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                                 int j, apages, abufsz;
 269                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 270                                 apages = abufsz / PAGE_SIZE;
 271                                 /* these references to pages must be invalidated
 272                                  * to prevent access in osd_bufs_put() */
 273                                 for (j = 0; j < apages; j++)
 274                                         lnb[i + j].lnb_page = NULL;
 275                                 dmu_return_arcbuf(lnb[i].lnb_data);
 276                                 atomic_dec(&osd->od_zerocopy_loan);
 277                         }
 278                 }
 279                 lnb[i].lnb_page = NULL;
 280                 lnb[i].lnb_data = NULL;
 281         }
 282
 283         return 0;
 284 }
 285
 286 static inline struct page *kmem_to_page(void *addr)
 287 {
 288         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 289         if (is_vmalloc_addr(addr))
 290                 return vmalloc_to_page(addr);
 291         else
 292                 return virt_to_page(addr);
 293 }
 294
 295 /**
 296  * Prepare buffers for read.
 297  *
 298  * The function maps the range described by \a off and \a len to \a lnb array.
 299  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 300  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 301  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 302  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 303  * A better implementation would just return the buffers (potentially unfilled)
 304  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 305  *
 306  * \param[in] env       environment
 307  * \param[in] obj       object
 308  * \param[in] off       offset in bytes
 309  * \param[in] len       the number of bytes to access
 310  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 311  *
 312  * \retval              0 for success
 313  * \retval              negative error number of failure
 314  */
 315 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 316                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 317 {
 318         struct osd_device *osd = osd_obj2dev(obj);
 319         unsigned long      start = cfs_time_current();
 320         int                rc, i, numbufs, npages = 0;
 321         dmu_buf_t        **dbp;
 322         ENTRY;
 323
 324         record_start_io(osd, READ, 0);
 325
 326         /* grab buffers for read:
 327          * OSD API let us to grab buffers first, then initiate IO(s)
 328          * so that all required IOs will be done in parallel, but at the
 329          * moment DMU doesn't provide us with a method to grab buffers.
 330          * If we discover this is a vital for good performance we
 331          * can get own replacement for dmu_buf_hold_array_by_bonus().
 332          */
 333         while (len > 0) {
 334                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 335                                                   off, len, TRUE, osd_0copy_tag,
 336                                                   &numbufs, &dbp);
 337                 if (unlikely(rc))
 338                         GOTO(err, rc);
 339
 340                 for (i = 0; i < numbufs; i++) {
 341                         int bufoff, tocpy, thispage;
 342                         void *dbf = dbp[i];
 343
 344                         LASSERT(len > 0);
 345
 346                         atomic_inc(&osd->od_zerocopy_pin);
 347
 348                         bufoff = off - dbp[i]->db_offset;
 349                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 350
 351                         /* kind of trick to differentiate dbuf vs. arcbuf */
 352                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 353                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 354
 355                         while (tocpy > 0) {
 356                                 thispage = PAGE_SIZE;
 357                                 thispage -= bufoff & (PAGE_SIZE - 1);
 358                                 thispage = min(tocpy, thispage);
 359
 360                                 lnb->lnb_rc = 0;
 361                                 lnb->lnb_file_offset = off;
 362                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 363                                 lnb->lnb_len = thispage;
 364                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 365                                                              bufoff);
 366                                 /* mark just a single slot: we need this
 367                                  * reference to dbuf to be released once */
 368                                 lnb->lnb_data = dbf;
 369                                 dbf = NULL;
 370
 371                                 tocpy -= thispage;
 372                                 len -= thispage;
 373                                 bufoff += thispage;
 374                                 off += thispage;
 375
 376                                 npages++;
 377                                 lnb++;
 378                         }
 379
 380                         /* steal dbuf so dmu_buf_rele_array() can't release
 381                          * it */
 382                         dbp[i] = NULL;
 383                 }
 384
 385                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 386         }
 387
 388         record_end_io(osd, READ, cfs_time_current() - start,
 389                       npages * PAGE_SIZE, npages);
 390
 391         RETURN(npages);
 392
 393 err:
 394         LASSERT(rc < 0);
 395         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 396         RETURN(rc);
 397 }
 398
 399 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 400 {
 401         arc_buf_t *abuf;
 402
 403         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 404         if (unlikely(!abuf))
 405                 return ERR_PTR(-ENOMEM);
 406
 407 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 408         /**
 409          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 410          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 411          * setting up page array for RDMA transfer. See LU-9305.
