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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd-zfs/osd_io.c
  37  *
  38  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  39  * Author: Mike Pershin <tappro@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSD
  43
  44 #include <lustre_ver.h>
  45 #include <libcfs/libcfs.h>
  46 #include <obd_support.h>
  47 #include <lustre_net.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_disk.h>
  51 #include <lustre_fid.h>
  52 #include <lustre/lustre_idl.h>  /* LLOG_MIN_CHUNK_SIZE definition */
  53
  54 #include "osd_internal.h"
  55
  56 #include <sys/dnode.h>
  57 #include <sys/dbuf.h>
  58 #include <sys/spa.h>
  59 #include <sys/stat.h>
  60 #include <sys/zap.h>
  61 #include <sys/spa_impl.h>
  62 #include <sys/zfs_znode.h>
  63 #include <sys/dmu_tx.h>
  64 #include <sys/dmu_objset.h>
  65 #include <sys/dsl_prop.h>
  66 #include <sys/sa_impl.h>
  67 #include <sys/txg.h>
  68
  69 static char *osd_zerocopy_tag = "zerocopy";
  70
  71
  72 static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
  73 {
  74         struct obd_histogram *h = osd->od_brw_stats.hist;
  75
  76         if (rw == READ) {
  77                 atomic_inc(&osd->od_r_in_flight);
  78                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
  79                                  atomic_read(&osd->od_r_in_flight));
  80                 lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
  81
  82         } else {
  83                 atomic_inc(&osd->od_w_in_flight);
  84                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
  85                                  atomic_read(&osd->od_w_in_flight));
  86                 lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
  87
  88         }
  89 }
  90
  91 static void record_end_io(struct osd_device *osd, int rw,
  92                           unsigned long elapsed, int disksize, int npages)
  93 {
  94         struct obd_histogram *h = osd->od_brw_stats.hist;
  95
  96         if (rw == READ) {
  97                 atomic_dec(&osd->od_r_in_flight);
  98                 lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
  99                 if (disksize > 0)
 100                         lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
 101                 if (elapsed)
 102                         lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
 103
 104         } else {
 105                 atomic_dec(&osd->od_w_in_flight);
 106                 lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
 107                 if (disksize > 0)
 108                         lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
 109                 if (elapsed)
 110                         lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
 111         }
 112 }
 113
 114 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 115                         struct lu_buf *buf, loff_t *pos)
 116 {
 117         struct osd_object *obj  = osd_dt_obj(dt);
 118         struct osd_device *osd = osd_obj2dev(obj);
 119         uint64_t           old_size;
 120         int                size = buf->lb_len;
 121         int                rc;
 122         unsigned long      start;
 123
 124         LASSERT(dt_object_exists(dt));
 125         LASSERT(obj->oo_db);
 126
 127         start = cfs_time_current();
 128
 129         read_lock(&obj->oo_attr_lock);
 130         old_size = obj->oo_attr.la_size;
 131         read_unlock(&obj->oo_attr_lock);
 132
 133         if (*pos + size > old_size) {
 134                 if (old_size < *pos)
 135                         return 0;
 136                 else
 137                         size = old_size - *pos;
 138         }
 139
 140         record_start_io(osd, READ, 0);
 141
 142         rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
 143                         buf->lb_buf, DMU_READ_PREFETCH);
 144
 145         record_end_io(osd, READ, cfs_time_current() - start, size,
 146                       size >> PAGE_SHIFT);
 147         if (rc == 0) {
 148                 rc = size;
 149                 *pos += size;
 150         }
 151         return rc;
 152 }
 153
 154 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 155                                 const struct lu_buf *buf, loff_t pos,
 156                                 struct thandle *th)
 157 {
 158         struct osd_object  *obj  = osd_dt_obj(dt);
 159         struct osd_device  *osd = osd_obj2dev(obj);
 160         struct osd_thandle *oh;
 161         uint64_t            oid;
 162         ENTRY;
 163
 164         oh = container_of0(th, struct osd_thandle, ot_super);
 165
 166         /* in some cases declare can race with creation (e.g. llog)
 167          * and we need to wait till object is initialized. notice
 168          * LOHA_EXISTs is supposed to be the last step in the
 169          * initialization */
 170
 171         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 172         if (dt_object_exists(dt))
 173                 oid = obj->oo_db->db_object;
 174         else
 175                 oid = DMU_NEW_OBJECT;
