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 #define DEBUG_SUBSYSTEM S_OSD
  45
  46 #include <lustre_ver.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd_support.h>
  49 #include <lustre_net.h>
  50 #include <obd.h>
  51 #include <obd_class.h>
  52 #include <lustre_disk.h>
  53 #include <lustre_fid.h>
  54 #include <lustre/lustre_idl.h>  /* LLOG_CHUNK_SIZE definition */
  55
  56 #include "osd_internal.h"
  57
  58 #include <sys/dnode.h>
  59 #include <sys/dbuf.h>
  60 #include <sys/spa.h>
  61 #include <sys/stat.h>
  62 #include <sys/zap.h>
  63 #include <sys/spa_impl.h>
  64 #include <sys/zfs_znode.h>
  65 #include <sys/dmu_tx.h>
  66 #include <sys/dmu_objset.h>
  67 #include <sys/dsl_prop.h>
  68 #include <sys/sa_impl.h>
  69 #include <sys/txg.h>
  70
  71 static char *osd_zerocopy_tag = "zerocopy";
  72
  73 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
  74                         struct lu_buf *buf, loff_t *pos,
  75                         struct lustre_capa *capa)
  76 {
  77         struct osd_object *obj  = osd_dt_obj(dt);
  78         struct osd_device *osd = osd_obj2dev(obj);
  79         uint64_t           old_size;
  80         int                size = buf->lb_len;
  81         int                rc;
  82
  83         LASSERT(dt_object_exists(dt));
  84         LASSERT(obj->oo_db);
  85
  86         read_lock(&obj->oo_attr_lock);
  87         old_size = obj->oo_attr.la_size;
  88         read_unlock(&obj->oo_attr_lock);
  89
  90         if (*pos + size > old_size) {
  91                 if (old_size < *pos)
  92                         return 0;
  93                 else
  94                         size = old_size - *pos;
  95         }
  96
  97         rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
  98                         buf->lb_buf, DMU_READ_PREFETCH);
  99         if (rc == 0) {
 100                 rc = size;
 101                 *pos += size;
 102
 103                 /* XXX: workaround for bug in HEAD: fsfilt_ldiskfs_read() returns
 104                  * requested number of bytes, not actually read ones */
 105                 if (S_ISLNK(obj->oo_dt.do_lu.lo_header->loh_attr))
 106                         rc = buf->lb_len;
 107         }
 108         return rc;
 109 }
 110
 111 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 112                                 const struct lu_buf *buf, loff_t pos,
 113                                 struct thandle *th)
 114 {
 115         struct osd_object  *obj  = osd_dt_obj(dt);
 116         struct osd_device  *osd = osd_obj2dev(obj);
 117         struct osd_thandle *oh;
 118         uint64_t            oid;
 119         ENTRY;
 120
 121         oh = container_of0(th, struct osd_thandle, ot_super);
 122
 123         /* in some cases declare can race with creation (e.g. llog)
 124          * and we need to wait till object is initialized. notice
 125          * LOHA_EXISTs is supposed to be the last step in the
 126          * initialization */
 127
 128         /* declare possible size change. notice we can't check
 129          * current size here as another thread can change it */
 130
 131         if (dt_object_exists(dt)) {
 132                 LASSERT(obj->oo_db);
 133                 oid = obj->oo_db->db_object;
 134
 135                 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 136         } else {
 137                 oid = DMU_NEW_OBJECT;
 138                 dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
 139         }
 140
 141         /* XXX: we still miss for append declaration support in ZFS
 142          *      -1 means append which is used by llog mostly, llog
 143          *      can grow upto LLOG_CHUNK_SIZE*8 records */
 144         if (pos == -1)
 145                 pos = max_t(loff_t, 256 * 8 * LLOG_CHUNK_SIZE,
 146                             obj->oo_attr.la_size + (2 << 20));
 147         dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
 148
 149         /* dt_declare_write() is usually called for system objects, such
 150          * as llog or last_rcvd files. We needn't enforce quota on those
 151          * objects, so always set the lqi_space as 0. */
 152         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 153                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 154                                  false));
 155 }
 156
 157 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 158                         const struct lu_buf *buf, loff_t *pos,
