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, 2013, 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 inline struct page *kmem_to_page(void *addr)
 246 {
 247         if (is_vmalloc_addr(addr))
 248                 return vmalloc_to_page(addr);
 249         else
 250                 return virt_to_page(addr);
 251 }
 252
 253 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 254                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 255 {
 256         struct osd_device *osd = osd_obj2dev(obj);
 257         dmu_buf_t        **dbp;
 258         int                rc, i, numbufs, npages = 0;
 259         ENTRY;
 260
 261         /* grab buffers for read:
 262          * OSD API let us to grab buffers first, then initiate IO(s)
 263          * so that all required IOs will be done in parallel, but at the
 264          * moment DMU doesn't provide us with a method to grab buffers.
 265          * If we discover this is a vital for good performance we
 266          * can get own replacement for dmu_buf_hold_array_by_bonus().
 267          */
 268         while (len > 0) {
 269                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 270                                                   osd_zerocopy_tag, &numbufs,
 271                                                   &dbp);
 272                 if (unlikely(rc))
 273                         GOTO(err, rc);
 274
 275                 for (i = 0; i < numbufs; i++) {
 276                         int bufoff, tocpy, thispage;
 277                         void *dbf = dbp[i];
 278
 279                         LASSERT(len > 0);
 280
 281                         cfs_atomic_inc(&osd->od_zerocopy_pin);
 282
 283                         bufoff = off - dbp[i]->db_offset;
 284                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 285
 286                         /* kind of trick to differentiate dbuf vs. arcbuf */
 287                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 288                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 289
 290                         while (tocpy > 0) {
 291                                 thispage = PAGE_CACHE_SIZE;
 292                                 thispage -= bufoff & (PAGE_CACHE_SIZE - 1);
 293                                 thispage = min(tocpy, thispage);
 294
 295                                 lnb->rc = 0;
 296                                 lnb->lnb_file_offset = off;
 297                                 lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK;
 298                                 lnb->len = thispage;
 299                                 lnb->page = kmem_to_page(dbp[i]->db_data +
 300                                                                 bufoff);
 301                                 /* mark just a single slot: we need this
 302                                  * reference to dbuf to be release once */
 303                                 lnb->dentry = dbf;
 304                                 dbf = NULL;
 305
 306                                 tocpy -= thispage;
 307                                 len -= thispage;
 308                                 bufoff += thispage;
 309                                 off += thispage;
 310
 311                                 npages++;
 312                                 lnb++;
 313                         }
 314
 315                         /* steal dbuf so dmu_buf_rele_array() cant release it */
 316                         dbp[i] = NULL;
 317                 }
 318
 319                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 320         }
 321
 322         RETURN(npages);
 323
 324 err:
 325         LASSERT(rc < 0);
 326         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 327         RETURN(rc);
 328 }
 329
 330 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 331                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 332 {
 333         struct osd_device *osd = osd_obj2dev(obj);
 334         int                plen, off_in_block, sz_in_block;
 335         int                i = 0, npages = 0;
 336         arc_buf_t         *abuf;
 337         uint32_t           bs;
 338         uint64_t           dummy;
 339         ENTRY;
 340
 341         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 342
 343         /*
 344          * currently only full blocks are subject to zerocopy approach:
 345          * so that we're sure nobody is trying to update the same block
 346          */
 347         while (len > 0) {
 348                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 349
 350                 off_in_block = off & (bs - 1);
 351                 sz_in_block = min_t(int, bs - off_in_block, len);
 352
 353                 if (sz_in_block == bs) {
 354                         /* full block, try to use zerocopy */
 355
 356                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 357                         if (unlikely(abuf == NULL))
 358                                 GOTO(out_err, -ENOMEM);
 359
 360                         cfs_atomic_inc(&osd->od_zerocopy_loan);
 361
 362                         /* go over pages arcbuf contains, put them as
 363                          * local niobufs for ptlrpc's bulks */
 364                         while (sz_in_block > 0) {
 365                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 366
 367                                 lnb[i].lnb_file_offset = off;
 368                                 lnb[i].lnb_page_offset = 0;
 369                                 lnb[i].len = plen;
 370                                 lnb[i].rc = 0;
 371                                 if (sz_in_block == bs)
 372                                         lnb[i].dentry = (void *)abuf;
 373                                 else
 374                                         lnb[i].dentry = NULL;
 375
 376                                 /* this one is not supposed to fail */
