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, 2015, 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_zerocopy_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_db);
 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_db->db_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_db->db_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_db);
 203
 204         LASSERT(th != NULL);
 205         oh = container_of0(th, struct osd_thandle, ot_super);
 206
 207         record_start_io(osd, WRITE, 0);
 208
 209         dmu_write(osd->od_os, obj->oo_db->db_object, offset,
 210                 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
 211         write_lock(&obj->oo_attr_lock);
 212         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 213                 obj->oo_attr.la_size = offset + buf->lb_len;
 214                 write_unlock(&obj->oo_attr_lock);
 215                 /* osd_object_sa_update() will be copying directly from oo_attr
 216                  * into dbuf.  any update within a single txg will copy the
 217                  * most actual */
 218                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 219                                         &obj->oo_attr.la_size, 8, oh);
 220                 if (unlikely(rc))
 221                         GOTO(out, rc);
 222         } else {
 223                 write_unlock(&obj->oo_attr_lock);
 224         }
 225
 226         *pos += buf->lb_len;
 227         rc = buf->lb_len;
 228
 229 out:
 230         record_end_io(osd, WRITE, 0, buf->lb_len,
 231                       buf->lb_len >> PAGE_SHIFT);
 232
 233         RETURN(rc);
 234 }
 235
 236 /*
 237  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 238  *      purposes as it's not very well defined ...
 239  *      instead I use the lowest bit of the address so that:
 240  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 241  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 242  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 243  *
 244  *      bzzz, to blame
 245  */
 246 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 247                         struct niobuf_local *lnb, int npages)
 248 {
 249         struct osd_object *obj  = osd_dt_obj(dt);
 250         struct osd_device *osd = osd_obj2dev(obj);
 251         unsigned long      ptr;
 252         int                i;
 253
 254         LASSERT(dt_object_exists(dt));
 255         LASSERT(obj->oo_db);
 256
 257         for (i = 0; i < npages; i++) {
 258                 if (lnb[i].lnb_page == NULL)
 259                         continue;
 260                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 261                         /* this is anonymous page allocated for copy-write */
 262                         lnb[i].lnb_page->mapping = NULL;
 263                         __free_page(lnb[i].lnb_page);
 264                         atomic_dec(&osd->od_zerocopy_alloc);
 265                 } else {
 266                         /* see comment in osd_bufs_get_read() */
 267                         ptr = (unsigned long)lnb[i].lnb_data;
 268                         if (ptr & 1UL) {
 269                                 ptr &= ~1UL;
 270                                 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
 271                                 atomic_dec(&osd->od_zerocopy_pin);
 272                         } else if (lnb[i].lnb_data != NULL) {
 273                                 dmu_return_arcbuf(lnb[i].lnb_data);
 274                                 atomic_dec(&osd->od_zerocopy_loan);
 275                         }
 276                 }
 277                 lnb[i].lnb_page = NULL;
 278                 lnb[i].lnb_data = NULL;
 279         }
 280
 281         return 0;
 282 }
 283
 284 static inline struct page *kmem_to_page(void *addr)
 285 {
 286         if (is_vmalloc_addr(addr))
 287                 return vmalloc_to_page(addr);
 288         else
 289                 return virt_to_page(addr);
 290 }
 291
 292 /**
 293  * Prepare buffers for read.
 294  *
 295  * The function maps the range described by \a off and \a len to \a lnb array.
 296  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 297  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 298  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 299  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 300  * A better implementation would just return the buffers (potentially unfilled)
 301  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 302  *
 303  * \param[in] env       environment
 304  * \param[in] obj       object
 305  * \param[in] off       offset in bytes
 306  * \param[in] len       the number of bytes to access
 307  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 308  *
 309  * \retval              0 for success
 310  * \retval              negative error number of failure
 311  */
 312 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 313                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 314 {
 315         struct osd_device *osd = osd_obj2dev(obj);
 316         unsigned long      start = cfs_time_current();
 317         int                rc, i, numbufs, npages = 0;
 318         dmu_buf_t        **dbp;
 319         ENTRY;
 320
 321         record_start_io(osd, READ, 0);
 322
 323         /* grab buffers for read:
 324          * OSD API let us to grab buffers first, then initiate IO(s)
 325          * so that all required IOs will be done in parallel, but at the
 326          * moment DMU doesn't provide us with a method to grab buffers.
 327          * If we discover this is a vital for good performance we
 328          * can get own replacement for dmu_buf_hold_array_by_bonus().
