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  * Copyright (c) 2012, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd-zfs/osd_io.c
  37  *
  38  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  39  * Author: Mike Pershin <tappro@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSD
  43
  44 #include <lustre_ver.h>
  45 #include <libcfs/libcfs.h>
  46 #include <obd_support.h>
  47 #include <lustre_net.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_disk.h>
  51 #include <lustre_fid.h>
  52 #include <lustre/lustre_idl.h>  /* LLOG_MIN_CHUNK_SIZE definition */
  53
  54 #include "osd_internal.h"
  55
  56 #include <sys/dnode.h>
  57 #include <sys/dbuf.h>
  58 #include <sys/spa.h>
  59 #include <sys/stat.h>
  60 #include <sys/zap.h>
  61 #include <sys/spa_impl.h>
  62 #include <sys/zfs_znode.h>
  63 #include <sys/dmu_tx.h>
  64 #include <sys/dmu_objset.h>
  65 #include <sys/dsl_prop.h>
  66 #include <sys/sa_impl.h>
  67 #include <sys/txg.h>
  68
  69 static char *osd_zerocopy_tag = "zerocopy";
  70
  71
  72 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  73 {
  74         struct obd_histogram *h = osd->od_brw_stats.hist;
  75
  76         if (rw == READ) {
  77                 atomic_inc(&osd->od_r_in_flight);
  78                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  79                                  atomic_read(&osd->od_r_in_flight));
  80                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  81
  82         } else {
  83                 atomic_inc(&osd->od_w_in_flight);
  84                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  85                                  atomic_read(&osd->od_w_in_flight));
  86                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  87
  88         }
  89 }
  90
  91 static void record_end_io(struct osd_device *osd, int rw,
  92                           unsigned long elapsed, int disksize, int npages)
  93 {
  94         struct obd_histogram *h = osd->od_brw_stats.hist;
  95
  96         if (rw == READ) {
  97                 atomic_dec(&osd->od_r_in_flight);
  98                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  99                 if (disksize > 0)
 100                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
 101                 if (elapsed)
 102                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
 103
 104         } else {
 105                 atomic_dec(&osd->od_w_in_flight);
 106                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 107                 if (disksize > 0)
 108                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 109                 if (elapsed)
 110                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 111         }
 112 }
 113
 114 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 115                         struct lu_buf *buf, loff_t *pos)
 116 {
 117         struct osd_object *obj  = osd_dt_obj(dt);
 118         struct osd_device *osd = osd_obj2dev(obj);
 119         uint64_t           old_size;
 120         int                size = buf->lb_len;
 121         int                rc;
 122         unsigned long      start;
 123
 124         LASSERT(dt_object_exists(dt));
 125         LASSERT(obj->oo_db);
 126
 127         start = cfs_time_current();
 128
 129         read_lock(&obj->oo_attr_lock);
 130         old_size = obj->oo_attr.la_size;
 131         read_unlock(&obj->oo_attr_lock);
 132
 133         if (*pos + size > old_size) {
 134                 if (old_size < *pos)
 135                         return 0;
 136                 else
 137                         size = old_size - *pos;
 138         }
 139
 140         record_start_io(osd, READ, 0);
 141
 142         rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
 143                         buf->lb_buf, DMU_READ_PREFETCH);
 144
 145         record_end_io(osd, READ, cfs_time_current() - start, size,
 146                       size >> PAGE_CACHE_SHIFT);
 147         if (rc == 0) {
 148                 rc = size;
 149                 *pos += size;
 150         }