 377                                 lnb[i].page = kmem_to_page(abuf->b_data +
 378                                                         off_in_block);
 379                                 LASSERT(lnb[i].page);
 380
 381                                 lprocfs_counter_add(osd->od_stats,
 382                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 383
 384                                 sz_in_block -= plen;
 385                                 len -= plen;
 386                                 off += plen;
 387                                 off_in_block += plen;
 388                                 i++;
 389                                 npages++;
 390                         }
 391                 } else {
 392                         if (off_in_block == 0 && len < bs &&
 393                                         off + len >= obj->oo_attr.la_size)
 394                                 lprocfs_counter_add(osd->od_stats,
 395                                                 LPROC_OSD_TAIL_IO, 1);
 396
 397                         /* can't use zerocopy, allocate temp. buffers */
 398                         while (sz_in_block > 0) {
 399                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 400
 401                                 lnb[i].lnb_file_offset = off;
 402                                 lnb[i].lnb_page_offset = 0;
 403                                 lnb[i].len = plen;
 404                                 lnb[i].rc = 0;
 405                                 lnb[i].dentry = NULL;
 406
 407                                 lnb[i].page = alloc_page(OSD_GFP_IO);
 408                                 if (unlikely(lnb[i].page == NULL))
 409                                         GOTO(out_err, -ENOMEM);
 410
 411                                 LASSERT(lnb[i].page->mapping == NULL);
 412                                 lnb[i].page->mapping = (void *)obj;
 413
 414                                 cfs_atomic_inc(&osd->od_zerocopy_alloc);
 415                                 lprocfs_counter_add(osd->od_stats,
 416                                                 LPROC_OSD_COPY_IO, 1);
 417
 418                                 sz_in_block -= plen;
 419                                 len -= plen;
 420                                 off += plen;
 421                                 i++;
 422                                 npages++;
 423                         }
 424                 }
 425         }
 426
 427         RETURN(npages);
 428
 429 out_err:
 430         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 431         RETURN(-ENOMEM);
 432 }
 433
 434 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 435                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 436                         int rw, struct lustre_capa *capa)
 437 {
 438         struct osd_object *obj  = osd_dt_obj(dt);
 439         int                rc;
 440
 441         LASSERT(dt_object_exists(dt));
 442         LASSERT(obj->oo_db);
 443
 444         if (rw == 0)
 445                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 446         else
 447                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 448
 449         return rc;
 450 }
 451
 452 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 453                         struct niobuf_local *lnb, int npages)
 454 {
 455         struct osd_object *obj = osd_dt_obj(dt);
 456
 457         LASSERT(dt_object_exists(dt));
 458         LASSERT(obj->oo_db);
 459
 460         return 0;
 461 }
 462
 463 /* Return number of blocks that aren't mapped in the [start, start + size]
 464  * region */
 465 static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
 466                                 uint32_t size)
 467 {
 468         dmu_buf_impl_t  *dbi = (dmu_buf_impl_t *)obj->oo_db;
 469         dmu_buf_impl_t  *db;
 470         dnode_t         *dn;
 471         uint32_t         blkshift;
 472         uint64_t         end, blkid;
 473         int              rc;
 474         ENTRY;
 475
 476         DB_DNODE_ENTER(dbi);
 477         dn = DB_DNODE(dbi);
 478
 479         if (dn->dn_maxblkid == 0) {
 480                 if (start + size <= dn->dn_datablksz)
 481                         GOTO(out, size = 0);
 482                 if (start < dn->dn_datablksz)
 483                         start = dn->dn_datablksz;
 484                 /* assume largest block size */
 485                 blkshift = SPA_MAXBLOCKSHIFT;
 486         } else {
 487                 /* blocksize can't change */
 488                 blkshift = dn->dn_datablkshift;
 489         }
 490
 491         /* compute address of last block */
 492         end = (start + size - 1) >> blkshift;
 493         /* align start on block boundaries */
 494         start >>= blkshift;
 495
 496         /* size is null, can't be mapped */
 497         if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
 498                 GOTO(out, size = (end - start + 1) << blkshift);
 499
 500         /* beyond EOF, can't be mapped */
 501         if (start > dn->dn_maxblkid)
 502                 GOTO(out, size = (end - start + 1) << blkshift);
 503
 504         size = 0;
 505         for (blkid = start; blkid <= end; blkid++) {
 506                 if (blkid == dn->dn_maxblkid)
 507                         /* this one is mapped for sure */
 508                         continue;
 509                 if (blkid > dn->dn_maxblkid) {
 510                         size += (end - blkid + 1) << blkshift;
 511                         GOTO(out, size);
 512                 }
 513
 514                 rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
 515                 if (rc) {
 516                         /* for ENOENT (block not mapped) and any other errors,
 517                          * assume the block isn't mapped */
 518                         size += 1 << blkshift;
 519                         continue;
 520                 }
 521                 dbuf_rele(db, FTAG);
 522         }
 523
