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 dbuf_set_pending_evict(dmu_buf_t *db)
  68 {
  69         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
  70         dbi->db_pending_evict = TRUE;
  71 }
  72
  73 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  74 {
  75         struct obd_histogram *h = osd->od_brw_stats.hist;
  76
  77         if (rw == READ) {
  78                 atomic_inc(&osd->od_r_in_flight);
  79                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  80                                  atomic_read(&osd->od_r_in_flight));
  81                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  82
  83         } else {
  84                 atomic_inc(&osd->od_w_in_flight);
  85                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  86                                  atomic_read(&osd->od_w_in_flight));
  87                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  88
  89         }
  90 }
  91
  92 static void record_end_io(struct osd_device *osd, int rw,
  93                           unsigned long elapsed, int disksize, int npages)
  94 {
  95         struct obd_histogram *h = osd->od_brw_stats.hist;
  96
  97         if (rw == READ) {
  98                 atomic_dec(&osd->od_r_in_flight);
  99                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
 100                 if (disksize > 0)
 101                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
 102                 if (elapsed)
 103                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
 104
 105         } else {
 106                 atomic_dec(&osd->od_w_in_flight);
 107                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 108                 if (disksize > 0)
 109                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 110                 if (elapsed)
 111                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 112         }
 113 }
 114
 115 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 116                           struct lu_buf *buf, loff_t *pos, size_t *size)
 117 {
 118         struct osd_object *obj = osd_dt_obj(dt);
 119         uint64_t old_size;
 120         int rc;
 121
 122         LASSERT(dt_object_exists(dt));
 123         LASSERT(obj->oo_dn);
 124
 125         read_lock(&obj->oo_attr_lock);
 126         old_size = obj->oo_attr.la_size;
 127         read_unlock(&obj->oo_attr_lock);
 128
 129         if (*pos + *size > old_size) {
 130                 if (old_size < *pos)
 131                         return 0;
 132
 133                 *size = old_size - *pos;
 134         }
 135
 136         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 137                           buf->lb_buf, DMU_READ_PREFETCH);
 138         if (!rc) {
 139                 rc = *size;
 140                 *pos += *size;
 141         }
 142
 143         return rc;
 144 }
 145
 146 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 147                         struct lu_buf *buf, loff_t *pos)
 148 {
 149         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 150         size_t size = buf->lb_len;
 151         ktime_t start;
 152         s64 delta_ms;
 153         int rc;
 154
 155         start = ktime_get();
 156         record_start_io(osd, READ, 0);
 157         rc = __osd_read(env, dt, buf, pos, &size);
 158         delta_ms = ktime_ms_delta(ktime_get(), start);
 159         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 160
 161         return rc;
 162 }
 163
 164 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 165                                          struct dt_object *dt,
 166                                          struct lu_buf *buf, loff_t *pos)
 167 {
 168         size_t size = buf->lb_len;
 169
 170         return __osd_read(env, dt, buf, pos, &size);
 171 }
 172
 173 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 174                                 const struct lu_buf *buf, loff_t pos,
 175                                 struct thandle *th)
 176 {
 177         struct osd_object  *obj  = osd_dt_obj(dt);
 178         struct osd_device  *osd = osd_obj2dev(obj);
 179         struct osd_thandle *oh;
 180         uint64_t            oid;
 181         ENTRY;
 182
 183         oh = container_of0(th, struct osd_thandle, ot_super);
 184
 185         /* in some cases declare can race with creation (e.g. llog)
 186          * and we need to wait till object is initialized. notice
 187          * LOHA_EXISTs is supposed to be the last step in the
 188          * initialization */
 189
 190         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 191         if (dt_object_exists(dt))
 192                 oid = obj->oo_dn->dn_object;
 193         else
 194                 oid = DMU_NEW_OBJECT;
 195
 196         /* XXX: we still miss for append declaration support in ZFS
 197          *      -1 means append which is used by llog mostly, llog
 198          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 199         if (pos == -1)
 200                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 201                             obj->oo_attr.la_size + (2 << 20));
 202         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 203
 204         /* dt_declare_write() is usually called for system objects, such
 205          * as llog or last_rcvd files. We needn't enforce quota on those
 206          * objects, so always set the lqi_space as 0. */
 207         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 208                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 209                                  0, oh, NULL, OSD_QID_BLK));
 210 }
 211
 212 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 213                         const struct lu_buf *buf, loff_t *pos,
 214                          struct thandle *th)
 215 {
 216         struct osd_object  *obj  = osd_dt_obj(dt);
