lustre/osd-zfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd-zfs/osd_io.c
  33  *
  34  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  35  * Author: Mike Pershin <tappro@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSD
  39
  40 #include <lustre_ver.h>
  41 #include <libcfs/libcfs.h>
  42 #include <obd_support.h>
  43 #include <lustre_net.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_disk.h>
  47 #include <lustre_fid.h>
  48 #include <lustre/lustre_idl.h>  /* LLOG_MIN_CHUNK_SIZE definition */
  49
  50 #include "osd_internal.h"
  51
  52 #include <sys/dnode.h>
  53 #include <sys/dbuf.h>
  54 #include <sys/spa.h>
  55 #include <sys/stat.h>
  56 #include <sys/zap.h>
  57 #include <sys/spa_impl.h>
  58 #include <sys/zfs_znode.h>
  59 #include <sys/dmu_tx.h>
  60 #include <sys/dmu_objset.h>
  61 #include <sys/dsl_prop.h>
  62 #include <sys/sa_impl.h>
  63 #include <sys/txg.h>
  64
  65 static char *osd_0copy_tag = "zerocopy";
  66
  67
  68 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  69 {
  70         struct obd_histogram *h = osd->od_brw_stats.hist;
  71
  72         if (rw == READ) {
  73                 atomic_inc(&osd->od_r_in_flight);
  74                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  75                                  atomic_read(&osd->od_r_in_flight));
  76                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  77
  78         } else {
  79                 atomic_inc(&osd->od_w_in_flight);
  80                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  81                                  atomic_read(&osd->od_w_in_flight));
  82                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  83
  84         }
  85 }
  86
  87 static void record_end_io(struct osd_device *osd, int rw,
  88                           unsigned long elapsed, int disksize, int npages)
  89 {
  90         struct obd_histogram *h = osd->od_brw_stats.hist;
  91
  92         if (rw == READ) {
  93                 atomic_dec(&osd->od_r_in_flight);
  94                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  95                 if (disksize > 0)
  96                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
  97                 if (elapsed)
  98                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
  99
 100         } else {
 101                 atomic_dec(&osd->od_w_in_flight);
 102                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 103                 if (disksize > 0)
 104                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 105                 if (elapsed)
 106                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 107         }
 108 }
 109
 110 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 111                         struct lu_buf *buf, loff_t *pos)
 112 {
 113         struct osd_object *obj  = osd_dt_obj(dt);
 114         struct osd_device *osd = osd_obj2dev(obj);
 115         uint64_t           old_size;
 116         int                size = buf->lb_len;
 117         int                rc;
 118         unsigned long      start;
 119
 120         LASSERT(dt_object_exists(dt));
 121         LASSERT(obj->oo_dn);
 122
 123         start = cfs_time_current();
 124
 125         read_lock(&obj->oo_attr_lock);
 126         old_size = obj->oo_attr.la_size;
 127         read_unlock(&obj->oo_attr_lock);
 128
 129         if (*pos + size > old_size) {
 130                 if (old_size < *pos)
 131                         return 0;
 132                 else
 133                         size = old_size - *pos;
 134         }
 135
 136         record_start_io(osd, READ, 0);
 137
 138         rc = osd_dmu_read(osd, obj->oo_dn, *pos, size, buf->lb_buf,
 139                           DMU_READ_PREFETCH);
 140
 141         record_end_io(osd, READ, cfs_time_current() - start, size,
 142                       size >> PAGE_SHIFT);
 143         if (rc == 0) {
 144                 rc = size;
 145                 *pos += size;
 146         }
 147         return rc;
 148 }
 149
 150 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 151                                 const struct lu_buf *buf, loff_t pos,
 152                                 struct thandle *th)
 153 {
 154         struct osd_object  *obj  = osd_dt_obj(dt);
 155         struct osd_device  *osd = osd_obj2dev(obj);
 156         struct osd_thandle *oh;
 157         uint64_t            oid;
 158         ENTRY;
 159
 160         oh = container_of0(th, struct osd_thandle, ot_super);
 161
 162         /* in some cases declare can race with creation (e.g. llog)
 163          * and we need to wait till object is initialized. notice
 164          * LOHA_EXISTs is supposed to be the last step in the
 165          * initialization */
 166
 167         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 168         if (dt_object_exists(dt))
 169                 oid = obj->oo_dn->dn_object;
 170         else
