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 <libcfs/libcfs.h>
  41 #include <obd_support.h>
  42 #include <lustre_net.h>
  43 #include <obd.h>
  44 #include <obd_class.h>
  45 #include <lustre_disk.h>
  46 #include <lustre_fid.h>
  47
  48 #include "osd_internal.h"
  49
  50 #include <sys/dnode.h>
  51 #include <sys/dbuf.h>
  52 #include <sys/spa.h>
  53 #include <sys/stat.h>
  54 #include <sys/zap.h>
  55 #include <sys/spa_impl.h>
  56 #include <sys/zfs_znode.h>
  57 #include <sys/dmu_tx.h>
  58 #include <sys/dmu_objset.h>
  59 #include <sys/dsl_prop.h>
  60 #include <sys/sa_impl.h>
  61 #include <sys/txg.h>
  62
  63 static char *osd_0copy_tag = "zerocopy";
  64
  65
  66 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  67 {
  68         struct obd_histogram *h = osd->od_brw_stats.hist;
  69
  70         if (rw == READ) {
  71                 atomic_inc(&osd->od_r_in_flight);
  72                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  73                                  atomic_read(&osd->od_r_in_flight));
  74                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  75
  76         } else {
  77                 atomic_inc(&osd->od_w_in_flight);
  78                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  79                                  atomic_read(&osd->od_w_in_flight));
  80                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  81
  82         }
  83 }
  84
  85 static void record_end_io(struct osd_device *osd, int rw,
  86                           unsigned long elapsed, int disksize, int npages)
  87 {
  88         struct obd_histogram *h = osd->od_brw_stats.hist;
  89
  90         if (rw == READ) {
  91                 atomic_dec(&osd->od_r_in_flight);
  92                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  93                 if (disksize > 0)
  94                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
  95                 if (elapsed)
  96                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
  97
  98         } else {
  99                 atomic_dec(&osd->od_w_in_flight);
 100                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 101                 if (disksize > 0)
 102                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 103                 if (elapsed)
 104                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 105         }
 106 }
 107
 108 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 109                         struct lu_buf *buf, loff_t *pos)
 110 {
 111         struct osd_object *obj  = osd_dt_obj(dt);
 112         struct osd_device *osd = osd_obj2dev(obj);
 113         uint64_t           old_size;
 114         int                size = buf->lb_len;
 115         int                rc;
 116         unsigned long      start;
 117
 118         LASSERT(dt_object_exists(dt));
 119         LASSERT(obj->oo_dn);
 120
 121         start = cfs_time_current();
 122
 123         read_lock(&obj->oo_attr_lock);
 124         old_size = obj->oo_attr.la_size;
 125         read_unlock(&obj->oo_attr_lock);
 126
 127         if (*pos + size > old_size) {
 128                 if (old_size < *pos)
 129                         return 0;
 130                 else
 131                         size = old_size - *pos;
 132         }
 133
 134         record_start_io(osd, READ, 0);
 135
 136         rc = osd_dmu_read(osd, obj->oo_dn, *pos, size, buf->lb_buf,
 137                           DMU_READ_PREFETCH);
 138
 139         record_end_io(osd, READ, cfs_time_current() - start, size,
 140                       size >> PAGE_SHIFT);
 141         if (rc == 0) {
 142                 rc = size;
 143                 *pos += size;
 144         }
 145         return rc;
 146 }
 147
 148 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 149                                 const struct lu_buf *buf, loff_t pos,
 150                                 struct thandle *th)
 151 {
 152         struct osd_object  *obj  = osd_dt_obj(dt);
 153         struct osd_device  *osd = osd_obj2dev(obj);
 154         struct osd_thandle *oh;
 155         uint64_t            oid;
 156         ENTRY;
 157
 158         oh = container_of0(th, struct osd_thandle, ot_super);
 159
 160         /* in some cases declare can race with creation (e.g. llog)
 161          * and we need to wait till object is initialized. notice
 162          * LOHA_EXISTs is supposed to be the last step in the
 163          * initialization */
 164
 165         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 166         if (dt_object_exists(dt))
 167                 oid = obj->oo_dn->dn_object;
 168         else
 169                 oid = DMU_NEW_OBJECT;
 170
 171         /* XXX: we still miss for append declaration support in ZFS
 172          *      -1 means append which is used by llog mostly, llog
 173          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 174         if (pos == -1)
