos;
 164         int                 rc;
 165         ENTRY;
 166
 167         LASSERT(dt_object_exists(dt));
 168         LASSERT(obj->oo_db);
 169
 170         LASSERT(th != NULL);
 171         oh = container_of0(th, struct osd_thandle, ot_super);
 172
 173         dmu_write(osd->od_objset.os, obj->oo_db->db_object, offset,
 174                 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
 175         write_lock(&obj->oo_attr_lock);
 176         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 177                 obj->oo_attr.la_size = offset + buf->lb_len;
 178                 write_unlock(&obj->oo_attr_lock);
 179                 /* osd_object_sa_update() will be copying directly from oo_attr
 180                  * into dbuf.  any update within a single txg will copy the
 181                  * most actual */
 182                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
 183                                         &obj->oo_attr.la_size, 8, oh);
 184                 if (unlikely(rc))
 185                         GOTO(out, rc);
 186         } else {
 187                 write_unlock(&obj->oo_attr_lock);
 188         }
 189
 190         *pos += buf->lb_len;
 191         rc = buf->lb_len;
 192
 193 out:
 194         RETURN(rc);
 195 }
 196
 197 /*
 198  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 199  *      purposes as it's not very well defined ...
 200  *      instead I use the lowest bit of the address so that:
 201  *        arc buffer:  .lnb_obj = abuf          (arc we loan for write)
 202  *        dbuf buffer: .lnb_obj = dbuf | 1      (dbuf we get for read)
 203  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 204  *
 205  *      bzzz, to blame
 206  */
 207 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 208                         struct niobuf_local *lnb, int npages)
 209 {
 210         struct osd_object *obj  = osd_dt_obj(dt);
 211         struct osd_device *osd = osd_obj2dev(obj);
 212         unsigned long      ptr;
 213         int                i;
 214
 215         LASSERT(dt_object_exists(dt));
 216         LASSERT(obj->oo_db);
 217
 218         for (i = 0; i < npages; i++) {
 219                 if (lnb[i].page == NULL)
 220                         continue;
 221                 if (lnb[i].page->mapping == (void *)obj) {
 222                         /* this is anonymous page allocated for copy-write */
 223                         lnb[i].page->mapping = NULL;
 224                         __free_page(lnb[i].page);
 225                         cfs_atomic_dec(&osd->od_zerocopy_alloc);
 226                 } else {
 227                         /* see comment in osd_bufs_get_read() */
 228                         ptr = (unsigned long)lnb[i].dentry;
 229                         if (ptr & 1UL) {
 230                                 ptr &= ~1UL;
 231                                 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
 232                                 cfs_atomic_dec(&osd->od_zerocopy_pin);
 233                         } else if (lnb[i].dentry != NULL) {
 234                                 dmu_return_arcbuf((void *)lnb[i].dentry);
 235                                 cfs_atomic_dec(&osd->od_zerocopy_loan);
 236                         }
 237                 }
 238                 lnb[i].page = NULL;
 239                 lnb[i].dentry = NULL;
 240         }
 241
 242         return 0;
 243 }
 244
 245 static struct page *kmem_to_page(void *addr)
 246 {
 247         struct page *page;
 248
 249         if (kmem_virt(addr))
 250                 page = vmalloc_to_page(addr);
 251         else
 252                 page = virt_to_page(addr);
 253
 254         return page;
 255 }
 256
 257 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 258                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 259 {
 260         struct osd_device *osd = osd_obj2dev(obj);
 261         dmu_buf_t        **dbp;
 262         int                rc, i, numbufs, npages = 0;
 263         ENTRY;
 264
 265         /* grab buffers for read:
 266          * OSD API let us to grab buffers first, then initiate IO(s)
 267          * so that all required IOs will be done in parallel, but at the
 268          * moment DMU doesn't provide us with a method to grab buffers.
 269          * If we discover this is a vital for good performance we
 270          * can get own replacement for dmu_buf_hold_array_by_bonus().