 176
 177         /* XXX: we still miss for append declaration support in ZFS
 178          *      -1 means append which is used by llog mostly, llog
 179          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 180         if (pos == -1)
 181                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 182                             obj->oo_attr.la_size + (2 << 20));
 183         dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
 184
 185         /* dt_declare_write() is usually called for system objects, such
 186          * as llog or last_rcvd files. We needn't enforce quota on those
 187          * objects, so always set the lqi_space as 0. */
 188         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 189                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 190                                  false));
 191 }
 192
 193 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 194                         const struct lu_buf *buf, loff_t *pos,
 195                         struct thandle *th, int ignore_quota)
 196 {
 197         struct osd_object  *obj  = osd_dt_obj(dt);
 198         struct osd_device  *osd = osd_obj2dev(obj);
 199         struct osd_thandle *oh;
 200         uint64_t            offset = *pos;
 201         int                 rc;
 202
 203         ENTRY;
 204
 205         LASSERT(dt_object_exists(dt));
 206         LASSERT(obj->oo_db);
 207
 208         LASSERT(th != NULL);
 209         oh = container_of0(th, struct osd_thandle, ot_super);
 210
 211         record_start_io(osd, WRITE, 0);
 212
 213         dmu_write(osd->od_os, obj->oo_db->db_object, offset,
 214                 (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
 215         write_lock(&obj->oo_attr_lock);
 216         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 217                 obj->oo_attr.la_size = offset + buf->lb_len;
 218                 write_unlock(&obj->oo_attr_lock);
 219                 /* osd_object_sa_update() will be copying directly from oo_attr
 220                  * into dbuf.  any update within a single txg will copy the
 221                  * most actual */
 222                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 223                                         &obj->oo_attr.la_size, 8, oh);
 224                 if (unlikely(rc))
 225                         GOTO(out, rc);
 226         } else {
 227                 write_unlock(&obj->oo_attr_lock);
 228         }
 229
 230         *pos += buf->lb_len;
 231         rc = buf->lb_len;
 232
 233 out:
 234         record_end_io(osd, WRITE, 0, buf->lb_len,
 235                       buf->lb_len >> PAGE_SHIFT);
 236
 237         RETURN(rc);
 238 }
 239
 240 /*
 241  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 242  *      purposes as it's not very well defined ...
 243  *      instead I use the lowest bit of the address so that:
 244  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 245  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 246  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 247  *
 248  *      bzzz, to blame
 249  */
 250 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 251                         struct niobuf_local *lnb, int npages)
 252 {
 253         struct osd_object *obj  = osd_dt_obj(dt);
 254         struct osd_device *osd = osd_obj2dev(obj);
 255         unsigned long      ptr;
 256         int                i;
 257
 258         LASSERT(dt_object_exists(dt));
 259         LASSERT(obj->oo_db);
 260
 261         for (i = 0; i < npages; i++) {
 262                 if (lnb[i].lnb_page == NULL)
 263                         continue;
 264                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 265                         /* this is anonymous page allocated for copy-write */
 266                         lnb[i].lnb_page->mapping = NULL;
 267                         __free_page(lnb[i].lnb_page);
 268                         atomic_dec(&osd->od_zerocopy_alloc);
 269                 } else {
 270                         /* see comment in osd_bufs_get_read() */
 271                         ptr = (unsigned long)lnb[i].lnb_data;
 272                         if (ptr & 1UL) {
 273                                 ptr &= ~1UL;
 274                                 dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
 275                                 atomic_dec(&osd->od_zerocopy_pin);
 276                         } else if (lnb[i].lnb_data != NULL) {
 277                                 dmu_return_arcbuf(lnb[i].lnb_data);
 278                                 atomic_dec(&osd->od_zerocopy_loan);
 279                         }
 280                 }
 281                 lnb[i].lnb_page = NULL;
 282                 lnb[i].lnb_data = NULL;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static inline struct page *kmem_to_page(void *addr)
 289 {
 290         if (is_vmalloc_addr(addr))
 291                 return vmalloc_to_page(addr);
 292         else
 293                 return virt_to_page(addr);
 294 }
 295
 296 /**
 297  * Prepare buffers for read.