 329          */
 330         while (len > 0) {
 331                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 332                                                   osd_zerocopy_tag, &numbufs,
 333                                                   &dbp);
 334                 if (unlikely(rc))
 335                         GOTO(err, rc);
 336
 337                 for (i = 0; i < numbufs; i++) {
 338                         int bufoff, tocpy, thispage;
 339                         void *dbf = dbp[i];
 340
 341                         LASSERT(len > 0);
 342
 343                         atomic_inc(&osd->od_zerocopy_pin);
 344
 345                         bufoff = off - dbp[i]->db_offset;
 346                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 347
 348                         /* kind of trick to differentiate dbuf vs. arcbuf */
 349                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 350                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 351
 352                         while (tocpy > 0) {
 353                                 thispage = PAGE_SIZE;
 354                                 thispage -= bufoff & (PAGE_SIZE - 1);
 355                                 thispage = min(tocpy, thispage);
 356
 357                                 lnb->lnb_rc = 0;
 358                                 lnb->lnb_file_offset = off;
 359                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 360                                 lnb->lnb_len = thispage;
 361                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 362                                                              bufoff);
 363                                 /* mark just a single slot: we need this
 364                                  * reference to dbuf to be released once */
 365                                 lnb->lnb_data = dbf;
 366                                 dbf = NULL;
 367
 368                                 tocpy -= thispage;
 369                                 len -= thispage;
 370                                 bufoff += thispage;
 371                                 off += thispage;
 372
 373                                 npages++;
 374                                 lnb++;
 375                         }
 376
 377                         /* steal dbuf so dmu_buf_rele_array() can't release
 378                          * it */
 379                         dbp[i] = NULL;
 380                 }
 381
 382                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 383         }
 384
 385         record_end_io(osd, READ, cfs_time_current() - start,
 386                       npages * PAGE_SIZE, npages);
 387
 388         RETURN(npages);
 389
 390 err:
 391         LASSERT(rc < 0);
 392         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 393         RETURN(rc);
 394 }
 395
 396 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 397                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 398 {
 399         struct osd_device *osd = osd_obj2dev(obj);
 400         int                plen, off_in_block, sz_in_block;
 401         int                rc, i = 0, npages = 0;
 402         arc_buf_t         *abuf;
 403         uint32_t           bs;
 404         uint64_t           dummy;
 405         ENTRY;
 406
 407         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 408
 409         /*
 410          * currently only full blocks are subject to zerocopy approach:
 411          * so that we're sure nobody is trying to update the same block
 412          */
 413         while (len > 0) {
 414                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 415
 416                 off_in_block = off & (bs - 1);
 417                 sz_in_block = min_t(int, bs - off_in_block, len);
 418
 419                 if (sz_in_block == bs) {
 420                         /* full block, try to use zerocopy */
 421
 422                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 423                         if (unlikely(abuf == NULL))
 424                                 GOTO(out_err, rc = -ENOMEM);
 425
 426                         atomic_inc(&osd->od_zerocopy_loan);
 427
 428                         /* go over pages arcbuf contains, put them as
 429                          * local niobufs for ptlrpc's bulks */
 430                         while (sz_in_block > 0) {
 431                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 432
 433                                 lnb[i].lnb_file_offset = off;
 434                                 lnb[i].lnb_page_offset = 0;
 435                                 lnb[i].lnb_len = plen;
 436                                 lnb[i].lnb_rc = 0;
 437                                 if (sz_in_block == bs)
 438                                         lnb[i].lnb_data = abuf;
 439                                 else
 440                                         lnb[i].lnb_data = NULL;
 441
 442                                 /* this one is not supposed to fail */
 443                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 444                                                         off_in_block);
 445                                 LASSERT(lnb[i].lnb_page);
 446
 447                                 lprocfs_counter_add(osd->od_stats,
 448                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 449
 450                                 sz_in_block -= plen;
 451                                 len -= plen;
 452                                 off += plen;