 151         return rc;
 152 }
 153
 154 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 155                                 const struct lu_buf *buf, loff_t pos,
 156                                 struct thandle *th)
 157 {
 158         struct osd_object  *obj  = osd_dt_obj(dt);
 159         struct osd_device  *osd = osd_obj2dev(obj);
 160         struct osd_thandle *oh;
 161         uint64_t            oid;
 162         ENTRY;
 163
 164         oh = container_of0(th, struct osd_thandle, ot_super);
 165
 166         /* in some cases declare can race with creation (e.g. llog)
 167          * and we need to wait till object is initialized. notice
 168          * LOHA_EXISTs is supposed to be the last step in the
 169          * initialization */
 170
 171         /* declare possible size change. notice we can't check
 172          * current size here as another thread can change it */
 173
 174         if (dt_object_exists(dt)) {
 175                 LASSERT(obj->oo_db);
 176                 oid = obj->oo_db->db_object;
 177
 178                 dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 179         } else {
 180                 oid = DMU_NEW_OBJECT;
 181                 dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
 182         }
 183
 184         /* XXX: we still miss for append declaration support in ZFS
 185          *      -1 means append which is used by llog mostly, llog
 186          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 187         if (pos == -1)
 188                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 189                             obj->oo_attr.la_size + (2 << 20));
 190         dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
 191
 192         /* dt_declare_write() is usually called for system objects, such
 193          * as llog or last_rcvd files. We needn't enforce quota on those
 194          * objects, so always set the lqi_space as 0. */
 195         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 196                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 197                                  false));
 198 }
 199
 200 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 201                         const struct lu_buf *buf, loff_t *pos,
 202                         struct thandle *th, int ignore_quota)
 203 {
 204         struct osd_object  *obj  = osd_dt_obj(dt);
 205         struct osd_device  *osd = osd_obj2dev(obj);
 206         struct osd_thandle *oh;
 207         uint64_t            offset = *pos;
 208         int                 rc;
 209
 210         ENTRY;
 211
 212         LASSERT(dt_object_exists(dt));
 213         LASSERT(obj->oo_db);
 214
 215         LASSERT(th != NULL);
 216         oh = container_of0(th, struct osd_thandle, ot_super);
 217
 218         record_start_io(osd, WRITE, 0);
 219
 220         dmu_write(osd->od_os, obj->oo_db->db_object, offset,
 221                 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
 222         write_lock(&obj->oo_attr_lock);
 223         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 224                 obj->oo_attr.la_size = offset + buf->lb_len;
 225                 write_unlock(&obj->oo_attr_lock);
 226                 /* osd_object_sa_update() will be copying directly from oo_attr
 227                  * into dbuf.  any update within a single txg will copy the
 228                  * most actual */
 229                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 230                                         &obj->oo_attr.la_size, 8, oh);
 231                 if (unlikely(rc))
 232                         GOTO(out, rc);
 233         } else {
 234                 write_unlock(&obj->oo_attr_lock);
 235         }
 236
 237         *pos += buf->lb_len;
 238         rc = buf->lb_len;
 239
 240 out:
 241         record_end_io(osd, WRITE, 0, buf->lb_len,
 242                       buf->lb_len >> PAGE_CACHE_SHIFT);
 243
 244         RETURN(rc);
 245 }
 246
 247 /*
 248  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 249  *      purposes as it's not very well defined ...