 298  *
 299  * The function maps the range described by \a off and \a len to \a lnb array.
 300  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 301  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 302  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 303  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 304  * A better implementation would just return the buffers (potentially unfilled)
 305  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 306  *
 307  * \param[in] env       environment
 308  * \param[in] obj       object
 309  * \param[in] off       offset in bytes
 310  * \param[in] len       the number of bytes to access
 311  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 312  *
 313  * \retval              0 for success
 314  * \retval              negative error number of failure
 315  */
 316 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 317                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 318 {
 319         struct osd_device *osd = osd_obj2dev(obj);
 320         unsigned long      start = cfs_time_current();
 321         int                rc, i, numbufs, npages = 0;
 322         dmu_buf_t        **dbp;
 323         ENTRY;
 324
 325         record_start_io(osd, READ, 0);
 326
 327         /* grab buffers for read:
 328          * OSD API let us to grab buffers first, then initiate IO(s)
 329          * so that all required IOs will be done in parallel, but at the
 330          * moment DMU doesn't provide us with a method to grab buffers.
 331          * If we discover this is a vital for good performance we
 332          * can get own replacement for dmu_buf_hold_array_by_bonus().
 333          */
 334         while (len > 0) {
 335                 rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
 336                                                   osd_zerocopy_tag, &numbufs,
 337                                                   &dbp);
 338                 if (unlikely(rc))
 339                         GOTO(err, rc);
 340
 341                 for (i = 0; i < numbufs; i++) {
 342                         int bufoff, tocpy, thispage;
 343                         void *dbf = dbp[i];
 344
 345                         LASSERT(len > 0);
 346
 347                         atomic_inc(&osd->od_zerocopy_pin);
 348
 349                         bufoff = off - dbp[i]->db_offset;
 350                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 351
 352                         /* kind of trick to differentiate dbuf vs. arcbuf */
 353                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 354                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 355
 356                         while (tocpy > 0) {
 357                                 thispage = PAGE_SIZE;
 358                                 thispage -= bufoff & (PAGE_SIZE - 1);
 359                                 thispage = min(tocpy, thispage);
 360
 361                                 lnb->lnb_rc = 0;
 362                                 lnb->lnb_file_offset = off;
 363                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 364                                 lnb->lnb_len = thispage;
 365                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 366                                                              bufoff);
 367                                 /* mark just a single slot: we need this
 368                                  * reference to dbuf to be released once */
 369                                 lnb->lnb_data = dbf;
 370                                 dbf = NULL;
 371
 372                                 tocpy -= thispage;
 373                                 len -= thispage;
 374                                 bufoff += thispage;
 375                                 off += thispage;
 376
 377                                 npages++;
 378                                 lnb++;
 379                         }
 380
 381                         /* steal dbuf so dmu_buf_rele_array() can't release
 382                          * it */
 383                         dbp[i] = NULL;
 384                 }
 385
 386                 dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
 387         }
 388
 389         record_end_io(osd, READ, cfs_time_current() - start,
 390                       npages * PAGE_SIZE, npages);
 391
 392         RETURN(npages);
 393
 394 err:
 395         LASSERT(rc < 0);
 396         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 397         RETURN(rc);
 398 }
 399
 400 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 401                                 loff_t off, ssize_t len, struct niobuf_local *lnb)
 402 {
 403         struct osd_device *osd = osd_obj2dev(obj);
 404         int                plen, off_in_block, sz_in_block;
 405         int                rc, i = 0, npages = 0;
 406         arc_buf_t         *abuf;
 407         uint32_t           bs;
 408         uint64_t           dummy;
 409         ENTRY;
 410
 411         dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
 412
 413         /*
 414          * currently only full blocks are subject to zerocopy approach:
 415          * so that we're sure nobody is trying to update the same block
 416          */
 417         while (len > 0) {
 418                 LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
 419
 420                 off_in_block = off & (bs - 1);
 421                 sz_in_block = min_t(int, bs - off_in_block, len);
 422
 423                 if (sz_in_block == bs) {
