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 #include <lustre_quota.h>
  50
  51 #include "osd_internal.h"
  52
  53 #include <sys/dnode.h>
  54 #include <sys/dbuf.h>
  55 #include <sys/spa.h>
  56 #include <sys/stat.h>
  57 #include <sys/zap.h>
  58 #include <sys/spa_impl.h>
  59 #include <sys/zfs_znode.h>
  60 #include <sys/dmu_tx.h>
  61 #include <sys/dmu_objset.h>
  62 #include <sys/dsl_prop.h>
  63 #include <sys/sa_impl.h>
  64 #include <sys/txg.h>
  65
  66 static char *osd_0copy_tag = "zerocopy";
  67
  68
  69 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  70 {
  71         struct obd_histogram *h = osd->od_brw_stats.hist;
  72
  73         if (rw == READ) {
  74                 atomic_inc(&osd->od_r_in_flight);
  75                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  76                                  atomic_read(&osd->od_r_in_flight));
  77                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  78
  79         } else {
  80                 atomic_inc(&osd->od_w_in_flight);
  81                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  82                                  atomic_read(&osd->od_w_in_flight));
  83                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  84
  85         }
  86 }
  87
  88 static void record_end_io(struct osd_device *osd, int rw,
  89                           unsigned long elapsed, int disksize, int npages)
  90 {
  91         struct obd_histogram *h = osd->od_brw_stats.hist;
  92
  93         if (rw == READ) {
  94                 atomic_dec(&osd->od_r_in_flight);
  95                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  96                 if (disksize > 0)
  97                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
  98                 if (elapsed)
  99                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
 100
 101         } else {
 102                 atomic_dec(&osd->od_w_in_flight);
 103                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 104                 if (disksize > 0)
 105                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 106                 if (elapsed)
 107                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 108         }
 109 }
 110
 111 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 112                         struct lu_buf *buf, loff_t *pos)
 113 {
 114         struct osd_object *obj  = osd_dt_obj(dt);
 115         struct osd_device *osd = osd_obj2dev(obj);
 116         uint64_t           old_size;
 117         int                size = buf->lb_len;
 118         int                rc;
 119         unsigned long      start;
 120
 121         LASSERT(dt_object_exists(dt));
 122         LASSERT(obj->oo_dn);
 123
 124         start = cfs_time_current();
 125
 126         read_lock(&obj->oo_attr_lock);
 127         old_size = obj->oo_attr.la_size;
 128         read_unlock(&obj->oo_attr_lock);
 129
 130         if (*pos + size > old_size) {
 131                 if (old_size < *pos)
 132                         return 0;
 133                 else
 134                         size = old_size - *pos;
 135         }
 136
 137         record_start_io(osd, READ, 0);
 138
 139         rc = osd_dmu_read(osd, obj->oo_dn, *pos, size, buf->lb_buf,
 140                           DMU_READ_PREFETCH);
 141
 142         record_end_io(osd, READ, cfs_time_current() - start, size,
 143                       size >> PAGE_SHIFT);
 144         if (rc == 0) {
 145                 rc = size;
 146                 *pos += size;
 147         }
 148         return rc;
 149 }
 150
 151 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 152                                 const struct lu_buf *buf, loff_t pos,
 153                                 struct thandle *th)
 154 {
 155         struct osd_object  *obj  = osd_dt_obj(dt);
 156         struct osd_device  *osd = osd_obj2dev(obj);
 157         struct osd_thandle *oh;
 158         uint64_t            oid;
 159         ENTRY;
 160
 161         oh = container_of0(th, struct osd_thandle, ot_super);
 162
 163         /* in some cases declare can race with creation (e.g. llog)
 164          * and we need to wait till object is initialized. notice
 165          * LOHA_EXISTs is supposed to be the last step in the
 166          * initialization */
 167
 168         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 169         if (dt_object_exists(dt))
 170                 oid = obj->oo_dn->dn_object;
 171         else
 172                 oid = DMU_NEW_OBJECT;
 173
 174         /* XXX: we still miss for append declaration support in ZFS
 175          *      -1 means append which is used by llog mostly, llog
 176          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 177         if (pos == -1)
 178                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 179                             obj->oo_attr.la_size + (2 << 20));
 180         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 181