 453                                 off_in_block += plen;
 454                                 i++;
 455                                 npages++;
 456                         }
 457                 } else {
 458                         if (off_in_block == 0 && len < bs &&
 459                                         off + len >= obj->oo_attr.la_size)
 460                                 lprocfs_counter_add(osd->od_stats,
 461                                                 LPROC_OSD_TAIL_IO, 1);
 462
 463                         /* can't use zerocopy, allocate temp. buffers */
 464                         while (sz_in_block > 0) {
 465                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 466
 467                                 lnb[i].lnb_file_offset = off;
 468                                 lnb[i].lnb_page_offset = 0;
 469                                 lnb[i].lnb_len = plen;
 470                                 lnb[i].lnb_rc = 0;
 471                                 lnb[i].lnb_data = NULL;
 472
 473                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 474                                 if (unlikely(lnb[i].lnb_page == NULL))
 475                                         GOTO(out_err, rc = -ENOMEM);
 476
 477                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 478                                 lnb[i].lnb_page->mapping = (void *)obj;
 479
 480                                 atomic_inc(&osd->od_zerocopy_alloc);
 481                                 lprocfs_counter_add(osd->od_stats,
 482                                                 LPROC_OSD_COPY_IO, 1);
 483
 484                                 sz_in_block -= plen;
 485                                 len -= plen;
 486                                 off += plen;
 487                                 i++;
 488                                 npages++;
 489                         }
 490                 }
 491         }
 492
 493         RETURN(npages);
 494
 495 out_err:
 496         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 497         RETURN(rc);
 498 }
 499
 500 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 501                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 502                         int rw)
 503 {
 504         struct osd_object *obj  = osd_dt_obj(dt);
 505         int                rc;
 506
 507         LASSERT(dt_object_exists(dt));
 508         LASSERT(obj->oo_db);
 509
 510         if (rw == 0)
 511                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 512         else
 513                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 514
 515         return rc;
 516 }
 517
 518 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 519                         struct niobuf_local *lnb, int npages)
 520 {
 521         struct osd_object *obj = osd_dt_obj(dt);
 522
 523         LASSERT(dt_object_exists(dt));
 524         LASSERT(obj->oo_db);
 525
 526         return 0;
 527 }
 528
 529 static inline uint32_t osd_get_blocksz(struct osd_object *obj)
 530 {
 531         uint32_t blksz;
 532         u_longlong_t unused;
 533
 534         LASSERT(obj->oo_db);
 535
 536         dmu_object_size_from_db(obj->oo_db, &blksz, &unused);
 537         return blksz;
 538 }
 539
 540 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 541                                            uint64_t offset,
 542                                            uint32_t blksz)
 543 {
 544         LASSERT(blksz > 0);
 545
 546         size += offset % blksz;
 547
 548         if (likely(IS_PO2(blksz)))
 549                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 550
 551         size += blksz - 1;
 552         do_div(size, blksz);
 553         return size * blksz;
 554 }
 555
 556 static int osd_declare_write_commit(const struct lu_env *env,
 557                                     struct dt_object *dt,
 558                                     struct niobuf_local *lnb, int npages,
 559                                     struct thandle *th)
 560 {
 561         struct osd_object  *obj = osd_dt_obj(dt);
 562         struct osd_device  *osd = osd_obj2dev(obj);
 563         struct osd_thandle *oh;
 564         uint64_t            offset = 0;
 565         uint32_t            size = 0;
 566         uint32_t            blksz = osd_get_blocksz(obj);
 567         int                 i, rc, flags = 0;
 568         bool                ignore_quota = false, synced = false;
 569         long long           space = 0;
 570         struct page        *last_page = NULL;
 571         unsigned long       discont_pages = 0;
 572         ENTRY;
 573
 574         LASSERT(dt_object_exists(dt));
 575         LASSERT(obj->oo_db);
 576
 577         LASSERT(lnb);
 578         LASSERT(npages > 0);
 579
 580         oh = container_of0(th, struct osd_thandle, ot_super);
 581
 582         for (i = 0; i < npages; i++) {
 583                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 584                         ++discont_pages;
 585                 last_page = lnb[i].lnb_page;
 586                 if (lnb[i].lnb_rc)
 587                         /* ENOSPC, network RPC error, etc.
 588                          * We don't want to book space for pages which will be
 589                          * skipped in osd_write_commit(). Hence we skip pages
 590                          * with lnb_rc != 0 here too */
 591                         continue;
 592                 /* ignore quota for the whole request if any page is from
 593                  * client cache or written by root.