 171                 oid = DMU_NEW_OBJECT;
 172
 173         /* XXX: we still miss for append declaration support in ZFS
 174          *      -1 means append which is used by llog mostly, llog
 175          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 176         if (pos == -1)
 177                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 178                             obj->oo_attr.la_size + (2 << 20));
 179         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 180
 181         /* dt_declare_write() is usually called for system objects, such
 182          * as llog or last_rcvd files. We needn't enforce quota on those
 183          * objects, so always set the lqi_space as 0. */
 184         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 185                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 186                                  false));
 187 }
 188
 189 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 190                         const struct lu_buf *buf, loff_t *pos,
 191                         struct thandle *th, int ignore_quota)
 192 {
 193         struct osd_object  *obj  = osd_dt_obj(dt);
 194         struct osd_device  *osd = osd_obj2dev(obj);
 195         struct osd_thandle *oh;
 196         uint64_t            offset = *pos;
 197         int                 rc;
 198
 199         ENTRY;
 200
 201         LASSERT(dt_object_exists(dt));
 202         LASSERT(obj->oo_dn);
 203
 204         LASSERT(th != NULL);
 205         oh = container_of0(th, struct osd_thandle, ot_super);
 206
 207         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 208                       buf->lb_buf, oh->ot_tx);
 209         write_lock(&obj->oo_attr_lock);
 210         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 211                 obj->oo_attr.la_size = offset + buf->lb_len;
 212                 write_unlock(&obj->oo_attr_lock);
 213                 /* osd_object_sa_update() will be copying directly from oo_attr
 214                  * into dbuf.  any update within a single txg will copy the
 215                  * most actual */
 216                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 217                                         &obj->oo_attr.la_size, 8, oh);
 218                 if (unlikely(rc))
 219                         GOTO(out, rc);
 220         } else {
 221                 write_unlock(&obj->oo_attr_lock);
 222         }
 223
 224         *pos += buf->lb_len;
 225         rc = buf->lb_len;
 226
 227 out:
 228         RETURN(rc);
 229 }
 230
 231 /*
 232  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 233  *      purposes as it's not very well defined ...
 234  *      instead I use the lowest bit of the address so that:
 235  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 236  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 237  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 238  *
 239  *      bzzz, to blame
 240  */
 241 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 242                         struct niobuf_local *lnb, int npages)
 243 {
 244         struct osd_object *obj  = osd_dt_obj(dt);
 245         struct osd_device *osd = osd_obj2dev(obj);
 246         unsigned long      ptr;
 247         int                i;
 248
 249         LASSERT(dt_object_exists(dt));
 250         LASSERT(obj->oo_dn);
 251
 252         for (i = 0; i < npages; i++) {
 253                 if (lnb[i].lnb_page == NULL)
 254                         continue;
 255                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 256                         /* this is anonymous page allocated for copy-write */
 257                         lnb[i].lnb_page->mapping = NULL;
 258                         __free_page(lnb[i].lnb_page);
 259                         atomic_dec(&osd->od_zerocopy_alloc);
 260                 } else {
 261                         /* see comment in osd_bufs_get_read() */
 262                         ptr = (unsigned long)lnb[i].lnb_data;
 263                         if (ptr & 1UL) {
 264                                 ptr &= ~1UL;
 265                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 266                                 atomic_dec(&osd->od_zerocopy_pin);
 267                         } else if (lnb[i].lnb_data != NULL) {
 268                                 dmu_return_arcbuf(lnb[i].lnb_data);
 269                                 atomic_dec(&osd->od_zerocopy_loan);
 270                         }
 271                 }
 272                 lnb[i].lnb_page = NULL;
 273                 lnb[i].lnb_data = NULL;
 274         }
 275
 276         return 0;
 277 }
 278
 279 static inline struct page *kmem_to_page(void *addr)
 280 {
 281         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 282         if (is_vmalloc_addr(addr))
 283                 return vmalloc_to_page(addr);
 284         else
 285                 return virt_to_page(addr);
 286 }
 287
 288 /**
 289  * Prepare buffers for read.