 182         /* dt_declare_write() is usually called for system objects, such
 183          * as llog or last_rcvd files. We needn't enforce quota on those
 184          * objects, so always set the lqi_space as 0. */
 185         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 186                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 187                                  0, oh, NULL, OSD_QID_BLK));
 188 }
 189
 190 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 191                         const struct lu_buf *buf, loff_t *pos,
 192                         struct thandle *th, int ignore_quota)
 193 {
 194         struct osd_object  *obj  = osd_dt_obj(dt);
 195         struct osd_device  *osd = osd_obj2dev(obj);
 196         struct osd_thandle *oh;
 197         uint64_t            offset = *pos;
 198         int                 rc;
 199
 200         ENTRY;
 201
 202         LASSERT(dt_object_exists(dt));
 203         LASSERT(obj->oo_dn);
 204
 205         LASSERT(th != NULL);
 206         oh = container_of0(th, struct osd_thandle, ot_super);
 207
 208         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 209                       buf->lb_buf, oh->ot_tx);
 210         write_lock(&obj->oo_attr_lock);
 211         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 212                 obj->oo_attr.la_size = offset + buf->lb_len;
 213                 write_unlock(&obj->oo_attr_lock);
 214                 /* osd_object_sa_update() will be copying directly from oo_attr
 215                  * into dbuf.  any update within a single txg will copy the
 216                  * most actual */
 217                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 218                                         &obj->oo_attr.la_size, 8, oh);
 219                 if (unlikely(rc))
 220                         GOTO(out, rc);
 221         } else {
 222                 write_unlock(&obj->oo_attr_lock);
 223         }
 224
 225         *pos += buf->lb_len;
 226         rc = buf->lb_len;
 227
 228 out:
 229         RETURN(rc);
 230 }
 231
 232 /*
 233  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 234  *      purposes as it's not very well defined ...
 235  *      instead I use the lowest bit of the address so that:
 236  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 237  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 238  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 239  *
 240  *      bzzz, to blame
 241  */
 242 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 243                         struct niobuf_local *lnb, int npages)
 244 {
 245         struct osd_object *obj  = osd_dt_obj(dt);
 246         struct osd_device *osd = osd_obj2dev(obj);
 247         unsigned long      ptr;
 248         int                i;
 249
 250         LASSERT(dt_object_exists(dt));
 251         LASSERT(obj->oo_dn);
 252
 253         for (i = 0; i < npages; i++) {
 254                 if (lnb[i].lnb_page == NULL)
 255                         continue;
 256                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 257                         /* this is anonymous page allocated for copy-write */
 258                         lnb[i].lnb_page->mapping = NULL;
 259                         __free_page(lnb[i].lnb_page);
 260                         atomic_dec(&osd->od_zerocopy_alloc);
 261                 } else {
 262                         /* see comment in osd_bufs_get_read() */
 263                         ptr = (unsigned long)lnb[i].lnb_data;
 264                         if (ptr & 1UL) {
 265                                 ptr &= ~1UL;
 266                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 267                                 atomic_dec(&osd->od_zerocopy_pin);
 268                         } else if (lnb[i].lnb_data != NULL) {
 269                                 int j, apages, abufsz;
 270                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 271                                 apages = abufsz / PAGE_SIZE;
 272                                 /* these references to pages must be invalidated
 273                                  * to prevent access in osd_bufs_put() */
 274                                 for (j = 0; j < apages; j++)
 275                                         lnb[i + j].lnb_page = NULL;
 276                                 dmu_return_arcbuf(lnb[i].lnb_data);
 277                                 atomic_dec(&osd->od_zerocopy_loan);
 278                         }
 279                 }
 280                 lnb[i].lnb_page = NULL;
 281                 lnb[i].lnb_data = NULL;
 282         }
 283
 284         return 0;
 285 }
 286
 287 static inline struct page *kmem_to_page(void *addr)
 288 {
 289         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 290         if (is_vmalloc_addr(addr))
 291                 return vmalloc_to_page(addr);
 292         else
 293                 return virt_to_page(addr);
 294 }
 295
 296 /**
 297  * Prepare buffers for read.