 271          */
 272         while (len > 0) {
 273                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 274                                                 osd_zerocopy_tag, &numbufs,
 275                                                 &dbp);
 276                 LASSERT(rc == 0);
 277
 278                 for (i = 0; i < numbufs; i++) {
 279                         int bufoff, tocpy, thispage;
 280                         void *dbf = dbp[i];
 281
 282                         LASSERT(len > 0);
 283
 284                         cfs_atomic_inc(&osd->od_zerocopy_pin);
 285
 286                         bufoff = off - dbp[i]->db_offset;
 287                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 288
 289                         /* kind of trick to differentiate dbuf vs. arcbuf */
 290                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 291                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 292
 293                         while (tocpy > 0) {
 294                                 thispage = CFS_PAGE_SIZE;
 295                                 thispage -= bufoff & (CFS_PAGE_SIZE - 1);
 296                                 thispage = min(tocpy, thispage);
 297
 298                                 lnb->rc = 0;
 299                                 lnb->lnb_file_offset = off;
 300                                 lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK;
 301                                 lnb->len = thispage;
 302                                 lnb->page = kmem_to_page(dbp[i]->db_data +
 303                                                                 bufoff);
 304                                 /* mark just a single slot: we need this
 305                                  * reference to dbuf to be release once */
 306                                 lnb->dentry = dbf;
 307                                 dbf = NULL;
 308
 309                                 tocpy -= thispage;
 310                                 len -= thispage;
 311                                 bufoff += thispage;
 312                                 off += thispage;
 313
 314                                 npages++;
 315                                 lnb++;
 316                         }
 317
 318                         /* steal dbuf so dmu_buf_rele_array() cant release it */
 319                         dbp[i] = NULL;
 320                 }
 321
 322                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 323         }
 324
 325         RETURN(npages);
 326 }
 327
 328 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 329                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 330 {
 331         struct osd_device *osd = osd_obj2dev(obj);
 332         int                plen, off_in_block, sz_in_block;
 333         int                i = 0, npages = 0;
 334         arc_buf_t         *abuf;
 335         uint32_t           bs;
 336         uint64_t           dummy;
 337         ENTRY;
 338
 339         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 340
 341         /*
 342          * currently only full blocks are subject to zerocopy approach:
 343          * so that we're sure nobody is trying to update the same block
 344          */
 345         while (len > 0) {
 346                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 347
 348                 off_in_block = off & (bs - 1);
 349                 sz_in_block = min_t(int, bs - off_in_block, len);
 350
 351                 if (sz_in_block == bs) {
 352                         /* full block, try to use zerocopy */
 353
 354                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 355                         if (unlikely(abuf == NULL))
 356                                 GOTO(out_err, -ENOMEM);
 357
 358                         cfs_atomic_inc(&osd->od_zerocopy_loan);
 359
 360                         /* go over pages arcbuf contains, put them as
 361                          * local niobufs for ptlrpc's bulks */
 362                         while (sz_in_block > 0) {
 363                                 plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
 364
 365                                 lnb[i].lnb_file_offset = off;
 366                                 lnb[i].lnb_page_offset = 0;
 367                                 lnb[i].len = plen;
 368                                 lnb[i].rc = 0;
 369                                 if (sz_in_block == bs)
 370                                         lnb[i].dentry = (void *)abuf;
 371                                 else
 372                                         lnb[i].dentry = NULL;
 373
 374                                 /* this one is not supposed to fail */
 375                                 lnb[i].page = kmem_to_page(abuf->b_data +
 376                                                         off_in_block);
 377                                 LASSERT(lnb[i].page);
 378
 379                                 lprocfs_counter_add(osd->od_stats,
 380                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 381
 382                                 sz_in_block -= plen;
 383                                 len -= plen;
 384                                 off += plen;
 385                                 off_in_block += plen;
 386                                 i++;
 387                                 npages++;
 388                         }
 389                 } else {
 390                         if (off_in_block == 0 && len < bs &&
 391                                         off + len >= obj->oo_attr.la_size)
 392                                 lprocfs_counter_add(osd->od_stats,
 393                                                 LPROC_OSD_TAIL_IO, 1);
 394
 395                         /* can't use zerocopy, allocate temp. buffers */
 396                         while (sz_in_block > 0) {
 397                                 plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