 159                         struct thandle *th, struct lustre_capa *capa,
 160                         int ignore_quota)
 161 {
 162         struct osd_object  *obj  = osd_dt_obj(dt);
 163         struct osd_device  *osd = osd_obj2dev(obj);
 164         struct osd_thandle *oh;
 165         uint64_t            offset = *pos;
 166         int                 rc;
 167         ENTRY;
 168
 169         LASSERT(dt_object_exists(dt));
 170         LASSERT(obj->oo_db);
 171
 172         LASSERT(th != NULL);
 173         oh = container_of0(th, struct osd_thandle, ot_super);
 174
 175         dmu_write(osd->od_os, obj->oo_db->db_object, offset,
 176                 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
 177         write_lock(&obj->oo_attr_lock);
 178         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 179                 obj->oo_attr.la_size = offset + buf->lb_len;
 180                 write_unlock(&obj->oo_attr_lock);
 181                 /* osd_object_sa_update() will be copying directly from oo_attr
 182                  * into dbuf.  any update within a single txg will copy the
 183                  * most actual */
 184                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 185                                         &obj->oo_attr.la_size, 8, oh);
 186                 if (unlikely(rc))
 187                         GOTO(out, rc);
 188         } else {
 189                 write_unlock(&obj->oo_attr_lock);
 190         }
 191
 192         *pos += buf->lb_len;
 193         rc = buf->lb_len;
 194
 195 out:
 196         RETURN(rc);
 197 }
 198
 199 /*
 200  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 201  *      purposes as it's not very well defined ...
 202  *      instead I use the lowest bit of the address so that:
 203  *        arc buffer:  .lnb_obj = abuf          (arc we loan for write)
 204  *        dbuf buffer: .lnb_obj = dbuf | 1      (dbuf we get for read)
 205  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 206  *
 207  *      bzzz, to blame
 208  */
 209 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 210                         struct niobuf_local *lnb, int npages)
 211 {
 212         struct osd_object *obj  = osd_dt_obj(dt);
 213         struct osd_device *osd = osd_obj2dev(obj);
 214         unsigned long      ptr;
 215         int                i;
 216
 217         LASSERT(dt_object_exists(dt));
 218         LASSERT(obj->oo_db);
 219
 220         for (i = 0; i < npages; i++) {
 221                 if (lnb[i].lnb_page == NULL)
 222                         continue;
 223                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 224                         /* this is anonymous page allocated for copy-write */
 225                         lnb[i].lnb_page->mapping = NULL;
 226                         __free_page(lnb[i].lnb_page);
 227                         atomic_dec(&osd->od_zerocopy_alloc);
 228                 } else {
 229                         /* see comment in osd_bufs_get_read() */
 230                         ptr = (unsigned long)lnb[i].lnb_data;
 231                         if (ptr & 1UL) {
 232                                 ptr &= ~1UL;
 233                                 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
 234                                 atomic_dec(&osd->od_zerocopy_pin);
 235                         } else if (lnb[i].lnb_data != NULL) {
 236                                 dmu_return_arcbuf(lnb[i].lnb_data);
 237                                 atomic_dec(&osd->od_zerocopy_loan);
 238                         }
 239                 }
 240                 lnb[i].lnb_page = NULL;
 241                 lnb[i].lnb_data = NULL;
 242         }
 243
 244         return 0;
 245 }
 246
 247 static inline struct page *kmem_to_page(void *addr)
 248 {
 249         if (is_vmalloc_addr(addr))
 250                 return vmalloc_to_page(addr);
 251         else
 252                 return virt_to_page(addr);
 253 }
 254
 255 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 256                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 257 {
 258         struct osd_device *osd = osd_obj2dev(obj);
 259         dmu_buf_t        **dbp;
 260         int                rc, i, numbufs, npages = 0;
 261         ENTRY;
 262
 263         /* grab buffers for read:
 264          * OSD API let us to grab buffers first, then initiate IO(s)
 265          * so that all required IOs will be done in parallel, but at the
 266          * moment DMU doesn't provide us with a method to grab buffers.