 250  *      instead I use the lowest bit of the address so that:
 251  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 252  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 253  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 254  *
 255  *      bzzz, to blame
 256  */
 257 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 258                         struct niobuf_local *lnb, int npages)
 259 {
 260         struct osd_object *obj  = osd_dt_obj(dt);
 261         struct osd_device *osd = osd_obj2dev(obj);
 262         unsigned long      ptr;
 263         int                i;
 264
 265         LASSERT(dt_object_exists(dt));
 266         LASSERT(obj->oo_db);
 267
 268         for (i = 0; i < npages; i++) {
 269                 if (lnb[i].lnb_page == NULL)
 270                         continue;
 271                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 272                         /* this is anonymous page allocated for copy-write */
 273                         lnb[i].lnb_page->mapping = NULL;
 274                         __free_page(lnb[i].lnb_page);
 275                         atomic_dec(&osd->od_zerocopy_alloc);
 276                 } else {
 277                         /* see comment in osd_bufs_get_read() */
 278                         ptr = (unsigned long)lnb[i].lnb_data;
 279                         if (ptr & 1UL) {
 280                                 ptr &= ~1UL;
 281                                 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
 282                                 atomic_dec(&osd->od_zerocopy_pin);
 283                         } else if (lnb[i].lnb_data != NULL) {
 284                                 dmu_return_arcbuf(lnb[i].lnb_data);
 285                                 atomic_dec(&osd->od_zerocopy_loan);
 286                         }
 287                 }
 288                 lnb[i].lnb_page = NULL;
 289                 lnb[i].lnb_data = NULL;
 290         }
 291
 292         return 0;
 293 }
 294
 295 static inline struct page *kmem_to_page(void *addr)
 296 {
 297         if (is_vmalloc_addr(addr))
 298                 return vmalloc_to_page(addr);
 299         else
 300                 return virt_to_page(addr);
 301 }
 302
 303 /**
 304  * Prepare buffers for read.
 305  *
 306  * The function maps the range described by \a off and \a len to \a lnb array.
 307  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 308  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 309  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 310  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 311  * A better implementation would just return the buffers (potentially unfilled)
 312  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 313  *
 314  * \param[in] env       environment
 315  * \param[in] obj       object
 316  * \param[in] off       offset in bytes
 317  * \param[in] len       the number of bytes to access
 318  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 319  *
 320  * \retval              0 for success
 321  * \retval              negative error number of failure
 322  */
 323 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 324                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 325 {
 326         struct osd_device *osd = osd_obj2dev(obj);
 327         unsigned long      start = cfs_time_current();
 328         int                rc, i, numbufs, npages = 0;
 329         dmu_buf_t        **dbp;
 330         ENTRY;
 331
 332         record_start_io(osd, READ, 0);
 333
 334         /* grab buffers for read:
 335          * OSD API let us to grab buffers first, then initiate IO(s)
 336          * so that all required IOs will be done in parallel, but at the
 337          * moment DMU doesn't provide us with a method to grab buffers.
 338          * If we discover this is a vital for good performance we
 339          * can get own replacement for dmu_buf_hold_array_by_bonus().
 340          */
 341         while (len > 0) {
 342                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 343                                                   osd_zerocopy_tag, &numbufs,
 344                                                   &dbp);
 345                 if (unlikely(rc))
 346                         GOTO(err, rc);
 347
 348                 for (i = 0; i < numbufs; i++) {
 349                         int bufoff, tocpy, thispage;
 350                         void *dbf = dbp[i];
 351
 352                         LASSERT(len > 0);
 353
 354                         atomic_inc(&osd->od_zerocopy_pin);
 355
 356                         bufoff = off - dbp[i]->db_offset;
 357                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 358
 359                         /* kind of trick to differentiate dbuf vs. arcbuf */
 360                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 361                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 362
 363                         while (tocpy > 0) {
 364                                 thispage = PAGE_CACHE_SIZE;
 365                                 thispage -= bufoff & (PAGE_CACHE_SIZE - 1);
 366                                 thispage = min(tocpy, thispage);
 367
 368                                 lnb->lnb_rc = 0;
 369                                 lnb->lnb_file_offset = off;
 370                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 371                                 lnb->lnb_len = thispage;