 175                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 176                             obj->oo_attr.la_size + (2 << 20));
 177         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 178
 179         /* dt_declare_write() is usually called for system objects, such
 180          * as llog or last_rcvd files. We needn't enforce quota on those
 181          * objects, so always set the lqi_space as 0. */
 182         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 183                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 184                                  false));
 185 }
 186
 187 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 188                         const struct lu_buf *buf, loff_t *pos,
 189                         struct thandle *th, int ignore_quota)
 190 {
 191         struct osd_object  *obj  = osd_dt_obj(dt);
 192         struct osd_device  *osd = osd_obj2dev(obj);
 193         struct osd_thandle *oh;
 194         uint64_t            offset = *pos;
 195         int                 rc;
 196
 197         ENTRY;
 198
 199         LASSERT(dt_object_exists(dt));
 200         LASSERT(obj->oo_dn);
 201
 202         LASSERT(th != NULL);
 203         oh = container_of0(th, struct osd_thandle, ot_super);
 204
 205         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 206                       buf->lb_buf, oh->ot_tx);
 207         write_lock(&obj->oo_attr_lock);
 208         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 209                 obj->oo_attr.la_size = offset + buf->lb_len;
 210                 write_unlock(&obj->oo_attr_lock);
 211                 /* osd_object_sa_update() will be copying directly from oo_attr
 212                  * into dbuf.  any update within a single txg will copy the
 213                  * most actual */
 214                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 215                                         &obj->oo_attr.la_size, 8, oh);
 216                 if (unlikely(rc))
 217                         GOTO(out, rc);
 218         } else {
 219                 write_unlock(&obj->oo_attr_lock);
 220         }
 221
 222         *pos += buf->lb_len;
 223         rc = buf->lb_len;
 224
 225 out:
 226         RETURN(rc);
 227 }
 228
 229 /*
 230  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 231  *      purposes as it's not very well defined ...
 232  *      instead I use the lowest bit of the address so that:
 233  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 234  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 235  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 236  *
 237  *      bzzz, to blame
 238  */
 239 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 240                         struct niobuf_local *lnb, int npages)
 241 {
 242         struct osd_object *obj  = osd_dt_obj(dt);
 243         struct osd_device *osd = osd_obj2dev(obj);
 244         unsigned long      ptr;
 245         int                i;
 246
 247         LASSERT(dt_object_exists(dt));
 248         LASSERT(obj->oo_dn);
 249
 250         for (i = 0; i < npages; i++) {
 251                 if (lnb[i].lnb_page == NULL)
 252                         continue;
 253                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 254                         /* this is anonymous page allocated for copy-write */
 255                         lnb[i].lnb_page->mapping = NULL;
 256                         __free_page(lnb[i].lnb_page);
 257                         atomic_dec(&osd->od_zerocopy_alloc);
 258                 } else {
 259                         /* see comment in osd_bufs_get_read() */
 260                         ptr = (unsigned long)lnb[i].lnb_data;
 261                         if (ptr & 1UL) {
 262                                 ptr &= ~1UL;
 263                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 264                                 atomic_dec(&osd->od_zerocopy_pin);
 265                         } else if (lnb[i].lnb_data != NULL) {
 266                                 int j, apages, abufsz;
 267                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 268                                 apages = abufsz / PAGE_SIZE;
 269                                 /* these references to pages must be invalidated
 270                                  * to prevent access in osd_bufs_put() */
 271                                 for (j = 0; j < apages; j++)
 272                                         lnb[i + j].lnb_page = 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         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 287         if (is_vmalloc_addr(addr))
 288                 return vmalloc_to_page(addr);
 289         else
 290                 return virt_to_page(addr);
 291 }
 292
 293 /**
 294  * Prepare buffers for read.
 295  *
 296  * The function maps the range described by \a off and \a len to \a lnb array.
