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 static void dbuf_set_pending_evict(dmu_buf_t *db)
  67 {
  68         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
  69         dbi->db_pending_evict = TRUE;
  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         else
  99                 atomic_dec(&osd->od_w_in_flight);
 100
 101         lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
 102         if (disksize > 0)
 103                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
 104         if (elapsed)
 105                 lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
 106 }
 107
 108 static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
 109                           struct lu_buf *buf, loff_t *pos, size_t *size)
 110 {
 111         struct osd_object *obj = osd_dt_obj(dt);
 112         uint64_t old_size;
 113         int rc;
 114
 115         LASSERT(dt_object_exists(dt));
 116         LASSERT(obj->oo_dn);
 117
 118         read_lock(&obj->oo_attr_lock);
 119         old_size = obj->oo_attr.la_size;
 120         read_unlock(&obj->oo_attr_lock);
 121
 122         if (*pos + *size > old_size) {
 123                 if (old_size < *pos)
 124                         return 0;
 125
 126                 *size = old_size - *pos;
 127         }
 128
 129         rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
 130                           buf->lb_buf, DMU_READ_PREFETCH);
 131         if (!rc) {
 132                 rc = *size;
 133                 *pos += *size;
 134         }
 135
 136         return rc;
 137 }
 138
 139 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 140                         struct lu_buf *buf, loff_t *pos)
 141 {
 142         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 143         size_t size = buf->lb_len;
 144         ktime_t start;
 145         s64 delta_ms;
 146         int rc;
 147
 148         start = ktime_get();
 149         record_start_io(osd, READ, 0);
 150         rc = __osd_read(env, dt, buf, pos, &size);
 151         delta_ms = ktime_ms_delta(ktime_get(), start);
 152         record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
 153
 154         return rc;
 155 }
 156
 157 static inline ssize_t osd_read_no_record(const struct lu_env *env,
 158                                          struct dt_object *dt,
 159                                          struct lu_buf *buf, loff_t *pos)
 160 {
 161         size_t size = buf->lb_len;
 162
 163         return __osd_read(env, dt, buf, pos, &size);
 164 }
 165
 166 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 167                                 const struct lu_buf *buf, loff_t pos,
 168                                 struct thandle *th)
 169 {
 170         struct osd_object  *obj  = osd_dt_obj(dt);
 171         struct osd_device  *osd = osd_obj2dev(obj);
 172         struct osd_thandle *oh;
 173         uint64_t            oid;
 174         ENTRY;
 175
 176         oh = container_of(th, struct osd_thandle, ot_super);
 177
 178         /* in some cases declare can race with creation (e.g. llog)
 179          * and we need to wait till object is initialized. notice
 180          * LOHA_EXISTs is supposed to be the last step in the
 181          * initialization */
 182
 183         /* size change (in dnode) will be declared by dmu_tx_hold_write() */
 184         if (dt_object_exists(dt))
 185                 oid = obj->oo_dn->dn_object;
 186         else
 187                 oid = DMU_NEW_OBJECT;
 188
 189         /* XXX: we still miss for append declaration support in ZFS
 190          *      -1 means append which is used by llog mostly, llog
 191          *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
 192         if (pos == -1)
 193                 pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
 194                             obj->oo_attr.la_size + (2 << 20));
 195         osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, pos, buf->lb_len);
 196
 197         /* dt_declare_write() is usually called for system objects, such
 198          * as llog or last_rcvd files. We needn't enforce quota on those
 199          * objects, so always set the lqi_space as 0. */
 200         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 201                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 202                                  0, oh, NULL, OSD_QID_BLK));
 203 }
 204
 205 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
 206                         const struct lu_buf *buf, loff_t *pos,
 207                          struct thandle *th)
 208 {
 209         struct osd_object  *obj  = osd_dt_obj(dt);
 210         struct osd_device  *osd = osd_obj2dev(obj);
 211         struct osd_thandle *oh;
 212         uint64_t            offset = *pos;
 213         int                 rc;
 214
 215         ENTRY;
 216
 217         LASSERT(dt_object_exists(dt));
 218         LASSERT(obj->oo_dn);
 219
 220         LASSERT(th != NULL);
 221         oh = container_of(th, struct osd_thandle, ot_super);
 222
 223         osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
 224                       buf->lb_buf, oh->ot_tx);
 225         write_lock(&obj->oo_attr_lock);
 226         if (obj->oo_attr.la_size < offset + buf->lb_len) {
 227                 obj->oo_attr.la_size = offset + buf->lb_len;
 228                 write_unlock(&obj->oo_attr_lock);
 229                 /* osd_object_sa_update() will be copying directly from oo_attr
 230                  * into dbuf.  any update within a single txg will copy the
 231                  * most actual */
 232                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 233                                         &obj->oo_attr.la_size, 8, oh);
 234                 if (unlikely(rc))
 235                         GOTO(out, rc);
 236         } else {
 237                 write_unlock(&obj->oo_attr_lock);
 238         }
 239
 240         *pos += buf->lb_len;
 241         rc = buf->lb_len;
 242
 243 out:
 244         RETURN(rc);
 245 }
 246
 247 /*
 248  * XXX: for the moment I don't want to use lnb_flags for osd-internal
 249  *      purposes as it's not very well defined ...
