lustre/osc/osc_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * Implementation of cl_page for OSC layer.
  40  *
  41  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_OSC
  45
  46 #include "osc_cl_internal.h"
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Comment out osc_page_protected because it may sleep inside the
  54  * the client_obd_list_lock.
  55  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  56  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  57  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  58  */
  59 #if 0
  60 static int osc_page_is_dlocked(const struct lu_env *env,
  61                                const struct osc_page *opg,
  62                                enum cl_lock_mode mode, int pending, int unref)
  63 {
  64         struct cl_page         *page;
  65         struct osc_object      *obj;
  66         struct osc_thread_info *info;
  67         struct ldlm_res_id     *resname;
  68         struct lustre_handle   *lockh;
  69         ldlm_policy_data_t     *policy;
  70         ldlm_mode_t             dlmmode;
  71         int                     flags;
  72
  73         cfs_might_sleep();
  74
  75         info = osc_env_info(env);
  76         resname = &info->oti_resname;
  77         policy = &info->oti_policy;
  78         lockh = &info->oti_handle;
  79         page = opg->ops_cl.cpl_page;
  80         obj = cl2osc(opg->ops_cl.cpl_obj);
  81
  82         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  83         if (pending)
  84                 flags |= LDLM_FL_CBPENDING;
  85
  86         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  87         osc_lock_build_res(env, obj, resname);
  88         osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
  89         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  90                               dlmmode, &flags, NULL, lockh, unref);
  91 }
  92
  93 /**
  94  * Checks an invariant that a page in the cache is covered by a lock, as
  95  * needed.
  96  */
  97 static int osc_page_protected(const struct lu_env *env,
  98                               const struct osc_page *opg,
  99                               enum cl_lock_mode mode, int unref)
 100 {
 101         struct cl_object_header *hdr;
 102         struct cl_lock          *scan;
 103         struct cl_page          *page;
 104         struct cl_lock_descr    *descr;
 105         int result;
 106
 107         LINVRNT(!opg->ops_temp);
 108
 109         page = opg->ops_cl.cpl_page;
 110         if (page->cp_owner != NULL &&
 111             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 112                 /*
 113                  * If IO is done without locks (liblustre, or lloop), lock is
 114                  * not required.
 115                  */
 116                 result = 1;
 117         else
 118                 /* otherwise check for a DLM lock */
 119         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 120         if (result == 0) {
 121                 /* maybe this page is a part of a lockless io? */
 122                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 123                 descr = &osc_env_info(env)->oti_descr;
 124                 descr->cld_mode = mode;
 125                 descr->cld_start = page->cp_index;
 126                 descr->cld_end   = page->cp_index;
 127                 cfs_spin_lock(&hdr->coh_lock_guard);
 128                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 129                         /*
 130                          * Lock-less sub-lock has to be either in HELD state
 131                          * (when io is actively going on), or in CACHED state,
 132                          * when top-lock is being unlocked:
 133                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 134                          */
 135                         if ((scan->cll_state == CLS_HELD ||
 136                              scan->cll_state == CLS_CACHED) &&
 137                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 138                                 struct osc_lock *olck;
 139
 140                                 olck = osc_lock_at(scan);
 141                                 result = osc_lock_is_lockless(olck);
 142                                 break;
 143                         }
 144                 }
 145                 cfs_spin_unlock(&hdr->coh_lock_guard);
 146         }
 147         return result;
 148 }
 149 #else
 150 static int osc_page_protected(const struct lu_env *env,
 151                               const struct osc_page *opg,
 152                               enum cl_lock_mode mode, int unref)
 153 {
 154         return 1;
 155 }
 156 #endif
 157
 158 /*****************************************************************************
 159  *
 160  * Page operations.
 161  *
 162  */
 163 static void osc_page_fini(const struct lu_env *env,
 164                           struct cl_page_slice *slice)
 165 {
 166         struct osc_page *opg = cl2osc_page(slice);
 167         CDEBUG(D_TRACE, "%p\n", opg);
 168         LASSERT(opg->ops_lock == NULL);
 169         OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
 170 }
 171
 172 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 173 {
 174         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 175
 176         LASSERT(!opg->ops_transfer_pinned);
 177         cl_page_get(page);
 178         lu_ref_add_atomic(&page->cp_reference, label, page);
 179         opg->ops_transfer_pinned = 1;
 180 }
 181
 182 static void osc_page_transfer_put(const struct lu_env *env,
 183                                   struct osc_page *opg)
 184 {
 185         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 186
 187         if (opg->ops_transfer_pinned) {
 188                 lu_ref_del(&page->cp_reference, "transfer", page);
 189                 opg->ops_transfer_pinned = 0;
 190                 cl_page_put(env, page);
 191         }
 192 }
 193
 194 /**
 195  * This is called once for every page when it is submitted for a transfer
 196  * either opportunistic (osc_page_cache_add()), or immediate
 197  * (osc_page_submit()).