 372                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 373                                                              bufoff);
 374                                 /* mark just a single slot: we need this
 375                                  * reference to dbuf to be released once */
 376                                 lnb->lnb_data = dbf;
 377                                 dbf = NULL;
 378
 379                                 tocpy -= thispage;
 380                                 len -= thispage;
 381                                 bufoff += thispage;
 382                                 off += thispage;
 383
 384                                 npages++;
 385                                 lnb++;
 386                         }
 387
 388                         /* steal dbuf so dmu_buf_rele_array() can't release
 389                          * it */
 390                         dbp[i] = NULL;
 391                 }
 392
 393                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 394         }
 395
 396         record_end_io(osd, READ, cfs_time_current() - start,
 397                       npages * PAGE_SIZE, npages);
 398
 399         RETURN(npages);
 400
 401 err:
 402         LASSERT(rc < 0);
 403         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 404         RETURN(rc);
 405 }
 406
 407 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 408                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 409 {
 410         struct osd_device *osd = osd_obj2dev(obj);
 411         int                plen, off_in_block, sz_in_block;
 412         int                rc, i = 0, npages = 0;
 413         arc_buf_t         *abuf;
 414         uint32_t           bs;
 415         uint64_t           dummy;
 416         ENTRY;
 417
 418         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 419
 420         /*
 421          * currently only full blocks are subject to zerocopy approach:
 422          * so that we're sure nobody is trying to update the same block
 423          */
 424         while (len > 0) {
 425                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 426
 427                 off_in_block = off & (bs - 1);
 428                 sz_in_block = min_t(int, bs - off_in_block, len);
 429
 430                 if (sz_in_block == bs) {
 431                         /* full block, try to use zerocopy */
 432
 433                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 434                         if (unlikely(abuf == NULL))
 435                                 GOTO(out_err, rc = -ENOMEM);
 436
 437                         atomic_inc(&osd->od_zerocopy_loan);
 438
 439                         /* go over pages arcbuf contains, put them as
 440                          * local niobufs for ptlrpc's bulks */
 441                         while (sz_in_block > 0) {
 442                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 443
 444                                 lnb[i].lnb_file_offset = off;
 445                                 lnb[i].lnb_page_offset = 0;
 446                                 lnb[i].lnb_len = plen;
 447                                 lnb[i].lnb_rc = 0;
 448                                 if (sz_in_block == bs)
 449                                         lnb[i].lnb_data = abuf;
 450                                 else
 451                                         lnb[i].lnb_data = NULL;
 452
 453                                 /* this one is not supposed to fail */
 454                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 455                                                         off_in_block);
 456                                 LASSERT(lnb[i].lnb_page);
 457
 458                                 lprocfs_counter_add(osd->od_stats,
 459                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 460
 461                                 sz_in_block -= plen;
 462                                 len -= plen;
 463                                 off += plen;
 464                                 off_in_block += plen;
 465                                 i++;
 466                                 npages++;
 467                         }
 468                 } else {
 469                         if (off_in_block == 0 && len < bs &&
 470                                         off + len >= obj->oo_attr.la_size)
 471                                 lprocfs_counter_add(osd->od_stats,
 472                                                 LPROC_OSD_TAIL_IO, 1);
 473
 474                         /* can't use zerocopy, allocate temp. buffers */
 475                         while (sz_in_block > 0) {
 476                                 plen = min_t(int, sz_in_block, PAGE_CACHE_SIZE);
 477
 478                                 lnb[i].lnb_file_offset = off;
 479                                 lnb[i].lnb_page_offset = 0;
 480                                 lnb[i].lnb_len = plen;
 481                                 lnb[i].lnb_rc = 0;
 482                                 lnb[i].lnb_data = NULL;
 483
 484                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 485                                 if (unlikely(lnb[i].lnb_page == NULL))
 486                                         GOTO(out_err, rc = -ENOMEM);
 487
 488                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 489                                 lnb[i].lnb_page->mapping = (void *)obj;
 490
 491                                 atomic_inc(&osd->od_zerocopy_alloc);