 297  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 298  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 299  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 300  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 301  * A better implementation would just return the buffers (potentially unfilled)
 302  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 303  *
 304  * \param[in] env       environment
 305  * \param[in] obj       object
 306  * \param[in] off       offset in bytes
 307  * \param[in] len       the number of bytes to access
 308  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 309  *
 310  * \retval              0 for success
 311  * \retval              negative error number of failure
 312  */
 313 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 314                              loff_t off, ssize_t len, struct niobuf_local *lnb)
 315 {
 316         struct osd_device *osd = osd_obj2dev(obj);
 317         unsigned long      start = cfs_time_current();
 318         int                rc, i, numbufs, npages = 0;
 319         dmu_buf_t        **dbp;
 320         ENTRY;
 321
 322         record_start_io(osd, READ, 0);
 323
 324         /* grab buffers for read:
 325          * OSD API let us to grab buffers first, then initiate IO(s)
 326          * so that all required IOs will be done in parallel, but at the
 327          * moment DMU doesn't provide us with a method to grab buffers.
 328          * If we discover this is a vital for good performance we
 329          * can get own replacement for dmu_buf_hold_array_by_bonus().
 330          */
 331         while (len > 0) {
 332                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 333                                                   off, len, TRUE, osd_0copy_tag,
 334                                                   &numbufs, &dbp);
 335                 if (unlikely(rc))
 336                         GOTO(err, rc);
 337
 338                 for (i = 0; i < numbufs; i++) {
 339                         int bufoff, tocpy, thispage;
 340                         void *dbf = dbp[i];
 341
 342                         LASSERT(len > 0);
 343
 344                         atomic_inc(&osd->od_zerocopy_pin);
 345
 346                         bufoff = off - dbp[i]->db_offset;
 347                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 348
 349                         /* kind of trick to differentiate dbuf vs. arcbuf */
 350                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 351                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 352
 353                         while (tocpy > 0) {
 354                                 thispage = PAGE_SIZE;
 355                                 thispage -= bufoff & (PAGE_SIZE - 1);
 356                                 thispage = min(tocpy, thispage);
 357
 358                                 lnb->lnb_rc = 0;
 359                                 lnb->lnb_file_offset = off;
 360                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 361                                 lnb->lnb_len = thispage;
 362                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 363                                                              bufoff);
 364                                 /* mark just a single slot: we need this
 365                                  * reference to dbuf to be released once */
 366                                 lnb->lnb_data = dbf;
 367                                 dbf = NULL;
 368
 369                                 tocpy -= thispage;
 370                                 len -= thispage;
 371                                 bufoff += thispage;
 372                                 off += thispage;
 373
 374                                 npages++;
 375                                 lnb++;
 376                         }
 377
 378                         /* steal dbuf so dmu_buf_rele_array() can't release
 379                          * it */
 380                         dbp[i] = NULL;
 381                 }
 382
 383                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 384         }
 385
 386         record_end_io(osd, READ, cfs_time_current() - start,
 387                       npages * PAGE_SIZE, npages);
 388
 389         RETURN(npages);
 390
 391 err:
 392         LASSERT(rc < 0);
 393         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 394         RETURN(rc);
 395 }
 396
 397 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 398 {
 399         arc_buf_t *abuf;
 400
 401         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 402         if (unlikely(!abuf))
 403                 return ERR_PTR(-ENOMEM);
 404
 405 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 406         /**
 407          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 408          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 409          * setting up page array for RDMA transfer. See LU-9305.