 524         GOTO(out, size);
 525 out:
 526         DB_DNODE_EXIT(dbi);
 527         return size;
 528 }
 529
 530 static int osd_declare_write_commit(const struct lu_env *env,
 531                                 struct dt_object *dt,
 532                                 struct niobuf_local *lnb, int npages,
 533                                 struct thandle *th)
 534 {
 535         struct osd_object  *obj = osd_dt_obj(dt);
 536         struct osd_device  *osd = osd_obj2dev(obj);
 537         struct osd_thandle *oh;
 538         uint64_t            offset = 0;
 539         uint32_t            size = 0;
 540         int                 i, rc, flags = 0;
 541         bool                ignore_quota = false, synced = false;
 542         long long           space = 0;
 543         ENTRY;
 544
 545         LASSERT(dt_object_exists(dt));
 546         LASSERT(obj->oo_db);
 547
 548         LASSERT(lnb);
 549         LASSERT(npages > 0);
 550
 551         oh = container_of0(th, struct osd_thandle, ot_super);
 552
 553         for (i = 0; i < npages; i++) {
 554                 if (lnb[i].rc)
 555                         /* ENOSPC, network RPC error, etc.
 556                          * We don't want to book space for pages which will be
 557                          * skipped in osd_write_commit(). Hence we skip pages
 558                          * with lnb_rc != 0 here too */
 559                         continue;
 560                 /* ignore quota for the whole request if any page is from
 561                  * client cache or written by root.
 562                  *
 563                  * XXX once we drop the 1.8 client support, the checking
 564                  * for whether page is from cache can be simplified as:
 565                  * !(lnb[i].flags & OBD_BRW_SYNC)
 566                  *
 567                  * XXX we could handle this on per-lnb basis as done by
 568                  * grant. */
 569                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 570                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 571                     OBD_BRW_FROM_GRANT)
 572                         ignore_quota = true;
 573                 if (size == 0) {
 574                         /* first valid lnb */
 575                         offset = lnb[i].lnb_file_offset;
 576                         size = lnb[i].len;
 577                         continue;
 578                 }
 579                 if (offset + size == lnb[i].lnb_file_offset) {
 580                         /* this lnb is contiguous to the previous one */
 581                         size += lnb[i].len;
 582                         continue;
 583                 }
 584
 585                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
 586
 587                 /* estimating space that will be consumed by a write is rather
 588                  * complicated with ZFS. As a consequence, we don't account for
 589                  * indirect blocks and quota overrun will be adjusted once the
 590                  * operation is committed, if required. */
 591                 space += osd_count_not_mapped(obj, offset, size);
 592
 593                 offset = lnb->lnb_file_offset;
 594                 size = lnb->len;
 595         }
 596
 597         if (size) {
 598                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
 599                 space += osd_count_not_mapped(obj, offset, size);
 600         }
 601
 602         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 603
 604         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 605
 606         /* backend zfs filesystem might be configured to store multiple data
 607          * copies */
 608         space  *= osd->od_objset.os->os_copies;
 609         space   = toqb(space);
 610         CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
 611                "space\n", npages, space);
 612
 613 retry:
 614         /* acquire quota space if needed */
 615         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 616                                obj->oo_attr.la_gid, space, oh, true, &flags,
 617                                ignore_quota);
 618
 619         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 620                 dt_sync(env, th->th_dev);
 621                 synced = true;
 622                 CDEBUG(D_QUOTA, "retry after sync\n");
 623                 flags = 0;
 624                 goto retry;
 625         }
 626
 627         /* we need only to store the overquota flags in the first lnb for
 628          * now, once we support multiple objects BRW, this code needs be
 629          * revised. */
 630         if (flags & QUOTA_FL_OVER_USRQUOTA)
 631                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 632         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 633                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 634
 635         RETURN(rc);
 636 }
 637
 638 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 639                         struct niobuf_local *lnb, int npages,
 640                         struct thandle *th)
 641 {
 642         struct osd_object  *obj  = osd_dt_obj(dt);
 643         struct osd_device  *osd = osd_obj2dev(obj);
 644         udmu_objset_t      *uos = &osd->od_objset;
 645         struct osd_thandle *oh;
 646         uint64_t            new_size = 0;
 647         int                 i, rc = 0;
 648         ENTRY;
 649
 650         LASSERT(dt_object_exists(dt));
 651         LASSERT(obj->oo_db);
 652
 653         LASSERT(th != NULL);
 654         oh = container_of0(th, struct osd_thandle, ot_super);
 655
 656         for (i = 0; i < npages; i++) {
 657                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 658                         (unsigned) lnb[i].len,
 659                         (unsigned) lnb[i].lnb_file_offset);
 660
 661                 if (lnb[i].rc) {
 662                         /* ENOSPC, network RPC error, etc.