 290  *
 291  * The function maps the range described by \a off and \a len to \a lnb array.
 292  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 293  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 294  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 295  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 296  * A better implementation would just return the buffers (potentially unfilled)
 297  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 298  *
 299  * \param[in] env       environment
 300  * \param[in] obj       object
 301  * \param[in] off       offset in bytes
 302  * \param[in] len       the number of bytes to access
 303  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 304  *
 305  * \retval              0 for success
 306  * \retval              negative error number of failure
 307  */
 308 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 309                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 310 {
 311         struct osd_device *osd = osd_obj2dev(obj);
 312         unsigned long      start = cfs_time_current();
 313         int                rc, i, numbufs, npages = 0;
 314         dmu_buf_t        **dbp;
 315         ENTRY;
 316
 317         record_start_io(osd, READ, 0);
 318
 319         /* grab buffers for read:
 320          * OSD API let us to grab buffers first, then initiate IO(s)
 321          * so that all required IOs will be done in parallel, but at the
 322          * moment DMU doesn't provide us with a method to grab buffers.
 323          * If we discover this is a vital for good performance we
 324          * can get own replacement for dmu_buf_hold_array_by_bonus().
 325          */
 326         while (len > 0) {
 327                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 328                                                   off, len, TRUE, osd_0copy_tag,
 329                                                   &numbufs, &dbp);
 330                 if (unlikely(rc))
 331                         GOTO(err, rc);
 332
 333                 for (i = 0; i < numbufs; i++) {
 334                         int bufoff, tocpy, thispage;
 335                         void *dbf = dbp[i];
 336
 337                         LASSERT(len > 0);
 338
 339                         atomic_inc(&osd->od_zerocopy_pin);
 340
 341                         bufoff = off - dbp[i]->db_offset;
 342                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 343
 344                         /* kind of trick to differentiate dbuf vs. arcbuf */
 345                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 346                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 347
 348                         while (tocpy > 0) {
 349                                 thispage = PAGE_SIZE;
 350                                 thispage -= bufoff & (PAGE_SIZE - 1);
 351                                 thispage = min(tocpy, thispage);
 352
 353                                 lnb->lnb_rc = 0;
 354                                 lnb->lnb_file_offset = off;
 355                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 356                                 lnb->lnb_len = thispage;
 357                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 358                                                              bufoff);
 359                                 /* mark just a single slot: we need this
 360                                  * reference to dbuf to be released once */
 361                                 lnb->lnb_data = dbf;
 362                                 dbf = NULL;
 363
 364                                 tocpy -= thispage;
 365                                 len -= thispage;
 366                                 bufoff += thispage;
 367                                 off += thispage;
 368
 369                                 npages++;
 370                                 lnb++;
 371                         }
 372
 373                         /* steal dbuf so dmu_buf_rele_array() can't release
 374                          * it */
 375                         dbp[i] = NULL;
 376                 }
 377
 378                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 379         }
 380
 381         record_end_io(osd, READ, cfs_time_current() - start,
 382                       npages * PAGE_SIZE, npages);
 383
 384         RETURN(npages);
 385
 386 err:
 387         LASSERT(rc < 0);
 388         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 389         RETURN(rc);
 390 }
 391
 392 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 393 {
 394         arc_buf_t *abuf;
 395
 396         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 397         if (unlikely(!abuf))
 398                 return ERR_PTR(-ENOMEM);
 399
 400 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 401         /**
 402          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 403          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 404          * setting up page array for RDMA transfer. See LU-9305.