 492                                 lprocfs_counter_add(osd->od_stats,
 493                                                 LPROC_OSD_COPY_IO, 1);
 494
 495                                 sz_in_block -= plen;
 496                                 len -= plen;
 497                                 off += plen;
 498                                 i++;
 499                                 npages++;
 500                         }
 501                 }
 502         }
 503
 504         RETURN(npages);
 505
 506 out_err:
 507         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 508         RETURN(rc);
 509 }
 510
 511 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 512                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 513                         int rw)
 514 {
 515         struct osd_object *obj  = osd_dt_obj(dt);
 516         int                rc;
 517
 518         LASSERT(dt_object_exists(dt));
 519         LASSERT(obj->oo_db);
 520
 521         if (rw == 0)
 522                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 523         else
 524                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 525
 526         return rc;
 527 }
 528
 529 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 530                         struct niobuf_local *lnb, int npages)
 531 {
 532         struct osd_object *obj = osd_dt_obj(dt);
 533
 534         LASSERT(dt_object_exists(dt));
 535         LASSERT(obj->oo_db);
 536
 537         return 0;
 538 }
 539
 540 /* Return number of blocks that aren't mapped in the [start, start + size]
 541  * region */
 542 static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
 543                                 uint32_t size)
 544 {
 545         dmu_buf_impl_t  *dbi = (dmu_buf_impl_t *)obj->oo_db;
 546         dmu_buf_impl_t  *db;
 547         dnode_t         *dn;
 548         uint32_t         blkshift;
 549         uint64_t         end, blkid;
 550         int              rc;
 551         ENTRY;
 552
 553         DB_DNODE_ENTER(dbi);
 554         dn = DB_DNODE(dbi);
 555
 556         if (dn->dn_maxblkid == 0) {
 557                 if (start + size <= dn->dn_datablksz)
 558                         GOTO(out, size = 0);
 559                 if (start < dn->dn_datablksz)
 560                         start = dn->dn_datablksz;
 561                 /* assume largest block size */
 562                 blkshift = osd_spa_maxblockshift(
 563                         dmu_objset_spa(osd_obj2dev(obj)->od_os));
 564         } else {
 565                 /* blocksize can't change */
 566                 blkshift = dn->dn_datablkshift;
 567         }
 568
 569         /* compute address of last block */
 570         end = (start + size - 1) >> blkshift;
 571         /* align start on block boundaries */
 572         start >>= blkshift;
 573
 574         /* size is null, can't be mapped */
 575         if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
 576                 GOTO(out, size = (end - start + 1) << blkshift);
 577
 578         /* beyond EOF, can't be mapped */
 579         if (start > dn->dn_maxblkid)
 580                 GOTO(out, size = (end - start + 1) << blkshift);
 581
 582         size = 0;
 583         for (blkid = start; blkid <= end; blkid++) {
 584                 if (blkid == dn->dn_maxblkid)
 585                         /* this one is mapped for sure */
 586                         continue;
 587                 if (blkid > dn->dn_maxblkid) {
 588                         size += (end - blkid + 1) << blkshift;
 589                         GOTO(out, size);
 590                 }
 591
 592                 rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
 593                 if (rc) {
 594                         /* for ENOENT (block not mapped) and any other errors,
 595                          * assume the block isn't mapped */
 596                         size += 1 << blkshift;
 597                         continue;
 598                 }
 599                 dbuf_rele(db, FTAG);
 600         }
 601
 602         GOTO(out, size);
 603 out:
 604         DB_DNODE_EXIT(dbi);
 605         return size;
 606 }
 607
 608 static int osd_declare_write_commit(const struct lu_env *env,
 609                                 struct dt_object *dt,
 610                                 struct niobuf_local *lnb, int npages,
 611                                 struct thandle *th)
 612 {
 613         struct osd_object  *obj = osd_dt_obj(dt);
 614         struct osd_device  *osd = osd_obj2dev(obj);
 615         struct osd_thandle *oh;
 616         uint64_t            offset = 0;
 617         uint32_t            size = 0;
 618         int                 i, rc, flags = 0;
 619         bool                ignore_quota = false, synced = false;
 620         long long           space = 0;
 621         struct page        *last_page = NULL;
 622         unsigned long       discont_pages = 0;
 623         ENTRY;
 624
 625         LASSERT(dt_object_exists(dt));
 626         LASSERT(obj->oo_db);
 627
 628         LASSERT(lnb);
 629         LASSERT(npages > 0);
 630
 631         oh = container_of0(th, struct osd_thandle, ot_super);
 632
 633         for (i = 0; i < npages; i++) {
 634                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 635                         ++discont_pages;
 636                 last_page = lnb[i].lnb_page;
 637                 if (lnb[i].lnb_rc)
 638                         /* ENOSPC, network RPC error, etc.