 424                         /* full block, try to use zerocopy */
 425
 426                         abuf = dmu_request_arcbuf(obj->oo_db, bs);
 427                         if (unlikely(abuf == NULL))
 428                                 GOTO(out_err, rc = -ENOMEM);
 429
 430                         atomic_inc(&osd->od_zerocopy_loan);
 431
 432                         /* go over pages arcbuf contains, put them as
 433                          * local niobufs for ptlrpc's bulks */
 434                         while (sz_in_block > 0) {
 435                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 436
 437                                 lnb[i].lnb_file_offset = off;
 438                                 lnb[i].lnb_page_offset = 0;
 439                                 lnb[i].lnb_len = plen;
 440                                 lnb[i].lnb_rc = 0;
 441                                 if (sz_in_block == bs)
 442                                         lnb[i].lnb_data = abuf;
 443                                 else
 444                                         lnb[i].lnb_data = NULL;
 445
 446                                 /* this one is not supposed to fail */
 447                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 448                                                         off_in_block);
 449                                 LASSERT(lnb[i].lnb_page);
 450
 451                                 lprocfs_counter_add(osd->od_stats,
 452                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 453
 454                                 sz_in_block -= plen;
 455                                 len -= plen;
 456                                 off += plen;
 457                                 off_in_block += plen;
 458                                 i++;
 459                                 npages++;
 460                         }
 461                 } else {
 462                         if (off_in_block == 0 && len < bs &&
 463                                         off + len >= obj->oo_attr.la_size)
 464                                 lprocfs_counter_add(osd->od_stats,
 465                                                 LPROC_OSD_TAIL_IO, 1);
 466
 467                         /* can't use zerocopy, allocate temp. buffers */
 468                         while (sz_in_block > 0) {
 469                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 470
 471                                 lnb[i].lnb_file_offset = off;
 472                                 lnb[i].lnb_page_offset = 0;
 473                                 lnb[i].lnb_len = plen;
 474                                 lnb[i].lnb_rc = 0;
 475                                 lnb[i].lnb_data = NULL;
 476
 477                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 478                                 if (unlikely(lnb[i].lnb_page == NULL))
 479                                         GOTO(out_err, rc = -ENOMEM);
 480
 481                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 482                                 lnb[i].lnb_page->mapping = (void *)obj;
 483
 484                                 atomic_inc(&osd->od_zerocopy_alloc);
 485                                 lprocfs_counter_add(osd->od_stats,
 486                                                 LPROC_OSD_COPY_IO, 1);
 487
 488                                 sz_in_block -= plen;
 489                                 len -= plen;
 490                                 off += plen;
 491                                 i++;
 492                                 npages++;
 493                         }
 494                 }
 495         }
 496
 497         RETURN(npages);
 498
 499 out_err:
 500         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 501         RETURN(rc);
 502 }
 503
 504 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 505                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 506                         int rw)
 507 {
 508         struct osd_object *obj  = osd_dt_obj(dt);
 509         int                rc;
 510
 511         LASSERT(dt_object_exists(dt));
 512         LASSERT(obj->oo_db);
 513
 514         if (rw == 0)
 515                 rc = osd_bufs_get_read(env, obj, offset, len, lnb);
 516         else
 517                 rc = osd_bufs_get_write(env, obj, offset, len, lnb);
 518
 519         return rc;
 520 }
 521
 522 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 523                         struct niobuf_local *lnb, int npages)
 524 {
 525         struct osd_object *obj = osd_dt_obj(dt);
 526
 527         LASSERT(dt_object_exists(dt));
 528         LASSERT(obj->oo_db);
 529
 530         return 0;
 531 }
 532
 533 static inline uint32_t osd_get_blocksz(struct osd_object *obj)
 534 {
 535         uint32_t blksz;
 536         u_longlong_t unused;
 537
 538         LASSERT(obj->oo_db);
 539
 540         dmu_object_size_from_db(obj->oo_db, &blksz, &unused);
 541         return blksz;
 542 }
 543
 544 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 545                                            uint64_t offset,
 546                                            uint32_t blksz)
 547 {
 548         LASSERT(blksz > 0);
 549
 550         size += offset % blksz;
 551
 552         if (likely(IS_PO2(blksz)))
 553                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 554
 555         size += blksz - 1;
 556         do_div(size, blksz);
 557         return size * blksz;
 558 }
 559
 560 static int osd_declare_write_commit(const struct lu_env *env,