 267          * If we discover this is a vital for good performance we
 268          * can get own replacement for dmu_buf_hold_array_by_bonus().
 269          */
 270         while (len > 0) {
 271                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 272                                                   osd_zerocopy_tag, &numbufs,
 273                                                   &dbp);
 274                 if (unlikely(rc))
 275                         GOTO(err, rc);
 276
 277                 for (i = 0; i < numbufs; i++) {
 278                         int bufoff, tocpy, thispage;
 279                         void *dbf = dbp[i];
 280
 281                         LASSERT(len > 0);
 282
 283                         atomic_inc(&osd->od_zerocopy_pin);
 284
 285                         bufoff = off - dbp[i]->db_offset;
 286                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 287
 288                         /* kind of trick to differentiate dbuf vs. arcbuf */
 289                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 290                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 291
 292                         while (tocpy > 0) {
 293                                 thispage = PAGE_CACHE_SIZE;
 294                                 thispage -= bufoff & (PAGE_CACHE_SIZE - 1);
 295                                 thispage = min(tocpy, thispage);
 296
 297                                 lnb->lnb_rc = 0;
 298                                 lnb->lnb_file_offset = off;
 299                                 lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK;
 300                                 lnb->lnb_len = thispage;
 301                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 302                                                              bufoff);
 303                                 /* mark just a single slot: we need this
 304                                  * reference to dbuf to be release once */
 305                                 lnb->lnb_data = dbf;
 306                                 dbf = NULL;
 307
 308                                 tocpy -= thispage;
 309                                 len -= thispage;
 310                                 bufoff += thispage;
 311                                 off += thispage;
 312
 313                                 npages++;
 314                                 lnb++;
 315                         }
 316
 317                         /* steal dbuf so dmu_buf_rele_array() cant release it */
 318                         dbp[i] = NULL;
 319                 }
 320
 321                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 322         }
 323
 324         RETURN(npages);
 325
 326 err:
 327         LASSERT(rc < 0);
 328         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 329         RETURN(rc);
 330 }
 331
 332 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 333                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 334 {
 335         struct osd_device *osd = osd_obj2dev(obj);
 336         int                plen, off_in_block, sz_in_block;
 337         int                rc, i = 0, npages = 0;
 338         arc_buf_t         *abuf;
 339         uint32_t           bs;
 340         uint64_t           dummy;
 341         ENTRY;
 342
 343         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 344
 345         /*
 346          * currently only full blocks are subject to zerocopy approach:
 347          * so that we're sure nobody is trying to update the same block
 348          */
 349         while (len > 0) {
 350                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 351
 352                 off_in_block = off & (bs - 1);
 353                 sz_in_block = min_t(int, bs - off_in_block, len);
 354
 355                 if (sz_in_block == bs) {
 356                         /* full block, try to use zerocopy */
 357
 358                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 359                         if (unlikely(abuf == NULL))
 360                                 GOTO(out_err, rc = -ENOMEM);
 361
 362                         atomic_inc(&osd->od_zerocopy_loan);
 363
 364                         /* go over pages arcbuf contains, put them as
 365                          * local niobufs for ptlrpc's bulks */
 366                         while (sz_in_block > 0) {
 367                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 368
 369                                 lnb[i].lnb_file_offset = off;
 370                                 lnb[i].lnb_page_offset = 0;
 371                                 lnb[i].lnb_len = plen;
 372                                 lnb[i].lnb_rc = 0;
 373                                 if (sz_in_block == bs)
 374                                         lnb[i].lnb_data = abuf;
 375                                 else
 376                                         lnb[i].lnb_data = NULL;
 377
 378                                 /* this one is not supposed to fail */
 379                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 380                                                         off_in_block);
 381                                 LASSERT(lnb[i].lnb_page);
 382
 383                                 lprocfs_counter_add(osd->od_stats,
 384                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 385
 386                                 sz_in_block -= plen;
 387                                 len -= plen;
 388                                 off += plen;
 389                                 off_in_block += plen;
 390                                 i++;
 391                                 npages++;