 298  *
 299  * The function maps the range described by \a off and \a len to \a lnb array.
 300  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 301  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 302  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 303  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 304  * A better implementation would just return the buffers (potentially unfilled)
 305  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 306  *
 307  * \param[in] env       environment
 308  * \param[in] obj       object
 309  * \param[in] off       offset in bytes
 310  * \param[in] len       the number of bytes to access
 311  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 312  *
 313  * \retval              0 for success
 314  * \retval              negative error number of failure
 315  */
 316 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 317                              loff_t off, ssize_t len, struct niobuf_local *lnb)
 318 {
 319         struct osd_device *osd = osd_obj2dev(obj);
 320         unsigned long      start = cfs_time_current();
 321         int                rc, i, numbufs, npages = 0;
 322         dmu_buf_t        **dbp;
 323         ENTRY;
 324
 325         record_start_io(osd, READ, 0);
 326
 327         /* grab buffers for read:
 328          * OSD API let us to grab buffers first, then initiate IO(s)
 329          * so that all required IOs will be done in parallel, but at the
 330          * moment DMU doesn't provide us with a method to grab buffers.
 331          * If we discover this is a vital for good performance we
 332          * can get own replacement for dmu_buf_hold_array_by_bonus().
 333          */
 334         while (len > 0) {
 335                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 336                                                   off, len, TRUE, osd_0copy_tag,
 337                                                   &numbufs, &dbp);
 338                 if (unlikely(rc))
 339                         GOTO(err, rc);
 340
 341                 for (i = 0; i < numbufs; i++) {
 342                         int bufoff, tocpy, thispage;
 343                         void *dbf = dbp[i];
 344
 345                         LASSERT(len > 0);
 346
 347                         atomic_inc(&osd->od_zerocopy_pin);
 348
 349                         bufoff = off - dbp[i]->db_offset;
 350                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 351
 352                         /* kind of trick to differentiate dbuf vs. arcbuf */
 353                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 354                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 355
 356                         while (tocpy > 0) {
 357                                 thispage = PAGE_SIZE;
 358                                 thispage -= bufoff & (PAGE_SIZE - 1);
 359                                 thispage = min(tocpy, thispage);
 360
 361                                 lnb->lnb_rc = 0;
 362                                 lnb->lnb_file_offset = off;
 363                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 364                                 lnb->lnb_len = thispage;
 365                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 366                                                              bufoff);
 367                                 /* mark just a single slot: we need this
 368                                  * reference to dbuf to be released once */
 369                                 lnb->lnb_data = dbf;
 370                                 dbf = NULL;
 371
 372                                 tocpy -= thispage;
 373                                 len -= thispage;
 374                                 bufoff += thispage;
 375                                 off += thispage;
 376
 377                                 npages++;
 378                                 lnb++;
 379                         }
 380
 381                         /* steal dbuf so dmu_buf_rele_array() can't release
 382                          * it */
 383                         dbp[i] = NULL;
 384                 }
 385
 386                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 387         }
 388
 389         record_end_io(osd, READ, cfs_time_current() - start,
 390                       npages * PAGE_SIZE, npages);
 391
 392         RETURN(npages);
 393
 394 err:
 395         LASSERT(rc < 0);
 396         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 397         RETURN(rc);
 398 }
 399
 400 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 401 {
 402         arc_buf_t *abuf;
 403
 404         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 405         if (unlikely(!abuf))
 406                 return ERR_PTR(-ENOMEM);
 407
 408 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 409         /**
 410          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 411          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 412          * setting up page array for RDMA transfer. See LU-9305.