 250  *      instead I use the lowest bit of the address so that:
 251  *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
 252  *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
 253  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
 254  *
 255  *      bzzz, to blame
 256  */
 257 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 258                         struct niobuf_local *lnb, int npages)
 259 {
 260         struct osd_object *obj  = osd_dt_obj(dt);
 261         struct osd_device *osd = osd_obj2dev(obj);
 262         unsigned long      ptr;
 263         int                i;
 264
 265         LASSERT(dt_object_exists(dt));
 266         LASSERT(obj->oo_dn);
 267
 268         for (i = 0; i < npages; i++) {
 269                 if (lnb[i].lnb_page == NULL)
 270                         continue;
 271                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 272                         /* this is anonymous page allocated for copy-write */
 273                         lnb[i].lnb_page->mapping = NULL;
 274                         __free_page(lnb[i].lnb_page);
 275                         atomic_dec(&osd->od_zerocopy_alloc);
 276                 } else {
 277                         /* see comment in osd_bufs_get_read() */
 278                         ptr = (unsigned long)lnb[i].lnb_data;
 279                         if (ptr & 1UL) {
 280                                 ptr &= ~1UL;
 281                                 dmu_buf_rele((void *)ptr, osd_0copy_tag);
 282                                 atomic_dec(&osd->od_zerocopy_pin);
 283                         } else if (lnb[i].lnb_data != NULL) {
 284                                 int j, apages, abufsz;
 285                                 abufsz = arc_buf_size(lnb[i].lnb_data);
 286                                 apages = abufsz >> PAGE_SHIFT;
 287                                 /* these references to pages must be invalidated
 288                                  * to prevent access in osd_bufs_put() */
 289                                 for (j = 0; j < apages; j++)
 290                                         lnb[i + j].lnb_page = NULL;
 291                                 dmu_return_arcbuf(lnb[i].lnb_data);
 292                                 atomic_dec(&osd->od_zerocopy_loan);
 293                         }
 294                 }
 295                 lnb[i].lnb_page = NULL;
 296                 lnb[i].lnb_data = NULL;
 297         }
 298
 299         return 0;
 300 }
 301
 302 static inline struct page *kmem_to_page(void *addr)
 303 {
 304         LASSERT(!((unsigned long)addr & ~PAGE_MASK));
 305         if (is_vmalloc_addr(addr))
 306                 return vmalloc_to_page(addr);
 307         else
 308                 return virt_to_page(addr);
 309 }
 310
 311 /**
 312  * Prepare buffers for read.
 313  *
 314  * The function maps the range described by \a off and \a len to \a lnb array.
 315  * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
 316  * we fill \a lnb array with the pages storing ARC buffers. Notice the current
 317  * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
 318  * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
 319  * A better implementation would just return the buffers (potentially unfilled)
 320  * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
 321  *
 322  * \param[in] env       environment
 323  * \param[in] obj       object
 324  * \param[in] off       offset in bytes
 325  * \param[in] len       the number of bytes to access
 326  * \param[out] lnb      array of local niobufs pointing to the buffers with data
 327  *
 328  * \retval              0 for success
 329  * \retval              negative error number of failure
 330  */
 331 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 332                              loff_t off, ssize_t len, struct niobuf_local *lnb,
 333                              int maxlnb)
 334 {
 335         struct osd_device *osd = osd_obj2dev(obj);
 336         int rc, i, numbufs, npages = 0, drop_cache = 0;
 337         ktime_t start = ktime_get();
 338         dmu_buf_t **dbp;
 339         s64 delta_ms;
 340
 341         ENTRY;
 342         record_start_io(osd, READ, 0);
 343
 344         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
 345                 drop_cache = 1;
 346
 347         /* grab buffers for read:
 348          * OSD API let us to grab buffers first, then initiate IO(s)
 349          * so that all required IOs will be done in parallel, but at the
 350          * moment DMU doesn't provide us with a method to grab buffers.
 351          * If we discover this is a vital for good performance we
 352          * can get own replacement for dmu_buf_hold_array_by_bonus().