 392                         }
 393                 } else {
 394                         if (off_in_block == 0 && len < bs &&
 395                                         off + len >= obj->oo_attr.la_size)
 396                                 lprocfs_counter_add(osd->od_stats,
 397                                                 LPROC_OSD_TAIL_IO, 1);
 398
 399                         /* can't use zerocopy, allocate temp. buffers */
 400                         while (sz_in_block > 0) {
 401                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 402
 403                                 lnb[i].lnb_file_offset = off;
 404                                 lnb[i].lnb_page_offset = 0;
 405                                 lnb[i].lnb_len = plen;
 406                                 lnb[i].lnb_rc = 0;
 407                                 lnb[i].lnb_data = NULL;
 408
 409                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 410                                 if (unlikely(lnb[i].lnb_page == NULL))
 411                                         GOTO(out_err, rc = -ENOMEM);
 412
 413                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 414                                 lnb[i].lnb_page->mapping = (void *)obj;
 415
 416                                 atomic_inc(&osd->od_zerocopy_alloc);
 417                                 lprocfs_counter_add(osd->od_stats,
 418                                                 LPROC_OSD_COPY_IO, 1);
 419
 420                                 sz_in_block -= plen;
 421                                 len -= plen;
 422                                 off += plen;
 423                                 i++;
 424                                 npages++;
 425                         }
 426                 }
 427         }
 428
 429         RETURN(npages);
 430
 431 out_err:
 432         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 433         RETURN(rc);
 434 }
 435
 436 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 437                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 438                         int rw, struct lustre_capa *capa)
 439 {
 440         struct osd_object *obj  = osd_dt_obj(dt);
 441         int                rc;
 442
 443         LASSERT(dt_object_exists(dt));
 444         LASSERT(obj->oo_db);
 445
 446         if (rw == 0)
 447                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 448         else
 449                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 450
 451         return rc;
 452 }
 453
 454 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 455                         struct niobuf_local *lnb, int npages)
 456 {
 457         struct osd_object *obj = osd_dt_obj(dt);
 458
 459         LASSERT(dt_object_exists(dt));
 460         LASSERT(obj->oo_db);
 461
 462         return 0;
 463 }
 464
 465 /* Return number of blocks that aren't mapped in the [start, start + size]
 466  * region */
 467 static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
 468                                 uint32_t size)
 469 {
 470         dmu_buf_impl_t  *dbi = (dmu_buf_impl_t *)obj->oo_db;
 471         dmu_buf_impl_t  *db;
 472         dnode_t         *dn;
 473         uint32_t         blkshift;
 474         uint64_t         end, blkid;
 475         int              rc;
 476         ENTRY;
 477
 478         DB_DNODE_ENTER(dbi);
 479         dn = DB_DNODE(dbi);
 480
 481         if (dn->dn_maxblkid == 0) {
 482                 if (start + size <= dn->dn_datablksz)
 483                         GOTO(out, size = 0);
 484                 if (start < dn->dn_datablksz)
 485                         start = dn->dn_datablksz;
 486                 /* assume largest block size */
 487                 blkshift = SPA_MAXBLOCKSHIFT;
 488         } else {
 489                 /* blocksize can't change */
 490                 blkshift = dn->dn_datablkshift;
 491         }
 492
 493         /* compute address of last block */
 494         end = (start + size - 1) >> blkshift;
 495         /* align start on block boundaries */
 496         start >>= blkshift;
 497
 498         /* size is null, can't be mapped */
 499         if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
 500                 GOTO(out, size = (end - start + 1) << blkshift);
 501
 502         /* beyond EOF, can't be mapped */
 503         if (start > dn->dn_maxblkid)
 504                 GOTO(out, size = (end - start + 1) << blkshift);
 505
 506         size = 0;
 507         for (blkid = start; blkid <= end; blkid++) {
 508                 if (blkid == dn->dn_maxblkid)
 509                         /* this one is mapped for sure */
 510                         continue;
 511                 if (blkid > dn->dn_maxblkid) {
 512                         size += (end - blkid + 1) << blkshift;
 513                         GOTO(out, size);
 514                 }
 515
 516                 rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
 517                 if (rc) {
 518                         /* for ENOENT (block not mapped) and any other errors,
 519                          * assume the block isn't mapped */
 520                         size += 1 << blkshift;
 521                         continue;
 522                 }
 523                 dbuf_rele(db, FTAG);
 524         }
 525
 526         GOTO(out, size);
 527 out:
 528         DB_DNODE_EXIT(dbi);
 529         return size;
 530 }
 531
 532 static int osd_declare_write_commit(const struct lu_env *env,
 533                                 struct dt_object *dt,
 534                                 struct niobuf_local *lnb, int npages,
 535                                 struct thandle *th)