 412          */
 413         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 414                 dmu_return_arcbuf(abuf);
 415                 return NULL;
 416         }
 417 #endif
 418
 419         return abuf;
 420 }
 421
 422 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 423                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 424 {
 425         struct osd_device *osd = osd_obj2dev(obj);
 426         int                plen, off_in_block, sz_in_block;
 427         int                rc, i = 0, npages = 0;
 428         dnode_t *dn = obj->oo_dn;
 429         arc_buf_t *abuf;
 430         uint32_t bs = dn->dn_datablksz;
 431         ENTRY;
 432
 433         /*
 434          * currently only full blocks are subject to zerocopy approach:
 435          * so that we're sure nobody is trying to update the same block
 436          */
 437         while (len > 0) {
 438                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 439
 440                 off_in_block = off & (bs - 1);
 441                 sz_in_block = min_t(int, bs - off_in_block, len);
 442
 443                 abuf = NULL;
 444                 if (sz_in_block == bs) {
 445                         /* full block, try to use zerocopy */
 446                         abuf = osd_request_arcbuf(dn, bs);
 447                         if (unlikely(IS_ERR(abuf)))
 448                                 GOTO(out_err, rc = PTR_ERR(abuf));
 449                 }
 450
 451                 if (abuf != NULL) {
 452                         atomic_inc(&osd->od_zerocopy_loan);
 453
 454                         /* go over pages arcbuf contains, put them as
 455                          * local niobufs for ptlrpc's bulks */
 456                         while (sz_in_block > 0) {
 457                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 458
 459                                 lnb[i].lnb_file_offset = off;
 460                                 lnb[i].lnb_page_offset = 0;
 461                                 lnb[i].lnb_len = plen;
 462                                 lnb[i].lnb_rc = 0;
 463                                 if (sz_in_block == bs)
 464                                         lnb[i].lnb_data = abuf;
 465                                 else
 466                                         lnb[i].lnb_data = NULL;
 467
 468                                 /* this one is not supposed to fail */
 469                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 470                                                         off_in_block);
 471                                 LASSERT(lnb[i].lnb_page);
 472
 473                                 lprocfs_counter_add(osd->od_stats,
 474                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 475
 476                                 sz_in_block -= plen;
 477                                 len -= plen;
 478                                 off += plen;
 479                                 off_in_block += plen;
 480                                 i++;
 481                                 npages++;
 482                         }
 483                 } else {
 484                         if (off_in_block == 0 && len < bs &&
 485                                         off + len >= obj->oo_attr.la_size)
 486                                 lprocfs_counter_add(osd->od_stats,
 487                                                 LPROC_OSD_TAIL_IO, 1);
 488
 489                         /* can't use zerocopy, allocate temp. buffers */
 490                         while (sz_in_block > 0) {
 491                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 492
 493                                 lnb[i].lnb_file_offset = off;
 494                                 lnb[i].lnb_page_offset = 0;
 495                                 lnb[i].lnb_len = plen;
 496                                 lnb[i].lnb_rc = 0;
 497                                 lnb[i].lnb_data = NULL;
 498
 499                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 500                                 if (unlikely(lnb[i].lnb_page == NULL))
 501                                         GOTO(out_err, rc = -ENOMEM);
 502
 503                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 504                                 lnb[i].lnb_page->mapping = (void *)obj;
 505
 506                                 atomic_inc(&osd->od_zerocopy_alloc);
 507                                 lprocfs_counter_add(osd->od_stats,
 508                                                 LPROC_OSD_COPY_IO, 1);
 509
 510                                 sz_in_block -= plen;
 511                                 len -= plen;
 512                                 off += plen;
 513                                 i++;
 514                                 npages++;
 515                         }
 516                 }
 517         }
 518
 519         RETURN(npages);
 520
 521 out_err:
 522         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 523         RETURN(rc);
 524 }
 525
 526 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 527                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 528                         int rw)
 529 {
 530         struct osd_object *obj  = osd_dt_obj(dt);
 531         int                rc;
 532
 533         LASSERT(dt_object_exists(dt));
 534         LASSERT(obj->oo_dn);
 535
 536         if (rw == 0)
 537                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 538         else
 539                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 540
 541         return rc;
 542 }
 543
 544 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 545                         struct niobuf_local *lnb, int npages)
 546 {
 547         struct osd_object *obj = osd_dt_obj(dt);
 548
 549         LASSERT(dt_object_exists(dt));
 550         LASSERT(obj->oo_dn);
 551
 552         return 0;
 553 }
 554
