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, 2017, 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 #include <lustre_quota.h>
  48
  49 #include "osd_internal.h"
  50
  51 #include <sys/dnode.h>
  52 #include <sys/dbuf.h>
  53 #include <sys/spa.h>
  54 #include <sys/stat.h>
  55 #include <sys/zap.h>
  56 #include <sys/spa_impl.h>
  57 #include <sys/zfs_znode.h>
  58 #include <sys/dmu_tx.h>
  59 #include <sys/dmu_objset.h>
  60 #include <sys/dsl_prop.h>
  61 #include <sys/sa_impl.h>
  62 #include <sys/txg.h>
  63
  64 static char *osd_0copy_tag = "zerocopy";
  65
  66
  67 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  68 {
  69         struct obd_histogram *h = osd->od_brw_stats.hist;
  70
  71         if (rw == READ) {
  72                 atomic_inc(&osd->od_r_in_flight);
  73                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  74                                  atomic_read(&osd->od_r_in_flight));
  75                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  76
  77         } else {
  78                 atomic_inc(&osd->od_w_in_flight);
  79                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  80                                  atomic_read(&osd->od_w_in_flight));
  81                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  82
  83         }
  84 }
  85
  86 static void record_end_io(struct osd_device *osd, int rw,
  87                           unsigned long elapsed, int disksize, int npages)
  88 {
  89         struct obd_histogram *h = osd->od_brw_stats.hist;
  90
  91         if (rw == READ) {
  92                 atomic_dec(&osd->od_r_in_flight);
  93                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  94                 if (disksize > 0)
  95                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
  96                 if (elapsed)
  97                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
  98
  99         } else {
 100                 atomic_dec(&osd->od_w_in_flight);
 101                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 102                 if (disksize > 0)
 103                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 104                 if (elapsed)
 105                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 106         }
 107 }
 108
 109 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 110                           struct lu_buf *buf, loff_t *pos, size_t *size)
 111 {
 112         struct osd_object *obj = osd_dt_obj(dt);
 113         uint64_t old_size;
 114         int rc;
 115
 116         LASSERT(dt_object_exists(dt));
 117         LASSERT(obj->oo_dn);
 118
 119         read_lock(&obj->oo_attr_lock);
 120         old_size = obj->oo_attr.la_size;
 121         read_unlock(&obj->oo_attr_lock);
 122
 123         if (*pos + *size > old_size) {
 124                 if (old_size < *pos)
 125                         return 0;
 126
 127                 *size = old_size - *pos;
 128         }
 129
 130         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 131                           buf->lb_buf, DMU_READ_PREFETCH);
 132         if (!rc) {
 133                 rc = *size;
 134                 *pos += *size;
 135         }
 136
 137         return rc;
 138 }
 139
 140 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 141                         struct lu_buf *buf, loff_t *pos)
 142 {
 143         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 144         size_t size = buf->lb_len;
 145         ktime_t start;
 146         s64 delta_ms;
 147         int rc;
 148
 149         start = ktime_get();
 150         record_start_io(osd, READ, 0);
 151         rc = __osd_read(env, dt, buf, pos, &size);
 152         delta_ms = ktime_ms_delta(ktime_get(), start);
 153         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 154
 155         return rc;
 156 }
 157
 158 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 159                                          struct dt_object *dt,
 160                                          struct lu_buf *buf, loff_t *pos)
 161 {
 162         size_t size = buf->lb_len;
 163
 164         return __osd_read(env, dt, buf, pos, &size);
 165 }
 166
 167 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 168                                 const struct lu_buf *buf, loff_t pos,
 169                                 struct thandle *th)
 170 {
 171         struct osd_object  *obj  = osd_dt_obj(dt);
 172         struct osd_device  *osd = osd_obj2dev(obj);
 173         struct osd_thandle *oh;
 174         uint64_t            oid;
 175         ENTRY;
 176
 177         oh = container_of0(th, struct osd_thandle, ot_super);
 178
 179         /* in some cases declare can race with creation (e.g. llog)
 180          * and we need to wait till object is initialized. notice
 181          * LOHA_EXISTs is supposed to be the last step in the
 182          * initialization */
 183
 184         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 185         if (dt_object_exists(dt))
 186                 oid = obj->oo_dn->dn_object;
 187         else
 188                 oid = DMU_NEW_OBJECT;
 189
 190         /* XXX: we still miss for append declaration support in ZFS
 191          *      -1 means append which is used by llog mostly, llog