 413          */
 414         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 415                 dmu_return_arcbuf(abuf);
 416                 return NULL;
 417         }
 418 #endif
 419
 420         return abuf;
 421 }
 422
 423 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 424                               loff_t off, ssize_t len, struct niobuf_local *lnb)
 425 {
 426         struct osd_device *osd = osd_obj2dev(obj);
 427         int                plen, off_in_block, sz_in_block;
 428         int                rc, i = 0, npages = 0;
 429         dnode_t *dn = obj->oo_dn;
 430         arc_buf_t *abuf;
 431         uint32_t bs = dn->dn_datablksz;
 432         ENTRY;
 433
 434         /*
 435          * currently only full blocks are subject to zerocopy approach:
 436          * so that we're sure nobody is trying to update the same block
 437          */
 438         while (len > 0) {
 439                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 440
 441                 off_in_block = off & (bs - 1);
 442                 sz_in_block = min_t(int, bs - off_in_block, len);
 443
 444                 abuf = NULL;
 445                 if (sz_in_block == bs) {
 446                         /* full block, try to use zerocopy */
 447                         abuf = osd_request_arcbuf(dn, bs);
 448                         if (unlikely(IS_ERR(abuf)))
 449                                 GOTO(out_err, rc = PTR_ERR(abuf));
 450                 }
 451
 452                 if (abuf != NULL) {
 453                         atomic_inc(&osd->od_zerocopy_loan);
 454
 455                         /* go over pages arcbuf contains, put them as
 456                          * local niobufs for ptlrpc's bulks */
 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                                 if (sz_in_block == bs)
 465                                         lnb[i].lnb_data = abuf;
 466                                 else
 467                                         lnb[i].lnb_data = NULL;
 468
 469                                 /* this one is not supposed to fail */
 470                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 471                                                         off_in_block);
 472                                 LASSERT(lnb[i].lnb_page);
 473
 474                                 lprocfs_counter_add(osd->od_stats,
 475                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 476
 477                                 sz_in_block -= plen;
 478                                 len -= plen;
 479                                 off += plen;
 480                                 off_in_block += plen;
 481                                 i++;
 482                                 npages++;
 483                         }
 484                 } else {
 485                         if (off_in_block == 0 && len < bs &&
 486                                         off + len >= obj->oo_attr.la_size)
 487                                 lprocfs_counter_add(osd->od_stats,
 488                                                 LPROC_OSD_TAIL_IO, 1);
 489
 490                         /* can't use zerocopy, allocate temp. buffers */
 491                         while (sz_in_block > 0) {
 492                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 493
 494                                 lnb[i].lnb_file_offset = off;
 495                                 lnb[i].lnb_page_offset = 0;
 496                                 lnb[i].lnb_len = plen;
 497                                 lnb[i].lnb_rc = 0;
 498                                 lnb[i].lnb_data = NULL;
 499
 500                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 501                                 if (unlikely(lnb[i].lnb_page == NULL))
 502                                         GOTO(out_err, rc = -ENOMEM);
 503
 504                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 505                                 lnb[i].lnb_page->mapping = (void *)obj;
 506
 507                                 atomic_inc(&osd->od_zerocopy_alloc);
 508                                 lprocfs_counter_add(osd->od_stats,
 509                                                 LPROC_OSD_COPY_IO, 1);
 510
 511                                 sz_in_block -= plen;
 512                                 len -= plen;
 513                                 off += plen;
 514                                 i++;
 515                                 npages++;
 516                         }
 517                 }
 518         }
 519
 520         RETURN(npages);
 521
 522 out_err:
 523         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 524         RETURN(rc);
 525 }
 526
 527 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 528                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 529                         enum dt_bufs_type rw)
 530 {
 531         struct osd_object *obj  = osd_dt_obj(dt);
 532         int                rc;
 533
 534         LASSERT(dt_object_exists(dt));
 535         LASSERT(obj->oo_dn);
 536
 537         if (rw & DT_BUFS_TYPE_WRITE)
 538                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 539         else
 540                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 541
 542         return rc;
 543 }
 544
 545 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 546                         struct niobuf_local *lnb, int npages)
 547 {
 548         struct osd_object *obj = osd_dt_obj(dt);
 549
 550         LASSERT(dt_object_exists(dt));
 551         LASSERT(obj->oo_dn);
 552
 553         return 0;
 554 }
 555
 556 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 557                                            uint64_t offset,
 558                                            uint32_t blksz)
 559 {
 560         LASSERT(blksz > 0);
 561
 562         size += offset % blksz;
 563