 217         struct osd_device  *osd = osd_obj2dev(obj);
 218         struct osd_thandle *oh;
 219         uint64_t            offset = *pos;
 220         int                 rc;
 221
 222         ENTRY;
 223
 224         LASSERT(dt_object_exists(dt));
 225         LASSERT(obj->oo_dn);
 226
 227         LASSERT(th != NULL);
 228         oh = container_of0(th, struct osd_thandle, ot_super);
 229
 230         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 231                       buf->lb_buf, oh->ot_tx);
 232         write_lock(&obj->oo_attr_lock);
 233         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 234                 obj->oo_attr.la_size = offset + buf->lb_len;
 235                 write_unlock(&obj->oo_attr_lock);
 236                 /* osd_object_sa_update() will be copying directly from oo_attr
 237                  * into dbuf.  any update within a single txg will copy the
 238                  * most actual */
 239                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 240                                         &obj->oo_attr.la_size, 8, oh);
 241                 if (unlikely(rc))
 242                         GOTO(out, rc);
 243         } else {
 244                 write_unlock(&obj->oo_attr_lock);
 245         }
 246
 247         *pos += buf->lb_len;
 248         rc = buf->lb_len;
 249
 250 out:
 251         RETURN(rc);
 252 }
 253
 254 /*
 255  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 256  *      purposes as it's not very well defined ...
 257  *      instead I use the lowest bit of the address so that:
 258  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 259  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 260  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 261  *
 262  *      bzzz, to blame
 263  */
 264 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 265                         struct niobuf_local *lnb, int npages)
 266 {
 267         struct osd_object *obj  = osd_dt_obj(dt);
 268         struct osd_device *osd = osd_obj2dev(obj);
 269         unsigned long      ptr;
 270         int                i;
 271
 272         LASSERT(dt_object_exists(dt));
 273         LASSERT(obj->oo_dn);
 274
 275         for (i = 0; i < npages; i++) {
 276                 if (lnb[i].lnb_page == NULL)
 277                         continue;
 278                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 279                         /* this is anonymous page allocated for copy-write */
 280                         lnb[i].lnb_page->mapping = NULL;
 281                         __free_page(lnb[i].lnb_page);
 282                         atomic_dec(&osd->od_zerocopy_alloc);
 283                 } else {
 284                         /* see comment in osd_bufs_get_read() */
 285                         ptr = (unsigned long)lnb[i].lnb_data;
 286                         if (ptr & 1UL) {
 287                                 ptr &= ~1UL;
 288                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 289                                 atomic_dec(&osd->od_zerocopy_pin);
 290                         } else if (lnb[i].lnb_data != NULL) {
 291                                 int j, apages, abufsz;
 292                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 293                                 apages = abufsz / PAGE_SIZE;
 294                                 /* these references to pages must be invalidated
 295                                  * to prevent access in osd_bufs_put() */
 296                                 for (j = 0; j < apages; j++)
 297                                         lnb[i + j].lnb_page = NULL;
 298                                 dmu_return_arcbuf(lnb[i].lnb_data);
 299                                 atomic_dec(&osd->od_zerocopy_loan);
 300                         }
 301                 }
 302                 lnb[i].lnb_page = NULL;
 303                 lnb[i].lnb_data = NULL;
 304         }
 305
 306         return 0;
 307 }
 308
 309 static inline struct page *kmem_to_page(void *addr)
 310 {
 311         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 312         if (is_vmalloc_addr(addr))
 313                 return vmalloc_to_page(addr);
 314         else
 315                 return virt_to_page(addr);
 316 }
 317
 318 /**
 319  * Prepare buffers for read.
 320  *
 321  * The function maps the range described by \a off and \a len to \a lnb array.
 322  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 323  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 324  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 325  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 326  * A better implementation would just return the buffers (potentially unfilled)
 327  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 328  *
 329  * \param[in] env       environment
 330  * \param[in] obj       object
 331  * \param[in] off       offset in bytes
 332  * \param[in] len       the number of bytes to access
 333  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 334  *
 335  * \retval              0 for success
 336  * \retval              negative error number of failure
 337  */
 338 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 339                              loff_t off, ssize_t len, struct niobuf_local *lnb)
 340 {
 341         struct osd_device *osd = osd_obj2dev(obj);
 342         int rc, i, numbufs, npages = 0, drop_cache = 0;
 343         ktime_t start = ktime_get();
 344         dmu_buf_t **dbp;
 345         s64 delta_ms;
 346
 347         ENTRY;
 348         record_start_io(osd, READ, 0);
 349
 350         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
 351                 drop_cache = 1;
 352
 353         /* grab buffers for read:
 354          * OSD API let us to grab buffers first, then initiate IO(s)
 355          * so that all required IOs will be done in parallel, but at the
 356          * moment DMU doesn't provide us with a method to grab buffers.