 594                  *
 595                  * XXX once we drop the 1.8 client support, the checking
 596                  * for whether page is from cache can be simplified as:
 597                  * !(lnb[i].flags & OBD_BRW_SYNC)
 598                  *
 599                  * XXX we could handle this on per-lnb basis as done by
 600                  * grant. */
 601                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 602                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 603                     OBD_BRW_FROM_GRANT)
 604                         ignore_quota = true;
 605                 if (size == 0) {
 606                         /* first valid lnb */
 607                         offset = lnb[i].lnb_file_offset;
 608                         size = lnb[i].lnb_len;
 609                         continue;
 610                 }
 611                 if (offset + size == lnb[i].lnb_file_offset) {
 612                         /* this lnb is contiguous to the previous one */
 613                         size += lnb[i].lnb_len;
 614                         continue;
 615                 }
 616
 617                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 618                                   offset, size);
 619                 /* Estimating space to be consumed by a write is rather
 620                  * complicated with ZFS. As a consequence, we don't account for
 621                  * indirect blocks and just use as a rough estimate the worse
 622                  * case where the old space is being held by a snapshot. Quota
 623                  * overrun will be adjusted once the operation is committed, if
 624                  * required. */
 625                 space += osd_roundup2blocksz(size, offset, blksz);
 626
 627                 offset = lnb[i].lnb_file_offset;
 628                 size = lnb[i].lnb_len;
 629         }
 630
 631         if (size) {
 632                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 633                                   offset, size);
 634                 space += osd_roundup2blocksz(size, offset, blksz);
 635         }
 636
 637         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 638
 639         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 640
 641         /* backend zfs filesystem might be configured to store multiple data
 642          * copies */
 643         space  *= osd->od_os->os_copies;
 644         space   = toqb(space);
 645         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 646                npages, space);
 647
 648         record_start_io(osd, WRITE, discont_pages);
 649 retry:
 650         /* acquire quota space if needed */
 651         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 652                                obj->oo_attr.la_gid, space, oh, true, &flags,
 653                                ignore_quota);
 654
 655         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 656                 dt_sync(env, th->th_dev);
 657                 synced = true;
 658                 CDEBUG(D_QUOTA, "retry after sync\n");
 659                 flags = 0;
 660                 goto retry;
 661         }
 662
 663         /* we need only to store the overquota flags in the first lnb for
 664          * now, once we support multiple objects BRW, this code needs be
 665          * revised. */
 666         if (flags & QUOTA_FL_OVER_USRQUOTA)
 667                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 668         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 669                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 670
 671         RETURN(rc);
 672 }
 673
 674 /**
 675  * Policy to grow ZFS block size by write pattern.
 676  * For sequential write, it grows block size gradually until it reaches the
 677  * maximum blocksize the dataset can support. Otherwise, it will pick a
 678  * a block size by the writing region of this I/O.
 679  */
 680 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 681                               uint64_t start, uint64_t end)
 682 {
 683         struct osd_device       *osd = osd_obj2dev(obj);
 684         dmu_buf_impl_t          *db = (dmu_buf_impl_t *)obj->oo_db;
 685         dnode_t                 *dn;
 686         uint32_t                 blksz;
 687         int                      rc = 0;
 688
 689         ENTRY;
 690
 691         DB_DNODE_ENTER(db);
 692         dn = DB_DNODE(db);
 693
 694         if (dn->dn_maxblkid > 0) /* can't change block size */
 695                 GOTO(out, rc);
 696
 697         if (dn->dn_datablksz >= osd->od_max_blksz)
 698                 GOTO(out, rc);
 699
 700         down_write(&obj->oo_guard);
 701
 702         blksz = dn->dn_datablksz;
 703         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 704                 GOTO(out_unlock, rc);
 705
 706         /* now ZFS can support up to 16MB block size, and if the write
 707          * is sequential, it just increases the block size gradually */
 708         if (start <= blksz) { /* sequential */
 709                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 710         } else { /* sparse, pick a block size by write region */
 711                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 712                                         end - start);
 713         }
 714
 715         if (!is_power_of_2(blksz))
 716                 blksz = size_roundup_power2(blksz);
 717
 718         if (blksz > dn->dn_datablksz) {
 719                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 720                                                blksz, 0, oh->ot_tx);
 721                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 722                 if (rc < 0)
 723                         CDEBUG(D_INODE, "object "DFID": change block size"
 724                                "%u -> %u error rc = %d\n",
 725                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 726                                dn->dn_datablksz, blksz, rc);
 727         }
 728         EXIT;
 729 out_unlock:
 730         up_write(&obj->oo_guard);
 731 out:
 732         DB_DNODE_EXIT(db);
 733         return rc;
 734 }
 735
 736 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 737                         struct niobuf_local *lnb, int npages,
 738                         struct thandle *th)
 739 {
 740         struct osd_object  *obj  = osd_dt_obj(dt);
 741         struct osd_device  *osd = osd_obj2dev(obj);
 742         struct osd_thandle *oh;
 743         uint64_t            new_size = 0;
 744         int                 i, rc = 0;
 745         unsigned long      iosize = 0;
 746         ENTRY;
 747
 748         LASSERT(dt_object_exists(dt));
 749         LASSERT(obj->oo_db);
 750
 751         LASSERT(th != NULL);
 752         oh = container_of0(th, struct osd_thandle, ot_super);
 753
 754         /* adjust block size. Assume the buffers are sorted. */
 755         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 756                                  lnb[npages - 1].lnb_file_offset +
 757                                  lnb[npages - 1].lnb_len);
 758         for (i = 0; i < npages; i++) {
 759                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 760                         (unsigned) lnb[i].lnb_len,
 761                         (unsigned) lnb[i].lnb_file_offset);
 762
 763                 if (lnb[i].lnb_rc) {
 764                         /* ENOSPC, network RPC error, etc.