 536 {
 537         struct osd_object  *obj = osd_dt_obj(dt);
 538         struct osd_device  *osd = osd_obj2dev(obj);
 539         struct osd_thandle *oh;
 540         uint64_t            offset = 0;
 541         uint32_t            size = 0;
 542         int                 i, rc, flags = 0;
 543         bool                ignore_quota = false, synced = false;
 544         long long           space = 0;
 545         ENTRY;
 546
 547         LASSERT(dt_object_exists(dt));
 548         LASSERT(obj->oo_db);
 549
 550         LASSERT(lnb);
 551         LASSERT(npages > 0);
 552
 553         oh = container_of0(th, struct osd_thandle, ot_super);
 554
 555         for (i = 0; i < npages; i++) {
 556                 if (lnb[i].lnb_rc)
 557                         /* ENOSPC, network RPC error, etc.
 558                          * We don't want to book space for pages which will be
 559                          * skipped in osd_write_commit(). Hence we skip pages
 560                          * with lnb_rc != 0 here too */
 561                         continue;
 562                 /* ignore quota for the whole request if any page is from
 563                  * client cache or written by root.
 564                  *
 565                  * XXX once we drop the 1.8 client support, the checking
 566                  * for whether page is from cache can be simplified as:
 567                  * !(lnb[i].flags & OBD_BRW_SYNC)
 568                  *
 569                  * XXX we could handle this on per-lnb basis as done by
 570                  * grant. */
 571                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 572                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 573                     OBD_BRW_FROM_GRANT)
 574                         ignore_quota = true;
 575                 if (size == 0) {
 576                         /* first valid lnb */
 577                         offset = lnb[i].lnb_file_offset;
 578                         size = lnb[i].lnb_len;
 579                         continue;
 580                 }
 581                 if (offset + size == lnb[i].lnb_file_offset) {
 582                         /* this lnb is contiguous to the previous one */
 583                         size += lnb[i].lnb_len;
 584                         continue;
 585                 }
 586
 587                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 588                                   offset, size);
 589                 /* estimating space that will be consumed by a write is rather
 590                  * complicated with ZFS. As a consequence, we don't account for
 591                  * indirect blocks and quota overrun will be adjusted once the
 592                  * operation is committed, if required. */
 593                 space += osd_count_not_mapped(obj, offset, size);
 594
 595                 offset = lnb[i].lnb_file_offset;
 596                 size = lnb[i].lnb_len;
 597         }
 598
 599         if (size) {
 600                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 601                                   offset, size);
 602                 space += osd_count_not_mapped(obj, offset, size);
 603         }
 604
 605         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 606
 607         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 608
 609         /* backend zfs filesystem might be configured to store multiple data
 610          * copies */
 611         space  *= osd->od_os->os_copies;
 612         space   = toqb(space);
 613         CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
 614                "space\n", npages, space);
 615
 616 retry:
 617         /* acquire quota space if needed */
 618         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 619                                obj->oo_attr.la_gid, space, oh, true, &flags,
 620                                ignore_quota);
 621
 622         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 623                 dt_sync(env, th->th_dev);
 624                 synced = true;
 625                 CDEBUG(D_QUOTA, "retry after sync\n");
 626                 flags = 0;
 627                 goto retry;
 628         }
 629
 630         /* we need only to store the overquota flags in the first lnb for
 631          * now, once we support multiple objects BRW, this code needs be
 632          * revised. */
 633         if (flags & QUOTA_FL_OVER_USRQUOTA)
 634                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 635         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 636                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 637
 638         RETURN(rc);
 639 }
 640
 641 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 642                         struct niobuf_local *lnb, int npages,
 643                         struct thandle *th)
 644 {
 645         struct osd_object  *obj  = osd_dt_obj(dt);
 646         struct osd_device  *osd = osd_obj2dev(obj);
 647         struct osd_thandle *oh;
 648         uint64_t            new_size = 0;
 649         int                 i, rc = 0;
 650         ENTRY;
 651
 652         LASSERT(dt_object_exists(dt));
 653         LASSERT(obj->oo_db);
 654
 655         LASSERT(th != NULL);
 656         oh = container_of0(th, struct osd_thandle, ot_super);
 657
 658         for (i = 0; i < npages; i++) {
 659                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 660                         (unsigned) lnb[i].lnb_len,
 661                         (unsigned) lnb[i].lnb_file_offset);
 662
 663                 if (lnb[i].lnb_rc) {
 664                         /* ENOSPC, network RPC error, etc.