 353          */
 354         while (len > 0 &&
 355                (obj->oo_dn->dn_datablkshift != 0 ||
 356                 off < obj->oo_dn->dn_datablksz)) {
 357                 if (obj->oo_dn->dn_datablkshift == 0 &&
 358                     off + len > obj->oo_dn->dn_datablksz)
 359                         len = obj->oo_dn->dn_datablksz - off;
 360
 361                 dbp = NULL;
 362                 if (unlikely(npages >= maxlnb))
 363                         GOTO(err, rc = -EOVERFLOW);
 364
 365                 rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
 366                                                   off, len, TRUE, osd_0copy_tag,
 367                                                   &numbufs, &dbp);
 368                 if (unlikely(rc))
 369                         GOTO(err, rc);
 370
 371                 for (i = 0; i < numbufs; i++) {
 372                         int bufoff, tocpy, thispage;
 373                         void *dbf = dbp[i];
 374
 375                         LASSERT(len > 0);
 376
 377                         atomic_inc(&osd->od_zerocopy_pin);
 378
 379                         bufoff = off - dbp[i]->db_offset;
 380                         tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
 381
 382                         /* kind of trick to differentiate dbuf vs. arcbuf */
 383                         LASSERT(((unsigned long)dbp[i] & 1) == 0);
 384                         dbf = (void *) ((unsigned long)dbp[i] | 1);
 385
 386                         while (tocpy > 0) {
 387                                 if (unlikely(npages >= maxlnb))
 388                                         GOTO(err, rc = -EOVERFLOW);
 389
 390                                 thispage = PAGE_SIZE;
 391                                 thispage -= bufoff & (PAGE_SIZE - 1);
 392                                 thispage = min(tocpy, thispage);
 393
 394                                 lnb->lnb_rc = 0;
 395                                 lnb->lnb_file_offset = off;
 396                                 lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
 397                                 lnb->lnb_len = thispage;
 398                                 lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
 399                                                              bufoff);
 400                                 /* mark just a single slot: we need this
 401                                  * reference to dbuf to be released once */
 402                                 lnb->lnb_data = dbf;
 403                                 dbf = NULL;
 404
 405                                 tocpy -= thispage;
 406                                 len -= thispage;
 407                                 bufoff += thispage;
 408                                 off += thispage;
 409
 410                                 npages++;
 411                                 lnb++;
 412                         }
 413
 414                         if (drop_cache)
 415                                 dbuf_set_pending_evict(dbp[i]);
 416
 417                         /* steal dbuf so dmu_buf_rele_array() can't release
 418                          * it */
 419                         dbp[i] = NULL;
 420                 }
 421
 422                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 423         }
 424
 425         delta_ms = ktime_ms_delta(ktime_get(), start);
 426         record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
 427
 428         RETURN(npages);
 429
 430 err:
 431         LASSERT(rc < 0);
 432         if (dbp)
 433                 dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 434         osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
 435         RETURN(rc);
 436 }
 437
 438 static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
 439 {
 440         arc_buf_t *abuf;
 441
 442         abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
 443         if (unlikely(!abuf))
 444                 return ERR_PTR(-ENOMEM);
 445
 446 #if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
 447         /**
 448          * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
 449          * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
 450          * setting up page array for RDMA transfer. See LU-9305.
 451          */
 452         if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
 453                 dmu_return_arcbuf(abuf);
 454                 return NULL;
 455         }
 456 #endif
 457
 458         return abuf;
 459 }
 460
 461 static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 462                               loff_t off, ssize_t len, struct niobuf_local *lnb,
 463                               int maxlnb)
 464 {
 465         struct osd_device *osd = osd_obj2dev(obj);
 466         int                poff, plen, off_in_block, sz_in_block;
 467         int                rc, i = 0, npages = 0;
 468         dnode_t *dn = obj->oo_dn;
 469         arc_buf_t *abuf;
 470         uint32_t bs = dn->dn_datablksz;
 471         ENTRY;
 472
 473         /*
 474          * currently only full blocks are subject to zerocopy approach:
 475          * so that we're sure nobody is trying to update the same block
 476          */
 477         while (len > 0) {
 478                 if (unlikely(npages >= maxlnb))
 479                         GOTO(out_err, rc = -EOVERFLOW);
 480
 481                 off_in_block = off & (bs - 1);
 482                 sz_in_block = min_t(int, bs - off_in_block, len);
 483
 484                 abuf = NULL;
 485                 if (sz_in_block == bs) {
 486                         /* full block, try to use zerocopy */
 487                         abuf = osd_request_arcbuf(dn, bs);
 488                         if (unlikely(IS_ERR(abuf)))
 489                                 GOTO(out_err, rc = PTR_ERR(abuf));
 490                 }
 491
 492                 if (abuf != NULL) {
 493                         atomic_inc(&osd->od_zerocopy_loan);
 494
 495                         /* go over pages arcbuf contains, put them as
 496                          * local niobufs for ptlrpc's bulks */
 497                         while (sz_in_block > 0) {
 498                                 plen = min_t(int, sz_in_block, PAGE_SIZE);
 499
 500                                 if (unlikely(npages >= maxlnb))
 501                                         GOTO(out_err, rc = -EOVERFLOW);
 502
 503                                 lnb[i].lnb_file_offset = off;