 564         if (likely(is_power_of_2(blksz)))
 565                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 566
 567         size += blksz - 1;
 568         do_div(size, blksz);
 569         return size * blksz;
 570 }
 571
 572 static int osd_declare_write_commit(const struct lu_env *env,
 573                                     struct dt_object *dt,
 574                                     struct niobuf_local *lnb, int npages,
 575                                     struct thandle *th)
 576 {
 577         struct osd_object  *obj = osd_dt_obj(dt);
 578         struct osd_device  *osd = osd_obj2dev(obj);
 579         struct osd_thandle *oh;
 580         uint64_t            offset = 0;
 581         uint32_t            size = 0;
 582         uint32_t blksz = obj->oo_dn->dn_datablksz;
 583         int                 i, rc, flags = 0;
 584         bool synced = false;
 585         long long           space = 0;
 586         struct page        *last_page = NULL;
 587         unsigned long       discont_pages = 0;
 588         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 589         ENTRY;
 590
 591         LASSERT(dt_object_exists(dt));
 592         LASSERT(obj->oo_dn);
 593
 594         LASSERT(lnb);
 595         LASSERT(npages > 0);
 596
 597         oh = container_of0(th, struct osd_thandle, ot_super);
 598
 599         for (i = 0; i < npages; i++) {
 600                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 601                         ++discont_pages;
 602                 last_page = lnb[i].lnb_page;
 603                 if (lnb[i].lnb_rc)
 604                         /* ENOSPC, network RPC error, etc.
 605                          * We don't want to book space for pages which will be
 606                          * skipped in osd_write_commit(). Hence we skip pages
 607                          * with lnb_rc != 0 here too */
 608                         continue;
 609                 /* ignore quota for the whole request if any page is from
 610                  * client cache or written by root.
 611                  *
 612                  * XXX once we drop the 1.8 client support, the checking
 613                  * for whether page is from cache can be simplified as:
 614                  * !(lnb[i].flags & OBD_BRW_SYNC)
 615                  *
 616                  * XXX we could handle this on per-lnb basis as done by
 617                  * grant. */
 618                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 619                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 620                     OBD_BRW_FROM_GRANT)
 621                         declare_flags |= OSD_QID_FORCE;
 622
 623                 if (size == 0) {
 624                         /* first valid lnb */
 625                         offset = lnb[i].lnb_file_offset;
 626                         size = lnb[i].lnb_len;
 627                         continue;
 628                 }
 629                 if (offset + size == lnb[i].lnb_file_offset) {
 630                         /* this lnb is contiguous to the previous one */
 631                         size += lnb[i].lnb_len;
 632                         continue;
 633                 }
 634
 635                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 636                                   obj->oo_dn, offset, size);
 637                 /* Estimating space to be consumed by a write is rather
 638                  * complicated with ZFS. As a consequence, we don't account for
 639                  * indirect blocks and just use as a rough estimate the worse
 640                  * case where the old space is being held by a snapshot. Quota
 641                  * overrun will be adjusted once the operation is committed, if
 642                  * required. */
 643                 space += osd_roundup2blocksz(size, offset, blksz);
 644
 645                 offset = lnb[i].lnb_file_offset;
 646                 size = lnb[i].lnb_len;
 647         }
 648
 649         if (size) {
 650                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 651                                   offset, size);
 652                 space += osd_roundup2blocksz(size, offset, blksz);
 653         }
 654
 655         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 656
 657         /* backend zfs filesystem might be configured to store multiple data
 658          * copies */
 659         space  *= osd->od_os->os_copies;
 660         space   = toqb(space);
 661         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 662                npages, space);
 663
 664         record_start_io(osd, WRITE, discont_pages);
 665 retry:
 666         /* acquire quota space if needed */
 667         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 668                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 669                                space, oh, &flags, declare_flags);
 670
 671         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 672                 dt_sync(env, th->th_dev);
 673                 synced = true;
 674                 CDEBUG(D_QUOTA, "retry after sync\n");
 675                 flags = 0;
 676                 goto retry;
 677         }
 678
 679         /* we need only to store the overquota flags in the first lnb for
 680          * now, once we support multiple objects BRW, this code needs be
 681          * revised. */
 682         if (flags & QUOTA_FL_OVER_USRQUOTA)
 683                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 684         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 685                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 686 #ifdef ZFS_PROJINHERIT
 687         if (flags & QUOTA_FL_OVER_PRJQUOTA)
 688                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 689 #endif
 690
 691         RETURN(rc);
 692 }
 693
 694 /**
 695  * Policy to grow ZFS block size by write pattern.