 357          * If we discover this is a vital for good performance we
 358          * can get own replacement for dmu_buf_hold_array_by_bonus().
 359          */
 360         while (len > 0 &&
 361                (obj->oo_dn->dn_datablkshift != 0 ||
 362                 off < obj->oo_dn->dn_datablksz)) {
 363                 if (obj->oo_dn->dn_datablkshift == 0 &&
 364                     off + len > obj->oo_dn->dn_datablksz)
 365                         len = obj->oo_dn->dn_datablksz - off;
 366
 367                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 368                                                   off, len, TRUE, osd_0copy_tag,
 369                                                   &numbufs, &dbp);
 370                 if (unlikely(rc))
 371                         GOTO(err, rc);
 372
 373                 for (i = 0; i < numbufs; i++) {
 374                         int bufoff, tocpy, thispage;
 375                         void *dbf = dbp[i];
 376
 377                         LASSERT(len > 0);
 378
 379                         atomic_inc(&osd->od_zerocopy_pin);
 380
 381                         bufoff = off - dbp[i]->db_offset;
 382                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 383
 384                         /* kind of trick to differentiate dbuf vs. arcbuf */
 385                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 386                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 387
 388                         while (tocpy > 0) {
 389                                 thispage = PAGE_SIZE;
 390                                 thispage -= bufoff & (PAGE_SIZE - 1);
 391                                 thispage = min(tocpy, thispage);
 392
 393                                 lnb->lnb_rc = 0;
 394                                 lnb->lnb_file_offset = off;
 395                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 396                                 lnb->lnb_len = thispage;
 397                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 398                                                              bufoff);
 399                                 /* mark just a single slot: we need this
 400                                  * reference to dbuf to be released once */
 401                                 lnb->lnb_data = dbf;
 402                                 dbf = NULL;
 403
 404                                 tocpy -= thispage;
 405                                 len -= thispage;
 406                                 bufoff += thispage;
 407                                 off += thispage;
 408
 409                                 npages++;
 410                                 lnb++;
 411                         }
 412
 413                         if (drop_cache)
 414                                 dbuf_set_pending_evict(dbp[i]);
 415
 416                         /* steal dbuf so dmu_buf_rele_array() can't release
 417                          * it */
 418                         dbp[i] = NULL;
 419                 }
 420
 421                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 422         }
 423
 424         delta_ms = ktime_ms_delta(ktime_get(), start);
 425         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 426
 427         RETURN(npages);
 428
 429 err:
 430         LASSERT(rc < 0);
 431         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 432         RETURN(rc);
 433 }
 434
 435 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 436 {
 437         arc_buf_t *abuf;
 438
 439         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 440         if (unlikely(!abuf))
 441                 return ERR_PTR(-ENOMEM);
 442
 443 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 444         /**
 445          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 446          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 447          * setting up page array for RDMA transfer. See LU-9305.