 192          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 193         if (pos == -1)
 194                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 195                             obj->oo_attr.la_size + (2 << 20));
 196         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 197
 198         /* dt_declare_write() is usually called for system objects, such
 199          * as llog or last_rcvd files. We needn't enforce quota on those
 200          * objects, so always set the lqi_space as 0. */
 201         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 202                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 203                                  0, oh, NULL, OSD_QID_BLK));
 204 }
 205
 206 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 207                         const struct lu_buf *buf, loff_t *pos,
 208                         struct thandle *th, int ignore_quota)
 209 {
 210         struct osd_object  *obj  = osd_dt_obj(dt);
 211         struct osd_device  *osd = osd_obj2dev(obj);
 212         struct osd_thandle *oh;
 213         uint64_t            offset = *pos;
 214         int                 rc;
 215
 216         ENTRY;
 217
 218         LASSERT(dt_object_exists(dt));
 219         LASSERT(obj->oo_dn);
 220
 221         LASSERT(th != NULL);
 222         oh = container_of0(th, struct osd_thandle, ot_super);
 223
 224         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 225                       buf->lb_buf, oh->ot_tx);
 226         write_lock(&obj->oo_attr_lock);
 227         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 228                 obj->oo_attr.la_size = offset + buf->lb_len;
 229                 write_unlock(&obj->oo_attr_lock);
 230                 /* osd_object_sa_update() will be copying directly from oo_attr
 231                  * into dbuf.  any update within a single txg will copy the
 232                  * most actual */
 233                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 234                                         &obj->oo_attr.la_size, 8, oh);
 235                 if (unlikely(rc))
 236                         GOTO(out, rc);
 237         } else {
 238                 write_unlock(&obj->oo_attr_lock);
 239         }
 240
 241         *pos += buf->lb_len;
 242         rc = buf->lb_len;
 243
 244 out:
 245         RETURN(rc);
 246 }
 247
 248 /*
 249  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 250  *      purposes as it's not very well defined ...
 251  *      instead I use the lowest bit of the address so that:
 252  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 253  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 254  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 255  *
 256  *      bzzz, to blame
 257  */
 258 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 259                         struct niobuf_local *lnb, int npages)
 260 {
 261         struct osd_object *obj  = osd_dt_obj(dt);
 262         struct osd_device *osd = osd_obj2dev(obj);
 263         unsigned long      ptr;
 264         int                i;
 265
 266         LASSERT(dt_object_exists(dt));
 267         LASSERT(obj->oo_dn);
 268
 269         for (i = 0; i < npages; i++) {
 270                 if (lnb[i].lnb_page == NULL)
 271                         continue;
 272                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 273                         /* this is anonymous page allocated for copy-write */
 274                         lnb[i].lnb_page->mapping = NULL;
 275                         __free_page(lnb[i].lnb_page);
 276                         atomic_dec(&osd->od_zerocopy_alloc);
 277                 } else {
 278                         /* see comment in osd_bufs_get_read() */
 279                         ptr = (unsigned long)lnb[i].lnb_data;
 280                         if (ptr & 1UL) {
 281                                 ptr &= ~1UL;
 282                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 283                                 atomic_dec(&osd->od_zerocopy_pin);
 284                         } else if (lnb[i].lnb_data != NULL) {
 285                                 int j, apages, abufsz;
 286                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 287                                 apages = abufsz / PAGE_SIZE;
 288                                 /* these references to pages must be invalidated
 289                                  * to prevent access in osd_bufs_put() */
 290                                 for (j = 0; j < apages; j++)
 291                                         lnb[i + j].lnb_page = NULL;
 292                                 dmu_return_arcbuf(lnb[i].lnb_data);
 293                                 atomic_dec(&osd->od_zerocopy_loan);
 294                         }
 295                 }
 296                 lnb[i].lnb_page = NULL;
 297                 lnb[i].lnb_data = NULL;
 298         }
 299
 300         return 0;
 301 }
 302
 303 static inline struct page *kmem_to_page(void *addr)
 304 {
 305         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 306         if (is_vmalloc_addr(addr))
 307                 return vmalloc_to_page(addr);
 308         else
 309                 return virt_to_page(addr);
 310 }
 311
 312 /**
 313  * Prepare buffers for read.