 765                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 766                          * so we skip this page if lnb_rc is set to -ENOSPC */
 767                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 768                                 PFID(lu_object_fid(&dt->do_lu)), i,
 769                                 lnb[i].lnb_rc);
 770                         continue;
 771                 }
 772
 773                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 774                         dmu_write(osd->od_os, obj->oo_db->db_object,
 775                                 lnb[i].lnb_file_offset, lnb[i].lnb_len,
 776                                 kmap(lnb[i].lnb_page), oh->ot_tx);
 777                         kunmap(lnb[i].lnb_page);
 778                 } else if (lnb[i].lnb_data) {
 779                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 780                         /* buffer loaned for zerocopy, try to use it.
 781                          * notice that dmu_assign_arcbuf() is smart
 782                          * enough to recognize changed blocksize
 783                          * in this case it fallbacks to dmu_write() */
 784                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 785                                           lnb[i].lnb_data, oh->ot_tx);
 786                         /* drop the reference, otherwise osd_put_bufs()
 787                          * will be releasing it - bad! */
 788                         lnb[i].lnb_data = NULL;
 789                         atomic_dec(&osd->od_zerocopy_loan);
 790                 }
 791
 792                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 793                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 794                 iosize += lnb[i].lnb_len;
 795         }
 796
 797         if (unlikely(new_size == 0)) {
 798                 /* no pages to write, no transno is needed */
 799                 th->th_local = 1;
 800                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 801                  * since ofd_commitrw_write() retries several times on ENOSPC */
 802                 record_end_io(osd, WRITE, 0, 0, 0);
 803                 RETURN(0);
 804         }
 805
 806         write_lock(&obj->oo_attr_lock);
 807         if (obj->oo_attr.la_size < new_size) {
 808                 obj->oo_attr.la_size = new_size;
 809                 write_unlock(&obj->oo_attr_lock);
 810                 /* osd_object_sa_update() will be copying directly from
 811                  * oo_attr into dbuf. any update within a single txg will copy
 812                  * the most actual */
 813                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 814                                           &obj->oo_attr.la_size, 8, oh);
 815         } else {
 816                 write_unlock(&obj->oo_attr_lock);
 817         }
 818
 819         record_end_io(osd, WRITE, 0, iosize, npages);
 820
 821         RETURN(rc);
 822 }
 823
 824 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 825                         struct niobuf_local *lnb, int npages)
 826 {
 827         struct osd_object *obj  = osd_dt_obj(dt);
 828         int                i;
 829         loff_t             eof;
 830
 831         LASSERT(dt_object_exists(dt));
 832         LASSERT(obj->oo_db);
 833
 834         read_lock(&obj->oo_attr_lock);
 835         eof = obj->oo_attr.la_size;
 836         read_unlock(&obj->oo_attr_lock);
 837
 838         for (i = 0; i < npages; i++) {
 839                 if (unlikely(lnb[i].lnb_rc < 0))
 840                         continue;
 841
 842                 lnb[i].lnb_rc = lnb[i].lnb_len;
 843
 844                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 845                         if (eof <= lnb[i].lnb_file_offset)
 846                                 lnb[i].lnb_rc = 0;
 847                         else
 848                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 849
 850                         /* all subsequent rc should be 0 */
 851                         while (++i < npages)
 852                                 lnb[i].lnb_rc = 0;
 853                         break;
 854                 }
 855         }
 856
 857         return 0;
 858 }
 859
 860 /*
 861  * Punch/truncate an object
 862  *
 863  *      IN:     db  - dmu_buf of the object to free data in.