 410          */
 411         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 412                 dmu_return_arcbuf(abuf);
 413                 return NULL;
 414         }
 415 #endif
 416
 417         return abuf;
 418 }
 419
 420 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 421                               loff_t off, ssize_t len, struct niobuf_local *lnb)
 422 {
 423         struct osd_device *osd = osd_obj2dev(obj);
 424         int                plen, off_in_block, sz_in_block;
 425         int                rc, i = 0, npages = 0;
 426         dnode_t *dn = obj->oo_dn;
 427         arc_buf_t *abuf;
 428         uint32_t bs = dn->dn_datablksz;
 429         ENTRY;
 430
 431         /*
 432          * currently only full blocks are subject to zerocopy approach:
 433          * so that we're sure nobody is trying to update the same block
 434          */
 435         while (len > 0) {
 436                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 437
 438                 off_in_block = off & (bs - 1);
 439                 sz_in_block = min_t(int, bs - off_in_block, len);
 440
 441                 abuf = NULL;
 442                 if (sz_in_block == bs) {
 443                         /* full block, try to use zerocopy */
 444                         abuf = osd_request_arcbuf(dn, bs);
 445                         if (unlikely(IS_ERR(abuf)))
 446                                 GOTO(out_err, rc = PTR_ERR(abuf));
 447                 }
 448
 449                 if (abuf != NULL) {
 450                         atomic_inc(&osd->od_zerocopy_loan);
 451
 452                         /* go over pages arcbuf contains, put them as
 453                          * local niobufs for ptlrpc's bulks */
 454                         while (sz_in_block > 0) {
 455                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 456
 457                                 lnb[i].lnb_file_offset = off;
 458                                 lnb[i].lnb_page_offset = 0;
 459                                 lnb[i].lnb_len = plen;
 460                                 lnb[i].lnb_rc = 0;
 461                                 if (sz_in_block == bs)
 462                                         lnb[i].lnb_data = abuf;
 463                                 else
 464                                         lnb[i].lnb_data = NULL;
 465
 466                                 /* this one is not supposed to fail */
 467                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 468                                                         off_in_block);
 469                                 LASSERT(lnb[i].lnb_page);
 470
 471                                 lprocfs_counter_add(osd->od_stats,
 472                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 473
 474                                 sz_in_block -= plen;
 475                                 len -= plen;
 476                                 off += plen;
 477                                 off_in_block += plen;
 478                                 i++;
 479                                 npages++;
 480                         }
 481                 } else {
 482                         if (off_in_block == 0 && len < bs &&
 483                                         off + len >= obj->oo_attr.la_size)
 484                                 lprocfs_counter_add(osd->od_stats,
 485                                                 LPROC_OSD_TAIL_IO, 1);
 486
 487                         /* can't use zerocopy, allocate temp. buffers */
 488                         while (sz_in_block > 0) {
 489                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 490
 491                                 lnb[i].lnb_file_offset = off;
 492                                 lnb[i].lnb_page_offset = 0;
 493                                 lnb[i].lnb_len = plen;
 494                                 lnb[i].lnb_rc = 0;
 495                                 lnb[i].lnb_data = NULL;
 496
 497                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 498                                 if (unlikely(lnb[i].lnb_page == NULL))
 499                                         GOTO(out_err, rc = -ENOMEM);
 500
 501                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 502                                 lnb[i].lnb_page->mapping = (void *)obj;
 503
 504                                 atomic_inc(&osd->od_zerocopy_alloc);
 505                                 lprocfs_counter_add(osd->od_stats,
 506                                                 LPROC_OSD_COPY_IO, 1);
 507
 508                                 sz_in_block -= plen;
 509                                 len -= plen;
 510                                 off += plen;
 511                                 i++;
 512                                 npages++;
 513                         }
 514                 }
 515         }
 516
 517         RETURN(npages);
 518
 519 out_err:
 520         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 521         RETURN(rc);
 522 }
 523
 524 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 525                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 526                         enum dt_bufs_type rw)
 527 {
 528         struct osd_object *obj  = osd_dt_obj(dt);
 529         int                rc;
 530
 531         LASSERT(dt_object_exists(dt));
 532         LASSERT(obj->oo_dn);
 533
 534         if (rw & DT_BUFS_TYPE_WRITE)
 535                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 536         else
 537                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 538
 539         return rc;
 540 }
 541
 542 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 543                         struct niobuf_local *lnb, int npages)
 544 {
 545         struct osd_object *obj = osd_dt_obj(dt);
 546
 547         LASSERT(dt_object_exists(dt));
 548         LASSERT(obj->oo_dn);
 549
 550         return 0;
 551 }
 552