 405          */
 406         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 407                 dmu_return_arcbuf(abuf);
 408                 return NULL;
 409         }
 410 #endif
 411
 412         return abuf;
 413 }
 414
 415 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 416                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 417 {
 418         struct osd_device *osd = osd_obj2dev(obj);
 419         int                plen, off_in_block, sz_in_block;
 420         int                rc, i = 0, npages = 0;
 421         dnode_t *dn = obj->oo_dn;
 422         arc_buf_t *abuf;
 423         uint32_t bs = dn->dn_datablksz;
 424         ENTRY;
 425
 426         /*
 427          * currently only full blocks are subject to zerocopy approach:
 428          * so that we're sure nobody is trying to update the same block
 429          */
 430         while (len > 0) {
 431                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 432
 433                 off_in_block = off & (bs - 1);
 434                 sz_in_block = min_t(int, bs - off_in_block, len);
 435
 436                 abuf = NULL;
 437                 if (sz_in_block == bs) {
 438                         /* full block, try to use zerocopy */
 439                         abuf = osd_request_arcbuf(dn, bs);
 440                         if (unlikely(IS_ERR(abuf)))
 441                                 GOTO(out_err, rc = PTR_ERR(abuf));
 442                 }
 443
 444                 if (abuf != NULL) {
 445                         atomic_inc(&osd->od_zerocopy_loan);
 446
 447                         /* go over pages arcbuf contains, put them as
 448                          * local niobufs for ptlrpc's bulks */
 449                         while (sz_in_block > 0) {
 450                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 451
 452                                 lnb[i].lnb_file_offset = off;
 453                                 lnb[i].lnb_page_offset = 0;
 454                                 lnb[i].lnb_len = plen;
 455                                 lnb[i].lnb_rc = 0;
 456                                 if (sz_in_block == bs)
 457                                         lnb[i].lnb_data = abuf;
 458                                 else
 459                                         lnb[i].lnb_data = NULL;
 460
 461                                 /* this one is not supposed to fail */
 462                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 463                                                         off_in_block);
 464                                 LASSERT(lnb[i].lnb_page);
 465
 466                                 lprocfs_counter_add(osd->od_stats,
 467                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 468
 469                                 sz_in_block -= plen;
 470                                 len -= plen;
 471                                 off += plen;
 472                                 off_in_block += plen;
 473                                 i++;
 474                                 npages++;
 475                         }
 476                 } else {
 477                         if (off_in_block == 0 && len < bs &&
 478                                         off + len >= obj->oo_attr.la_size)
 479                                 lprocfs_counter_add(osd->od_stats,
 480                                                 LPROC_OSD_TAIL_IO, 1);
 481
 482                         /* can't use zerocopy, allocate temp. buffers */
 483                         while (sz_in_block > 0) {
 484                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 485
 486                                 lnb[i].lnb_file_offset = off;
 487                                 lnb[i].lnb_page_offset = 0;
 488                                 lnb[i].lnb_len = plen;
 489                                 lnb[i].lnb_rc = 0;
 490                                 lnb[i].lnb_data = NULL;
 491
 492                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 493                                 if (unlikely(lnb[i].lnb_page == NULL))
 494                                         GOTO(out_err, rc = -ENOMEM);
 495
 496                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 497                                 lnb[i].lnb_page->mapping = (void *)obj;
 498
 499                                 atomic_inc(&osd->od_zerocopy_alloc);
 500                                 lprocfs_counter_add(osd->od_stats,
 501                                                 LPROC_OSD_COPY_IO, 1);
 502
 503                                 sz_in_block -= plen;
 504                                 len -= plen;
 505                                 off += plen;
 506                                 i++;
 507                                 npages++;
 508                         }
 509                 }
 510         }
 511
 512         RETURN(npages);
 513
 514 out_err:
 515         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 516         RETURN(rc);
 517 }
 518
 519 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 520                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 521                         int rw)
 522 {
 523         struct osd_object *obj  = osd_dt_obj(dt);
 524         int                rc;
 525
 526         LASSERT(dt_object_exists(dt));
 527         LASSERT(obj->oo_dn);
 528
 529         if (rw == 0)
 530                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 531         else
 532                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 533
 534         return rc;
 535 }
 536
 537 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 538                         struct niobuf_local *lnb, int npages)
 539 {
 540         struct osd_object *obj = osd_dt_obj(dt);
 541
 542         LASSERT(dt_object_exists(dt));
 543         LASSERT(obj->oo_dn);
 544
 545         return 0;
 546 }
 547