 398
 399                                 lnb[i].lnb_file_offset = off;
 400                                 lnb[i].lnb_page_offset = 0;
 401                                 lnb[i].len = plen;
 402                                 lnb[i].rc = 0;
 403                                 lnb[i].dentry = NULL;
 404
 405                                 lnb[i].page = alloc_page(OSD_GFP_IO);
 406                                 if (unlikely(lnb[i].page == NULL))
 407                                         GOTO(out_err, -ENOMEM);
 408
 409                                 LASSERT(lnb[i].page->mapping == NULL);
 410                                 lnb[i].page->mapping = (void *)obj;
 411
 412                                 cfs_atomic_inc(&osd->od_zerocopy_alloc);
 413                                 lprocfs_counter_add(osd->od_stats,
 414                                                 LPROC_OSD_COPY_IO, 1);
 415
 416                                 sz_in_block -= plen;
 417                                 len -= plen;
 418                                 off += plen;
 419                                 i++;
 420                                 npages++;
 421                         }
 422                 }
 423         }
 424
 425         RETURN(npages);
 426
 427 out_err:
 428         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 429         RETURN(-ENOMEM);
 430 }
 431
 432 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 433                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 434                         int rw, struct lustre_capa *capa)
 435 {
 436         struct osd_object *obj  = osd_dt_obj(dt);
 437         int                rc;
 438
 439         LASSERT(dt_object_exists(dt));
 440         LASSERT(obj->oo_db);
 441
 442         if (rw == 0)
 443                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 444         else
 445                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 446
 447         return rc;
 448 }
 449
 450 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 451                         struct niobuf_local *lnb, int npages)
 452 {
 453         struct osd_object *obj = osd_dt_obj(dt);
 454
 455         LASSERT(dt_object_exists(dt));
 456         LASSERT(obj->oo_db);
 457
 458         return 0;
 459 }
 460
 461 /* Return number of blocks that aren't mapped in the [start, start + size]
 462  * region */
 463 static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
 464                                 uint32_t size)
 465 {
 466         dmu_buf_impl_t  *dbi = (dmu_buf_impl_t *)obj->oo_db;
 467         dmu_buf_impl_t  *db;
 468         dnode_t         *dn;
 469         uint32_t         blkshift;
 470         uint64_t         end, blkid;
 471         int              rc;
 472         ENTRY;
 473
 474         DB_DNODE_ENTER(dbi);
 475         dn = DB_DNODE(dbi);
 476
 477         if (dn->dn_maxblkid == 0) {
 478                 if (start + size <= dn->dn_datablksz)
 479                         GOTO(out, size = 0);
 480                 if (start < dn->dn_datablksz)
 481                         start = dn->dn_datablksz;
 482                 /* assume largest block size */
 483                 blkshift = SPA_MAXBLOCKSHIFT;
 484         } else {
 485                 /* blocksize can't change */
 486                 blkshift = dn->dn_datablkshift;
 487         }
 488
 489         /* compute address of last block */
 490         end = (start + size - 1) >> blkshift;
 491         /* align start on block boundaries */
 492         start >>= blkshift;
 493
 494         /* size is null, can't be mapped */
 495         if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
 496                 GOTO(out, size = (end - start + 1) << blkshift);
 497
 498         /* beyond EOF, can't be mapped */
 499         if (start > dn->dn_maxblkid)
 500                 GOTO(out, size = (end - start + 1) << blkshift);
 501
 502         size = 0;
 503         for (blkid = start; blkid <= end; blkid++) {
 504                 if (blkid == dn->dn_maxblkid)
 505                         /* this one is mapped for sure */
 506                         continue;
 507                 if (blkid > dn->dn_maxblkid) {
 508                         size += (end - blkid + 1) << blkshift;
 509                         GOTO(out, size);
 510                 }
 511
 512                 rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
 513                 if (rc) {
 514                         /* for ENOENT (block not mapped) and any other errors,
 515                          * assume the block isn't mapped */
 516                         size += 1 << blkshift;
 517                         continue;
 518                 }
 519                 dbuf_rele(db, FTAG);
 520         }
 521
 522         GOTO(out, size);
 523 out:
 524         DB_DNODE_EXIT(dbi);
 525         return size;
 526 }
 527
 528 static int osd_declare_write_commit(const struct lu_env *env,
 529                                 struct dt_object *dt,
 530                                 struct niobuf_local *lnb, int npages,
 531                                 struct thandle *th)
 532 {
 533         struct osd_object  *obj = osd_dt_obj(dt);
 534         struct osd_device  *osd = osd_obj2dev(obj);
 535         struct osd_thandle *oh;
 536         uint64_t            offset = 0;
 537         uint32_t            size = 0;
 538         int                 i, rc, flags = 0;
 539         bool                ignore_quota = false, synced = false;
 540         long long           space = 0;
 541         ENTRY;
 542
 543         LASSERT(dt_object_exists(dt));
 544         LASSERT(obj->oo_db);
 545
 546         LASSERT(lnb);
 547         LASSERT(npages > 0);
 548
 549         oh = container_of0(th, struct osd_thandle, ot_super);
 550
 551         for (i = 0; i < npages; i++) {
 552                 if (lnb[i].rc)
 553                         /* ENOSPC, network RPC error, etc.