 553 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 554                                            uint64_t offset,
 555                                            uint32_t blksz)
 556 {
 557         LASSERT(blksz > 0);
 558
 559         size += offset % blksz;
 560
 561         if (likely(is_power_of_2(blksz)))
 562                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 563
 564         size += blksz - 1;
 565         do_div(size, blksz);
 566         return size * blksz;
 567 }
 568
 569 static int osd_declare_write_commit(const struct lu_env *env,
 570                                     struct dt_object *dt,
 571                                     struct niobuf_local *lnb, int npages,
 572                                     struct thandle *th)
 573 {
 574         struct osd_object  *obj = osd_dt_obj(dt);
 575         struct osd_device  *osd = osd_obj2dev(obj);
 576         struct osd_thandle *oh;
 577         uint64_t            offset = 0;
 578         uint32_t            size = 0;
 579         uint32_t blksz = obj->oo_dn->dn_datablksz;
 580         int                 i, rc, flags = 0;
 581         bool                ignore_quota = false, synced = false;
 582         long long           space = 0;
 583         struct page        *last_page = NULL;
 584         unsigned long       discont_pages = 0;
 585         ENTRY;
 586
 587         LASSERT(dt_object_exists(dt));
 588         LASSERT(obj->oo_dn);
 589
 590         LASSERT(lnb);
 591         LASSERT(npages > 0);
 592
 593         oh = container_of0(th, struct osd_thandle, ot_super);
 594
 595         for (i = 0; i < npages; i++) {
 596                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 597                         ++discont_pages;
 598                 last_page = lnb[i].lnb_page;
 599                 if (lnb[i].lnb_rc)
 600                         /* ENOSPC, network RPC error, etc.
 601                          * We don't want to book space for pages which will be
 602                          * skipped in osd_write_commit(). Hence we skip pages
 603                          * with lnb_rc != 0 here too */
 604                         continue;
 605                 /* ignore quota for the whole request if any page is from
 606                  * client cache or written by root.
 607                  *
 608                  * XXX once we drop the 1.8 client support, the checking
 609                  * for whether page is from cache can be simplified as:
 610                  * !(lnb[i].flags & OBD_BRW_SYNC)
 611                  *
 612                  * XXX we could handle this on per-lnb basis as done by
 613                  * grant. */
 614                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 615                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 616                     OBD_BRW_FROM_GRANT)
 617                         ignore_quota = true;
 618                 if (size == 0) {
 619                         /* first valid lnb */
 620                         offset = lnb[i].lnb_file_offset;
 621                         size = lnb[i].lnb_len;
 622                         continue;
 623                 }
 624                 if (offset + size == lnb[i].lnb_file_offset) {
 625                         /* this lnb is contiguous to the previous one */
 626                         size += lnb[i].lnb_len;
 627                         continue;
 628                 }
 629
 630                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 631                                   obj->oo_dn, offset, size);
 632                 /* Estimating space to be consumed by a write is rather
 633                  * complicated with ZFS. As a consequence, we don't account for
 634                  * indirect blocks and just use as a rough estimate the worse
 635                  * case where the old space is being held by a snapshot. Quota
 636                  * overrun will be adjusted once the operation is committed, if
 637                  * required. */
 638                 space += osd_roundup2blocksz(size, offset, blksz);
 639
 640                 offset = lnb[i].lnb_file_offset;
 641                 size = lnb[i].lnb_len;
 642         }
 643
 644         if (size) {
 645                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 646                                   offset, size);
 647                 space += osd_roundup2blocksz(size, offset, blksz);
 648         }
 649
 650         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 651
 652         /* backend zfs filesystem might be configured to store multiple data
 653          * copies */
 654         space  *= osd->od_os->os_copies;
 655         space   = toqb(space);
 656         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 657                npages, space);
 658
 659         record_start_io(osd, WRITE, discont_pages);
 660 retry:
 661         /* acquire quota space if needed */
 662         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 663                                obj->oo_attr.la_gid, space, oh, true, &flags,
 664                                ignore_quota);
 665
 666         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 667                 dt_sync(env, th->th_dev);
 668                 synced = true;
 669                 CDEBUG(D_QUOTA, "retry after sync\n");
 670                 flags = 0;
 671                 goto retry;
 672         }
 673
 674         /* we need only to store the overquota flags in the first lnb for
 675          * now, once we support multiple objects BRW, this code needs be
 676          * revised. */
 677         if (flags & QUOTA_FL_OVER_USRQUOTA)
 678                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 679         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 680                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 681
 682         RETURN(rc);
 683 }
 684
 685 /**
 686  * Policy to grow ZFS block size by write pattern.