 448          */
 449         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 450                 dmu_return_arcbuf(abuf);
 451                 return NULL;
 452         }
 453 #endif
 454
 455         return abuf;
 456 }
 457
 458 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 459                               loff_t off, ssize_t len, struct niobuf_local *lnb)
 460 {
 461         struct osd_device *osd = osd_obj2dev(obj);
 462         int                poff, plen, off_in_block, sz_in_block;
 463         int                rc, i = 0, npages = 0;
 464         dnode_t *dn = obj->oo_dn;
 465         arc_buf_t *abuf;
 466         uint32_t bs = dn->dn_datablksz;
 467         ENTRY;
 468
 469         /*
 470          * currently only full blocks are subject to zerocopy approach:
 471          * so that we're sure nobody is trying to update the same block
 472          */
 473         while (len > 0) {
 474                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 475
 476                 off_in_block = off & (bs - 1);
 477                 sz_in_block = min_t(int, bs - off_in_block, len);
 478
 479                 abuf = NULL;
 480                 if (sz_in_block == bs) {
 481                         /* full block, try to use zerocopy */
 482                         abuf = osd_request_arcbuf(dn, bs);
 483                         if (unlikely(IS_ERR(abuf)))
 484                                 GOTO(out_err, rc = PTR_ERR(abuf));
 485                 }
 486
 487                 if (abuf != NULL) {
 488                         atomic_inc(&osd->od_zerocopy_loan);
 489
 490                         /* go over pages arcbuf contains, put them as
 491                          * local niobufs for ptlrpc's bulks */
 492                         while (sz_in_block > 0) {
 493                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 494
 495                                 lnb[i].lnb_file_offset = off;
 496                                 lnb[i].lnb_page_offset = 0;
 497                                 lnb[i].lnb_len = plen;
 498                                 lnb[i].lnb_rc = 0;
 499                                 if (sz_in_block == bs)
 500                                         lnb[i].lnb_data = abuf;
 501                                 else
 502                                         lnb[i].lnb_data = NULL;
 503
 504                                 /* this one is not supposed to fail */
 505                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 506                                                         off_in_block);
 507                                 LASSERT(lnb[i].lnb_page);
 508
 509                                 lprocfs_counter_add(osd->od_stats,
 510                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 511
 512                                 sz_in_block -= plen;
 513                                 len -= plen;
 514                                 off += plen;
 515                                 off_in_block += plen;
 516                                 i++;
 517                                 npages++;
 518                         }
 519                 } else {
 520                         if (off_in_block == 0 && len < bs &&
 521                                         off + len >= obj->oo_attr.la_size)
 522                                 lprocfs_counter_add(osd->od_stats,
 523                                                 LPROC_OSD_TAIL_IO, 1);
 524
 525                         /* can't use zerocopy, allocate temp. buffers */
 526                         poff = off & (PAGE_SIZE - 1);
 527                         while (sz_in_block > 0) {
 528                                 plen = min_t(int, poff + sz_in_block,
 529                                              PAGE_SIZE);
 530                                 plen -= poff;
 531
 532                                 lnb[i].lnb_file_offset = off;
 533                                 lnb[i].lnb_page_offset = poff;
 534                                 poff = 0;
 535
 536                                 lnb[i].lnb_len = plen;
 537                                 lnb[i].lnb_rc = 0;
 538                                 lnb[i].lnb_data = NULL;
 539
 540                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 541                                 if (unlikely(lnb[i].lnb_page == NULL))
 542                                         GOTO(out_err, rc = -ENOMEM);
 543
 544                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 545                                 lnb[i].lnb_page->mapping = (void *)obj;
 546
 547                                 atomic_inc(&osd->od_zerocopy_alloc);
 548                                 lprocfs_counter_add(osd->od_stats,
 549                                                 LPROC_OSD_COPY_IO, 1);
 550
 551                                 sz_in_block -= plen;
 552                                 len -= plen;
 553                                 off += plen;
 554                                 i++;
 555                                 npages++;
 556                         }
 557                 }
 558         }
 559
 560         RETURN(npages);
 561
 562 out_err:
 563         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 564         RETURN(rc);
 565 }
 566
 567 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 568                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 569                         enum dt_bufs_type rw)
 570 {
 571         struct osd_object *obj  = osd_dt_obj(dt);
 572         int                rc;
 573
 574         LASSERT(dt_object_exists(dt));
 575         LASSERT(obj->oo_dn);
 576
 577         if (rw & DT_BUFS_TYPE_WRITE)
 578                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 579         else
 580                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 581
 582         return rc;
 583 }
 584
 585 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 586                         struct niobuf_local *lnb, int npages)
 587 {
 588         struct osd_object *obj = osd_dt_obj(dt);
 589
 590         LASSERT(dt_object_exists(dt));
 591         LASSERT(obj->oo_dn);
 592
 593         return 0;
 594 }
 595
 596 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 597                                            uint64_t offset,
 598                                            uint32_t blksz)
 599 {
 600         LASSERT(blksz > 0);
 601
 602         size += offset % blksz;
 603
 604         if (likely(is_power_of_2(blksz)))
 605                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 606
 607         size += blksz - 1;
 608         do_div(size, blksz);
 609         return size * blksz;
 610 }
 611
 612 static int osd_declare_write_commit(const struct lu_env *env,
 613                                     struct dt_object *dt,
 614                                     struct niobuf_local *lnb, int npages,
 615                                     struct thandle *th)
 616 {
 617         struct osd_object  *obj = osd_dt_obj(dt);
 618         struct osd_device  *osd = osd_obj2dev(obj);
 619         struct osd_thandle *oh;
 620         uint64_t            offset = 0;
 621         uint32_t            size = 0;
 622         uint32_t blksz = obj->oo_dn->dn_datablksz;
 623         int                 i, rc, flags = 0;
 624         bool synced = false;
 625         long long           space = 0;
 626         struct page        *last_page = NULL;
 627         unsigned long       discont_pages = 0;
 628         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 629         ENTRY;
 630
 631         LASSERT(dt_object_exists(dt));
 632         LASSERT(obj->oo_dn);
 633
 634         LASSERT(lnb);
 635         LASSERT(npages > 0);
 636
 637         oh = container_of0(th, struct osd_thandle, ot_super);
 638
 639         for (i = 0; i < npages; i++) {
 640                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 641                         ++discont_pages;
 642                 last_page = lnb[i].lnb_page;
 643                 if (lnb[i].lnb_rc)
 644                         /* ENOSPC, network RPC error, etc.