 639                          * We don't want to book space for pages which will be
 640                          * skipped in osd_write_commit(). Hence we skip pages
 641                          * with lnb_rc != 0 here too */
 642                         continue;
 643                 /* ignore quota for the whole request if any page is from
 644                  * client cache or written by root.
 645                  *
 646                  * XXX once we drop the 1.8 client support, the checking
 647                  * for whether page is from cache can be simplified as:
 648                  * !(lnb[i].flags & OBD_BRW_SYNC)
 649                  *
 650                  * XXX we could handle this on per-lnb basis as done by
 651                  * grant. */
 652                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 653                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 654                     OBD_BRW_FROM_GRANT)
 655                         ignore_quota = true;
 656                 if (size == 0) {
 657                         /* first valid lnb */
 658                         offset = lnb[i].lnb_file_offset;
 659                         size = lnb[i].lnb_len;
 660                         continue;
 661                 }
 662                 if (offset + size == lnb[i].lnb_file_offset) {
 663                         /* this lnb is contiguous to the previous one */
 664                         size += lnb[i].lnb_len;
 665                         continue;
 666                 }
 667
 668                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 669                                   offset, size);
 670                 /* estimating space that will be consumed by a write is rather
 671                  * complicated with ZFS. As a consequence, we don't account for
 672                  * indirect blocks and quota overrun will be adjusted once the
 673                  * operation is committed, if required. */
 674                 space += osd_count_not_mapped(obj, offset, size);
 675
 676                 offset = lnb[i].lnb_file_offset;
 677                 size = lnb[i].lnb_len;
 678         }
 679
 680         if (size) {
 681                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 682                                   offset, size);
 683                 space += osd_count_not_mapped(obj, offset, size);
 684         }
 685
 686         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 687
 688         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 689
 690         /* backend zfs filesystem might be configured to store multiple data
 691          * copies */
 692         space  *= osd->od_os->os_copies;
 693         space   = toqb(space);
 694         CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
 695                "space\n", npages, space);
 696
 697         record_start_io(osd, WRITE, discont_pages);
 698 retry:
 699         /* acquire quota space if needed */
 700         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 701                                obj->oo_attr.la_gid, space, oh, true, &flags,
 702                                ignore_quota);
 703
 704         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 705                 dt_sync(env, th->th_dev);
 706                 synced = true;
 707                 CDEBUG(D_QUOTA, "retry after sync\n");
 708                 flags = 0;
 709                 goto retry;
 710         }
 711
 712         /* we need only to store the overquota flags in the first lnb for
 713          * now, once we support multiple objects BRW, this code needs be
 714          * revised. */
 715         if (flags & QUOTA_FL_OVER_USRQUOTA)
 716                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 717         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 718                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 719
 720         RETURN(rc);
 721 }
 722
 723 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 724                         struct niobuf_local *lnb, int npages,
 725                         struct thandle *th)
 726 {
 727         struct osd_object  *obj  = osd_dt_obj(dt);
 728         struct osd_device  *osd = osd_obj2dev(obj);
 729         struct osd_thandle *oh;
 730         uint64_t            new_size = 0;
 731         int                 i, rc = 0;
 732         unsigned long      iosize = 0;
 733         ENTRY;
 734
 735         LASSERT(dt_object_exists(dt));
 736         LASSERT(obj->oo_db);
 737
 738         LASSERT(th != NULL);
 739         oh = container_of0(th, struct osd_thandle, ot_super);
 740
 741         for (i = 0; i < npages; i++) {
 742                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 743                         (unsigned) lnb[i].lnb_len,
 744                         (unsigned) lnb[i].lnb_file_offset);
 745
 746                 if (lnb[i].lnb_rc) {
 747                         /* ENOSPC, network RPC error, etc.
