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 = -dmu_read(osd->od_os, obj->oo_dn->dn_object, *pos, size,
 139                         buf->lb_buf, 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         dmu_tx_hold_write(oh->ot_tx, oid, 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         dmu_write(osd->od_os, obj->oo_dn->dn_object, offset,
 208                 (uint64_t)buf->lb_len, 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         if (is_vmalloc_addr(addr))
 282                 return vmalloc_to_page(addr);
 283         else
 284                 return virt_to_page(addr);
 285 }
 286
 287 /**
 288  * Prepare buffers for read.
 289  *
 290  * The function maps the range described by \a off and \a len to \a lnb array.
 291  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 292  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 293  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 294  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 295  * A better implementation would just return the buffers (potentially unfilled)
 296  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 297  *
 298  * \param[in] env       environment
 299  * \param[in] obj       object
 300  * \param[in] off       offset in bytes
 301  * \param[in] len       the number of bytes to access
 302  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 303  *
 304  * \retval              0 for success
 305  * \retval              negative error number of failure
 306  */
 307 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 308                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 309 {
 310         struct osd_device *osd = osd_obj2dev(obj);
 311         unsigned long      start = cfs_time_current();
 312         int                rc, i, numbufs, npages = 0;
 313         dmu_buf_t        **dbp;
 314         ENTRY;
 315
 316         record_start_io(osd, READ, 0);
 317
 318         /* grab buffers for read:
 319          * OSD API let us to grab buffers first, then initiate IO(s)
 320          * so that all required IOs will be done in parallel, but at the
 321          * moment DMU doesn't provide us with a method to grab buffers.
 322          * If we discover this is a vital for good performance we
 323          * can get own replacement for dmu_buf_hold_array_by_bonus().
 324          */
 325         while (len > 0) {
 326                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 327                                                   off, len, TRUE, osd_0copy_tag,
 328                                                   &numbufs, &dbp);
 329                 if (unlikely(rc))
 330                         GOTO(err, rc);
 331
 332                 for (i = 0; i < numbufs; i++) {
 333                         int bufoff, tocpy, thispage;
 334                         void *dbf = dbp[i];
 335
 336                         LASSERT(len > 0);
 337
 338                         atomic_inc(&osd->od_zerocopy_pin);
 339
 340                         bufoff = off - dbp[i]->db_offset;
 341                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 342
 343                         /* kind of trick to differentiate dbuf vs. arcbuf */
 344                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 345                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 346
 347                         while (tocpy > 0) {
 348                                 thispage = PAGE_SIZE;
 349                                 thispage -= bufoff & (PAGE_SIZE - 1);
 350                                 thispage = min(tocpy, thispage);
 351
 352                                 lnb->lnb_rc = 0;
 353                                 lnb->lnb_file_offset = off;
 354                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 355                                 lnb->lnb_len = thispage;
 356                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 357                                                              bufoff);
 358                                 /* mark just a single slot: we need this
 359                                  * reference to dbuf to be released once */
 360                                 lnb->lnb_data = dbf;
 361                                 dbf = NULL;
 362
 363                                 tocpy -= thispage;
 364                                 len -= thispage;
 365                                 bufoff += thispage;
 366                                 off += thispage;
 367
 368                                 npages++;
 369                                 lnb++;
 370                         }
 371
 372                         /* steal dbuf so dmu_buf_rele_array() can't release
 373                          * it */
 374                         dbp[i] = NULL;
 375                 }
 376
 377                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 378         }
 379
 380         record_end_io(osd, READ, cfs_time_current() - start,
 381                       npages * PAGE_SIZE, npages);
 382
 383         RETURN(npages);
 384
 385 err:
 386         LASSERT(rc < 0);
 387         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 388         RETURN(rc);
 389 }
 390
 391 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 392                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 393 {
 394         struct osd_device *osd = osd_obj2dev(obj);
 395         int                plen, off_in_block, sz_in_block;
 396         int                rc, i = 0, npages = 0;
 397         dnode_t *dn = obj->oo_dn;
 398         arc_buf_t *abuf;
 399         uint32_t bs = dn->dn_datablksz;