 645                          * We don't want to book space for pages which will be
 646                          * skipped in osd_write_commit(). Hence we skip pages
 647                          * with lnb_rc != 0 here too */
 648                         continue;
 649                 /* ignore quota for the whole request if any page is from
 650                  * client cache or written by root.
 651                  *
 652                  * XXX once we drop the 1.8 client support, the checking
 653                  * for whether page is from cache can be simplified as:
 654                  * !(lnb[i].flags & OBD_BRW_SYNC)
 655                  *
 656                  * XXX we could handle this on per-lnb basis as done by
 657                  * grant. */
 658                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 659                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 660                     OBD_BRW_FROM_GRANT)
 661                         declare_flags |= OSD_QID_FORCE;
 662
 663                 if (size == 0) {
 664                         /* first valid lnb */
 665                         offset = lnb[i].lnb_file_offset;
 666                         size = lnb[i].lnb_len;
 667                         continue;
 668                 }
 669                 if (offset + size == lnb[i].lnb_file_offset) {
 670                         /* this lnb is contiguous to the previous one */
 671                         size += lnb[i].lnb_len;
 672                         continue;
 673                 }
 674
 675                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 676                                   obj->oo_dn, offset, size);
 677                 /* Estimating space to be consumed by a write is rather
 678                  * complicated with ZFS. As a consequence, we don't account for
 679                  * indirect blocks and just use as a rough estimate the worse
 680                  * case where the old space is being held by a snapshot. Quota
 681                  * overrun will be adjusted once the operation is committed, if
 682                  * required. */
 683                 space += osd_roundup2blocksz(size, offset, blksz);
 684
 685                 offset = lnb[i].lnb_file_offset;
 686                 size = lnb[i].lnb_len;
 687         }
 688
 689         if (size) {
 690                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 691                                   offset, size);
 692                 space += osd_roundup2blocksz(size, offset, blksz);
 693         }
 694
 695         /* backend zfs filesystem might be configured to store multiple data
 696          * copies */
 697         space  *= osd->od_os->os_copies;
 698         space   = toqb(space);
 699         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 700                npages, space);
 701
 702         record_start_io(osd, WRITE, discont_pages);
 703 retry:
 704         /* acquire quota space if needed */
 705         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 706                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 707                                space, oh, &flags, declare_flags);
 708
 709         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 710                 dt_sync(env, th->th_dev);
 711                 synced = true;
 712                 CDEBUG(D_QUOTA, "retry after sync\n");
 713                 flags = 0;
 714                 goto retry;
 715         }
 716
 717         /* we need only to store the overquota flags in the first lnb for
 718          * now, once we support multiple objects BRW, this code needs be
 719          * revised. */
 720         if (flags & QUOTA_FL_OVER_USRQUOTA)
 721                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 722         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 723                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 724 #ifdef ZFS_PROJINHERIT
 725         if (flags & QUOTA_FL_OVER_PRJQUOTA)
 726                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 727 #endif
 728
 729         RETURN(rc);
 730 }
 731
 732 /**
 733  * Policy to grow ZFS block size by write pattern.
 734  * For sequential write, it grows block size gradually until it reaches the
 735  * maximum blocksize the dataset can support. Otherwise, it will pick a
 736  * a block size by the writing region of this I/O.