 663                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 664                          * so we skip this page if lnb_rc is set to -ENOSPC */
 665                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 666                                 PFID(lu_object_fid(&dt->do_lu)), i,
 667                                 lnb[i].rc);
 668                         continue;
 669                 }
 670
 671                 if (lnb[i].page->mapping == (void *)obj) {
 672                         dmu_write(osd->od_objset.os, obj->oo_db->db_object,
 673                                 lnb[i].lnb_file_offset, lnb[i].len,
 674                                 kmap(lnb[i].page), oh->ot_tx);
 675                         kunmap(lnb[i].page);
 676                 } else if (lnb[i].dentry) {
 677                         LASSERT(((unsigned long)lnb[i].dentry & 1) == 0);
 678                         /* buffer loaned for zerocopy, try to use it.
 679                          * notice that dmu_assign_arcbuf() is smart
 680                          * enough to recognize changed blocksize
 681                          * in this case it fallbacks to dmu_write() */
 682                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 683                                         (void *)lnb[i].dentry, oh->ot_tx);
 684                         /* drop the reference, otherwise osd_put_bufs()
 685                          * will be releasing it - bad! */
 686                         lnb[i].dentry = NULL;
 687                         cfs_atomic_dec(&osd->od_zerocopy_loan);
 688                 }
 689
 690                 if (new_size < lnb[i].lnb_file_offset + lnb[i].len)
 691                         new_size = lnb[i].lnb_file_offset + lnb[i].len;
 692         }
 693
 694         if (unlikely(new_size == 0)) {
 695                 /* no pages to write, no transno is needed */
 696                 th->th_local = 1;
 697                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 698                  * since ofd_commitrw_write() retries several times on ENOSPC */
 699                 RETURN(0);
 700         }
 701
 702         write_lock(&obj->oo_attr_lock);
 703         if (obj->oo_attr.la_size < new_size) {
 704                 obj->oo_attr.la_size = new_size;
 705                 write_unlock(&obj->oo_attr_lock);
 706                 /* osd_object_sa_update() will be copying directly from
 707                  * oo_attr into dbuf. any update within a single txg will copy
 708                  * the most actual */
 709                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
 710                                         &obj->oo_attr.la_size, 8, oh);
 711         } else {
 712                 write_unlock(&obj->oo_attr_lock);
 713         }
 714
 715         RETURN(rc);
 716 }
 717
 718 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 719                         struct niobuf_local *lnb, int npages)
 720 {
 721         struct osd_object *obj  = osd_dt_obj(dt);
 722         struct lu_buf      buf;
 723         loff_t             offset;
 724         int                i;
 725
 726         LASSERT(dt_object_exists(dt));
 727         LASSERT(obj->oo_db);
 728
 729         for (i = 0; i < npages; i++) {
 730                 buf.lb_buf = kmap(lnb[i].page);
 731                 buf.lb_len = lnb[i].len;
 732                 offset = lnb[i].lnb_file_offset;
 733
 734                 CDEBUG(D_OTHER, "read %u bytes at %u\n",
 735                         (unsigned) lnb[i].len,
 736                         (unsigned) lnb[i].lnb_file_offset);
 737                 lnb[i].rc = osd_read(env, dt, &buf, &offset, NULL);
 738                 kunmap(lnb[i].page);
 739
 740                 if (lnb[i].rc < buf.lb_len) {
 741                         /* all subsequent rc should be 0 */
 742                         while (++i < npages)
 743                                 lnb[i].rc = 0;
 744                         break;
 745                 }
 746         }
 747
 748         return 0;
 749 }
 750
 751 /*
 752  * Punch/truncate an object
 753  *
 754  *      IN:     db  - dmu_buf of the object to free data in.