 504                                 lnb[i].lnb_page_offset = 0;
 505                                 lnb[i].lnb_len = plen;
 506                                 lnb[i].lnb_rc = 0;
 507                                 if (sz_in_block == bs)
 508                                         lnb[i].lnb_data = abuf;
 509                                 else
 510                                         lnb[i].lnb_data = NULL;
 511
 512                                 /* this one is not supposed to fail */
 513                                 lnb[i].lnb_page = kmem_to_page(abuf->b_data +
 514                                                         off_in_block);
 515                                 LASSERT(lnb[i].lnb_page);
 516
 517                                 lprocfs_counter_add(osd->od_stats,
 518                                                 LPROC_OSD_ZEROCOPY_IO, 1);
 519
 520                                 sz_in_block -= plen;
 521                                 len -= plen;
 522                                 off += plen;
 523                                 off_in_block += plen;
 524                                 i++;
 525                                 npages++;
 526                         }
 527                 } else {
 528                         if (off_in_block == 0 && len < bs &&
 529                                         off + len >= obj->oo_attr.la_size)
 530                                 lprocfs_counter_add(osd->od_stats,
 531                                                 LPROC_OSD_TAIL_IO, 1);
 532
 533                         /* can't use zerocopy, allocate temp. buffers */
 534                         poff = off & (PAGE_SIZE - 1);
 535                         while (sz_in_block > 0) {
 536                                 plen = min_t(int, poff + sz_in_block,
 537                                              PAGE_SIZE);
 538                                 plen -= poff;
 539
 540                                 if (unlikely(npages >= maxlnb))
 541                                         GOTO(out_err, rc = -EOVERFLOW);
 542
 543                                 lnb[i].lnb_file_offset = off;
 544                                 lnb[i].lnb_page_offset = poff;
 545                                 poff = 0;
 546
 547                                 lnb[i].lnb_len = plen;
 548                                 lnb[i].lnb_rc = 0;
 549                                 lnb[i].lnb_data = NULL;
 550
 551                                 lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
 552                                 if (unlikely(lnb[i].lnb_page == NULL))
 553                                         GOTO(out_err, rc = -ENOMEM);
 554
 555                                 LASSERT(lnb[i].lnb_page->mapping == NULL);
 556                                 lnb[i].lnb_page->mapping = (void *)obj;
 557
 558                                 atomic_inc(&osd->od_zerocopy_alloc);
 559                                 lprocfs_counter_add(osd->od_stats,
 560                                                 LPROC_OSD_COPY_IO, 1);
 561
 562                                 sz_in_block -= plen;
 563                                 len -= plen;
 564                                 off += plen;
 565                                 i++;
 566                                 npages++;
 567                         }
 568                 }
 569         }
 570
 571         RETURN(npages);
 572
 573 out_err:
 574         osd_bufs_put(env, &obj->oo_dt, lnb, npages);
 575         RETURN(rc);
 576 }
 577
 578 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 579                         loff_t offset, ssize_t len, struct niobuf_local *lnb,
 580                         int maxlnb, enum dt_bufs_type rw)
 581 {
 582         struct osd_object *obj  = osd_dt_obj(dt);
 583         int                rc;
 584
 585         LASSERT(dt_object_exists(dt));
 586         LASSERT(obj->oo_dn);
 587
 588         if (rw & DT_BUFS_TYPE_WRITE)
 589                 rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
 590         else
 591                 rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
 592
 593         return rc;
 594 }
 595
 596 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 597                         struct niobuf_local *lnb, int npages)
 598 {
 599         struct osd_object *obj = osd_dt_obj(dt);
 600
 601         LASSERT(dt_object_exists(dt));
 602         LASSERT(obj->oo_dn);
 603
 604         return 0;
 605 }
 606
 607 static inline uint64_t osd_roundup2blocksz(uint64_t size,
 608                                            uint64_t offset,
 609                                            uint32_t blksz)
 610 {
 611         LASSERT(blksz > 0);
 612
 613         size += offset % blksz;
 614
 615         if (likely(is_power_of_2(blksz)))
 616                 return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 617
 618         size += blksz - 1;
 619         do_div(size, blksz);
 620         return size * blksz;
 621 }
 622
 623 static int osd_declare_write_commit(const struct lu_env *env,
 624                                     struct dt_object *dt,
 625                                     struct niobuf_local *lnb, int npages,
 626                                     struct thandle *th)
 627 {
 628         struct osd_object  *obj = osd_dt_obj(dt);
 629         struct osd_device  *osd = osd_obj2dev(obj);
 630         struct osd_thandle *oh;
 631         uint64_t            offset = 0;
 632         uint32_t            size = 0;
 633         uint32_t blksz = obj->oo_dn->dn_datablksz;
 634         int                 i, rc;
 635         bool synced = false;
 636         long long           space = 0;
 637         struct page        *last_page = NULL;
 638         unsigned long       discont_pages = 0;
 639         enum osd_quota_local_flags local_flags = 0;
 640         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
 641         ENTRY;
 642
 643         LASSERT(dt_object_exists(dt));
 644         LASSERT(obj->oo_dn);
 645
 646         LASSERT(lnb);
 647         LASSERT(npages > 0);
 648
 649         oh = container_of(th, struct osd_thandle, ot_super);
 650
 651         for (i = 0; i < npages; i++) {
 652                 if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
 653                         ++discont_pages;
 654                 last_page = lnb[i].lnb_page;
 655                 if (lnb[i].lnb_rc)
 656                         /* ENOSPC, network RPC error, etc.