 548 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 549                                            uint64_t offset,
 550                                            uint32_t blksz)
 551 {
 552         LASSERT(blksz > 0);
 553
 554         size += offset % blksz;
 555
 556         if (likely(is_power_of_2(blksz)))
 557                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 558
 559         size += blksz - 1;
 560         do_div(size, blksz);
 561         return size * blksz;
 562 }
 563
 564 static int osd_declare_write_commit(const struct lu_env *env,
 565                                     struct dt_object *dt,
 566                                     struct niobuf_local *lnb, int npages,
 567                                     struct thandle *th)
 568 {
 569         struct osd_object  *obj = osd_dt_obj(dt);
 570         struct osd_device  *osd = osd_obj2dev(obj);
 571         struct osd_thandle *oh;
 572         uint64_t            offset = 0;
 573         uint32_t            size = 0;
 574         uint32_t blksz = obj->oo_dn->dn_datablksz;
 575         int                 i, rc, flags = 0;
 576         bool                ignore_quota = false, synced = false;
 577         long long           space = 0;
 578         struct page        *last_page = NULL;
 579         unsigned long       discont_pages = 0;
 580         ENTRY;
 581
 582         LASSERT(dt_object_exists(dt));
 583         LASSERT(obj->oo_dn);
 584
 585         LASSERT(lnb);
 586         LASSERT(npages > 0);
 587
 588         oh = container_of0(th, struct osd_thandle, ot_super);
 589
 590         for (i = 0; i < npages; i++) {
 591                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 592                         ++discont_pages;
 593                 last_page = lnb[i].lnb_page;
 594                 if (lnb[i].lnb_rc)
 595                         /* ENOSPC, network RPC error, etc.
 596                          * We don't want to book space for pages which will be
 597                          * skipped in osd_write_commit(). Hence we skip pages
 598                          * with lnb_rc != 0 here too */
 599                         continue;
 600                 /* ignore quota for the whole request if any page is from
 601                  * client cache or written by root.
 602                  *
 603                  * XXX once we drop the 1.8 client support, the checking
 604                  * for whether page is from cache can be simplified as:
 605                  * !(lnb[i].flags & OBD_BRW_SYNC)
 606                  *
 607                  * XXX we could handle this on per-lnb basis as done by
 608                  * grant. */
 609                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 610                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 611                     OBD_BRW_FROM_GRANT)
 612                         ignore_quota = true;
 613                 if (size == 0) {
 614                         /* first valid lnb */
 615                         offset = lnb[i].lnb_file_offset;
 616                         size = lnb[i].lnb_len;
 617                         continue;
 618                 }
 619                 if (offset + size == lnb[i].lnb_file_offset) {
 620                         /* this lnb is contiguous to the previous one */
 621                         size += lnb[i].lnb_len;
 622                         continue;
 623                 }
 624
 625                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 626                                   obj->oo_dn, offset, size);
 627                 /* Estimating space to be consumed by a write is rather
 628                  * complicated with ZFS. As a consequence, we don't account for
 629                  * indirect blocks and just use as a rough estimate the worse
 630                  * case where the old space is being held by a snapshot. Quota
 631                  * overrun will be adjusted once the operation is committed, if
 632                  * required. */
 633                 space += osd_roundup2blocksz(size, offset, blksz);
 634
 635                 offset = lnb[i].lnb_file_offset;
 636                 size = lnb[i].lnb_len;
 637         }
 638
 639         if (size) {
 640                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 641                                   offset, size);
 642                 space += osd_roundup2blocksz(size, offset, blksz);
 643         }
 644
 645         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 646
 647         /* backend zfs filesystem might be configured to store multiple data
 648          * copies */
 649         space  *= osd->od_os->os_copies;
 650         space   = toqb(space);
 651         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 652                npages, space);
 653
 654         record_start_io(osd, WRITE, discont_pages);
 655 retry:
 656         /* acquire quota space if needed */
 657         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 658                                obj->oo_attr.la_gid, space, oh, true, &flags,
 659                                ignore_quota);
 660
 661         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 662                 dt_sync(env, th->th_dev);
 663                 synced = true;
 664                 CDEBUG(D_QUOTA, "retry after sync\n");
 665                 flags = 0;
 666                 goto retry;
 667         }
 668
 669         /* we need only to store the overquota flags in the first lnb for
 670          * now, once we support multiple objects BRW, this code needs be
 671          * revised. */
 672         if (flags & QUOTA_FL_OVER_USRQUOTA)
 673                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 674         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 675                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 676
 677         RETURN(rc);
 678 }
 679
 680 /**
 681  * Policy to grow ZFS block size by write pattern.