 555 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 556                                            uint64_t offset,
 557                                            uint32_t blksz)
 558 {
 559         LASSERT(blksz > 0);
 560
 561         size += offset % blksz;
 562
 563         if (likely(is_power_of_2(blksz)))
 564                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 565
 566         size += blksz - 1;
 567         do_div(size, blksz);
 568         return size * blksz;
 569 }
 570
 571 static int osd_declare_write_commit(const struct lu_env *env,
 572                                     struct dt_object *dt,
 573                                     struct niobuf_local *lnb, int npages,
 574                                     struct thandle *th)
 575 {
 576         struct osd_object  *obj = osd_dt_obj(dt);
 577         struct osd_device  *osd = osd_obj2dev(obj);
 578         struct osd_thandle *oh;
 579         uint64_t            offset = 0;
 580         uint32_t            size = 0;
 581         uint32_t blksz = obj->oo_dn->dn_datablksz;
 582         int                 i, rc, flags = 0;
 583         bool                ignore_quota = false, synced = false;
 584         long long           space = 0;
 585         struct page        *last_page = NULL;
 586         unsigned long       discont_pages = 0;
 587         ENTRY;
 588
 589         LASSERT(dt_object_exists(dt));
 590         LASSERT(obj->oo_dn);
 591
 592         LASSERT(lnb);
 593         LASSERT(npages > 0);
 594
 595         oh = container_of0(th, struct osd_thandle, ot_super);
 596
 597         for (i = 0; i < npages; i++) {
 598                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 599                         ++discont_pages;
 600                 last_page = lnb[i].lnb_page;
 601                 if (lnb[i].lnb_rc)
 602                         /* ENOSPC, network RPC error, etc.
 603                          * We don't want to book space for pages which will be
 604                          * skipped in osd_write_commit(). Hence we skip pages
 605                          * with lnb_rc != 0 here too */
 606                         continue;
 607                 /* ignore quota for the whole request if any page is from
 608                  * client cache or written by root.
 609                  *
 610                  * XXX once we drop the 1.8 client support, the checking
 611                  * for whether page is from cache can be simplified as:
 612                  * !(lnb[i].flags & OBD_BRW_SYNC)
 613                  *
 614                  * XXX we could handle this on per-lnb basis as done by
 615                  * grant. */
 616                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 617                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 618                     OBD_BRW_FROM_GRANT)
 619                         ignore_quota = true;
 620                 if (size == 0) {
 621                         /* first valid lnb */
 622                         offset = lnb[i].lnb_file_offset;
 623                         size = lnb[i].lnb_len;
 624                         continue;
 625                 }
 626                 if (offset + size == lnb[i].lnb_file_offset) {
 627                         /* this lnb is contiguous to the previous one */
 628                         size += lnb[i].lnb_len;
 629                         continue;
 630                 }
 631
 632                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 633                                   obj->oo_dn, offset, size);
 634                 /* Estimating space to be consumed by a write is rather
 635                  * complicated with ZFS. As a consequence, we don't account for
 636                  * indirect blocks and just use as a rough estimate the worse
 637                  * case where the old space is being held by a snapshot. Quota
 638                  * overrun will be adjusted once the operation is committed, if
 639                  * required. */
 640                 space += osd_roundup2blocksz(size, offset, blksz);
 641
 642                 offset = lnb[i].lnb_file_offset;
 643                 size = lnb[i].lnb_len;
 644         }
 645
 646         if (size) {
 647                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 648                                   offset, size);
 649                 space += osd_roundup2blocksz(size, offset, blksz);
 650         }
 651
 652         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 653
 654         /* backend zfs filesystem might be configured to store multiple data
 655          * copies */
 656         space  *= osd->od_os->os_copies;
 657         space   = toqb(space);
 658         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 659                npages, space);
 660
 661         record_start_io(osd, WRITE, discont_pages);
 662 retry:
 663         /* acquire quota space if needed */
 664         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 665                                obj->oo_attr.la_gid, space, oh, true, &flags,
 666                                ignore_quota);
 667
 668         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 669                 dt_sync(env, th->th_dev);
 670                 synced = true;
 671                 CDEBUG(D_QUOTA, "retry after sync\n");
 672                 flags = 0;
 673                 goto retry;
 674         }
 675
 676         /* we need only to store the overquota flags in the first lnb for
 677          * now, once we support multiple objects BRW, this code needs be
 678          * revised. */
 679         if (flags & QUOTA_FL_OVER_USRQUOTA)
 680                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 681         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 682                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 683
 684         RETURN(rc);
 685 }
 686
 687 /**
 688  * Policy to grow ZFS block size by write pattern.