 561                                     struct dt_object *dt,
 562                                     struct niobuf_local *lnb, int npages,
 563                                     struct thandle *th)
 564 {
 565         struct osd_object  *obj = osd_dt_obj(dt);
 566         struct osd_device  *osd = osd_obj2dev(obj);
 567         struct osd_thandle *oh;
 568         uint64_t            offset = 0;
 569         uint32_t            size = 0;
 570         uint32_t            blksz = osd_get_blocksz(obj);
 571         int                 i, rc, flags = 0;
 572         bool                ignore_quota = false, synced = false;
 573         long long           space = 0;
 574         struct page        *last_page = NULL;
 575         unsigned long       discont_pages = 0;
 576         ENTRY;
 577
 578         LASSERT(dt_object_exists(dt));
 579         LASSERT(obj->oo_db);
 580
 581         LASSERT(lnb);
 582         LASSERT(npages > 0);
 583
 584         oh = container_of0(th, struct osd_thandle, ot_super);
 585
 586         for (i = 0; i < npages; i++) {
 587                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 588                         ++discont_pages;
 589                 last_page = lnb[i].lnb_page;
 590                 if (lnb[i].lnb_rc)
 591                         /* ENOSPC, network RPC error, etc.
 592                          * We don't want to book space for pages which will be
 593                          * skipped in osd_write_commit(). Hence we skip pages
 594                          * with lnb_rc != 0 here too */
 595                         continue;
 596                 /* ignore quota for the whole request if any page is from
 597                  * client cache or written by root.
 598                  *
 599                  * XXX once we drop the 1.8 client support, the checking
 600                  * for whether page is from cache can be simplified as:
 601                  * !(lnb[i].flags & OBD_BRW_SYNC)
 602                  *
 603                  * XXX we could handle this on per-lnb basis as done by
 604                  * grant. */
 605                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 606                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 607                     OBD_BRW_FROM_GRANT)
 608                         ignore_quota = true;
 609                 if (size == 0) {
 610                         /* first valid lnb */
 611                         offset = lnb[i].lnb_file_offset;
 612                         size = lnb[i].lnb_len;
 613                         continue;
 614                 }
 615                 if (offset + size == lnb[i].lnb_file_offset) {
 616                         /* this lnb is contiguous to the previous one */
 617                         size += lnb[i].lnb_len;
 618                         continue;
 619                 }
 620
 621                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 622                                   offset, size);
 623                 /* Estimating space to be consumed by a write is rather
 624                  * complicated with ZFS. As a consequence, we don't account for
 625                  * indirect blocks and just use as a rough estimate the worse
 626                  * case where the old space is being held by a snapshot. Quota
 627                  * overrun will be adjusted once the operation is committed, if
 628                  * required. */
 629                 space += osd_roundup2blocksz(size, offset, blksz);
 630
 631                 offset = lnb[i].lnb_file_offset;
 632                 size = lnb[i].lnb_len;
 633         }
 634
 635         if (size) {
 636                 dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
 637                                   offset, size);
 638                 space += osd_roundup2blocksz(size, offset, blksz);
 639         }
 640
 641         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 642
 643         oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
 644
 645         /* backend zfs filesystem might be configured to store multiple data
 646          * copies */
 647         space  *= osd->od_os->os_copies;
 648         space   = toqb(space);
 649         CDEBUG(D_QUOTA, "writing %d pages, reserving "LPD64"K of quota space\n",
 650                npages, space);
 651
 652         record_start_io(osd, WRITE, discont_pages);
 653 retry:
 654         /* acquire quota space if needed */
 655         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 656                                obj->oo_attr.la_gid, space, oh, true, &flags,
 657                                ignore_quota);
 658
 659         if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
 660                 dt_sync(env, th->th_dev);
 661                 synced = true;
 662                 CDEBUG(D_QUOTA, "retry after sync\n");
 663                 flags = 0;
 664                 goto retry;
 665         }
 666
 667         /* we need only to store the overquota flags in the first lnb for
 668          * now, once we support multiple objects BRW, this code needs be
 669          * revised. */
 670         if (flags & QUOTA_FL_OVER_USRQUOTA)
 671                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 672         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 673                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 674
 675         RETURN(rc);
 676 }
 677
 678 /**
 679  * Policy to grow ZFS block size by write pattern.