 400         ENTRY;
 401
 402         /*
 403          * currently only full blocks are subject to zerocopy approach:
 404          * so that we're sure nobody is trying to update the same block
 405          */
 406         while (len > 0) {
 407                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 408
 409                 off_in_block = off & (bs - 1);
 410                 sz_in_block = min_t(int, bs - off_in_block, len);
 411
 412                 if (sz_in_block == bs) {
 413                         /* full block, try to use zerocopy */
 414
 415                         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 416                         if (unlikely(abuf == NULL))
 417                                 GOTO(out_err, rc = -ENOMEM);
 418
 419                         atomic_inc(&osd->od_zerocopy_loan);
 420
 421                         /* go over pages arcbuf contains, put them as
 422                          * local niobufs for ptlrpc's bulks */
 423                         while (sz_in_block > 0) {
 424                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 425
 426                                 lnb[i].lnb_file_offset = off;
 427                                 lnb[i].lnb_page_offset = 0;
 428                                 lnb[i].lnb_len = plen;
 429                                 lnb[i].lnb_rc = 0;
 430                                 if (sz_in_block == bs)
 431                                         lnb[i].lnb_data = abuf;
 432                                 else
 433                                         lnb[i].lnb_data = NULL;
 434
 435                                 /* this one is not supposed to fail */
 436                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 437                                                         off_in_block);
 438                                 LASSERT(lnb[i].lnb_page);
 439
 440                                 lprocfs_counter_add(osd->od_stats,
 441                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 442
 443                                 sz_in_block -= plen;
 444                                 len -= plen;
 445                                 off += plen;
 446                                 off_in_block += plen;
 447                                 i++;
 448                                 npages++;
 449                         }
 450                 } else {
 451                         if (off_in_block == 0 && len < bs &&
 452                                         off + len >= obj->oo_attr.la_size)
 453                                 lprocfs_counter_add(osd->od_stats,
 454                                                 LPROC_OSD_TAIL_IO, 1);
 455
 456                         /* can't use zerocopy, allocate temp. buffers */
 457                         while (sz_in_block > 0) {
 458                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 459
 460                                 lnb[i].lnb_file_offset = off;
 461                                 lnb[i].lnb_page_offset = 0;
 462                                 lnb[i].lnb_len = plen;
 463                                 lnb[i].lnb_rc = 0;
 464                                 lnb[i].lnb_data = NULL;
 465
 466                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 467                                 if (unlikely(lnb[i].lnb_page == NULL))
 468                                         GOTO(out_err, rc = -ENOMEM);
 469
 470                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 471                                 lnb[i].lnb_page->mapping = (void *)obj;
 472
 473                                 atomic_inc(&osd->od_zerocopy_alloc);
 474                                 lprocfs_counter_add(osd->od_stats,
 475                                                 LPROC_OSD_COPY_IO, 1);
 476
 477                                 sz_in_block -= plen;
 478                                 len -= plen;
 479                                 off += plen;
 480                                 i++;
 481                                 npages++;
 482                         }
 483                 }
 484         }
 485
 486         RETURN(npages);
 487
 488 out_err:
 489         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 490         RETURN(rc);
 491 }
 492
 493 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 494                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 495                         int rw)
 496 {
 497         struct osd_object *obj  = osd_dt_obj(dt);
 498         int                rc;
 499
 500         LASSERT(dt_object_exists(dt));
 501         LASSERT(obj->oo_dn);
 502
 503         if (rw == 0)
 504                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 505         else
 506                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 507
 508         return rc;
 509 }
 510
 511 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 512                         struct niobuf_local *lnb, int npages)
 513 {
 514         struct osd_object *obj = osd_dt_obj(dt);
 515
 516         LASSERT(dt_object_exists(dt));
 517         LASSERT(obj->oo_dn);
 518
 519         return 0;
 520 }
 521
 522 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 523                                            uint64_t offset,
 524                                            uint32_t blksz)
 525 {
 526         LASSERT(blksz > 0);
 527
 528         size += offset % blksz;
 529
 530         if (likely(IS_PO2(blksz)))
 531                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 532
 533         size += blksz - 1;