 554                          * We don't want to book space for pages which will be
 555                          * skipped in osd_write_commit(). Hence we skip pages
 556                          * with lnb_rc != 0 here too */
 557                         continue;
 558                 /* ignore quota for the whole request if any page is from
 559                  * client cache or written by root.
 560                  *
 561                  * XXX once we drop the 1.8 client support, the checking
 562                  * for whether page is from cache can be simplified as:
 563                  * !(lnb[i].flags & OBD_BRW_SYNC)
 564                  *
 565                  * XXX we could handle this on per-lnb basis as done by
 566                  * grant. */
 567                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 568                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 569                     OBD_BRW_FROM_GRANT)
 570                         ignore_quota = true;
 571                 if (size == 0) {
 572                         /* first valid lnb */
 573                         offset = lnb[i].lnb_file_offset;
 574                         size = lnb[i].len;
 575                         continue;
 576                 }
 577                 if (offset + size == lnb[i].lnb_file_offset) {
 578                         /* this lnb is contiguous to the previous one */
 579                         size += lnb[i].len;
 580                         continue;
 581                 }
 582
 583                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
 584
 585                 /* estimating space that will be consumed by a write is rather
 586                  * complicated with ZFS. As a consequence, we don't account for
 587                  * indirect blocks and quota overrun will be adjusted once the
 588                  * operation is committed, if required. */
 589                 space += osd_count_not_mapped(obj, offset, size);
 590
 591                 offset = lnb->lnb_file_offset;
 592                 size = lnb->len;
 593         }
 594
 595         if (size) {
 596                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
 597                 space += osd_count_not_mapped(obj, offset, size);
 598         }
 599
 600         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 601
 602         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 603
 604         /* backend zfs filesystem might be configured to store multiple data
 605          * copies */
 606         space  *= osd->od_objset.os->os_copies;
 607         space   = toqb(space);
 608         CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
 609                "space\n", npages, space);
 610
 611 retry:
 612         /* acquire quota space if needed */
 613         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 614                                obj->oo_attr.la_gid, space, oh, true, &flags,
 615                                ignore_quota);
 616
 617         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 618                 dt_sync(env, th->th_dev);
 619                 synced = true;
 620                 CDEBUG(D_QUOTA, "retry after sync\n");
 621                 flags = 0;
 622                 goto retry;
 623         }
 624
 625         /* we need only to store the overquota flags in the first lnb for
 626          * now, once we support multiple objects BRW, this code needs be
 627          * revised. */
 628         if (flags & QUOTA_FL_OVER_USRQUOTA)
 629                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 630         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 631                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 632
 633         RETURN(rc);
 634 }
 635
 636 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 637                         struct niobuf_local *lnb, int npages,
 638                         struct thandle *th)
 639 {
 640         struct osd_object  *obj  = osd_dt_obj(dt);
 641         struct osd_device  *osd = osd_obj2dev(obj);
 642         udmu_objset_t      *uos = &osd->od_objset;
 643         struct osd_thandle *oh;
 644         uint64_t            new_size = 0;
 645         int                 i, rc = 0;
 646         ENTRY;
 647
 648         LASSERT(dt_object_exists(dt));
 649         LASSERT(obj->oo_db);
 650
 651         LASSERT(th != NULL);
 652         oh = container_of0(th, struct osd_thandle, ot_super);
 653
 654         for (i = 0; i < npages; i++) {
 655                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 656                         (unsigned) lnb[i].len,
 657                         (unsigned) lnb[i].lnb_file_offset);
 658
 659                 if (lnb[i].rc) {
 660                         /* ENOSPC, network RPC error, etc.