 687  * For sequential write, it grows block size gradually until it reaches the
 688  * maximum blocksize the dataset can support. Otherwise, it will pick a
 689  * a block size by the writing region of this I/O.
 690  */
 691 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 692                               uint64_t start, uint64_t end)
 693 {
 694         struct osd_device       *osd = osd_obj2dev(obj);
 695         dnode_t *dn = obj->oo_dn;
 696         uint32_t                 blksz;
 697         int                      rc = 0;
 698
 699         ENTRY;
 700
 701         if (dn->dn_maxblkid > 0) /* can't change block size */
 702                 GOTO(out, rc);
 703
 704         if (dn->dn_datablksz >= osd->od_max_blksz)
 705                 GOTO(out, rc);
 706
 707         down_write(&obj->oo_guard);
 708
 709         blksz = dn->dn_datablksz;
 710         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 711                 GOTO(out_unlock, rc);
 712
 713         /* now ZFS can support up to 16MB block size, and if the write
 714          * is sequential, it just increases the block size gradually */
 715         if (start <= blksz) { /* sequential */
 716                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 717         } else { /* sparse, pick a block size by write region */
 718                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 719                                         end - start);
 720         }
 721
 722         if (!is_power_of_2(blksz))
 723                 blksz = size_roundup_power2(blksz);
 724
 725         if (blksz > dn->dn_datablksz) {
 726                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 727                                                blksz, 0, oh->ot_tx);
 728                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 729                 if (rc < 0)
 730                         CDEBUG(D_INODE, "object "DFID": change block size"
 731                                "%u -> %u error rc = %d\n",
 732                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 733                                dn->dn_datablksz, blksz, rc);
 734         }
 735         EXIT;
 736 out_unlock:
 737         up_write(&obj->oo_guard);
 738 out:
 739         return rc;
 740 }
 741
 742 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 743                         struct niobuf_local *lnb, int npages,
 744                         struct thandle *th)
 745 {
 746         struct osd_object  *obj  = osd_dt_obj(dt);
 747         struct osd_device  *osd = osd_obj2dev(obj);
 748         struct osd_thandle *oh;
 749         uint64_t            new_size = 0;
 750         int                 i, rc = 0;
 751         unsigned long      iosize = 0;
 752         ENTRY;
 753
 754         LASSERT(dt_object_exists(dt));
 755         LASSERT(obj->oo_dn);
 756
 757         LASSERT(th != NULL);
 758         oh = container_of0(th, struct osd_thandle, ot_super);
 759
 760         /* adjust block size. Assume the buffers are sorted. */
 761         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 762                                  lnb[npages - 1].lnb_file_offset +
 763                                  lnb[npages - 1].lnb_len);
 764
 765         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 766          * size and assigning arcbuf take place at the same time.
 767          *
 768          * Thread 1:
 769          * osd_write_commit()
 770          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 771          *   -> dmu_object_set_blocksize()
 772          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 773          *       write lock held
 774          *     -> dbuf_new_size()
 775          *      -> dmu_buf_will_dirty()
 776          *       -> dbuf_read()
 777          *        -> wait for the dbuf state to change
 778          * Thread 2:
 779          * osd_write_commit()
 780          *  -> dmu_assign_arcbuf()
 781          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 782          *    -> dbuf_dirty()
 783          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 784          *
 785          * By taking the read lock, it can avoid thread 2 to enter into the
 786          * critical section of assigning the arcbuf, while thread 1 is
 787          * changing the block size.
 788          */
 789         down_read(&obj->oo_guard);
 790         for (i = 0; i < npages; i++) {
 791                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 792                         (unsigned) lnb[i].lnb_len,
 793                         (unsigned) lnb[i].lnb_file_offset);
 794
 795                 if (lnb[i].lnb_rc) {
 796                         /* ENOSPC, network RPC error, etc.