 864  *              off - start of section to free.
 865  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 866  *
 867  *      RETURN: 0 if success
 868  *              error code if failure
 869  *
 870  * The transaction passed to this routine must have
 871  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 872  * called and then assigned to a transaction group.
 873  */
 874 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 875                                 uint64_t size, uint64_t off, uint64_t len)
 876 {
 877         int rc = 0;
 878
 879         /* Assert that the transaction has been assigned to a
 880            transaction group. */
 881         LASSERT(tx->tx_txg != 0);
 882         /*
 883          * Nothing to do if file already at desired length.
 884          */
 885         if (len == DMU_OBJECT_END && size == off)
 886                 return 0;
 887
 888         if (off < size)
 889                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 890
 891         return rc;
 892 }
 893
 894 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 895                         __u64 start, __u64 end, struct thandle *th)
 896 {
 897         struct osd_object  *obj = osd_dt_obj(dt);
 898         struct osd_device  *osd = osd_obj2dev(obj);
 899         struct osd_thandle *oh;
 900         __u64               len;
 901         int                 rc = 0;
 902         ENTRY;
 903
 904         LASSERT(dt_object_exists(dt));
 905         LASSERT(osd_invariant(obj));
 906
 907         LASSERT(th != NULL);
 908         oh = container_of0(th, struct osd_thandle, ot_super);
 909
 910         write_lock(&obj->oo_attr_lock);
 911         /* truncate */
 912         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 913                 len = DMU_OBJECT_END;
 914         else
 915                 len = end - start;
 916         write_unlock(&obj->oo_attr_lock);
 917
 918         rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
 919                                 obj->oo_attr.la_size, start, len);
 920         /* set new size */
 921         if (len == DMU_OBJECT_END) {
 922                 write_lock(&obj->oo_attr_lock);
 923                 obj->oo_attr.la_size = start;
 924                 write_unlock(&obj->oo_attr_lock);
 925                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 926                                           &obj->oo_attr.la_size, 8, oh);
 927         }
 928         RETURN(rc);
 929 }
 930
 931 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 932                         __u64 start, __u64 end, struct thandle *handle)
 933 {
 934         struct osd_object  *obj = osd_dt_obj(dt);
 935         struct osd_device  *osd = osd_obj2dev(obj);
 936         struct osd_thandle *oh;
 937         __u64               len;
 938         ENTRY;
 939
 940         oh = container_of0(handle, struct osd_thandle, ot_super);
 941
 942         read_lock(&obj->oo_attr_lock);
 943         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 944                 len = DMU_OBJECT_END;
 945         else
 946                 len = end - start;
 947
 948         /* declare we'll free some blocks ... */
 949         if (start < obj->oo_attr.la_size) {
 950                 read_unlock(&obj->oo_attr_lock);
 951                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 952         } else {
 953                 read_unlock(&obj->oo_attr_lock);
 954         }
 955
 956         /* ... and we'll modify size attribute */
 957         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 958
 959         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 960                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 961                                  false));
 962 }
 963
 964 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
 965                        __u64 start, __u64 end, enum lu_ladvise_type advice)
 966 {
 967         int     rc;
 968         ENTRY;
 969
 970         switch (advice) {
 971         default:
 972                 rc = -ENOTSUPP;
 973                 break;
 974         }
 975
 976         RETURN(rc);
 977 }
 978
 979 struct dt_body_operations osd_body_ops = {
 980         .dbo_read                       = osd_read,
 981         .dbo_declare_write              = osd_declare_write,
 982         .dbo_write                      = osd_write,
 983         .dbo_bufs_get                   = osd_bufs_get,
 984         .dbo_bufs_put                   = osd_bufs_put,
 985         .dbo_write_prep                 = osd_write_prep,
 986         .dbo_declare_write_commit       = osd_declare_write_commit,
 987         .dbo_write_commit               = osd_write_commit,
 988         .dbo_read_prep                  = osd_read_prep,
 989         .dbo_declare_punch              = osd_declare_punch,
 990         .dbo_punch                      = osd_punch,
 991         .dbo_ladvise                    = osd_ladvise,
 992 };