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