 797                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 798                          * so we skip this page if lnb_rc is set to -ENOSPC */
 799                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 800                                 PFID(lu_object_fid(&dt->do_lu)), i,
 801                                 lnb[i].lnb_rc);
 802                         continue;
 803                 }
 804
 805                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 806                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 807                 if (lnb[i].lnb_page == NULL)
 808                         continue;
 809
 810                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 811                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 812                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 813                                       oh->ot_tx);
 814                         kunmap(lnb[i].lnb_page);
 815                         iosize += lnb[i].lnb_len;
 816                 } else if (lnb[i].lnb_data) {
 817                         int j, apages, abufsz;
 818                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 819                         /* buffer loaned for zerocopy, try to use it.
 820                          * notice that dmu_assign_arcbuf() is smart
 821                          * enough to recognize changed blocksize
 822                          * in this case it fallbacks to dmu_write() */
 823                         abufsz = arc_buf_size(lnb[i].lnb_data);
 824                         LASSERT(abufsz & PAGE_MASK);
 825                         apages = abufsz / PAGE_SIZE;
 826                         LASSERT(i + apages <= npages);
 827                         /* these references to pages must be invalidated
 828                          * to prevent access in osd_bufs_put() */
 829                         for (j = 0; j < apages; j++)
 830                                 lnb[i + j].lnb_page = NULL;
 831                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 832                                           lnb[i].lnb_file_offset,
 833                                           lnb[i].lnb_data, oh->ot_tx);
 834                         /* drop the reference, otherwise osd_put_bufs()
 835                          * will be releasing it - bad! */
 836                         lnb[i].lnb_data = NULL;
 837                         atomic_dec(&osd->od_zerocopy_loan);
 838                         iosize += abufsz;
 839                 }
 840
 841         }
 842         up_read(&obj->oo_guard);
 843
 844         if (unlikely(new_size == 0)) {
 845                 /* no pages to write, no transno is needed */
 846                 th->th_local = 1;
 847                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 848                  * since ofd_commitrw_write() retries several times on ENOSPC */
 849                 record_end_io(osd, WRITE, 0, 0, 0);
 850                 RETURN(0);
 851         }
 852
 853         write_lock(&obj->oo_attr_lock);
 854         if (obj->oo_attr.la_size < new_size) {
 855                 obj->oo_attr.la_size = new_size;
 856                 write_unlock(&obj->oo_attr_lock);
 857                 /* osd_object_sa_update() will be copying directly from
 858                  * oo_attr into dbuf. any update within a single txg will copy
 859                  * the most actual */
 860                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 861                                           &obj->oo_attr.la_size, 8, oh);
 862         } else {
 863                 write_unlock(&obj->oo_attr_lock);
 864         }
 865
 866         record_end_io(osd, WRITE, 0, iosize, npages);
 867
 868         RETURN(rc);
 869 }
 870
 871 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 872                         struct niobuf_local *lnb, int npages)
 873 {
 874         struct osd_object *obj  = osd_dt_obj(dt);
 875         int                i;
 876         loff_t             eof;
 877
 878         LASSERT(dt_object_exists(dt));
 879         LASSERT(obj->oo_dn);
 880
 881         read_lock(&obj->oo_attr_lock);
 882         eof = obj->oo_attr.la_size;
 883         read_unlock(&obj->oo_attr_lock);
 884
 885         for (i = 0; i < npages; i++) {
 886                 if (unlikely(lnb[i].lnb_rc < 0))
 887                         continue;
 888
 889                 lnb[i].lnb_rc = lnb[i].lnb_len;
 890
 891                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 892                         if (eof <= lnb[i].lnb_file_offset)
 893                                 lnb[i].lnb_rc = 0;
 894                         else
 895                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 896
 897                         /* all subsequent rc should be 0 */
 898                         while (++i < npages)
 899                                 lnb[i].lnb_rc = 0;
 900                         break;
 901                 }
 902         }
 903
 904         return 0;
 905 }
 906
 907 /*
 908  * Punch/truncate an object
 909  *
 910  *      IN:     db  - dmu_buf of the object to free data in.