 680  * For sequential write, it grows block size gradually until it reaches the
 681  * maximum blocksize the dataset can support. Otherwise, it will pick a
 682  * a block size by the writing region of this I/O.
 683  */
 684 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 685                               uint64_t start, uint64_t end)
 686 {
 687         struct osd_device       *osd = osd_obj2dev(obj);
 688         dmu_buf_impl_t          *db = (dmu_buf_impl_t *)obj->oo_db;
 689         dnode_t                 *dn;
 690         uint32_t                 blksz;
 691         int                      rc = 0;
 692
 693         ENTRY;
 694
 695         DB_DNODE_ENTER(db);
 696         dn = DB_DNODE(db);
 697
 698         if (dn->dn_maxblkid > 0) /* can't change block size */
 699                 GOTO(out, rc);
 700
 701         if (dn->dn_datablksz >= osd->od_max_blksz)
 702                 GOTO(out, rc);
 703
 704         down_write(&obj->oo_guard);
 705
 706         blksz = dn->dn_datablksz;
 707         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 708                 GOTO(out_unlock, rc);
 709
 710         /* now ZFS can support up to 16MB block size, and if the write
 711          * is sequential, it just increases the block size gradually */
 712         if (start <= blksz) { /* sequential */
 713                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 714         } else { /* sparse, pick a block size by write region */
 715                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 716                                         end - start);
 717         }
 718
 719         if (!is_power_of_2(blksz))
 720                 blksz = size_roundup_power2(blksz);
 721
 722         if (blksz > dn->dn_datablksz) {
 723                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 724                                                blksz, 0, oh->ot_tx);
 725                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 726                 if (rc < 0)
 727                         CDEBUG(D_INODE, "object "DFID": change block size"
 728                                "%u -> %u error rc = %d\n",
 729                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 730                                dn->dn_datablksz, blksz, rc);
 731         }
 732         EXIT;
 733 out_unlock:
 734         up_write(&obj->oo_guard);
 735 out:
 736         DB_DNODE_EXIT(db);
 737         return rc;
 738 }
 739
 740 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 741                         struct niobuf_local *lnb, int npages,
 742                         struct thandle *th)
 743 {
 744         struct osd_object  *obj  = osd_dt_obj(dt);
 745         struct osd_device  *osd = osd_obj2dev(obj);
 746         struct osd_thandle *oh;
 747         uint64_t            new_size = 0;
 748         int                 i, rc = 0;
 749         unsigned long      iosize = 0;
 750         ENTRY;
 751
 752         LASSERT(dt_object_exists(dt));
 753         LASSERT(obj->oo_db);
 754
 755         LASSERT(th != NULL);
 756         oh = container_of0(th, struct osd_thandle, ot_super);
 757
 758         /* adjust block size. Assume the buffers are sorted. */
 759         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 760                                  lnb[npages - 1].lnb_file_offset +
 761                                  lnb[npages - 1].lnb_len);
 762         for (i = 0; i < npages; i++) {
 763                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 764                         (unsigned) lnb[i].lnb_len,
 765                         (unsigned) lnb[i].lnb_file_offset);
 766
 767                 if (lnb[i].lnb_rc) {
 768                         /* ENOSPC, network RPC error, etc.