 689  * For sequential write, it grows block size gradually until it reaches the
 690  * maximum blocksize the dataset can support. Otherwise, it will pick a
 691  * a block size by the writing region of this I/O.
 692  */
 693 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 694                               uint64_t start, uint64_t end)
 695 {
 696         struct osd_device       *osd = osd_obj2dev(obj);
 697         dnode_t *dn = obj->oo_dn;
 698         uint32_t                 blksz;
 699         int                      rc = 0;
 700
 701         ENTRY;
 702
 703         if (dn->dn_maxblkid > 0) /* can't change block size */
 704                 GOTO(out, rc);
 705
 706         if (dn->dn_datablksz >= osd->od_max_blksz)
 707                 GOTO(out, rc);
 708
 709         down_write(&obj->oo_guard);
 710
 711         blksz = dn->dn_datablksz;
 712         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 713                 GOTO(out_unlock, rc);
 714
 715         /* now ZFS can support up to 16MB block size, and if the write
 716          * is sequential, it just increases the block size gradually */
 717         if (start <= blksz) { /* sequential */
 718                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 719         } else { /* sparse, pick a block size by write region */
 720                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 721                                         end - start);
 722         }
 723
 724         if (!is_power_of_2(blksz))
 725                 blksz = size_roundup_power2(blksz);
 726
 727         if (blksz > dn->dn_datablksz) {
 728                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 729                                                blksz, 0, oh->ot_tx);
 730                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 731                 if (rc < 0)
 732                         CDEBUG(D_INODE, "object "DFID": change block size"
 733                                "%u -> %u error rc = %d\n",
 734                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 735                                dn->dn_datablksz, blksz, rc);
 736         }
 737         EXIT;
 738 out_unlock:
 739         up_write(&obj->oo_guard);
 740 out:
 741         return rc;
 742 }
 743
 744 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 745                         struct niobuf_local *lnb, int npages,
 746                         struct thandle *th)
 747 {
 748         struct osd_object  *obj  = osd_dt_obj(dt);
 749         struct osd_device  *osd = osd_obj2dev(obj);
 750         struct osd_thandle *oh;
 751         uint64_t            new_size = 0;
 752         int                 i, rc = 0;
 753         unsigned long      iosize = 0;
 754         ENTRY;
 755
 756         LASSERT(dt_object_exists(dt));
 757         LASSERT(obj->oo_dn);
 758
 759         LASSERT(th != NULL);
 760         oh = container_of0(th, struct osd_thandle, ot_super);
 761
 762         /* adjust block size. Assume the buffers are sorted. */
 763         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 764                                  lnb[npages - 1].lnb_file_offset +
 765                                  lnb[npages - 1].lnb_len);
 766
 767         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 768          * size and assigning arcbuf take place at the same time.
 769          *
 770          * Thread 1:
 771          * osd_write_commit()
 772          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 773          *   -> dmu_object_set_blocksize()
 774          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 775          *       write lock held
 776          *     -> dbuf_new_size()
 777          *      -> dmu_buf_will_dirty()
 778          *       -> dbuf_read()
 779          *        -> wait for the dbuf state to change
 780          * Thread 2:
 781          * osd_write_commit()
 782          *  -> dmu_assign_arcbuf()
 783          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 784          *    -> dbuf_dirty()
 785          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 786          *
 787          * By taking the read lock, it can avoid thread 2 to enter into the
 788          * critical section of assigning the arcbuf, while thread 1 is
 789          * changing the block size.