 748                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 749                          * so we skip this page if lnb_rc is set to -ENOSPC */
 750                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 751                                 PFID(lu_object_fid(&dt->do_lu)), i,
 752                                 lnb[i].lnb_rc);
 753                         continue;
 754                 }
 755
 756                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 757                         dmu_write(osd->od_os, obj->oo_db->db_object,
 758                                 lnb[i].lnb_file_offset, lnb[i].lnb_len,
 759                                 kmap(lnb[i].lnb_page), oh->ot_tx);
 760                         kunmap(lnb[i].lnb_page);
 761                 } else if (lnb[i].lnb_data) {
 762                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 763                         /* buffer loaned for zerocopy, try to use it.
 764                          * notice that dmu_assign_arcbuf() is smart
 765                          * enough to recognize changed blocksize
 766                          * in this case it fallbacks to dmu_write() */
 767                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 768                                           lnb[i].lnb_data, oh->ot_tx);
 769                         /* drop the reference, otherwise osd_put_bufs()
 770                          * will be releasing it - bad! */
 771                         lnb[i].lnb_data = NULL;
 772                         atomic_dec(&osd->od_zerocopy_loan);
 773                 }
 774
 775                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 776                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 777                 iosize += lnb[i].lnb_len;
 778         }
 779
 780         if (unlikely(new_size == 0)) {
 781                 /* no pages to write, no transno is needed */
 782                 th->th_local = 1;
 783                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 784                  * since ofd_commitrw_write() retries several times on ENOSPC */
 785                 record_end_io(osd, WRITE, 0, 0, 0);
 786                 RETURN(0);
 787         }
 788
 789         write_lock(&obj->oo_attr_lock);
 790         if (obj->oo_attr.la_size < new_size) {
 791                 obj->oo_attr.la_size = new_size;
 792                 write_unlock(&obj->oo_attr_lock);
 793                 /* osd_object_sa_update() will be copying directly from
 794                  * oo_attr into dbuf. any update within a single txg will copy
 795                  * the most actual */
 796                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 797                                           &obj->oo_attr.la_size, 8, oh);
 798         } else {
 799                 write_unlock(&obj->oo_attr_lock);
 800         }
 801
 802         record_end_io(osd, WRITE, 0, iosize, npages);
 803
 804         RETURN(rc);
 805 }
 806
 807 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 808                         struct niobuf_local *lnb, int npages)
 809 {
 810         struct osd_object *obj  = osd_dt_obj(dt);
 811         int                i;
 812         unsigned long      size = 0;
 813         loff_t             eof;
 814
 815         LASSERT(dt_object_exists(dt));
 816         LASSERT(obj->oo_db);
 817
 818         read_lock(&obj->oo_attr_lock);
 819         eof = obj->oo_attr.la_size;
 820         read_unlock(&obj->oo_attr_lock);
 821
 822         for (i = 0; i < npages; i++) {
 823                 if (unlikely(lnb[i].lnb_rc < 0))
 824                         continue;
 825
 826                 lnb[i].lnb_rc = lnb[i].lnb_len;
 827                 size += lnb[i].lnb_rc;
 828
 829                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > eof) {
 830                         lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 831                         if (lnb[i].lnb_rc < 0)
 832                                 lnb[i].lnb_rc = 0;
 833
 834                         /* all subsequent rc should be 0 */
 835                         while (++i < npages)
 836                                 lnb[i].lnb_rc = 0;
 837                         break;
 838                 }
 839         }
 840
 841         return 0;
 842 }
 843
 844 /*
 845  * Punch/truncate an object
 846  *
 847  *      IN:     db  - dmu_buf of the object to free data in.