 769                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 770                          * so we skip this page if lnb_rc is set to -ENOSPC */
 771                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 772                                 PFID(lu_object_fid(&dt->do_lu)), i,
 773                                 lnb[i].lnb_rc);
 774                         continue;
 775                 }
 776
 777                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 778                         dmu_write(osd->od_os, obj->oo_db->db_object,
 779                                 lnb[i].lnb_file_offset, lnb[i].lnb_len,
 780                                 kmap(lnb[i].lnb_page), oh->ot_tx);
 781                         kunmap(lnb[i].lnb_page);
 782                 } else if (lnb[i].lnb_data) {
 783                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 784                         /* buffer loaned for zerocopy, try to use it.
 785                          * notice that dmu_assign_arcbuf() is smart
 786                          * enough to recognize changed blocksize
 787                          * in this case it fallbacks to dmu_write() */
 788                         dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
 789                                           lnb[i].lnb_data, oh->ot_tx);
 790                         /* drop the reference, otherwise osd_put_bufs()
 791                          * will be releasing it - bad! */
 792                         lnb[i].lnb_data = NULL;
 793                         atomic_dec(&osd->od_zerocopy_loan);
 794                 }
 795
 796                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 797                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 798                 iosize += lnb[i].lnb_len;
 799         }
 800
 801         if (unlikely(new_size == 0)) {
 802                 /* no pages to write, no transno is needed */
 803                 th->th_local = 1;
 804                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 805                  * since ofd_commitrw_write() retries several times on ENOSPC */
 806                 record_end_io(osd, WRITE, 0, 0, 0);
 807                 RETURN(0);
 808         }
 809
 810         write_lock(&obj->oo_attr_lock);
 811         if (obj->oo_attr.la_size < new_size) {
 812                 obj->oo_attr.la_size = new_size;
 813                 write_unlock(&obj->oo_attr_lock);
 814                 /* osd_object_sa_update() will be copying directly from
 815                  * oo_attr into dbuf. any update within a single txg will copy
 816                  * the most actual */
 817                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 818                                           &obj->oo_attr.la_size, 8, oh);
 819         } else {
 820                 write_unlock(&obj->oo_attr_lock);
 821         }
 822
 823         record_end_io(osd, WRITE, 0, iosize, npages);
 824
 825         RETURN(rc);
 826 }
 827
 828 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 829                         struct niobuf_local *lnb, int npages)
 830 {
 831         struct osd_object *obj  = osd_dt_obj(dt);
 832         int                i;
 833         loff_t             eof;
 834
 835         LASSERT(dt_object_exists(dt));
 836         LASSERT(obj->oo_db);
 837
 838         read_lock(&obj->oo_attr_lock);
 839         eof = obj->oo_attr.la_size;
 840         read_unlock(&obj->oo_attr_lock);
 841
 842         for (i = 0; i < npages; i++) {
 843                 if (unlikely(lnb[i].lnb_rc < 0))
 844                         continue;
 845
 846                 lnb[i].lnb_rc = lnb[i].lnb_len;
 847
 848                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 849                         if (eof <= lnb[i].lnb_file_offset)
 850                                 lnb[i].lnb_rc = 0;
 851                         else
 852                                 lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 853
 854                         /* all subsequent rc should be 0 */
 855                         while (++i < npages)
 856                                 lnb[i].lnb_rc = 0;
 857                         break;
 858                 }
 859         }
 860
 861         return 0;
 862 }
 863
 864 /*
 865  * Punch/truncate an object
 866  *
 867  *      IN:     db  - dmu_buf of the object to free data in.
 868  *              off - start of section to free.
 869  *              len - length of section to free (DMU_OBJECT_END => to EOF).
 870  *
 871  *      RETURN: 0 if success
 872  *              error code if failure
 873  *
 874  * The transaction passed to this routine must have
 875  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
 876  * called and then assigned to a transaction group.