 661                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 662                          * so we skip this page if lnb_rc is set to -ENOSPC */
 663                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 664                                 PFID(lu_object_fid(&dt->do_lu)), i,
 665                                 lnb[i].rc);
 666                         continue;
 667                 }
 668
 669                 if (lnb[i].page->mapping == (void *)obj) {
 670                         dmu_write(osd->od_objset.os, obj->oo_db->db_object,
 671                                 lnb[i].lnb_file_offset, lnb[i].len,
 672                                 kmap(lnb[i].page), oh->ot_tx);
 673                         kunmap(lnb[i].page);
 674                 } else if (lnb[i].dentry) {
 675                         LASSERT(((unsigned long)lnb[i].dentry & 1) == 0);
 676                         /* buffer loaned for zerocopy, try to use it.
 677                          * notice that dmu_assign_arcbuf() is smart
 678                          * enough to recognize changed blocksize
 679                          * in this case it fallbacks to dmu_write() */
 680                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 681                                         (void *)lnb[i].dentry, oh->ot_tx);
 682                         /* drop the reference, otherwise osd_put_bufs()
 683                          * will be releasing it - bad! */
 684                         lnb[i].dentry = NULL;
 685                         cfs_atomic_dec(&osd->od_zerocopy_loan);
 686                 }
 687
 688                 if (new_size < lnb[i].lnb_file_offset + lnb[i].len)
 689                         new_size = lnb[i].lnb_file_offset + lnb[i].len;
 690         }
 691
 692         if (unlikely(new_size == 0)) {
 693                 /* no pages to write, no transno is needed */
 694                 th->th_local = 1;
 695                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 696                  * since ofd_commitrw_write() retries several times on ENOSPC */
 697                 RETURN(0);
 698         }
 699
 700         write_lock(&obj->oo_attr_lock);
 701         if (obj->oo_attr.la_size < new_size) {
 702                 obj->oo_attr.la_size = new_size;
 703                 write_unlock(&obj->oo_attr_lock);
 704                 /* osd_object_sa_update() will be copying directly from
 705                  * oo_attr into dbuf. any update within a single txg will copy
 706                  * the most actual */
 707                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
 708                                         &obj->oo_attr.la_size, 8, oh);
 709         } else {
 710                 write_unlock(&obj->oo_attr_lock);
 711         }
 712
 713         RETURN(rc);
 714 }
 715
 716 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 717                         struct niobuf_local *lnb, int npages)
 718 {
 719         struct osd_object *obj  = osd_dt_obj(dt);
 720         struct lu_buf      buf;
 721         loff_t             offset;
 722         int                i;
 723
 724         LASSERT(dt_object_exists(dt));
 725         LASSERT(obj->oo_db);
 726
 727         for (i = 0; i < npages; i++) {
 728                 buf.lb_buf = kmap(lnb[i].page);
 729                 buf.lb_len = lnb[i].len;
 730                 offset = lnb[i].lnb_file_offset;
 731
 732                 CDEBUG(D_OTHER, "read %u bytes at %u\n",
 733                         (unsigned) lnb[i].len,
 734                         (unsigned) lnb[i].lnb_file_offset);
 735                 lnb[i].rc = osd_read(env, dt, &buf, &offset, NULL);
 736                 kunmap(lnb[i].page);
 737
 738                 if (lnb[i].rc < buf.lb_len) {
 739                         /* all subsequent rc should be 0 */
 740                         while (++i < npages)
 741                                 lnb[i].rc = 0;
 742                         break;
 743                 }
 744         }
 745
 746         return 0;
 747 }
 748
 749 /*
 750  * Punch/truncate an object
 751  *
 752  *      IN:     db  - dmu_buf of the object to free data in.