 682  * For sequential write, it grows block size gradually until it reaches the
 683  * maximum blocksize the dataset can support. Otherwise, it will pick a
 684  * a block size by the writing region of this I/O.
 685  */
 686 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 687                               uint64_t start, uint64_t end)
 688 {
 689         struct osd_device       *osd = osd_obj2dev(obj);
 690         dnode_t *dn = obj->oo_dn;
 691         uint32_t                 blksz;
 692         int                      rc = 0;
 693
 694         ENTRY;
 695
 696         if (dn->dn_maxblkid > 0) /* can't change block size */
 697                 GOTO(out, rc);
 698
 699         if (dn->dn_datablksz >= osd->od_max_blksz)
 700                 GOTO(out, rc);
 701
 702         down_write(&obj->oo_guard);
 703
 704         blksz = dn->dn_datablksz;
 705         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 706                 GOTO(out_unlock, rc);
 707
 708         /* now ZFS can support up to 16MB block size, and if the write
 709          * is sequential, it just increases the block size gradually */
 710         if (start <= blksz) { /* sequential */
 711                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 712         } else { /* sparse, pick a block size by write region */
 713                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 714                                         end - start);
 715         }
 716
 717         if (!is_power_of_2(blksz))
 718                 blksz = size_roundup_power2(blksz);
 719
 720         if (blksz > dn->dn_datablksz) {
 721                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 722                                                blksz, 0, oh->ot_tx);
 723                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 724                 if (rc < 0)
 725                         CDEBUG(D_INODE, "object "DFID": change block size"
 726                                "%u -> %u error rc = %d\n",
 727                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 728                                dn->dn_datablksz, blksz, rc);
 729         }
 730         EXIT;
 731 out_unlock:
 732         up_write(&obj->oo_guard);
 733 out:
 734         return rc;
 735 }
 736
 737 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 738                         struct niobuf_local *lnb, int npages,
 739                         struct thandle *th)
 740 {
 741         struct osd_object  *obj  = osd_dt_obj(dt);
 742         struct osd_device  *osd = osd_obj2dev(obj);
 743         struct osd_thandle *oh;
 744         uint64_t            new_size = 0;
 745         int                 i, rc = 0;
 746         unsigned long      iosize = 0;
 747         ENTRY;
 748
 749         LASSERT(dt_object_exists(dt));
 750         LASSERT(obj->oo_dn);
 751
 752         LASSERT(th != NULL);
 753         oh = container_of0(th, struct osd_thandle, ot_super);
 754
 755         /* adjust block size. Assume the buffers are sorted. */
 756         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 757                                  lnb[npages - 1].lnb_file_offset +
 758                                  lnb[npages - 1].lnb_len);
 759
 760         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 761          * size and assigning arcbuf take place at the same time.
 762          *
 763          * Thread 1:
 764          * osd_write_commit()
 765          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 766          *   -> dmu_object_set_blocksize()
 767          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 768          *       write lock held
 769          *     -> dbuf_new_size()
 770          *      -> dmu_buf_will_dirty()
 771          *       -> dbuf_read()
 772          *        -> wait for the dbuf state to change
 773          * Thread 2:
 774          * osd_write_commit()
 775          *  -> dmu_assign_arcbuf()
 776          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 777          *    -> dbuf_dirty()
 778          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 779          *
 780          * By taking the read lock, it can avoid thread 2 to enter into the
 781          * critical section of assigning the arcbuf, while thread 1 is
 782          * changing the block size.