 534         do_div(size, blksz);
 535         return size * blksz;
 536 }
 537
 538 static int osd_declare_write_commit(const struct lu_env *env,
 539                                     struct dt_object *dt,
 540                                     struct niobuf_local *lnb, int npages,
 541                                     struct thandle *th)
 542 {
 543         struct osd_object  *obj = osd_dt_obj(dt);
 544         struct osd_device  *osd = osd_obj2dev(obj);
 545         struct osd_thandle *oh;
 546         uint64_t            offset = 0;
 547         uint32_t            size = 0;
 548         uint32_t blksz = obj->oo_dn->dn_datablksz;
 549         int                 i, rc, flags = 0;
 550         bool                ignore_quota = false, synced = false;
 551         long long           space = 0;
 552         struct page        *last_page = NULL;
 553         unsigned long       discont_pages = 0;
 554         ENTRY;
 555
 556         LASSERT(dt_object_exists(dt));
 557         LASSERT(obj->oo_dn);
 558
 559         LASSERT(lnb);
 560         LASSERT(npages > 0);
 561
 562         oh = container_of0(th, struct osd_thandle, ot_super);
 563
 564         for (i = 0; i < npages; i++) {
 565                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 566                         ++discont_pages;
 567                 last_page = lnb[i].lnb_page;
 568                 if (lnb[i].lnb_rc)
 569                         /* ENOSPC, network RPC error, etc.
 570                          * We don't want to book space for pages which will be
 571                          * skipped in osd_write_commit(). Hence we skip pages
 572                          * with lnb_rc != 0 here too */
 573                         continue;
 574                 /* ignore quota for the whole request if any page is from
 575                  * client cache or written by root.
 576                  *
 577                  * XXX once we drop the 1.8 client support, the checking
 578                  * for whether page is from cache can be simplified as:
 579                  * !(lnb[i].flags & OBD_BRW_SYNC)
 580                  *
 581                  * XXX we could handle this on per-lnb basis as done by
 582                  * grant. */
 583                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 584                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 585                     OBD_BRW_FROM_GRANT)
 586                         ignore_quota = true;
 587                 if (size == 0) {
 588                         /* first valid lnb */
 589                         offset = lnb[i].lnb_file_offset;
 590                         size = lnb[i].lnb_len;
 591                         continue;
 592                 }
 593                 if (offset + size == lnb[i].lnb_file_offset) {
 594                         /* this lnb is contiguous to the previous one */
 595                         size += lnb[i].lnb_len;
 596                         continue;
 597                 }
 598
 599                 dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 600                                   offset, size);
 601                 /* Estimating space to be consumed by a write is rather
 602                  * complicated with ZFS. As a consequence, we don't account for
 603                  * indirect blocks and just use as a rough estimate the worse
 604                  * case where the old space is being held by a snapshot. Quota
 605                  * overrun will be adjusted once the operation is committed, if
 606                  * required. */
 607                 space += osd_roundup2blocksz(size, offset, blksz);
 608
 609                 offset = lnb[i].lnb_file_offset;
 610                 size = lnb[i].lnb_len;
 611         }
 612
 613         if (size) {
 614                 dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 615                                   offset, size);
 616                 space += osd_roundup2blocksz(size, offset, blksz);
 617         }
 618
 619         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 620
 621         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 622
 623         /* backend zfs filesystem might be configured to store multiple data
 624          * copies */
 625         space  *= osd->od_os->os_copies;
 626         space   = toqb(space);
 627         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 628                npages, space);
 629
 630         record_start_io(osd, WRITE, discont_pages);
 631 retry:
 632         /* acquire quota space if needed */
 633         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 634                                obj->oo_attr.la_gid, space, oh, true, &flags,
 635                                ignore_quota);
 636
 637         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 638                 dt_sync(env, th->th_dev);
 639                 synced = true;
 640                 CDEBUG(D_QUOTA, "retry after sync\n");
 641                 flags = 0;
 642                 goto retry;
 643         }
 644
 645         /* we need only to store the overquota flags in the first lnb for
 646          * now, once we support multiple objects BRW, this code needs be
 647          * revised. */
 648         if (flags & QUOTA_FL_OVER_USRQUOTA)
 649                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 650         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 651                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 652
 653         RETURN(rc);
 654 }
 655
 656 /**
 657  * Policy to grow ZFS block size by write pattern.