 657                          * We don't want to book space for pages which will be
 658                          * skipped in osd_write_commit(). Hence we skip pages
 659                          * with lnb_rc != 0 here too */
 660                         continue;
 661                 /* ignore quota for the whole request if any page is from
 662                  * client cache or written by root.
 663                  *
 664                  * XXX once we drop the 1.8 client support, the checking
 665                  * for whether page is from cache can be simplified as:
 666                  * !(lnb[i].flags & OBD_BRW_SYNC)
 667                  *
 668                  * XXX we could handle this on per-lnb basis as done by
 669                  * grant. */
 670                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
 671                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 672                     OBD_BRW_FROM_GRANT)
 673                         declare_flags |= OSD_QID_FORCE;
 674
 675                 if (size == 0) {
 676                         /* first valid lnb */
 677                         offset = lnb[i].lnb_file_offset;
 678                         size = lnb[i].lnb_len;
 679                         continue;
 680                 }
 681                 if (offset + size == lnb[i].lnb_file_offset) {
 682                         /* this lnb is contiguous to the previous one */
 683                         size += lnb[i].lnb_len;
 684                         continue;
 685                 }
 686
 687                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
 688                                   obj->oo_dn, offset, size);
 689                 /* Estimating space to be consumed by a write is rather
 690                  * complicated with ZFS. As a consequence, we don't account for
 691                  * indirect blocks and just use as a rough estimate the worse
 692                  * case where the old space is being held by a snapshot. Quota
 693                  * overrun will be adjusted once the operation is committed, if
 694                  * required. */
 695                 space += osd_roundup2blocksz(size, offset, blksz);
 696
 697                 offset = lnb[i].lnb_file_offset;
 698                 size = lnb[i].lnb_len;
 699         }
 700
 701         if (size) {
 702                 osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
 703                                   offset, size);
 704                 space += osd_roundup2blocksz(size, offset, blksz);
 705         }
 706
 707         /* backend zfs filesystem might be configured to store multiple data
 708          * copies */
 709         space  *= osd->od_os->os_copies;
 710         space   = toqb(space);
 711         CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
 712                npages, space);
 713
 714         record_start_io(osd, WRITE, discont_pages);
 715 retry:
 716         /* acquire quota space if needed */
 717         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 718                                obj->oo_attr.la_gid, obj->oo_attr.la_projid,
 719                                space, oh, &local_flags, declare_flags);
 720
 721         if (!synced && rc == -EDQUOT &&
 722             (local_flags & QUOTA_FL_SYNC) != 0) {
 723                 dt_sync(env, th->th_dev);
 724                 synced = true;
 725                 CDEBUG(D_QUOTA, "retry after sync\n");
 726                 local_flags = 0;
 727                 goto retry;
 728         }
 729
 730         /* we need only to store the overquota flags in the first lnb for
 731          * now, once we support multiple objects BRW, this code needs be
 732          * revised. */
 733         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
 734                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
 735         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
 736                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 737 #ifdef ZFS_PROJINHERIT
 738         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
 739                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
 740 #endif
 741
 742         RETURN(rc);
 743 }
 744
 745 /**
 746  * Policy to grow ZFS block size by write pattern.
 747  * For sequential write, it grows block size gradually until it reaches the
 748  * maximum blocksize the dataset can support. Otherwise, it will pick a
 749  * a block size by the writing region of this I/O.