 783          */
 784         down_read(&obj->oo_guard);
 785         for (i = 0; i < npages; i++) {
 786                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 787                         (unsigned) lnb[i].lnb_len,
 788                         (unsigned) lnb[i].lnb_file_offset);
 789
 790                 if (lnb[i].lnb_rc) {
 791                         /* ENOSPC, network RPC error, etc.
 792                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 793                          * so we skip this page if lnb_rc is set to -ENOSPC */
 794                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 795                                 PFID(lu_object_fid(&dt->do_lu)), i,
 796                                 lnb[i].lnb_rc);
 797                         continue;
 798                 }
 799
 800                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 801                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 802                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 803                                       oh->ot_tx);
 804                         kunmap(lnb[i].lnb_page);
 805                 } else if (lnb[i].lnb_data) {
 806                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 807                         /* buffer loaned for zerocopy, try to use it.
 808                          * notice that dmu_assign_arcbuf() is smart
 809                          * enough to recognize changed blocksize
 810                          * in this case it fallbacks to dmu_write() */
 811                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 812                                           lnb[i].lnb_file_offset,
 813                                           lnb[i].lnb_data, oh->ot_tx);
 814                         /* drop the reference, otherwise osd_put_bufs()
 815                          * will be releasing it - bad! */
 816                         lnb[i].lnb_data = NULL;
 817                         atomic_dec(&osd->od_zerocopy_loan);
 818                 }
 819
 820                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 821                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 822                 iosize += lnb[i].lnb_len;
 823         }
 824         up_read(&obj->oo_guard);
 825
 826         if (unlikely(new_size == 0)) {
 827                 /* no pages to write, no transno is needed */
 828                 th->th_local = 1;
 829                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 830                  * since ofd_commitrw_write() retries several times on ENOSPC */
 831                 record_end_io(osd, WRITE, 0, 0, 0);
 832                 RETURN(0);
 833         }
 834
 835         write_lock(&obj->oo_attr_lock);
 836         if (obj->oo_attr.la_size < new_size) {
 837                 obj->oo_attr.la_size = new_size;
 838                 write_unlock(&obj->oo_attr_lock);
 839                 /* osd_object_sa_update() will be copying directly from
 840                  * oo_attr into dbuf. any update within a single txg will copy
 841                  * the most actual */
 842                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 843                                           &obj->oo_attr.la_size, 8, oh);
 844         } else {
 845                 write_unlock(&obj->oo_attr_lock);
 846         }
 847
 848         record_end_io(osd, WRITE, 0, iosize, npages);
 849
 850         RETURN(rc);
 851 }
 852
 853 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 854                         struct niobuf_local *lnb, int npages)
 855 {
 856         struct osd_object *obj  = osd_dt_obj(dt);
 857         int                i;
 858         loff_t             eof;
 859
 860         LASSERT(dt_object_exists(dt));
 861         LASSERT(obj->oo_dn);
 862
 863         read_lock(&obj->oo_attr_lock);
 864         eof = obj->oo_attr.la_size;
 865         read_unlock(&obj->oo_attr_lock);
 866
 867         for (i = 0; i < npages; i++) {
 868                 if (unlikely(lnb[i].lnb_rc < 0))
 869                         continue;
 870
 871                 lnb[i].lnb_rc = lnb[i].lnb_len;
 872
 873                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 874                         if (eof <= lnb[i].lnb_file_offset)
 875                                 lnb[i].lnb_rc = 0;
 876                         else
 877                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 878
 879                         /* all subsequent rc should be 0 */
 880                         while (++i < npages)
 881                                 lnb[i].lnb_rc = 0;
 882                         break;
 883                 }
 884         }
 885
 886         return 0;
 887 }
 888
 889 /*
 890  * Punch/truncate an object
 891  *
 892  *      IN:     db  - dmu_buf of the object to free data in.
 893  *              off - start of section to free.