 911  *              off - start of section to free.
 912  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 913  *
 914  *      RETURN: 0 if success
 915  *              error code if failure
 916  *
 917  * The transaction passed to this routine must have
 918  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 919  * called and then assigned to a transaction group.
 920  */
 921 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 922                                 uint64_t size, uint64_t off, uint64_t len)
 923 {
 924         int rc = 0;
 925
 926         /* Assert that the transaction has been assigned to a
 927            transaction group. */
 928         LASSERT(tx->tx_txg != 0);
 929         /*
 930          * Nothing to do if file already at desired length.
 931          */
 932         if (len == DMU_OBJECT_END && size == off)
 933                 return 0;
 934
 935         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 936         if (off < size)
 937                 dmu_free_range(os, dn->dn_object, off, len, tx);
 938
 939         return rc;
 940 }
 941
 942 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 943                         __u64 start, __u64 end, struct thandle *th)
 944 {
 945         struct osd_object  *obj = osd_dt_obj(dt);
 946         struct osd_device  *osd = osd_obj2dev(obj);
 947         struct osd_thandle *oh;
 948         __u64               len;
 949         int                 rc = 0;
 950         ENTRY;
 951
 952         LASSERT(dt_object_exists(dt));
 953         LASSERT(osd_invariant(obj));
 954
 955         LASSERT(th != NULL);
 956         oh = container_of0(th, struct osd_thandle, ot_super);
 957
 958         write_lock(&obj->oo_attr_lock);
 959         /* truncate */
 960         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 961                 len = DMU_OBJECT_END;
 962         else
 963                 len = end - start;
 964         write_unlock(&obj->oo_attr_lock);
 965
 966         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 967                                 obj->oo_attr.la_size, start, len);
 968         /* set new size */
 969         if (len == DMU_OBJECT_END) {
 970                 write_lock(&obj->oo_attr_lock);
 971                 obj->oo_attr.la_size = start;
 972                 write_unlock(&obj->oo_attr_lock);
 973                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 974                                           &obj->oo_attr.la_size, 8, oh);
 975         }
 976         RETURN(rc);
 977 }
 978
 979 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 980                         __u64 start, __u64 end, struct thandle *handle)
 981 {
 982         struct osd_object  *obj = osd_dt_obj(dt);
 983         struct osd_device  *osd = osd_obj2dev(obj);
 984         struct osd_thandle *oh;
 985         __u64               len;
 986         ENTRY;
 987
 988         oh = container_of0(handle, struct osd_thandle, ot_super);
 989
 990         read_lock(&obj->oo_attr_lock);
 991         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 992                 len = DMU_OBJECT_END;
 993         else
 994                 len = end - start;
 995
 996         /* declare we'll free some blocks ... */
 997         if (start < obj->oo_attr.la_size) {
 998                 read_unlock(&obj->oo_attr_lock);
 999                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1000         } else {
1001                 read_unlock(&obj->oo_attr_lock);
1002         }
1003
1004         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1005                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
1006                                  false));
1007 }
1008
1009 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1010                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1011 {
1012         int     rc;
1013         ENTRY;
1014
1015         switch (advice) {
1016         default:
1017                 rc = -ENOTSUPP;
1018                 break;
1019         }
1020
1021         RETURN(rc);
1022 }
1023
1024 struct dt_body_operations osd_body_ops = {
1025         .dbo_read                       = osd_read,
1026         .dbo_declare_write              = osd_declare_write,
1027         .dbo_write                      = osd_write,
1028         .dbo_bufs_get                   = osd_bufs_get,
1029         .dbo_bufs_put                   = osd_bufs_put,
1030         .dbo_write_prep                 = osd_write_prep,
1031         .dbo_declare_write_commit       = osd_declare_write_commit,
1032         .dbo_write_commit               = osd_write_commit,
1033         .dbo_read_prep                  = osd_read_prep,
1034         .dbo_declare_punch              = osd_declare_punch,
1035         .dbo_punch                      = osd_punch,
1036         .dbo_ladvise                    = osd_ladvise,
1037 };