 737  */
 738 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 739                               uint64_t start, uint64_t end)
 740 {
 741         struct osd_device       *osd = osd_obj2dev(obj);
 742         dnode_t *dn = obj->oo_dn;
 743         uint32_t                 blksz;
 744         int                      rc = 0;
 745
 746         ENTRY;
 747
 748         if (dn->dn_maxblkid > 0) /* can't change block size */
 749                 GOTO(out, rc);
 750
 751         if (dn->dn_datablksz >= osd->od_max_blksz)
 752                 GOTO(out, rc);
 753
 754         down_write(&obj->oo_guard);
 755
 756         blksz = dn->dn_datablksz;
 757         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 758                 GOTO(out_unlock, rc);
 759
 760         /* now ZFS can support up to 16MB block size, and if the write
 761          * is sequential, it just increases the block size gradually */
 762         if (start <= blksz) { /* sequential */
 763                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 764         } else { /* sparse, pick a block size by write region */
 765                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 766                                         end - start);
 767         }
 768
 769         if (!is_power_of_2(blksz))
 770                 blksz = size_roundup_power2(blksz);
 771
 772         if (blksz > dn->dn_datablksz) {
 773                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 774                                                blksz, 0, oh->ot_tx);
 775                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 776                 if (rc < 0)
 777                         CDEBUG(D_INODE, "object "DFID": change block size"
 778                                "%u -> %u error rc = %d\n",
 779                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 780                                dn->dn_datablksz, blksz, rc);
 781         }
 782         EXIT;
 783 out_unlock:
 784         up_write(&obj->oo_guard);
 785 out:
 786         return rc;
 787 }
 788
 789 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 790                                         loff_t off, ssize_t len)
 791 {
 792         dmu_buf_t **dbp;
 793         int i, rc, numbufs;
 794
 795         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 796                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 797         if (unlikely(rc))
 798                 return;
 799
 800         for (i = 0; i < numbufs; i++)
 801                 dbuf_set_pending_evict(dbp[i]);
 802
 803         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 804 }
 805
 806 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 807                         struct niobuf_local *lnb, int npages,
 808                         struct thandle *th)
 809 {
 810         struct osd_object  *obj  = osd_dt_obj(dt);
 811         struct osd_device  *osd = osd_obj2dev(obj);
 812         struct osd_thandle *oh;
 813         uint64_t            new_size = 0;
 814         int                 i, abufsz, rc = 0, drop_cache = 0;
 815         unsigned long      iosize = 0;
 816         ENTRY;
 817
 818         LASSERT(dt_object_exists(dt));
 819         LASSERT(obj->oo_dn);
 820
 821         LASSERT(th != NULL);
 822         oh = container_of0(th, struct osd_thandle, ot_super);
 823
 824         /* adjust block size. Assume the buffers are sorted. */
 825         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 826                                  lnb[npages - 1].lnb_file_offset +
 827                                  lnb[npages - 1].lnb_len);
 828
 829         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 830             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 831             osd->od_readcache_max_filesize)
 832                 drop_cache = 1;
 833
 834         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 835                 RETURN(-ENOSPC);
 836
 837         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 838          * size and assigning arcbuf take place at the same time.
 839          *
 840          * Thread 1:
 841          * osd_write_commit()
 842          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 843          *   -> dmu_object_set_blocksize()
 844          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 845          *       write lock held
 846          *     -> dbuf_new_size()
 847          *      -> dmu_buf_will_dirty()
 848          *       -> dbuf_read()
 849          *        -> wait for the dbuf state to change
 850          * Thread 2:
 851          * osd_write_commit()
 852          *  -> dmu_assign_arcbuf()
 853          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 854          *    -> dbuf_dirty()
 855          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 856          *
 857          * By taking the read lock, it can avoid thread 2 to enter into the
 858          * critical section of assigning the arcbuf, while thread 1 is
 859          * changing the block size.
 860          */
 861         down_read(&obj->oo_guard);
 862         for (i = 0; i < npages; i++) {
 863                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 864                         (unsigned) lnb[i].lnb_len,
 865                         (unsigned) lnb[i].lnb_file_offset);
 866
 867                 if (lnb[i].lnb_rc) {
 868                         /* ENOSPC, network RPC error, etc.