 894  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 895  *
 896  *      RETURN: 0 if success
 897  *              error code if failure
 898  *
 899  * The transaction passed to this routine must have
 900  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 901  * called and then assigned to a transaction group.
 902  */
 903 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 904                                 uint64_t size, uint64_t off, uint64_t len)
 905 {
 906         int rc = 0;
 907
 908         /* Assert that the transaction has been assigned to a
 909            transaction group. */
 910         LASSERT(tx->tx_txg != 0);
 911         /*
 912          * Nothing to do if file already at desired length.
 913          */
 914         if (len == DMU_OBJECT_END && size == off)
 915                 return 0;
 916
 917         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 918         if (off < size)
 919                 dmu_free_range(os, dn->dn_object, off, len, tx);
 920
 921         return rc;
 922 }
 923
 924 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 925                         __u64 start, __u64 end, struct thandle *th)
 926 {
 927         struct osd_object  *obj = osd_dt_obj(dt);
 928         struct osd_device  *osd = osd_obj2dev(obj);
 929         struct osd_thandle *oh;
 930         __u64               len;
 931         int                 rc = 0;
 932         ENTRY;
 933
 934         LASSERT(dt_object_exists(dt));
 935         LASSERT(osd_invariant(obj));
 936
 937         LASSERT(th != NULL);
 938         oh = container_of0(th, struct osd_thandle, ot_super);
 939
 940         write_lock(&obj->oo_attr_lock);
 941         /* truncate */
 942         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 943                 len = DMU_OBJECT_END;
 944         else
 945                 len = end - start;
 946         write_unlock(&obj->oo_attr_lock);
 947
 948         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 949                                 obj->oo_attr.la_size, start, len);
 950         /* set new size */
 951         if (len == DMU_OBJECT_END) {
 952                 write_lock(&obj->oo_attr_lock);
 953                 obj->oo_attr.la_size = start;
 954                 write_unlock(&obj->oo_attr_lock);
 955                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 956                                           &obj->oo_attr.la_size, 8, oh);
 957         }
 958         RETURN(rc);
 959 }
 960
 961 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 962                         __u64 start, __u64 end, struct thandle *handle)
 963 {
 964         struct osd_object  *obj = osd_dt_obj(dt);
 965         struct osd_device  *osd = osd_obj2dev(obj);
 966         struct osd_thandle *oh;
 967         __u64               len;
 968         ENTRY;
 969
 970         oh = container_of0(handle, struct osd_thandle, ot_super);
 971
 972         read_lock(&obj->oo_attr_lock);
 973         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 974                 len = DMU_OBJECT_END;
 975         else
 976                 len = end - start;
 977
 978         /* declare we'll free some blocks ... */
 979         if (start < obj->oo_attr.la_size) {
 980                 read_unlock(&obj->oo_attr_lock);
 981                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
 982         } else {
 983                 read_unlock(&obj->oo_attr_lock);
 984         }
 985
 986         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 987                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 988                                  false));
 989 }
 990
 991 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
 992                        __u64 start, __u64 end, enum lu_ladvise_type advice)
 993 {
 994         int     rc;
 995         ENTRY;
 996
 997         switch (advice) {
 998         default:
 999                 rc = -ENOTSUPP;
1000                 break;
1001         }
1002
1003         RETURN(rc);
1004 }
1005
1006 struct dt_body_operations osd_body_ops = {
1007         .dbo_read                       = osd_read,
1008         .dbo_declare_write              = osd_declare_write,
1009         .dbo_write                      = osd_write,
1010         .dbo_bufs_get                   = osd_bufs_get,
1011         .dbo_bufs_put                   = osd_bufs_put,
1012         .dbo_write_prep                 = osd_write_prep,
1013         .dbo_declare_write_commit       = osd_declare_write_commit,
1014         .dbo_write_commit               = osd_write_commit,
1015         .dbo_read_prep                  = osd_read_prep,
1016         .dbo_declare_punch              = osd_declare_punch,
1017         .dbo_punch                      = osd_punch,
1018         .dbo_ladvise                    = osd_ladvise,
1019 };