 665                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 666                          * so we skip this page if lnb_rc is set to -ENOSPC */
 667                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 668                                 PFID(lu_object_fid(&dt->do_lu)), i,
 669                                 lnb[i].lnb_rc);
 670                         continue;
 671                 }
 672
 673                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 674                         dmu_write(osd->od_os, obj->oo_db->db_object,
 675                                 lnb[i].lnb_file_offset, lnb[i].lnb_len,
 676                                 kmap(lnb[i].lnb_page), oh->ot_tx);
 677                         kunmap(lnb[i].lnb_page);
 678                 } else if (lnb[i].lnb_data) {
 679                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 680                         /* buffer loaned for zerocopy, try to use it.
 681                          * notice that dmu_assign_arcbuf() is smart
 682                          * enough to recognize changed blocksize
 683                          * in this case it fallbacks to dmu_write() */
 684                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 685                                           lnb[i].lnb_data, oh->ot_tx);
 686                         /* drop the reference, otherwise osd_put_bufs()
 687                          * will be releasing it - bad! */
 688                         lnb[i].lnb_data = NULL;
 689                         atomic_dec(&osd->od_zerocopy_loan);
 690                 }
 691
 692                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 693                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 694         }
 695
 696         if (unlikely(new_size == 0)) {
 697                 /* no pages to write, no transno is needed */
 698                 th->th_local = 1;
 699                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 700                  * since ofd_commitrw_write() retries several times on ENOSPC */
 701                 RETURN(0);
 702         }
 703
 704         write_lock(&obj->oo_attr_lock);
 705         if (obj->oo_attr.la_size < new_size) {
 706                 obj->oo_attr.la_size = new_size;
 707                 write_unlock(&obj->oo_attr_lock);
 708                 /* osd_object_sa_update() will be copying directly from
 709                  * oo_attr into dbuf. any update within a single txg will copy
 710                  * the most actual */
 711                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 712                                           &obj->oo_attr.la_size, 8, oh);
 713         } else {
 714                 write_unlock(&obj->oo_attr_lock);
 715         }
 716
 717         RETURN(rc);
 718 }
 719
 720 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 721                         struct niobuf_local *lnb, int npages)
 722 {
 723         struct osd_object *obj  = osd_dt_obj(dt);
 724         loff_t              size;
 725         int                 i;
 726
 727         LASSERT(dt_object_exists(dt));
 728         LASSERT(obj->oo_db);
 729
 730         read_lock(&obj->oo_attr_lock);
 731         size = obj->oo_attr.la_size;
 732         read_unlock(&obj->oo_attr_lock);
 733
 734         for (i = 0; i < npages; i++) {
 735                 if (unlikely(lnb[i].lnb_rc < 0))
 736                         continue;
 737
 738                 lnb[i].lnb_rc = lnb[i].lnb_len;
 739
 740                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > size) {
 741                         lnb[i].lnb_rc = size - lnb[i].lnb_file_offset;
 742
 743                         /* all subsequent rc should be 0 */
 744                         while (++i < npages)
 745                                 lnb[i].lnb_rc = 0;
 746                         break;
 747                 }
 748         }
 749
 750         return 0;
 751 }
 752
 753 /*
 754  * Punch/truncate an object
 755  *
 756  *      IN:     db  - dmu_buf of the object to free data in.