 314  *
 315  * The function maps the range described by \a off and \a len to \a lnb array.
 316  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 317  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 318  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 319  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 320  * A better implementation would just return the buffers (potentially unfilled)
 321  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 322  *
 323  * \param[in] env       environment
 324  * \param[in] obj       object
 325  * \param[in] off       offset in bytes
 326  * \param[in] len       the number of bytes to access
 327  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 328  *
 329  * \retval              0 for success
 330  * \retval              negative error number of failure
 331  */
 332 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 333                              loff_t off, ssize_t len, struct niobuf_local *lnb)
 334 {
 335         struct osd_device *osd = osd_obj2dev(obj);
 336         int rc, i, numbufs, npages = 0;
 337         ktime_t start = ktime_get();
 338         dmu_buf_t **dbp;
 339         s64 delta_ms;
 340
 341         ENTRY;
 342         record_start_io(osd, READ, 0);
 343
 344         /* grab buffers for read:
 345          * OSD API let us to grab buffers first, then initiate IO(s)
 346          * so that all required IOs will be done in parallel, but at the
 347          * moment DMU doesn't provide us with a method to grab buffers.
 348          * If we discover this is a vital for good performance we
 349          * can get own replacement for dmu_buf_hold_array_by_bonus().
 350          */
 351         while (len > 0 &&
 352                (obj->oo_dn->dn_datablkshift != 0 ||
 353                 off < obj->oo_dn->dn_datablksz)) {
 354                 if (obj->oo_dn->dn_datablkshift == 0 &&
 355                     off + len > obj->oo_dn->dn_datablksz)
 356                         len = obj->oo_dn->dn_datablksz - off;
 357
 358                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 359                                                   off, len, TRUE, osd_0copy_tag,
 360                                                   &numbufs, &dbp);
 361                 if (unlikely(rc))
 362                         GOTO(err, rc);
 363
 364                 for (i = 0; i < numbufs; i++) {
 365                         int bufoff, tocpy, thispage;
 366                         void *dbf = dbp[i];
 367
 368                         LASSERT(len > 0);
 369
 370                         atomic_inc(&osd->od_zerocopy_pin);
 371
 372                         bufoff = off - dbp[i]->db_offset;
 373                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 374
 375                         /* kind of trick to differentiate dbuf vs. arcbuf */
 376                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 377                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 378
 379                         while (tocpy > 0) {
 380                                 thispage = PAGE_SIZE;
 381                                 thispage -= bufoff & (PAGE_SIZE - 1);
 382                                 thispage = min(tocpy, thispage);
 383
 384                                 lnb->lnb_rc = 0;
 385                                 lnb->lnb_file_offset = off;
 386                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 387                                 lnb->lnb_len = thispage;
 388                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 389                                                              bufoff);
 390                                 /* mark just a single slot: we need this
 391                                  * reference to dbuf to be released once */
 392                                 lnb->lnb_data = dbf;
 393                                 dbf = NULL;
 394
 395                                 tocpy -= thispage;
 396                                 len -= thispage;
 397                                 bufoff += thispage;
 398                                 off += thispage;
 399
 400                                 npages++;
 401                                 lnb++;
 402                         }
 403
 404                         /* steal dbuf so dmu_buf_rele_array() can't release
 405                          * it */
 406                         dbp[i] = NULL;
 407                 }
 408
 409                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 410         }
 411
 412         delta_ms = ktime_ms_delta(ktime_get(), start);
 413         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 414
 415         RETURN(npages);
 416
 417 err:
 418         LASSERT(rc < 0);
 419         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 420         RETURN(rc);
 421 }
 422
 423 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 424 {
 425         arc_buf_t *abuf;
 426
 427         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 428         if (unlikely(!abuf))
 429                 return ERR_PTR(-ENOMEM);
 430
 431 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 432         /**
 433          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 434          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 435          * setting up page array for RDMA transfer. See LU-9305.