 696  * For sequential write, it grows block size gradually until it reaches the
 697  * maximum blocksize the dataset can support. Otherwise, it will pick a
 698  * a block size by the writing region of this I/O.
 699  */
 700 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 701                               uint64_t start, uint64_t end)
 702 {
 703         struct osd_device       *osd = osd_obj2dev(obj);
 704         dnode_t *dn = obj->oo_dn;
 705         uint32_t                 blksz;
 706         int                      rc = 0;
 707
 708         ENTRY;
 709
 710         if (dn->dn_maxblkid > 0) /* can't change block size */
 711                 GOTO(out, rc);
 712
 713         if (dn->dn_datablksz >= osd->od_max_blksz)
 714                 GOTO(out, rc);
 715
 716         down_write(&obj->oo_guard);
 717
 718         blksz = dn->dn_datablksz;
 719         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 720                 GOTO(out_unlock, rc);
 721
 722         /* now ZFS can support up to 16MB block size, and if the write
 723          * is sequential, it just increases the block size gradually */
 724         if (start <= blksz) { /* sequential */
 725                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 726         } else { /* sparse, pick a block size by write region */
 727                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 728                                         end - start);
 729         }
 730
 731         if (!is_power_of_2(blksz))
 732                 blksz = size_roundup_power2(blksz);
 733
 734         if (blksz > dn->dn_datablksz) {
 735                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 736                                                blksz, 0, oh->ot_tx);
 737                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 738                 if (rc < 0)
 739                         CDEBUG(D_INODE, "object "DFID": change block size"
 740                                "%u -> %u error rc = %d\n",
 741                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 742                                dn->dn_datablksz, blksz, rc);
 743         }
 744         EXIT;
 745 out_unlock:
 746         up_write(&obj->oo_guard);
 747 out:
 748         return rc;
 749 }
 750
 751 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 752                         struct niobuf_local *lnb, int npages,
 753                         struct thandle *th)
 754 {
 755         struct osd_object  *obj  = osd_dt_obj(dt);
 756         struct osd_device  *osd = osd_obj2dev(obj);
 757         struct osd_thandle *oh;
 758         uint64_t            new_size = 0;
 759         int                 i, rc = 0;
 760         unsigned long      iosize = 0;
 761         ENTRY;
 762
 763         LASSERT(dt_object_exists(dt));
 764         LASSERT(obj->oo_dn);
 765
 766         LASSERT(th != NULL);
 767         oh = container_of0(th, struct osd_thandle, ot_super);
 768
 769         /* adjust block size. Assume the buffers are sorted. */
 770         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 771                                  lnb[npages - 1].lnb_file_offset +
 772                                  lnb[npages - 1].lnb_len);
 773
 774         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 775          * size and assigning arcbuf take place at the same time.
 776          *
 777          * Thread 1:
 778          * osd_write_commit()
 779          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 780          *   -> dmu_object_set_blocksize()
 781          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 782          *       write lock held
 783          *     -> dbuf_new_size()
 784          *      -> dmu_buf_will_dirty()
 785          *       -> dbuf_read()
 786          *        -> wait for the dbuf state to change
 787          * Thread 2:
 788          * osd_write_commit()
 789          *  -> dmu_assign_arcbuf()
 790          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 791          *    -> dbuf_dirty()
 792          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 793          *
 794          * By taking the read lock, it can avoid thread 2 to enter into the
 795          * critical section of assigning the arcbuf, while thread 1 is
 796          * changing the block size.
 797          */
 798         down_read(&obj->oo_guard);
 799         for (i = 0; i < npages; i++) {
 800                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 801                         (unsigned) lnb[i].lnb_len,
 802                         (unsigned) lnb[i].lnb_file_offset);
 803
 804                 if (lnb[i].lnb_rc) {
 805                         /* ENOSPC, network RPC error, etc.