 753  *              off - start of section to free.
 754  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 755  *
 756  *      RETURN: 0 if success
 757  *              error code if failure
 758  *
 759  * The transaction passed to this routine must have
 760  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 761  * called and then assigned to a transaction group.
 762  */
 763 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 764                                 uint64_t size, uint64_t off, uint64_t len)
 765 {
 766         int rc = 0;
 767
 768         /* Assert that the transaction has been assigned to a
 769            transaction group. */
 770         LASSERT(tx->tx_txg != 0);
 771         /*
 772          * Nothing to do if file already at desired length.
 773          */
 774         if (len == DMU_OBJECT_END && size == off)
 775                 return 0;
 776
 777         if (off < size)
 778                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 779
 780         return rc;
 781 }
 782
 783 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 784                         __u64 start, __u64 end, struct thandle *th,
 785                         struct lustre_capa *capa)
 786 {
 787         struct osd_object  *obj = osd_dt_obj(dt);
 788         struct osd_device  *osd = osd_obj2dev(obj);
 789         udmu_objset_t      *uos = &osd->od_objset;
 790         struct osd_thandle *oh;
 791         __u64               len;
 792         int                 rc = 0;
 793         ENTRY;
 794
 795         LASSERT(dt_object_exists(dt));
 796         LASSERT(osd_invariant(obj));
 797
 798         LASSERT(th != NULL);
 799         oh = container_of0(th, struct osd_thandle, ot_super);
 800
 801         write_lock(&obj->oo_attr_lock);
 802         /* truncate */
 803         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 804                 len = DMU_OBJECT_END;
 805         else
 806                 len = end - start;
 807         write_unlock(&obj->oo_attr_lock);
 808
 809         rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx,
 810                                 obj->oo_attr.la_size, start, len);
 811         /* set new size */
 812         if (len == DMU_OBJECT_END) {
 813                 write_lock(&obj->oo_attr_lock);
 814                 obj->oo_attr.la_size = start;
 815                 write_unlock(&obj->oo_attr_lock);
 816                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
 817                                         &obj->oo_attr.la_size, 8, oh);
 818         }
 819         RETURN(rc);
 820 }
 821
 822 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 823                         __u64 start, __u64 end, struct thandle *handle)
 824 {
 825         struct osd_object  *obj = osd_dt_obj(dt);
 826         struct osd_device  *osd = osd_obj2dev(obj);
 827         struct osd_thandle *oh;
 828         __u64               len;
 829         ENTRY;
 830
 831         oh = container_of0(handle, struct osd_thandle, ot_super);
 832
 833         read_lock(&obj->oo_attr_lock);
 834         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 835                 len = DMU_OBJECT_END;
 836         else
 837                 len = end - start;
 838
 839         /* declare we'll free some blocks ... */
 840         if (start < obj->oo_attr.la_size) {
 841                 read_unlock(&obj->oo_attr_lock);
 842                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 843         } else {
 844                 read_unlock(&obj->oo_attr_lock);
 845         }
 846
 847         /* ... and we'll modify size attribute */
 848         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 849
 850         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 851                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 852                                  false));
 853 }
 854
 855
 856 struct dt_body_operations osd_body_ops = {
 857         .dbo_read                       = osd_read,
 858         .dbo_declare_write              = osd_declare_write,
 859         .dbo_write                      = osd_write,
 860         .dbo_bufs_get                   = osd_bufs_get,
 861         .dbo_bufs_put                   = osd_bufs_put,
 862         .dbo_write_prep                 = osd_write_prep,
 863         .dbo_declare_write_commit       = osd_declare_write_commit,
 864         .dbo_write_commit               = osd_write_commit,
 865         .dbo_read_prep                  = osd_read_prep,
 866         .do_declare_punch               = osd_declare_punch,
 867         .do_punch                       = osd_punch,
 868 };
 869