 755  *              off - start of section to free.
 756  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 757  *
 758  *      RETURN: 0 if success
 759  *              error code if failure
 760  *
 761  * The transaction passed to this routine must have
 762  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 763  * called and then assigned to a transaction group.
 764  */
 765 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 766                                 uint64_t size, uint64_t off, uint64_t len)
 767 {
 768         int rc = 0;
 769
 770         /* Assert that the transaction has been assigned to a
 771            transaction group. */
 772         LASSERT(tx->tx_txg != 0);
 773         /*
 774          * Nothing to do if file already at desired length.
 775          */
 776         if (len == DMU_OBJECT_END && size == off)
 777                 return 0;
 778
 779         if (off < size)
 780                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 781
 782         return rc;
 783 }
 784
 785 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 786                         __u64 start, __u64 end, struct thandle *th,
 787                         struct lustre_capa *capa)
 788 {
 789         struct osd_object  *obj = osd_dt_obj(dt);
 790         struct osd_device  *osd = osd_obj2dev(obj);
 791         udmu_objset_t      *uos = &osd->od_objset;
 792         struct osd_thandle *oh;
 793         __u64               len;
 794         int                 rc = 0;
 795         ENTRY;
 796
 797         LASSERT(dt_object_exists(dt));
 798         LASSERT(osd_invariant(obj));
 799
 800         LASSERT(th != NULL);
 801         oh = container_of0(th, struct osd_thandle, ot_super);
 802
 803         write_lock(&obj->oo_attr_lock);
 804         /* truncate */
 805         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 806                 len = DMU_OBJECT_END;
 807         else
 808                 len = end - start;
 809         write_unlock(&obj->oo_attr_lock);
 810
 811         rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx,
 812                                 obj->oo_attr.la_size, start, len);
 813         /* set new size */
 814         if (len == DMU_OBJECT_END) {
 815                 write_lock(&obj->oo_attr_lock);
 816                 obj->oo_attr.la_size = start;
 817                 write_unlock(&obj->oo_attr_lock);
 818                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
 819                                         &obj->oo_attr.la_size, 8, oh);
 820         }
 821         RETURN(rc);
 822 }
 823
 824 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 825                         __u64 start, __u64 end, struct thandle *handle)
 826 {
 827         struct osd_object  *obj = osd_dt_obj(dt);
 828         struct osd_device  *osd = osd_obj2dev(obj);
 829         struct osd_thandle *oh;
 830         __u64               len;
 831         ENTRY;
 832
 833         oh = container_of0(handle, struct osd_thandle, ot_super);
 834
 835         read_lock(&obj->oo_attr_lock);
 836         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 837                 len = DMU_OBJECT_END;
 838         else
 839                 len = end - start;
 840
 841         /* declare we'll free some blocks ... */
 842         if (start < obj->oo_attr.la_size) {
 843                 read_unlock(&obj->oo_attr_lock);
 844                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 845         } else {
 846                 read_unlock(&obj->oo_attr_lock);
 847         }
 848
 849         /* ... and we'll modify size attribute */
 850         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 851
 852         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 853                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 854                                  false));
 855 }
 856
 857
 858 struct dt_body_operations osd_body_ops = {
 859         .dbo_read                       = osd_read,
 860         .dbo_declare_write              = osd_declare_write,
 861         .dbo_write                      = osd_write,
 862         .dbo_bufs_get                   = osd_bufs_get,
 863         .dbo_bufs_put                   = osd_bufs_put,
 864         .dbo_write_prep                 = osd_write_prep,
 865         .dbo_declare_write_commit       = osd_declare_write_commit,
 866         .dbo_write_commit               = osd_write_commit,
 867         .dbo_read_prep                  = osd_read_prep,
 868         .dbo_declare_punch              = osd_declare_punch,
 869         .dbo_punch                      = osd_punch,
 870 };
 871