 757  *              off - start of section to free.
 758  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 759  *
 760  *      RETURN: 0 if success
 761  *              error code if failure
 762  *
 763  * The transaction passed to this routine must have
 764  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 765  * called and then assigned to a transaction group.
 766  */
 767 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 768                                 uint64_t size, uint64_t off, uint64_t len)
 769 {
 770         int rc = 0;
 771
 772         /* Assert that the transaction has been assigned to a
 773            transaction group. */
 774         LASSERT(tx->tx_txg != 0);
 775         /*
 776          * Nothing to do if file already at desired length.
 777          */
 778         if (len == DMU_OBJECT_END && size == off)
 779                 return 0;
 780
 781         if (off < size)
 782                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 783
 784         return rc;
 785 }
 786
 787 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 788                         __u64 start, __u64 end, struct thandle *th,
 789                         struct lustre_capa *capa)
 790 {
 791         struct osd_object  *obj = osd_dt_obj(dt);
 792         struct osd_device  *osd = osd_obj2dev(obj);
 793         struct osd_thandle *oh;
 794         __u64               len;
 795         int                 rc = 0;
 796         ENTRY;
 797
 798         LASSERT(dt_object_exists(dt));
 799         LASSERT(osd_invariant(obj));
 800
 801         LASSERT(th != NULL);
 802         oh = container_of0(th, struct osd_thandle, ot_super);
 803
 804         write_lock(&obj->oo_attr_lock);
 805         /* truncate */
 806         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 807                 len = DMU_OBJECT_END;
 808         else
 809                 len = end - start;
 810         write_unlock(&obj->oo_attr_lock);
 811
 812         rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
 813                                 obj->oo_attr.la_size, start, len);
 814         /* set new size */
 815         if (len == DMU_OBJECT_END) {
 816                 write_lock(&obj->oo_attr_lock);
 817                 obj->oo_attr.la_size = start;
 818                 write_unlock(&obj->oo_attr_lock);
 819                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 820                                           &obj->oo_attr.la_size, 8, oh);
 821         }
 822         RETURN(rc);
 823 }
 824
 825 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 826                         __u64 start, __u64 end, struct thandle *handle)
 827 {
 828         struct osd_object  *obj = osd_dt_obj(dt);
 829         struct osd_device  *osd = osd_obj2dev(obj);
 830         struct osd_thandle *oh;
 831         __u64               len;
 832         ENTRY;
 833
 834         oh = container_of0(handle, struct osd_thandle, ot_super);
 835
 836         read_lock(&obj->oo_attr_lock);
 837         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 838                 len = DMU_OBJECT_END;
 839         else
 840                 len = end - start;
 841
 842         /* declare we'll free some blocks ... */
 843         if (start < obj->oo_attr.la_size) {
 844                 read_unlock(&obj->oo_attr_lock);
 845                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 846         } else {
 847                 read_unlock(&obj->oo_attr_lock);
 848         }
 849
 850         /* ... and we'll modify size attribute */
 851         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 852
 853         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 854                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 855                                  false));
 856 }
 857
 858
 859 struct dt_body_operations osd_body_ops = {
 860         .dbo_read                       = osd_read,
 861         .dbo_declare_write              = osd_declare_write,
 862         .dbo_write                      = osd_write,
 863         .dbo_bufs_get                   = osd_bufs_get,
 864         .dbo_bufs_put                   = osd_bufs_put,
 865         .dbo_write_prep                 = osd_write_prep,
 866         .dbo_declare_write_commit       = osd_declare_write_commit,
 867         .dbo_write_commit               = osd_write_commit,
 868         .dbo_read_prep                  = osd_read_prep,
 869         .dbo_declare_punch              = osd_declare_punch,
 870         .dbo_punch                      = osd_punch,
 871 };
 872