 790          */
 791         down_read(&obj->oo_guard);
 792         for (i = 0; i < npages; i++) {
 793                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 794                         (unsigned) lnb[i].lnb_len,
 795                         (unsigned) lnb[i].lnb_file_offset);
 796
 797                 if (lnb[i].lnb_rc) {
 798                         /* ENOSPC, network RPC error, etc.
 799                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 800                          * so we skip this page if lnb_rc is set to -ENOSPC */
 801                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 802                                 PFID(lu_object_fid(&dt->do_lu)), i,
 803                                 lnb[i].lnb_rc);
 804                         continue;
 805                 }
 806
 807                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 808                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 809                 if (lnb[i].lnb_page == NULL)
 810                         continue;
 811
 812                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 813                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 814                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 815                                       oh->ot_tx);
 816                         kunmap(lnb[i].lnb_page);
 817                         iosize += lnb[i].lnb_len;
 818                 } else if (lnb[i].lnb_data) {
 819                         int j, apages, abufsz;
 820                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 821                         /* buffer loaned for zerocopy, try to use it.
 822                          * notice that dmu_assign_arcbuf() is smart
 823                          * enough to recognize changed blocksize
 824                          * in this case it fallbacks to dmu_write() */
 825                         abufsz = arc_buf_size(lnb[i].lnb_data);
 826                         LASSERT(abufsz & PAGE_MASK);
 827                         apages = abufsz / PAGE_SIZE;
 828                         LASSERT(i + apages <= npages);
 829                         /* these references to pages must be invalidated
 830                          * to prevent access in osd_bufs_put() */
 831                         for (j = 0; j < apages; j++)
 832                                 lnb[i + j].lnb_page = NULL;
 833                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 834                                           lnb[i].lnb_file_offset,
 835                                           lnb[i].lnb_data, oh->ot_tx);
 836                         /* drop the reference, otherwise osd_put_bufs()
 837                          * will be releasing it - bad! */
 838                         lnb[i].lnb_data = NULL;
 839                         atomic_dec(&osd->od_zerocopy_loan);
 840                         iosize += abufsz;
 841                 }
 842
 843         }
 844         up_read(&obj->oo_guard);
 845
 846         if (unlikely(new_size == 0)) {
 847                 /* no pages to write, no transno is needed */
 848                 th->th_local = 1;
 849                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 850                  * since ofd_commitrw_write() retries several times on ENOSPC */
 851                 record_end_io(osd, WRITE, 0, 0, 0);
 852                 RETURN(0);
 853         }
 854
 855         write_lock(&obj->oo_attr_lock);
 856         if (obj->oo_attr.la_size < new_size) {
 857                 obj->oo_attr.la_size = new_size;
 858                 write_unlock(&obj->oo_attr_lock);
 859                 /* osd_object_sa_update() will be copying directly from
 860                  * oo_attr into dbuf. any update within a single txg will copy
 861                  * the most actual */
 862                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 863                                           &obj->oo_attr.la_size, 8, oh);
 864         } else {
 865                 write_unlock(&obj->oo_attr_lock);
 866         }
 867
 868         record_end_io(osd, WRITE, 0, iosize, npages);
 869
 870         RETURN(rc);
 871 }
 872
 873 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 874                         struct niobuf_local *lnb, int npages)
 875 {
 876         struct osd_object *obj  = osd_dt_obj(dt);
 877         int                i;
 878         loff_t             eof;
 879
 880         LASSERT(dt_object_exists(dt));
 881         LASSERT(obj->oo_dn);
 882
 883         read_lock(&obj->oo_attr_lock);
 884         eof = obj->oo_attr.la_size;
 885         read_unlock(&obj->oo_attr_lock);
 886
 887         for (i = 0; i < npages; i++) {
 888                 if (unlikely(lnb[i].lnb_rc < 0))
 889                         continue;
 890
 891                 lnb[i].lnb_rc = lnb[i].lnb_len;
 892
 893                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 894                         if (eof <= lnb[i].lnb_file_offset)
 895                                 lnb[i].lnb_rc = 0;
 896                         else
 897                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 898
 899                         /* all subsequent rc should be 0 */
 900                         while (++i < npages)
 901                                 lnb[i].lnb_rc = 0;
 902                         break;
 903                 }
 904         }
 905
 906         return 0;
 907 }
 908
 909 /*
 910  * Punch/truncate an object
 911  *
 912  *      IN:     db  - dmu_buf of the object to free data in.