 848  *              off - start of section to free.
 849  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 850  *
 851  *      RETURN: 0 if success
 852  *              error code if failure
 853  *
 854  * The transaction passed to this routine must have
 855  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 856  * called and then assigned to a transaction group.
 857  */
 858 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 859                                 uint64_t size, uint64_t off, uint64_t len)
 860 {
 861         int rc = 0;
 862
 863         /* Assert that the transaction has been assigned to a
 864            transaction group. */
 865         LASSERT(tx->tx_txg != 0);
 866         /*
 867          * Nothing to do if file already at desired length.
 868          */
 869         if (len == DMU_OBJECT_END && size == off)
 870                 return 0;
 871
 872         if (off < size)
 873                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 874
 875         return rc;
 876 }
 877
 878 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 879                         __u64 start, __u64 end, struct thandle *th)
 880 {
 881         struct osd_object  *obj = osd_dt_obj(dt);
 882         struct osd_device  *osd = osd_obj2dev(obj);
 883         struct osd_thandle *oh;
 884         __u64               len;
 885         int                 rc = 0;
 886         ENTRY;
 887
 888         LASSERT(dt_object_exists(dt));
 889         LASSERT(osd_invariant(obj));
 890
 891         LASSERT(th != NULL);
 892         oh = container_of0(th, struct osd_thandle, ot_super);
 893
 894         write_lock(&obj->oo_attr_lock);
 895         /* truncate */
 896         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 897                 len = DMU_OBJECT_END;
 898         else
 899                 len = end - start;
 900         write_unlock(&obj->oo_attr_lock);
 901
 902         rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
 903                                 obj->oo_attr.la_size, start, len);
 904         /* set new size */
 905         if (len == DMU_OBJECT_END) {
 906                 write_lock(&obj->oo_attr_lock);
 907                 obj->oo_attr.la_size = start;
 908                 write_unlock(&obj->oo_attr_lock);
 909                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 910                                           &obj->oo_attr.la_size, 8, oh);
 911         }
 912         RETURN(rc);
 913 }
 914
 915 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 916                         __u64 start, __u64 end, struct thandle *handle)
 917 {
 918         struct osd_object  *obj = osd_dt_obj(dt);
 919         struct osd_device  *osd = osd_obj2dev(obj);
 920         struct osd_thandle *oh;
 921         __u64               len;
 922         ENTRY;
 923
 924         oh = container_of0(handle, struct osd_thandle, ot_super);
 925
 926         read_lock(&obj->oo_attr_lock);
 927         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 928                 len = DMU_OBJECT_END;
 929         else
 930                 len = end - start;
 931
 932         /* declare we'll free some blocks ... */
 933         if (start < obj->oo_attr.la_size) {
 934                 read_unlock(&obj->oo_attr_lock);
 935                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 936         } else {
 937                 read_unlock(&obj->oo_attr_lock);
 938         }
 939
 940         /* ... and we'll modify size attribute */
 941         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 942
 943         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 944                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 945                                  false));
 946 }
 947
 948
 949 struct dt_body_operations osd_body_ops = {
 950         .dbo_read                       = osd_read,
 951         .dbo_declare_write              = osd_declare_write,
 952         .dbo_write                      = osd_write,
 953         .dbo_bufs_get                   = osd_bufs_get,
 954         .dbo_bufs_put                   = osd_bufs_put,
 955         .dbo_write_prep                 = osd_write_prep,
 956         .dbo_declare_write_commit       = osd_declare_write_commit,
 957         .dbo_write_commit               = osd_write_commit,
 958         .dbo_read_prep                  = osd_read_prep,
 959         .dbo_declare_punch              = osd_declare_punch,
 960         .dbo_punch                      = osd_punch,
 961 };