 750  */
 751 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
 752                               uint64_t start, uint64_t end)
 753 {
 754         struct osd_device       *osd = osd_obj2dev(obj);
 755         dnode_t *dn = obj->oo_dn;
 756         uint32_t                 blksz;
 757         int                      rc = 0;
 758
 759         ENTRY;
 760
 761         if (dn->dn_maxblkid > 0) /* can't change block size */
 762                 GOTO(out, rc);
 763
 764         if (dn->dn_datablksz >= osd->od_max_blksz)
 765                 GOTO(out, rc);
 766
 767         down_write(&obj->oo_guard);
 768
 769         blksz = dn->dn_datablksz;
 770         if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
 771                 GOTO(out_unlock, rc);
 772
 773         /* now ZFS can support up to 16MB block size, and if the write
 774          * is sequential, it just increases the block size gradually */
 775         if (start <= blksz) { /* sequential */
 776                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
 777         } else { /* sparse, pick a block size by write region */
 778                 blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
 779                                         end - start);
 780         }
 781
 782         if (!is_power_of_2(blksz))
 783                 blksz = size_roundup_power2(blksz);
 784
 785         if (blksz > dn->dn_datablksz) {
 786                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
 787                                                blksz, 0, oh->ot_tx);
 788                 LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
 789                 if (rc < 0)
 790                         CDEBUG(D_INODE, "object "DFID": change block size"
 791                                "%u -> %u error rc = %d\n",
 792                                PFID(lu_object_fid(&obj->oo_dt.do_lu)),
 793                                dn->dn_datablksz, blksz, rc);
 794         }
 795         EXIT;
 796 out_unlock:
 797         up_write(&obj->oo_guard);
 798 out:
 799         return rc;
 800 }
 801
 802 static void osd_evict_dbufs_after_write(struct osd_object *obj,
 803                                         loff_t off, ssize_t len)
 804 {
 805         dmu_buf_t **dbp;
 806         int i, rc, numbufs;
 807
 808         rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
 809                                           TRUE, osd_0copy_tag, &numbufs, &dbp);
 810         if (unlikely(rc))
 811                 return;
 812
 813         for (i = 0; i < numbufs; i++)
 814                 dbuf_set_pending_evict(dbp[i]);
 815
 816         dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
 817 }
 818
 819 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 820                         struct niobuf_local *lnb, int npages,
 821                         struct thandle *th, __u64 user_size)
 822 {
 823         struct osd_object  *obj  = osd_dt_obj(dt);
 824         struct osd_device  *osd = osd_obj2dev(obj);
 825         struct osd_thandle *oh;
 826         uint64_t            new_size = 0;
 827         int                 i, abufsz, rc = 0, drop_cache = 0;
 828         unsigned long      iosize = 0;
 829         ENTRY;
 830
 831         LASSERT(dt_object_exists(dt));
 832         LASSERT(obj->oo_dn);
 833
 834         LASSERT(th != NULL);
 835         oh = container_of(th, struct osd_thandle, ot_super);
 836
 837         /* adjust block size. Assume the buffers are sorted. */
 838         (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
 839                                  lnb[npages - 1].lnb_file_offset +
 840                                  lnb[npages - 1].lnb_len);
 841
 842         if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
 843             lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
 844             osd->od_readcache_max_filesize)
 845                 drop_cache = 1;
 846
 847         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
 848                 RETURN(-ENOSPC);
 849
 850         /* if la_size is already bigger than specified user_size,
 851          * ignore user_size
 852          */
 853         if (obj->oo_attr.la_size > user_size)
 854                 user_size = 0;
 855
 856         /* LU-8791: take oo_guard to avoid the deadlock that changing block
 857          * size and assigning arcbuf take place at the same time.
 858          *
 859          * Thread 1:
 860          * osd_write_commit()
 861          *  -> osd_grow_blocksize() with osd_object::oo_guard held
 862          *   -> dmu_object_set_blocksize()
 863          *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
 864          *       write lock held
 865          *     -> dbuf_new_size()
 866          *      -> dmu_buf_will_dirty()
 867          *       -> dbuf_read()
 868          *        -> wait for the dbuf state to change
 869          * Thread 2:
 870          * osd_write_commit()
 871          *  -> dmu_assign_arcbuf()
 872          *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
 873          *    -> dbuf_dirty()
 874          *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
 875          *
 876          * By taking the read lock, it can avoid thread 2 to enter into the
 877          * critical section of assigning the arcbuf, while thread 1 is
 878          * changing the block size.
 879          */
 880         down_read(&obj->oo_guard);
 881         for (i = 0; i < npages; i++) {
 882                 CDEBUG(D_INODE, "write %u bytes at %u\n",
 883                         (unsigned) lnb[i].lnb_len,
 884                         (unsigned) lnb[i].lnb_file_offset);
 885
 886                 if (lnb[i].lnb_rc) {
 887                         /* ENOSPC, network RPC error, etc.
 888                          * Unlike ldiskfs, zfs allocates new blocks on rewrite,
 889                          * so we skip this page if lnb_rc is set to -ENOSPC */
 890                         CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
 891                                 PFID(lu_object_fid(&dt->do_lu)), i,
 892                                 lnb[i].lnb_rc);
 893                         continue;
 894                 }
 895
 896                 if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
 897                         new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
 898                 if (lnb[i].lnb_page == NULL)
 899                         continue;
 900
 901                 if (lnb[i].lnb_page->mapping == (void *)obj) {
 902                         osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
 903                                       lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
 904                                       lnb[i].lnb_page_offset, oh->ot_tx);
 905                         kunmap(lnb[i].lnb_page);
 906                         iosize += lnb[i].lnb_len;
 907                         abufsz = lnb[i].lnb_len; /* to drop cache below */
 908                 } else if (lnb[i].lnb_data) {
 909                         int j, apages;
 910                         LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
 911                         /* buffer loaned for zerocopy, try to use it.