 436          */
 437         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 438                 dmu_return_arcbuf(abuf);
 439                 return NULL;
 440         }
 441 #endif
 442
 443         return abuf;
 444 }
 445
 446 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 447                               loff_t off, ssize_t len, struct niobuf_local *lnb)
 448 {
 449         struct osd_device *osd = osd_obj2dev(obj);
 450         int                plen, off_in_block, sz_in_block;
 451         int                rc, i = 0, npages = 0;
 452         dnode_t *dn = obj->oo_dn;
 453         arc_buf_t *abuf;
 454         uint32_t bs = dn->dn_datablksz;
 455         ENTRY;
 456
 457         /*
 458          * currently only full blocks are subject to zerocopy approach:
 459          * so that we're sure nobody is trying to update the same block
 460          */
 461         while (len > 0) {
 462                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 463
 464                 off_in_block = off & (bs - 1);
 465                 sz_in_block = min_t(int, bs - off_in_block, len);
 466
 467                 abuf = NULL;
 468                 if (sz_in_block == bs) {
 469                         /* full block, try to use zerocopy */
 470                         abuf = osd_request_arcbuf(dn, bs);
 471                         if (unlikely(IS_ERR(abuf)))
 472                                 GOTO(out_err, rc = PTR_ERR(abuf));
 473                 }
 474
 475                 if (abuf != NULL) {
 476                         atomic_inc(&osd->od_zerocopy_loan);
 477
 478                         /* go over pages arcbuf contains, put them as
 479                          * local niobufs for ptlrpc's bulks */
 480                         while (sz_in_block > 0) {
 481                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 482
 483                                 lnb[i].lnb_file_offset = off;
 484                                 lnb[i].lnb_page_offset = 0;
 485                                 lnb[i].lnb_len = plen;
 486                                 lnb[i].lnb_rc = 0;
 487                                 if (sz_in_block == bs)
 488                                         lnb[i].lnb_data = abuf;
 489                                 else
 490                                         lnb[i].lnb_data = NULL;
 491
 492                                 /* this one is not supposed to fail */
 493                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 494                                                         off_in_block);
 495                                 LASSERT(lnb[i].lnb_page);
 496
 497                                 lprocfs_counter_add(osd->od_stats,
 498                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 499
 500                                 sz_in_block -= plen;
 501                                 len -= plen;
 502                                 off += plen;
 503                                 off_in_block += plen;
 504                                 i++;
 505                                 npages++;
 506                         }
 507                 } else {
 508                         if (off_in_block == 0 && len < bs &&
 509                                         off + len >= obj->oo_attr.la_size)
 510                                 lprocfs_counter_add(osd->od_stats,
 511                                                 LPROC_OSD_TAIL_IO, 1);
 512
 513                         /* can't use zerocopy, allocate temp. buffers */
 514                         while (sz_in_block > 0) {
 515                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 516
 517                                 lnb[i].lnb_file_offset = off;
 518                                 lnb[i].lnb_page_offset = 0;
 519                                 lnb[i].lnb_len = plen;
 520                                 lnb[i].lnb_rc = 0;
 521                                 lnb[i].lnb_data = NULL;
 522
 523                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 524                                 if (unlikely(lnb[i].lnb_page == NULL))
 525                                         GOTO(out_err, rc = -ENOMEM);
 526
 527                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 528                                 lnb[i].lnb_page->mapping = (void *)obj;
 529
 530                                 atomic_inc(&osd->od_zerocopy_alloc);
 531                                 lprocfs_counter_add(osd->od_stats,
 532                                                 LPROC_OSD_COPY_IO, 1);
 533
 534                                 sz_in_block -= plen;
 535                                 len -= plen;
 536                                 off += plen;
 537                                 i++;
 538                                 npages++;
 539                         }
 540                 }
 541         }
 542
 543         RETURN(npages);
 544
 545 out_err:
 546         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 547         RETURN(rc);
 548 }
 549
 550 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 551                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 552                         enum dt_bufs_type rw)
 553 {
 554         struct osd_object *obj  = osd_dt_obj(dt);
 555         int                rc;
 556
 557         LASSERT(dt_object_exists(dt));
 558         LASSERT(obj->oo_dn);
 559
 560         if (rw & DT_BUFS_TYPE_WRITE)
 561                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 562         else
 563                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 564
 565         return rc;
 566 }
 567
 568 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 569                         struct niobuf_local *lnb, int npages)
 570 {
 571         struct osd_object *obj = osd_dt_obj(dt);
 572
 573         LASSERT(dt_object_exists(dt));
 574         LASSERT(obj->oo_dn);
 575
 576         return 0;
 577 }
 578
 579 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 580                                            uint64_t offset,