 658  * For sequential write, it grows block size gradually until it reaches the
 659  * maximum blocksize the dataset can support. Otherwise, it will pick a
 660  * a block size by the writing region of this I/O.
 661  */
 662 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 663                               uint64_t start, uint64_t end)
 664 {
 665         struct osd_device       *osd = osd_obj2dev(obj);
 666         dnode_t *dn = obj->oo_dn;
 667         uint32_t                 blksz;
 668         int                      rc = 0;
 669
 670         ENTRY;
 671
 672         if (dn->dn_maxblkid > 0) /* can't change block size */
 673                 GOTO(out, rc);
 674
 675         if (dn->dn_datablksz >= osd->od_max_blksz)
 676                 GOTO(out, rc);
 677
 678         down_write(&obj->oo_guard);
 679
 680         blksz = dn->dn_datablksz;
 681         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 682                 GOTO(out_unlock, rc);
 683
 684         /* now ZFS can support up to 16MB block size, and if the write
 685          * is sequential, it just increases the block size gradually */
 686         if (start <= blksz) { /* sequential */
 687                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 688         } else { /* sparse, pick a block size by write region */
 689                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 690                                         end - start);
 691         }
 692
 693         if (!is_power_of_2(blksz))
 694                 blksz = size_roundup_power2(blksz);
 695
 696         if (blksz > dn->dn_datablksz) {
 697                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 698                                                blksz, 0, oh->ot_tx);
 699                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 700                 if (rc < 0)
 701                         CDEBUG(D_INODE, "object "DFID": change block size"
 702                                "%u -> %u error rc = %d\n",
 703                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 704                                dn->dn_datablksz, blksz, rc);
 705         }
 706         EXIT;
 707 out_unlock:
 708         up_write(&obj->oo_guard);
 709 out:
 710         return rc;
 711 }
 712
 713 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 714                         struct niobuf_local *lnb, int npages,
 715                         struct thandle *th)
 716 {
 717         struct osd_object  *obj  = osd_dt_obj(dt);
 718         struct osd_device  *osd = osd_obj2dev(obj);
 719         struct osd_thandle *oh;
 720         uint64_t            new_size = 0;
 721         int                 i, rc = 0;
 722         unsigned long      iosize = 0;
 723         ENTRY;
 724
 725         LASSERT(dt_object_exists(dt));
 726         LASSERT(obj->oo_dn);
 727
 728         LASSERT(th != NULL);
 729         oh = container_of0(th, struct osd_thandle, ot_super);
 730
 731         /* adjust block size. Assume the buffers are sorted. */
 732         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 733                                  lnb[npages - 1].lnb_file_offset +
 734                                  lnb[npages - 1].lnb_len);
 735
 736         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 737          * size and assigning arcbuf take place at the same time.
 738          *
 739          * Thread 1:
 740          * osd_write_commit()
 741          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 742          *   -> dmu_object_set_blocksize()
 743          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 744          *       write lock held
 745          *     -> dbuf_new_size()
 746          *      -> dmu_buf_will_dirty()
 747          *       -> dbuf_read()
 748          *        -> wait for the dbuf state to change
 749          * Thread 2:
 750          * osd_write_commit()
 751          *  -> dmu_assign_arcbuf()
 752          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 753          *    -> dbuf_dirty()
 754          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 755          *
 756          * By taking the read lock, it can avoid thread 2 to enter into the
 757          * critical section of assigning the arcbuf, while thread 1 is
 758          * changing the block size.