 877  */
 878 static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
 879                                 uint64_t size, uint64_t off, uint64_t len)
 880 {
 881         int rc = 0;
 882
 883         /* Assert that the transaction has been assigned to a
 884            transaction group. */
 885         LASSERT(tx->tx_txg != 0);
 886         /*
 887          * Nothing to do if file already at desired length.
 888          */
 889         if (len == DMU_OBJECT_END && size == off)
 890                 return 0;
 891
 892         if (off < size)
 893                 rc = -dmu_free_range(os, db->db_object, off, len, tx);
 894
 895         return rc;
 896 }
 897
 898 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
 899                         __u64 start, __u64 end, struct thandle *th)
 900 {
 901         struct osd_object  *obj = osd_dt_obj(dt);
 902         struct osd_device  *osd = osd_obj2dev(obj);
 903         struct osd_thandle *oh;
 904         __u64               len;
 905         int                 rc = 0;
 906         ENTRY;
 907
 908         LASSERT(dt_object_exists(dt));
 909         LASSERT(osd_invariant(obj));
 910
 911         LASSERT(th != NULL);
 912         oh = container_of0(th, struct osd_thandle, ot_super);
 913
 914         write_lock(&obj->oo_attr_lock);
 915         /* truncate */
 916         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 917                 len = DMU_OBJECT_END;
 918         else
 919                 len = end - start;
 920         write_unlock(&obj->oo_attr_lock);
 921
 922         rc = __osd_object_punch(osd->od_os, obj->oo_db, oh->ot_tx,
 923                                 obj->oo_attr.la_size, start, len);
 924         /* set new size */
 925         if (len == DMU_OBJECT_END) {
 926                 write_lock(&obj->oo_attr_lock);
 927                 obj->oo_attr.la_size = start;
 928                 write_unlock(&obj->oo_attr_lock);
 929                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 930                                           &obj->oo_attr.la_size, 8, oh);
 931         }
 932         RETURN(rc);
 933 }
 934
 935 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
 936                         __u64 start, __u64 end, struct thandle *handle)
 937 {
 938         struct osd_object  *obj = osd_dt_obj(dt);
 939         struct osd_device  *osd = osd_obj2dev(obj);
 940         struct osd_thandle *oh;
 941         __u64               len;
 942         ENTRY;
 943
 944         oh = container_of0(handle, struct osd_thandle, ot_super);
 945
 946         read_lock(&obj->oo_attr_lock);
 947         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
 948                 len = DMU_OBJECT_END;
 949         else
 950                 len = end - start;
 951
 952         /* declare we'll free some blocks ... */
 953         if (start < obj->oo_attr.la_size) {
 954                 read_unlock(&obj->oo_attr_lock);
 955                 dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
 956         } else {
 957                 read_unlock(&obj->oo_attr_lock);
 958         }
 959
 960         /* ... and we'll modify size attribute */
 961         dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 962
 963         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 964                                  obj->oo_attr.la_gid, 0, oh, true, NULL,
 965                                  false));
 966 }
 967
 968 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
 969                        __u64 start, __u64 end, enum lu_ladvise_type advice)
 970 {
 971         int     rc;
 972         ENTRY;
 973
 974         switch (advice) {
 975         default:
 976                 rc = -ENOTSUPP;
 977                 break;
 978         }
 979
 980         RETURN(rc);
 981 }
 982
 983 struct dt_body_operations osd_body_ops = {
 984         .dbo_read                       = osd_read,
 985         .dbo_declare_write              = osd_declare_write,
 986         .dbo_write                      = osd_write,
 987         .dbo_bufs_get                   = osd_bufs_get,
 988         .dbo_bufs_put                   = osd_bufs_put,
 989         .dbo_write_prep                 = osd_write_prep,
 990         .dbo_declare_write_commit       = osd_declare_write_commit,
 991         .dbo_write_commit               = osd_write_commit,
 992         .dbo_read_prep                  = osd_read_prep,
 993         .dbo_declare_punch              = osd_declare_punch,
 994         .dbo_punch                      = osd_punch,
 995         .dbo_ladvise                    = osd_ladvise,
 996 };