 581                                            uint32_t blksz)
 582 {
 583         LASSERT(blksz > 0);
 584
 585         size += offset % blksz;
 586
 587         if (likely(is_power_of_2(blksz)))
 588                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 589
 590         size += blksz - 1;
 591         do_div(size, blksz);
 592         return size * blksz;
 593 }
 594
 595 static int osd_declare_write_commit(const struct lu_env *env,
 596                                     struct dt_object *dt,
 597                                     struct niobuf_local *lnb, int npages,
 598                                     struct thandle *th)
 599 {
 600         struct osd_object  *obj = osd_dt_obj(dt);
 601         struct osd_device  *osd = osd_obj2dev(obj);
 602         struct osd_thandle *oh;
 603         uint64_t            offset = 0;
 604         uint32_t            size = 0;
 605         uint32_t blksz = obj->oo_dn->dn_datablksz;
 606         int                 i, rc, flags = 0;
 607         bool synced = false;
 608         long long           space = 0;
 609         struct page        *last_page = NULL;
 610         unsigned long       discont_pages = 0;
 611         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 612         ENTRY;
 613
 614         LASSERT(dt_object_exists(dt));
 615         LASSERT(obj->oo_dn);
 616
 617         LASSERT(lnb);
 618         LASSERT(npages > 0);
 619
 620         oh = container_of0(th, struct osd_thandle, ot_super);
 621
 622         for (i = 0; i < npages; i++) {
 623                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 624                         ++discont_pages;
 625                 last_page = lnb[i].lnb_page;
 626                 if (lnb[i].lnb_rc)
 627                         /* ENOSPC, network RPC error, etc.
 628                          * We don't want to book space for pages which will be
 629                          * skipped in osd_write_commit(). Hence we skip pages
 630                          * with lnb_rc != 0 here too */
 631                         continue;
 632                 /* ignore quota for the whole request if any page is from
 633                  * client cache or written by root.
 634                  *
 635                  * XXX once we drop the 1.8 client support, the checking
 636                  * for whether page is from cache can be simplified as:
 637                  * !(lnb[i].flags & OBD_BRW_SYNC)
 638                  *
 639                  * XXX we could handle this on per-lnb basis as done by
 640                  * grant. */
 641                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 642                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 643                     OBD_BRW_FROM_GRANT)
 644                         declare_flags |= OSD_QID_FORCE;
 645
 646                 if (size == 0) {
 647                         /* first valid lnb */
 648                         offset = lnb[i].lnb_file_offset;
 649                         size = lnb[i].lnb_len;
 650                         continue;
 651                 }
 652                 if (offset + size == lnb[i].lnb_file_offset) {
 653                         /* this lnb is contiguous to the previous one */
 654                         size += lnb[i].lnb_len;
 655                         continue;
 656                 }
 657
 658                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 659                                   obj->oo_dn, offset, size);
 660                 /* Estimating space to be consumed by a write is rather
 661                  * complicated with ZFS. As a consequence, we don't account for
 662                  * indirect blocks and just use as a rough estimate the worse
 663                  * case where the old space is being held by a snapshot. Quota
 664                  * overrun will be adjusted once the operation is committed, if
 665                  * required. */
 666                 space += osd_roundup2blocksz(size, offset, blksz);
 667
 668                 offset = lnb[i].lnb_file_offset;
 669                 size = lnb[i].lnb_len;
 670         }
 671
 672         if (size) {
 673                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 674                                   offset, size);
 675                 space += osd_roundup2blocksz(size, offset, blksz);
 676         }
 677
 678         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 679
 680         /* backend zfs filesystem might be configured to store multiple data
 681          * copies */
 682         space  *= osd->od_os->os_copies;
 683         space   = toqb(space);
 684         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 685                npages, space);
 686
 687         record_start_io(osd, WRITE, discont_pages);
 688 retry:
 689         /* acquire quota space if needed */
 690         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 691                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 692                                space, oh, &flags, declare_flags);
 693
 694         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 695                 dt_sync(env, th->th_dev);
 696                 synced = true;
 697                 CDEBUG(D_QUOTA, "retry after sync\n");
 698                 flags = 0;
 699                 goto retry;
 700         }
 701
 702         /* we need only to store the overquota flags in the first lnb for
 703          * now, once we support multiple objects BRW, this code needs be
 704          * revised. */
 705         if (flags & QUOTA_FL_OVER_USRQUOTA)
 706                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 707         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 708                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 709 #ifdef ZFS_PROJINHERIT
 710         if (flags & QUOTA_FL_OVER_PRJQUOTA)
 711                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 712 #endif
 713
 714         RETURN(rc);
 715 }
 716
 717 /**
 718  * Policy to grow ZFS block size by write pattern.