 806                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 807                          * so we skip this page if lnb_rc is set to -ENOSPC */
 808                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 809                                 PFID(lu_object_fid(&dt->do_lu)), i,
 810                                 lnb[i].lnb_rc);
 811                         continue;
 812                 }
 813
 814                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 815                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 816                 if (lnb[i].lnb_page == NULL)
 817                         continue;
 818
 819                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 820                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 821                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 822                                       oh->ot_tx);
 823                         kunmap(lnb[i].lnb_page);
 824                         iosize += lnb[i].lnb_len;
 825                 } else if (lnb[i].lnb_data) {
 826                         int j, apages, abufsz;
 827                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 828                         /* buffer loaned for zerocopy, try to use it.
 829                          * notice that dmu_assign_arcbuf() is smart
 830                          * enough to recognize changed blocksize
 831                          * in this case it fallbacks to dmu_write() */
 832                         abufsz = arc_buf_size(lnb[i].lnb_data);
 833                         LASSERT(abufsz & PAGE_MASK);
 834                         apages = abufsz / PAGE_SIZE;
 835                         LASSERT(i + apages <= npages);
 836                         /* these references to pages must be invalidated
 837                          * to prevent access in osd_bufs_put() */
 838                         for (j = 0; j < apages; j++)
 839                                 lnb[i + j].lnb_page = NULL;
 840                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 841                                           lnb[i].lnb_file_offset,
 842                                           lnb[i].lnb_data, oh->ot_tx);
 843                         /* drop the reference, otherwise osd_put_bufs()
 844                          * will be releasing it - bad! */
 845                         lnb[i].lnb_data = NULL;
 846                         atomic_dec(&osd->od_zerocopy_loan);
 847                         iosize += abufsz;
 848                 }
 849
 850         }
 851         up_read(&obj->oo_guard);
 852
 853         if (unlikely(new_size == 0)) {
 854                 /* no pages to write, no transno is needed */
 855                 th->th_local = 1;
 856                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 857                  * since ofd_commitrw_write() retries several times on ENOSPC */
 858                 record_end_io(osd, WRITE, 0, 0, 0);
 859                 RETURN(0);
 860         }
 861
 862         write_lock(&obj->oo_attr_lock);
 863         if (obj->oo_attr.la_size < new_size) {
 864                 obj->oo_attr.la_size = new_size;
 865                 write_unlock(&obj->oo_attr_lock);
 866                 /* osd_object_sa_update() will be copying directly from
 867                  * oo_attr into dbuf. any update within a single txg will copy
 868                  * the most actual */
 869                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 870                                           &obj->oo_attr.la_size, 8, oh);
 871         } else {
 872                 write_unlock(&obj->oo_attr_lock);
 873         }
 874
 875         record_end_io(osd, WRITE, 0, iosize, npages);
 876
 877         RETURN(rc);
 878 }
 879
 880 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 881                         struct niobuf_local *lnb, int npages)
 882 {
 883         struct osd_object *obj  = osd_dt_obj(dt);
 884         int                i;
 885         loff_t             eof;
 886
 887         LASSERT(dt_object_exists(dt));
 888         LASSERT(obj->oo_dn);
 889
 890         read_lock(&obj->oo_attr_lock);
 891         eof = obj->oo_attr.la_size;
 892         read_unlock(&obj->oo_attr_lock);
 893
 894         for (i = 0; i < npages; i++) {
 895                 if (unlikely(lnb[i].lnb_rc < 0))
 896                         continue;
 897
 898                 lnb[i].lnb_rc = lnb[i].lnb_len;
 899
 900                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 901                         if (eof <= lnb[i].lnb_file_offset)
 902                                 lnb[i].lnb_rc = 0;
 903                         else
 904                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 905
 906                         /* all subsequent rc should be 0 */
 907                         while (++i < npages)
 908                                 lnb[i].lnb_rc = 0;
 909                         break;
 910                 }
 911         }
 912
 913         return 0;
 914 }
 915
 916 /*
 917  * Punch/truncate an object
 918  *
 919  *      IN:     db  - dmu_buf of the object to free data in.