 759          */
 760         down_read(&obj->oo_guard);
 761         for (i = 0; i < npages; i++) {
 762                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 763                         (unsigned) lnb[i].lnb_len,
 764                         (unsigned) lnb[i].lnb_file_offset);
 765
 766                 if (lnb[i].lnb_rc) {
 767                         /* ENOSPC, network RPC error, etc.
 768                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 769                          * so we skip this page if lnb_rc is set to -ENOSPC */
 770                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 771                                 PFID(lu_object_fid(&dt->do_lu)), i,
 772                                 lnb[i].lnb_rc);
 773                         continue;
 774                 }
 775
 776                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 777                         dmu_write(osd->od_os, obj->oo_dn->dn_object,
 778                                 lnb[i].lnb_file_offset, lnb[i].lnb_len,
 779                                 kmap(lnb[i].lnb_page), oh->ot_tx);
 780                         kunmap(lnb[i].lnb_page);
 781                 } else if (lnb[i].lnb_data) {
 782                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 783                         /* buffer loaned for zerocopy, try to use it.
 784                          * notice that dmu_assign_arcbuf() is smart
 785                          * enough to recognize changed blocksize
 786                          * in this case it fallbacks to dmu_write() */
 787                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 788                                           lnb[i].lnb_file_offset,
 789                                           lnb[i].lnb_data, oh->ot_tx);
 790                         /* drop the reference, otherwise osd_put_bufs()
 791                          * will be releasing it - bad! */
 792                         lnb[i].lnb_data = NULL;
 793                         atomic_dec(&osd->od_zerocopy_loan);
 794                 }
 795
 796                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 797                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 798                 iosize += lnb[i].lnb_len;
 799         }
 800         up_read(&obj->oo_guard);
 801
 802         if (unlikely(new_size == 0)) {
 803                 /* no pages to write, no transno is needed */
 804                 th->th_local = 1;
 805                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 806                  * since ofd_commitrw_write() retries several times on ENOSPC */
 807                 record_end_io(osd, WRITE, 0, 0, 0);
 808                 RETURN(0);
 809         }
 810
 811         write_lock(&obj->oo_attr_lock);
 812         if (obj->oo_attr.la_size < new_size) {
 813                 obj->oo_attr.la_size = new_size;
 814                 write_unlock(&obj->oo_attr_lock);
 815                 /* osd_object_sa_update() will be copying directly from
 816                  * oo_attr into dbuf. any update within a single txg will copy
 817                  * the most actual */
 818                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 819                                           &obj->oo_attr.la_size, 8, oh);
 820         } else {
 821                 write_unlock(&obj->oo_attr_lock);
 822         }
 823
 824         record_end_io(osd, WRITE, 0, iosize, npages);
 825
 826         RETURN(rc);
 827 }
 828
 829 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 830                         struct niobuf_local *lnb, int npages)
 831 {
 832         struct osd_object *obj  = osd_dt_obj(dt);
 833         int                i;
 834         loff_t             eof;
 835
 836         LASSERT(dt_object_exists(dt));
 837         LASSERT(obj->oo_dn);
 838
 839         read_lock(&obj->oo_attr_lock);
 840         eof = obj->oo_attr.la_size;
 841         read_unlock(&obj->oo_attr_lock);
 842
 843         for (i = 0; i < npages; i++) {
 844                 if (unlikely(lnb[i].lnb_rc < 0))
 845                         continue;
 846
 847                 lnb[i].lnb_rc = lnb[i].lnb_len;
 848
 849                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 850                         if (eof <= lnb[i].lnb_file_offset)
 851                                 lnb[i].lnb_rc = 0;
 852                         else
 853                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 854
 855                         /* all subsequent rc should be 0 */
 856                         while (++i < npages)
 857                                 lnb[i].lnb_rc = 0;
 858                         break;
 859                 }
 860         }
 861
 862         return 0;
 863 }
 864
 865 /*
 866  * Punch/truncate an object
 867  *
 868  *      IN:     db  - dmu_buf of the object to free data in.
 869  *              off - start of section to free.
 870  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 871  *
 872  *      RETURN: 0 if success
 873  *              error code if failure
 874  *
 875  * The transaction passed to this routine must have
 876  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 877  * called and then assigned to a transaction group.