 913  *              off - start of section to free.
 914  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 915  *
 916  *      RETURN: 0 if success
 917  *              error code if failure
 918  *
 919  * The transaction passed to this routine must have
 920  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 921  * called and then assigned to a transaction group.
 922  */
 923 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 924                                 uint64_t size, uint64_t off, uint64_t len)
 925 {
 926         int rc = 0;
 927
 928         /* Assert that the transaction has been assigned to a
 929            transaction group. */
 930         LASSERT(tx->tx_txg != 0);
 931         /*
 932          * Nothing to do if file already at desired length.
 933          */
 934         if (len == DMU_OBJECT_END && size == off)
 935                 return 0;
 936
 937         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 938         if (off < size)
 939                 dmu_free_range(os, dn->dn_object, off, len, tx);
 940
 941         return rc;
 942 }
 943
 944 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 945                         __u64 start, __u64 end, struct thandle *th)
 946 {
 947         struct osd_object  *obj = osd_dt_obj(dt);
 948         struct osd_device  *osd = osd_obj2dev(obj);
 949         struct osd_thandle *oh;
 950         __u64               len;
 951         int                 rc = 0;
 952         ENTRY;
 953
 954         LASSERT(dt_object_exists(dt));
 955         LASSERT(osd_invariant(obj));
 956
 957         LASSERT(th != NULL);
 958         oh = container_of0(th, struct osd_thandle, ot_super);
 959
 960         write_lock(&obj->oo_attr_lock);
 961         /* truncate */
 962         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 963                 len = DMU_OBJECT_END;
 964         else
 965                 len = end - start;
 966         write_unlock(&obj->oo_attr_lock);
 967
 968         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 969                                 obj->oo_attr.la_size, start, len);
 970         /* set new size */
 971         if (len == DMU_OBJECT_END) {
 972                 write_lock(&obj->oo_attr_lock);
 973                 obj->oo_attr.la_size = start;
 974                 write_unlock(&obj->oo_attr_lock);
 975                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 976                                           &obj->oo_attr.la_size, 8, oh);
 977         }
 978         RETURN(rc);
 979 }
 980
 981 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 982                         __u64 start, __u64 end, struct thandle *handle)
 983 {
 984         struct osd_object  *obj = osd_dt_obj(dt);
 985         struct osd_device  *osd = osd_obj2dev(obj);
 986         struct osd_thandle *oh;
 987         __u64               len;
 988         ENTRY;
 989
 990         oh = container_of0(handle, struct osd_thandle, ot_super);
 991
 992         read_lock(&obj->oo_attr_lock);
 993         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 994                 len = DMU_OBJECT_END;
 995         else
 996                 len = end - start;
 997
 998         /* declare we'll free some blocks ... */
 999         if (start < obj->oo_attr.la_size) {
1000                 read_unlock(&obj->oo_attr_lock);
1001                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1002         } else {
1003                 read_unlock(&obj->oo_attr_lock);
1004         }
1005
1006         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1007                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
1008                                  false));
1009 }
1010
1011 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1012                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1013 {
1014         int     rc;
1015         ENTRY;
1016
1017         switch (advice) {
1018         default:
1019                 rc = -ENOTSUPP;
1020                 break;
1021         }
1022
1023         RETURN(rc);
1024 }
1025
1026 struct dt_body_operations osd_body_ops = {
1027         .dbo_read                       = osd_read,
1028         .dbo_declare_write              = osd_declare_write,
1029         .dbo_write                      = osd_write,
1030         .dbo_bufs_get                   = osd_bufs_get,
1031         .dbo_bufs_put                   = osd_bufs_put,
1032         .dbo_write_prep                 = osd_write_prep,
1033         .dbo_declare_write_commit       = osd_declare_write_commit,
1034         .dbo_write_commit               = osd_write_commit,
1035         .dbo_read_prep                  = osd_read_prep,
1036         .dbo_declare_punch              = osd_declare_punch,
1037         .dbo_punch                      = osd_punch,
1038         .dbo_ladvise                    = osd_ladvise,
1039 };