 869                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 870                          * so we skip this page if lnb_rc is set to -ENOSPC */
 871                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 872                                 PFID(lu_object_fid(&dt->do_lu)), i,
 873                                 lnb[i].lnb_rc);
 874                         continue;
 875                 }
 876
 877                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 878                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 879                 if (lnb[i].lnb_page == NULL)
 880                         continue;
 881
 882                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 883                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 884                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page),
 885                                       oh->ot_tx);
 886                         kunmap(lnb[i].lnb_page);
 887                         iosize += lnb[i].lnb_len;
 888                         abufsz = lnb[i].lnb_len; /* to drop cache below */
 889                 } else if (lnb[i].lnb_data) {
 890                         int j, apages;
 891                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 892                         /* buffer loaned for zerocopy, try to use it.
 893                          * notice that dmu_assign_arcbuf() is smart
 894                          * enough to recognize changed blocksize
 895                          * in this case it fallbacks to dmu_write() */
 896                         abufsz = arc_buf_size(lnb[i].lnb_data);
 897                         LASSERT(abufsz & PAGE_MASK);
 898                         apages = abufsz / PAGE_SIZE;
 899                         LASSERT(i + apages <= npages);
 900                         /* these references to pages must be invalidated
 901                          * to prevent access in osd_bufs_put() */
 902                         for (j = 0; j < apages; j++)
 903                                 lnb[i + j].lnb_page = NULL;
 904                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 905                                           lnb[i].lnb_file_offset,
 906                                           lnb[i].lnb_data, oh->ot_tx);
 907                         /* drop the reference, otherwise osd_put_bufs()
 908                          * will be releasing it - bad! */
 909                         lnb[i].lnb_data = NULL;
 910                         atomic_dec(&osd->od_zerocopy_loan);
 911                         iosize += abufsz;
 912                 } else {
 913                         /* we don't want to deal with cache if nothing
 914                          * has been send to ZFS at this step */
 915                         continue;
 916                 }
 917
 918                 if (!drop_cache)
 919                         continue;
 920
 921                 /* we have to mark dbufs for eviction here because
 922                  * dmu_assign_arcbuf() may create a new dbuf for
 923                  * loaned abuf */
 924                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
 925                                             abufsz);
 926         }
 927         up_read(&obj->oo_guard);
 928
 929         if (unlikely(new_size == 0)) {
 930                 /* no pages to write, no transno is needed */
 931                 th->th_local = 1;
 932                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 933                  * since ofd_commitrw_write() retries several times on ENOSPC */
 934                 record_end_io(osd, WRITE, 0, 0, 0);
 935                 RETURN(0);
 936         }
 937
 938         write_lock(&obj->oo_attr_lock);
 939         if (obj->oo_attr.la_size < new_size) {
 940                 obj->oo_attr.la_size = new_size;
 941                 write_unlock(&obj->oo_attr_lock);
 942                 /* osd_object_sa_update() will be copying directly from
 943                  * oo_attr into dbuf. any update within a single txg will copy
 944                  * the most actual */
 945                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 946                                           &obj->oo_attr.la_size, 8, oh);
 947         } else {
 948                 write_unlock(&obj->oo_attr_lock);
 949         }
 950
 951         record_end_io(osd, WRITE, 0, iosize, npages);
 952
 953         RETURN(rc);
 954 }
 955
 956 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 957                         struct niobuf_local *lnb, int npages)
 958 {
 959         struct osd_object *obj  = osd_dt_obj(dt);
 960         int                i;
 961         loff_t             eof;
 962
 963         LASSERT(dt_object_exists(dt));
 964         LASSERT(obj->oo_dn);
 965
 966         read_lock(&obj->oo_attr_lock);
 967         eof = obj->oo_attr.la_size;
 968         read_unlock(&obj->oo_attr_lock);
 969
 970         for (i = 0; i < npages; i++) {
 971                 if (unlikely(lnb[i].lnb_rc < 0))
 972                         continue;
 973
 974                 lnb[i].lnb_rc = lnb[i].lnb_len;
 975
 976                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 977                         if (eof <= lnb[i].lnb_file_offset)
 978                                 lnb[i].lnb_rc = 0;
 979                         else
 980                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 981
 982                         /* all subsequent rc should be 0 */
 983                         while (++i < npages)
 984                                 lnb[i].lnb_rc = 0;
 985                         break;
 986                 }
 987         }
 988
 989         return 0;
 990 }
 991
 992 /*
 993  * Punch/truncate an object
 994  *
 995  *      IN:     db  - dmu_buf of the object to free data in.