 878  */
 879 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 880                                 uint64_t size, uint64_t off, uint64_t len)
 881 {
 882         int rc = 0;
 883
 884         /* Assert that the transaction has been assigned to a
 885            transaction group. */
 886         LASSERT(tx->tx_txg != 0);
 887         /*
 888          * Nothing to do if file already at desired length.
 889          */
 890         if (len == DMU_OBJECT_END && size == off)
 891                 return 0;
 892
 893         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 894         if (off < size)
 895                 dmu_free_range(os, dn->dn_object, off, len, tx);
 896
 897         return rc;
 898 }
 899
 900 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 901                         __u64 start, __u64 end, struct thandle *th)
 902 {
 903         struct osd_object  *obj = osd_dt_obj(dt);
 904         struct osd_device  *osd = osd_obj2dev(obj);
 905         struct osd_thandle *oh;
 906         __u64               len;
 907         int                 rc = 0;
 908         ENTRY;
 909
 910         LASSERT(dt_object_exists(dt));
 911         LASSERT(osd_invariant(obj));
 912
 913         LASSERT(th != NULL);
 914         oh = container_of0(th, struct osd_thandle, ot_super);
 915
 916         write_lock(&obj->oo_attr_lock);
 917         /* truncate */
 918         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 919                 len = DMU_OBJECT_END;
 920         else
 921                 len = end - start;
 922         write_unlock(&obj->oo_attr_lock);
 923
 924         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 925                                 obj->oo_attr.la_size, start, len);
 926         /* set new size */
 927         if (len == DMU_OBJECT_END) {
 928                 write_lock(&obj->oo_attr_lock);
 929                 obj->oo_attr.la_size = start;
 930                 write_unlock(&obj->oo_attr_lock);
 931                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 932                                           &obj->oo_attr.la_size, 8, oh);
 933         }
 934         RETURN(rc);
 935 }
 936
 937 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 938                         __u64 start, __u64 end, struct thandle *handle)
 939 {
 940         struct osd_object  *obj = osd_dt_obj(dt);
 941         struct osd_device  *osd = osd_obj2dev(obj);
 942         struct osd_thandle *oh;
 943         __u64               len;
 944         ENTRY;
 945
 946         oh = container_of0(handle, struct osd_thandle, ot_super);
 947
 948         read_lock(&obj->oo_attr_lock);
 949         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 950                 len = DMU_OBJECT_END;
 951         else
 952                 len = end - start;
 953
 954         /* declare we'll free some blocks ... */
 955         if (start < obj->oo_attr.la_size) {
 956                 read_unlock(&obj->oo_attr_lock);
 957                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
 958         } else {
 959                 read_unlock(&obj->oo_attr_lock);
 960         }
 961
 962         /* ... and we'll modify size attribute */
 963         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 964
 965         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 966                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 967                                  false));
 968 }
 969
 970 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
 971                        __u64 start, __u64 end, enum lu_ladvise_type advice)
 972 {
 973         int     rc;
 974         ENTRY;
 975
 976         switch (advice) {
 977         default:
 978                 rc = -ENOTSUPP;
 979                 break;
 980         }
 981
 982         RETURN(rc);
 983 }
 984
 985 struct dt_body_operations osd_body_ops = {
 986         .dbo_read                       = osd_read,
 987         .dbo_declare_write              = osd_declare_write,
 988         .dbo_write                      = osd_write,
 989         .dbo_bufs_get                   = osd_bufs_get,
 990         .dbo_bufs_put                   = osd_bufs_put,
 991         .dbo_write_prep                 = osd_write_prep,
 992         .dbo_declare_write_commit       = osd_declare_write_commit,
 993         .dbo_write_commit               = osd_write_commit,
 994         .dbo_read_prep                  = osd_read_prep,
 995         .dbo_declare_punch              = osd_declare_punch,
 996         .dbo_punch                      = osd_punch,
 997         .dbo_ladvise                    = osd_ladvise,
 998 };