 912                          * notice that dmu_assign_arcbuf() is smart
 913                          * enough to recognize changed blocksize
 914                          * in this case it fallbacks to dmu_write() */
 915                         abufsz = arc_buf_size(lnb[i].lnb_data);
 916                         LASSERT(abufsz & PAGE_MASK);
 917                         apages = abufsz >> PAGE_SHIFT;
 918                         LASSERT(i + apages <= npages);
 919                         /* these references to pages must be invalidated
 920                          * to prevent access in osd_bufs_put() */
 921                         for (j = 0; j < apages; j++)
 922                                 lnb[i + j].lnb_page = NULL;
 923                         dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
 924                                           lnb[i].lnb_file_offset,
 925                                           lnb[i].lnb_data, oh->ot_tx);
 926                         /* drop the reference, otherwise osd_put_bufs()
 927                          * will be releasing it - bad! */
 928                         lnb[i].lnb_data = NULL;
 929                         atomic_dec(&osd->od_zerocopy_loan);
 930                         iosize += abufsz;
 931                 } else {
 932                         /* we don't want to deal with cache if nothing
 933                          * has been send to ZFS at this step */
 934                         continue;
 935                 }
 936
 937                 if (!drop_cache)
 938                         continue;
 939
 940                 /* we have to mark dbufs for eviction here because
 941                  * dmu_assign_arcbuf() may create a new dbuf for
 942                  * loaned abuf */
 943                 osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
 944                                             abufsz);
 945         }
 946         up_read(&obj->oo_guard);
 947
 948         if (unlikely(new_size == 0)) {
 949                 /* no pages to write, no transno is needed */
 950                 th->th_local = 1;
 951                 /* it is important to return 0 even when all lnb_rc == -ENOSPC
 952                  * since ofd_commitrw_write() retries several times on ENOSPC */
 953                 record_end_io(osd, WRITE, 0, 0, 0);
 954                 RETURN(0);
 955         }
 956
 957         /* if file has grown, take user_size into account */
 958         if (user_size && new_size > user_size)
 959                 new_size = user_size;
 960         write_lock(&obj->oo_attr_lock);
 961         if (obj->oo_attr.la_size < new_size) {
 962                 obj->oo_attr.la_size = new_size;
 963                 write_unlock(&obj->oo_attr_lock);
 964                 /* osd_object_sa_update() will be copying directly from
 965                  * oo_attr into dbuf. any update within a single txg will copy
 966                  * the most actual */
 967                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
 968                                           &obj->oo_attr.la_size, 8, oh);
 969         } else {
 970                 write_unlock(&obj->oo_attr_lock);
 971         }
 972
 973         record_end_io(osd, WRITE, 0, iosize, npages);
 974
 975         RETURN(rc);
 976 }
 977
 978 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 979                         struct niobuf_local *lnb, int npages)
 980 {
 981         struct osd_object *obj  = osd_dt_obj(dt);
 982         int                i;
 983         loff_t             eof;
 984
 985         LASSERT(dt_object_exists(dt));
 986         LASSERT(obj->oo_dn);
 987
 988         read_lock(&obj->oo_attr_lock);
 989         eof = obj->oo_attr.la_size;
 990         read_unlock(&obj->oo_attr_lock);
 991
 992         for (i = 0; i < npages; i++) {
 993                 if (unlikely(lnb[i].lnb_rc < 0))
 994                         continue;
 995
 996                 lnb[i].lnb_rc = lnb[i].lnb_len;
 997
 998                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
 999                         /* send complete pages all the time */
1000                         if (eof <= lnb[i].lnb_file_offset)
1001                                 lnb[i].lnb_rc = 0;
1002
1003                         /* all subsequent rc should be 0 */
1004                         while (++i < npages)
1005                                 lnb[i].lnb_rc = 0;
1006                         break;
1007                 }
1008         }
1009
1010         return 0;
1011 }
1012
1013 /*
1014  * Punch/truncate an object
1015  *
1016  *      IN:     db  - dmu_buf of the object to free data in.
1017  *              off - start of section to free.
1018  *              len - length of section to free (DMU_OBJECT_END => to EOF).
1019  *
1020  *      RETURN: 0 if success
1021  *              error code if failure
1022  *
1023  * The transaction passed to this routine must have
1024  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
1025  * called and then assigned to a transaction group.
1026  */
1027 static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
1028                                 uint64_t size, uint64_t off, uint64_t len)
1029 {
1030         int rc = 0;
1031
1032         /* Assert that the transaction has been assigned to a
1033            transaction group. */
1034         LASSERT(tx->tx_txg != 0);
1035         /*
1036          * Nothing to do if file already at desired length.