 719  * For sequential write, it grows block size gradually until it reaches the
 720  * maximum blocksize the dataset can support. Otherwise, it will pick a
 721  * a block size by the writing region of this I/O.
 722  */
 723 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 724                               uint64_t start, uint64_t end)
 725 {
 726         struct osd_device       *osd = osd_obj2dev(obj);
 727         dnode_t *dn = obj->oo_dn;
 728         uint32_t                 blksz;
 729         int                      rc = 0;
 730
 731         ENTRY;
 732
 733         if (dn->dn_maxblkid > 0) /* can't change block size */
 734                 GOTO(out, rc);
 735
 736         if (dn->dn_datablksz >= osd->od_max_blksz)
 737                 GOTO(out, rc);
 738
 739         down_write(&obj->oo_guard);
 740
 741         blksz = dn->dn_datablksz;
 742         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 743                 GOTO(out_unlock, rc);
 744
 745         /* now ZFS can support up to 16MB block size, and if the write
 746          * is sequential, it just increases the block size gradually */
 747         if (start <= blksz) { /* sequential */
 748                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 749         } else { /* sparse, pick a block size by write region */
 750                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 751                                         end - start);
 752         }
 753
 754         if (!is_power_of_2(blksz))
 755                 blksz = size_roundup_power2(blksz);
 756
 757         if (blksz > dn->dn_datablksz) {
 758                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 759                                                blksz, 0, oh->ot_tx);
 760                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 761                 if (rc < 0)
 762                         CDEBUG(D_INODE, "object "DFID": change block size"
 763                                "%u -> %u error rc = %d\n",
 764                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 765                                dn->dn_datablksz, blksz, rc);
 766         }
 767         EXIT;
 768 out_unlock:
 769         up_write(&obj->oo_guard);
 770 out:
 771         return rc;
 772 }
 773
 774 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 775                         struct niobuf_local *lnb, int npages,
 776                         struct thandle *th)
 777 {
 778         struct osd_object  *obj  = osd_dt_obj(dt);
 779         struct osd_device  *osd = osd_obj2dev(obj);
 780         struct osd_thandle *oh;
 781         uint64_t            new_size = 0;
 782         int                 i, rc = 0;
 783         unsigned long      iosize = 0;
 784         ENTRY;
 785
 786         LASSERT(dt_object_exists(dt));
 787         LASSERT(obj->oo_dn);
 788
 789         LASSERT(th != NULL);
 790         oh = container_of0(th, struct osd_thandle, ot_super);
 791
 792         /* adjust block size. Assume the buffers are sorted. */
 793         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 794                                  lnb[npages - 1].lnb_file_offset +
 795                                  lnb[npages - 1].lnb_len);
 796
 797         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 798          * size and assigning arcbuf take place at the same time.
 799          *
 800          * Thread 1:
 801          * osd_write_commit()
 802          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 803          *   -> dmu_object_set_blocksize()
 804          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 805          *       write lock held
 806          *     -> dbuf_new_size()
 807          *      -> dmu_buf_will_dirty()
 808          *       -> dbuf_read()
 809          *        -> wait for the dbuf state to change
 810          * Thread 2:
 811          * osd_write_commit()
 812          *  -> dmu_assign_arcbuf()
 813          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 814          *    -> dbuf_dirty()
 815          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 816          *
 817          * By taking the read lock, it can avoid thread 2 to enter into the
 818          * critical section of assigning the arcbuf, while thread 1 is
 819          * changing the block size.
 820          */
 821         down_read(&obj->oo_guard);
 822         for (i = 0; i < npages; i++) {
 823                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 824                         (unsigned) lnb[i].lnb_len,
 825                         (unsigned) lnb[i].lnb_file_offset);
 826
 827                 if (lnb[i].lnb_rc) {
 828                         /* ENOSPC, network RPC error, etc.