 996  *              off - start of section to free.
 997  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 998  *
 999  *      RETURN: 0 if success
1000  *              error code if failure
1001  *
1002  * The transaction passed to this routine must have
1003  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1004  * called and then assigned to a transaction group.
1005  */
1006 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
1007                                 uint64_t size, uint64_t off, uint64_t len)
1008 {
1009         int rc = 0;
1010
1011         /* Assert that the transaction has been assigned to a
1012            transaction group. */
1013         LASSERT(tx->tx_txg != 0);
1014         /*
1015          * Nothing to do if file already at desired length.
1016          */
1017         if (len == DMU_OBJECT_END && size == off)
1018                 return 0;
1019
1020         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1021         if (off < size)
1022                 dmu_free_range(os, dn->dn_object, off, len, tx);
1023
1024         return rc;
1025 }
1026
1027 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1028                         __u64 start, __u64 end, struct thandle *th)
1029 {
1030         struct osd_object  *obj = osd_dt_obj(dt);
1031         struct osd_device  *osd = osd_obj2dev(obj);
1032         struct osd_thandle *oh;
1033         __u64               len;
1034         int                 rc = 0;
1035         ENTRY;
1036
1037         LASSERT(dt_object_exists(dt));
1038         LASSERT(osd_invariant(obj));
1039
1040         LASSERT(th != NULL);
1041         oh = container_of0(th, struct osd_thandle, ot_super);
1042
1043         write_lock(&obj->oo_attr_lock);
1044         /* truncate */
1045         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1046                 len = DMU_OBJECT_END;
1047         else
1048                 len = end - start;
1049         write_unlock(&obj->oo_attr_lock);
1050
1051         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
1052                                 obj->oo_attr.la_size, start, len);
1053         /* set new size */
1054         if (len == DMU_OBJECT_END) {
1055                 write_lock(&obj->oo_attr_lock);
1056                 obj->oo_attr.la_size = start;
1057                 write_unlock(&obj->oo_attr_lock);
1058                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1059                                           &obj->oo_attr.la_size, 8, oh);
1060         }
1061         RETURN(rc);
1062 }
1063
1064 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1065                         __u64 start, __u64 end, struct thandle *handle)
1066 {
1067         struct osd_object  *obj = osd_dt_obj(dt);
1068         struct osd_device  *osd = osd_obj2dev(obj);
1069         struct osd_thandle *oh;
1070         __u64               len;
1071         ENTRY;
1072
1073         oh = container_of0(handle, struct osd_thandle, ot_super);
1074
1075         read_lock(&obj->oo_attr_lock);
1076         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1077                 len = DMU_OBJECT_END;
1078         else
1079                 len = end - start;
1080
1081         /* declare we'll free some blocks ... */
1082         if (start < obj->oo_attr.la_size) {
1083                 read_unlock(&obj->oo_attr_lock);
1084                 dmu_tx_mark_netfree(oh->ot_tx);
1085                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1086         } else {
1087                 read_unlock(&obj->oo_attr_lock);
1088         }
1089
1090         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1091                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1092                                  0, oh, NULL, OSD_QID_BLK));
1093 }
1094
1095 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1096                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1097 {
1098         int     rc;
1099         ENTRY;
1100
1101         switch (advice) {
1102         default:
1103                 rc = -ENOTSUPP;
1104                 break;
1105         }
1106
1107         RETURN(rc);
1108 }
1109
1110 struct dt_body_operations osd_body_ops = {
1111         .dbo_read                       = osd_read,
1112         .dbo_declare_write              = osd_declare_write,
1113         .dbo_write                      = osd_write,
1114         .dbo_bufs_get                   = osd_bufs_get,
1115         .dbo_bufs_put                   = osd_bufs_put,
1116         .dbo_write_prep                 = osd_write_prep,
1117         .dbo_declare_write_commit       = osd_declare_write_commit,
1118         .dbo_write_commit               = osd_write_commit,
1119         .dbo_read_prep                  = osd_read_prep,
1120         .dbo_declare_punch              = osd_declare_punch,
1121         .dbo_punch                      = osd_punch,
1122         .dbo_ladvise                    = osd_ladvise,
1123 };
1124
1125 struct dt_body_operations osd_body_scrub_ops = {
1126         .dbo_read                       = osd_read_no_record,
1127         .dbo_declare_write              = osd_declare_write,
1128         .dbo_write                      = osd_write,
1129 };