 920  *              off - start of section to free.
 921  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 922  *
 923  *      RETURN: 0 if success
 924  *              error code if failure
 925  *
 926  * The transaction passed to this routine must have
 927  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 928  * called and then assigned to a transaction group.
 929  */
 930 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 931                                 uint64_t size, uint64_t off, uint64_t len)
 932 {
 933         int rc = 0;
 934
 935         /* Assert that the transaction has been assigned to a
 936            transaction group. */
 937         LASSERT(tx->tx_txg != 0);
 938         /*
 939          * Nothing to do if file already at desired length.
 940          */
 941         if (len == DMU_OBJECT_END && size == off)
 942                 return 0;
 943
 944         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 945         if (off < size)
 946                 dmu_free_range(os, dn->dn_object, off, len, tx);
 947
 948         return rc;
 949 }
 950
 951 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 952                         __u64 start, __u64 end, struct thandle *th)
 953 {
 954         struct osd_object  *obj = osd_dt_obj(dt);
 955         struct osd_device  *osd = osd_obj2dev(obj);
 956         struct osd_thandle *oh;
 957         __u64               len;
 958         int                 rc = 0;
 959         ENTRY;
 960
 961         LASSERT(dt_object_exists(dt));
 962         LASSERT(osd_invariant(obj));
 963
 964         LASSERT(th != NULL);
 965         oh = container_of0(th, struct osd_thandle, ot_super);
 966
 967         write_lock(&obj->oo_attr_lock);
 968         /* truncate */
 969         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 970                 len = DMU_OBJECT_END;
 971         else
 972                 len = end - start;
 973         write_unlock(&obj->oo_attr_lock);
 974
 975         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 976                                 obj->oo_attr.la_size, start, len);
 977         /* set new size */
 978         if (len == DMU_OBJECT_END) {
 979                 write_lock(&obj->oo_attr_lock);
 980                 obj->oo_attr.la_size = start;
 981                 write_unlock(&obj->oo_attr_lock);
 982                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 983                                           &obj->oo_attr.la_size, 8, oh);
 984         }
 985         RETURN(rc);
 986 }
 987
 988 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 989                         __u64 start, __u64 end, struct thandle *handle)
 990 {
 991         struct osd_object  *obj = osd_dt_obj(dt);
 992         struct osd_device  *osd = osd_obj2dev(obj);
 993         struct osd_thandle *oh;
 994         __u64               len;
 995         ENTRY;
 996
 997         oh = container_of0(handle, struct osd_thandle, ot_super);
 998
 999         read_lock(&obj->oo_attr_lock);
1000         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1001                 len = DMU_OBJECT_END;
1002         else
1003                 len = end - start;
1004
1005         /* declare we'll free some blocks ... */
1006         if (start < obj->oo_attr.la_size) {
1007                 read_unlock(&obj->oo_attr_lock);
1008                 dmu_tx_mark_netfree(oh->ot_tx);
1009                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1010         } else {
1011                 read_unlock(&obj->oo_attr_lock);
1012         }
1013
1014         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1015                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1016                                  0, oh, NULL, OSD_QID_BLK));
1017 }
1018
1019 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1020                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1021 {
1022         int     rc;
1023         ENTRY;
1024
1025         switch (advice) {
1026         default:
1027                 rc = -ENOTSUPP;
1028                 break;
1029         }
1030
1031         RETURN(rc);
1032 }
1033
1034 struct dt_body_operations osd_body_ops = {
1035         .dbo_read                       = osd_read,
1036         .dbo_declare_write              = osd_declare_write,
1037         .dbo_write                      = osd_write,
1038         .dbo_bufs_get                   = osd_bufs_get,
1039         .dbo_bufs_put                   = osd_bufs_put,
1040         .dbo_write_prep                 = osd_write_prep,
1041         .dbo_declare_write_commit       = osd_declare_write_commit,
1042         .dbo_write_commit               = osd_write_commit,
1043         .dbo_read_prep                  = osd_read_prep,
1044         .dbo_declare_punch              = osd_declare_punch,
1045         .dbo_punch                      = osd_punch,
1046         .dbo_ladvise                    = osd_ladvise,
1047 };