 829                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 830                          * so we skip this page if lnb_rc is set to -ENOSPC */
 831                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 832                                 PFID(lu_object_fid(&dt->do_lu)), i,
 833                                 lnb[i].lnb_rc);
 834                         continue;
 835                 }
 836
 837                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 838                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 839                 if (lnb[i].lnb_page == NULL)
 840                         continue;
 841
 842                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 843                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 844                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 845                                       oh->ot_tx);
 846                         kunmap(lnb[i].lnb_page);
 847                         iosize += lnb[i].lnb_len;
 848                 } else if (lnb[i].lnb_data) {
 849                         int j, apages, abufsz;
 850                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 851                         /* buffer loaned for zerocopy, try to use it.
 852                          * notice that dmu_assign_arcbuf() is smart
 853                          * enough to recognize changed blocksize
 854                          * in this case it fallbacks to dmu_write() */
 855                         abufsz = arc_buf_size(lnb[i].lnb_data);
 856                         LASSERT(abufsz & PAGE_MASK);
 857                         apages = abufsz / PAGE_SIZE;
 858                         LASSERT(i + apages <= npages);
 859                         /* these references to pages must be invalidated
 860                          * to prevent access in osd_bufs_put() */
 861                         for (j = 0; j < apages; j++)
 862                                 lnb[i + j].lnb_page = NULL;
 863                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 864                                           lnb[i].lnb_file_offset,
 865                                           lnb[i].lnb_data, oh->ot_tx);
 866                         /* drop the reference, otherwise osd_put_bufs()
 867                          * will be releasing it - bad! */
 868                         lnb[i].lnb_data = NULL;
 869                         atomic_dec(&osd->od_zerocopy_loan);
 870                         iosize += abufsz;
 871                 }
 872
 873         }
 874         up_read(&obj->oo_guard);
 875
 876         if (unlikely(new_size == 0)) {
 877                 /* no pages to write, no transno is needed */
 878                 th->th_local = 1;
 879                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 880                  * since ofd_commitrw_write() retries several times on ENOSPC */
 881                 record_end_io(osd, WRITE, 0, 0, 0);
 882                 RETURN(0);
 883         }
 884
 885         write_lock(&obj->oo_attr_lock);
 886         if (obj->oo_attr.la_size < new_size) {
 887                 obj->oo_attr.la_size = new_size;
 888                 write_unlock(&obj->oo_attr_lock);
 889                 /* osd_object_sa_update() will be copying directly from
 890                  * oo_attr into dbuf. any update within a single txg will copy
 891                  * the most actual */
 892                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 893                                           &obj->oo_attr.la_size, 8, oh);
 894         } else {
 895                 write_unlock(&obj->oo_attr_lock);
 896         }
 897
 898         record_end_io(osd, WRITE, 0, iosize, npages);
 899
 900         RETURN(rc);
 901 }
 902
 903 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 904                         struct niobuf_local *lnb, int npages)
 905 {
 906         struct osd_object *obj  = osd_dt_obj(dt);
 907         int                i;
 908         loff_t             eof;
 909
 910         LASSERT(dt_object_exists(dt));
 911         LASSERT(obj->oo_dn);
 912
 913         read_lock(&obj->oo_attr_lock);
 914         eof = obj->oo_attr.la_size;
 915         read_unlock(&obj->oo_attr_lock);
 916
 917         for (i = 0; i < npages; i++) {
 918                 if (unlikely(lnb[i].lnb_rc < 0))
 919                         continue;
 920
 921                 lnb[i].lnb_rc = lnb[i].lnb_len;
 922
 923                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 924                         if (eof <= lnb[i].lnb_file_offset)
 925                                 lnb[i].lnb_rc = 0;
 926                         else
 927                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 928
 929                         /* all subsequent rc should be 0 */
 930                         while (++i < npages)
 931                                 lnb[i].lnb_rc = 0;
 932                         break;
 933                 }
 934         }
 935
 936         return 0;
 937 }
 938
 939 /*
 940  * Punch/truncate an object
 941  *
 942  *      IN:     db  - dmu_buf of the object to free data in.
 943  *              off - start of section to free.
 944  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 945  *
 946  *      RETURN: 0 if success
 947  *              error code if failure
 948  *
 949  * The transaction passed to this routine must have
 950  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 951  * called and then assigned to a transaction group.