1037          */
1038         if (len == DMU_OBJECT_END && size == off)
1039                 return 0;
1040
1041         /* XXX: dnode_free_range() can be used to save on dnode lookup */
1042         if (off < size)
1043                 dmu_free_range(os, dn->dn_object, off, len, tx);
1044
1045         return rc;
1046 }
1047
1048 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1049                         __u64 start, __u64 end, struct thandle *th)
1050 {
1051         struct osd_object  *obj = osd_dt_obj(dt);
1052         struct osd_device  *osd = osd_obj2dev(obj);
1053         struct osd_thandle *oh;
1054         __u64               len;
1055         int                 rc = 0;
1056         ENTRY;
1057
1058         LASSERT(dt_object_exists(dt));
1059         LASSERT(osd_invariant(obj));
1060
1061         LASSERT(th != NULL);
1062         oh = container_of(th, struct osd_thandle, ot_super);
1063
1064         write_lock(&obj->oo_attr_lock);
1065         /* truncate */
1066         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1067                 len = DMU_OBJECT_END;
1068         else
1069                 len = end - start;
1070         write_unlock(&obj->oo_attr_lock);
1071
1072         rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
1073                                 obj->oo_attr.la_size, start, len);
1074         /* set new size */
1075         if (len == DMU_OBJECT_END) {
1076                 write_lock(&obj->oo_attr_lock);
1077                 obj->oo_attr.la_size = start;
1078                 write_unlock(&obj->oo_attr_lock);
1079                 rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
1080                                           &obj->oo_attr.la_size, 8, oh);
1081         }
1082         RETURN(rc);
1083 }
1084
1085 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1086                         __u64 start, __u64 end, struct thandle *handle)
1087 {
1088         struct osd_object  *obj = osd_dt_obj(dt);
1089         struct osd_device  *osd = osd_obj2dev(obj);
1090         struct osd_thandle *oh;
1091         __u64               len;
1092         ENTRY;
1093
1094         oh = container_of(handle, struct osd_thandle, ot_super);
1095
1096         read_lock(&obj->oo_attr_lock);
1097         if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
1098                 len = DMU_OBJECT_END;
1099         else
1100                 len = end - start;
1101
1102         /* declare we'll free some blocks ... */
1103         if (start < obj->oo_attr.la_size) {
1104                 read_unlock(&obj->oo_attr_lock);
1105                 dmu_tx_mark_netfree(oh->ot_tx);
1106                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
1107         } else {
1108                 read_unlock(&obj->oo_attr_lock);
1109         }
1110
1111         RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
1112                                  obj->oo_attr.la_gid, obj->oo_attr.la_projid,
1113                                  0, oh, NULL, OSD_QID_BLK));
1114 }
1115
1116 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1117                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1118 {
1119         int     rc;
1120         ENTRY;
1121
1122         switch (advice) {
1123         default:
1124                 rc = -ENOTSUPP;
1125                 break;
1126         }
1127
1128         RETURN(rc);
1129 }
1130
1131 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1132                          __u64 start, __u64 end, int mode, struct thandle *th)
1133 {
1134         int rc = -EOPNOTSUPP;
1135         ENTRY;
1136
1137          /*
1138           * space preallocation is not supported for ZFS
1139           * Returns -EOPNOTSUPP for now
1140           */
1141         RETURN(rc);
1142 }
1143
1144 static int osd_declare_fallocate(const struct lu_env *env,
1145                                  struct dt_object *dt, struct thandle *th)
1146 {
1147         int rc = -EOPNOTSUPP;
1148         ENTRY;
1149
1150          /*
1151           * space preallocation is not supported for ZFS
1152           * Returns -EOPNOTSUPP for now
1153           */
1154         RETURN(rc);
1155 }
1156
1157 struct dt_body_operations osd_body_ops = {
1158         .dbo_read                       = osd_read,
1159         .dbo_declare_write              = osd_declare_write,
1160         .dbo_write                      = osd_write,
1161         .dbo_bufs_get                   = osd_bufs_get,
1162         .dbo_bufs_put                   = osd_bufs_put,
1163         .dbo_write_prep                 = osd_write_prep,
1164         .dbo_declare_write_commit       = osd_declare_write_commit,
1165         .dbo_write_commit               = osd_write_commit,
1166         .dbo_read_prep                  = osd_read_prep,
1167         .dbo_declare_punch              = osd_declare_punch,
1168         .dbo_punch                      = osd_punch,
1169         .dbo_ladvise                    = osd_ladvise,
1170         .dbo_declare_fallocate          = osd_declare_fallocate,
1171         .dbo_fallocate                  = osd_fallocate,
1172 };
1173
1174 struct dt_body_operations osd_body_scrub_ops = {
1175         .dbo_read                       = osd_read_no_record,
1176         .dbo_declare_write              = osd_declare_write,
1177         .dbo_write                      = osd_write,
1178 };