 952  */
 953 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
 954                                 uint64_t size, uint64_t off, uint64_t len)
 955 {
 956         int rc = 0;
 957
 958         /* Assert that the transaction has been assigned to a
 959            transaction group. */
 960         LASSERT(tx->tx_txg != 0);
 961         /*
 962          * Nothing to do if file already at desired length.
 963          */
 964         if (len == DMU_OBJECT_END && size == off)
 965                 return 0;
 966
 967         /* XXX: dnode_free_range() can be used to save on dnode lookup */
 968         if (off < size)
 969                 dmu_free_range(os, dn->dn_object, off, len, tx);
 970
 971         return rc;
 972 }
 973
 974 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 975                         __u64 start, __u64 end, struct thandle *th)
 976 {
 977         struct osd_object  *obj = osd_dt_obj(dt);
 978         struct osd_device  *osd = osd_obj2dev(obj);
 979         struct osd_thandle *oh;
 980         __u64               len;
 981         int                 rc = 0;
 982         ENTRY;
 983
 984         LASSERT(dt_object_exists(dt));
 985         LASSERT(osd_invariant(obj));
 986
 987         LASSERT(th != NULL);
 988         oh = container_of0(th, struct osd_thandle, ot_super);
 989
 990         write_lock(&obj->oo_attr_lock);
 991         /* truncate */
 992         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 993                 len = DMU_OBJECT_END;
 994         else
 995                 len = end - start;
 996         write_unlock(&obj->oo_attr_lock);
 997
 998         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
 999                                 obj->oo_attr.la_size, start, len);
1000         /* set new size */
1001         if (len == DMU_OBJECT_END) {
1002                 write_lock(&obj->oo_attr_lock);
1003                 obj->oo_attr.la_size = start;
1004                 write_unlock(&obj->oo_attr_lock);
1005                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1006                                           &obj->oo_attr.la_size, 8, oh);
1007         }
1008         RETURN(rc);
1009 }
1010
1011 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1012                         __u64 start, __u64 end, struct thandle *handle)
1013 {
1014         struct osd_object  *obj = osd_dt_obj(dt);
1015         struct osd_device  *osd = osd_obj2dev(obj);
1016         struct osd_thandle *oh;
1017         __u64               len;
1018         ENTRY;
1019
1020         oh = container_of0(handle, struct osd_thandle, ot_super);
1021
1022         read_lock(&obj->oo_attr_lock);
1023         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1024                 len = DMU_OBJECT_END;
1025         else
1026                 len = end - start;
1027
1028         /* declare we'll free some blocks ... */
1029         if (start < obj->oo_attr.la_size) {
1030                 read_unlock(&obj->oo_attr_lock);
1031                 dmu_tx_mark_netfree(oh->ot_tx);
1032                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1033         } else {
1034                 read_unlock(&obj->oo_attr_lock);
1035         }
1036
1037         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1038                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1039                                  0, oh, NULL, OSD_QID_BLK));
1040 }
1041
1042 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1043                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1044 {
1045         int     rc;
1046         ENTRY;
1047
1048         switch (advice) {
1049         default:
1050                 rc = -ENOTSUPP;
1051                 break;
1052         }
1053
1054         RETURN(rc);
1055 }
1056
1057 struct dt_body_operations osd_body_ops = {
1058         .dbo_read                       = osd_read,
1059         .dbo_declare_write              = osd_declare_write,
1060         .dbo_write                      = osd_write,
1061         .dbo_bufs_get                   = osd_bufs_get,
1062         .dbo_bufs_put                   = osd_bufs_put,
1063         .dbo_write_prep                 = osd_write_prep,
1064         .dbo_declare_write_commit       = osd_declare_write_commit,
1065         .dbo_write_commit               = osd_write_commit,
1066         .dbo_read_prep                  = osd_read_prep,
1067         .dbo_declare_punch              = osd_declare_punch,
1068         .dbo_punch                      = osd_punch,
1069         .dbo_ladvise                    = osd_ladvise,
1070 };
1071
1072 struct dt_body_operations osd_body_scrub_ops = {
1073         .dbo_read                       = osd_read_no_record,
1074         .dbo_declare_write              = osd_declare_write,
1075         .dbo_write                      = osd_write,
1076 };