 198  */
 199 static void osc_page_transfer_add(const struct lu_env *env,
 200                                   struct osc_page *opg, enum cl_req_type crt)
 201 {
 202         struct osc_object *obj;
 203
 204         obj = cl2osc(opg->ops_cl.cpl_obj);
 205         cfs_spin_lock(&obj->oo_seatbelt);
 206         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 207         opg->ops_submitter = cfs_current();
 208         cfs_spin_unlock(&obj->oo_seatbelt);
 209 }
 210
 211 static int osc_page_cache_add(const struct lu_env *env,
 212                               const struct cl_page_slice *slice,
 213                               struct cl_io *unused)
 214 {
 215         struct osc_page   *opg = cl2osc_page(slice);
 216         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 217         int result;
 218         /* All cacheable IO is async-capable */
 219         int brw_flags = OBD_BRW_ASYNC;
 220         int noquota = 0;
 221
 222         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 223         ENTRY;
 224
 225         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
 226         brw_flags |= opg->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
 227         if (!client_is_remote(osc_export(obj)) &&
 228             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 229                 brw_flags |= OBD_BRW_NOQUOTA;
 230                 noquota = OBD_BRW_NOQUOTA;
 231         }
 232
 233         osc_page_transfer_get(opg, "transfer\0cache");
 234         result = osc_queue_async_io(env, osc_export(obj), NULL, obj->oo_oinfo,
 235                                     &opg->ops_oap, OBD_BRW_WRITE | noquota,
 236                                     opg->ops_from, opg->ops_to - opg->ops_from,
 237                                     brw_flags, 0);
 238         if (result != 0)
 239                 osc_page_transfer_put(env, opg);
 240         else
 241                 osc_page_transfer_add(env, opg, CRT_WRITE);
 242         RETURN(result);
 243 }
 244
 245 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 246                       pgoff_t start, pgoff_t end)
 247 {
 248         memset(policy, 0, sizeof *policy);
 249         policy->l_extent.start = cl_offset(obj, start);
 250         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 251 }
 252
 253 static int osc_page_addref_lock(const struct lu_env *env,
 254                                 struct osc_page *opg,
 255                                 struct cl_lock *lock)
 256 {
 257         struct osc_lock *olock;
 258         int              rc;
 259
 260         LASSERT(opg->ops_lock == NULL);
 261
 262         olock = osc_lock_at(lock);
 263         if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
 264                 cfs_atomic_dec(&olock->ols_pageref);
 265                 cl_lock_put(env, lock);
 266                 rc = 1;
 267         } else {
 268                 opg->ops_lock = lock;
 269                 rc = 0;
 270         }
 271         return rc;
 272 }
 273
 274 static void osc_page_putref_lock(const struct lu_env *env,
 275                                  struct osc_page *opg)
 276 {
 277         struct cl_lock  *lock = opg->ops_lock;
 278         struct osc_lock *olock;
 279
 280         LASSERT(lock != NULL);
 281         olock = osc_lock_at(lock);
 282
 283         cfs_atomic_dec(&olock->ols_pageref);
 284         opg->ops_lock = NULL;
 285
 286         /*
 287          * Note: usually this won't be the last reference of the lock, but if
 288          * it is, then all the lock_put do is at most just freeing some memory,
 289          * so it would be OK that caller is holding spinlocks.
 290          */
 291         LASSERT(cfs_atomic_read(&lock->cll_ref) > 1 || olock->ols_hold == 0);
 292         cl_lock_put(env, lock);
 293 }
 294
 295 static int osc_page_is_under_lock(const struct lu_env *env,
 296                                   const struct cl_page_slice *slice,
 297                                   struct cl_io *unused)
 298 {
 299         struct cl_lock *lock;
 300         int             result;
 301
 302         ENTRY;
 303         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 304                                NULL, 1, 0);
 305         if (lock != NULL &&
 306             osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 307                 result = -EBUSY;
 308         else
 309                 result = -ENODATA;
 310         RETURN(result);
 311 }
 312
 313 static void osc_page_disown(const struct lu_env *env,
 314                             const struct cl_page_slice *slice,
 315                             struct cl_io *io)
 316 {
 317         struct osc_page *opg = cl2osc_page(slice);
 318
 319         if (unlikely(opg->ops_lock))
 320                 osc_page_putref_lock(env, opg);
 321 }
 322
 323 static void osc_page_completion_read(const struct lu_env *env,
 324                                      const struct cl_page_slice *slice,
 325                                      int ioret)
 326 {
 327         struct osc_page *opg = cl2osc_page(slice);
 328
 329         if (likely(opg->ops_lock))
 330                 osc_page_putref_lock(env, opg);
 331 }
 332
 333 static int osc_page_fail(const struct lu_env *env,
 334                          const struct cl_page_slice *slice,
 335                          struct cl_io *unused)
 336 {
 337         /*
 338          * Cached read?
 339          */
 340         LBUG();
 341         return 0;
 342 }
 343
 344
 345 static const char *osc_list(cfs_list_t *head)
 346 {
 347         return cfs_list_empty(head) ? "-" : "+";
 348 }
 349
 350 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 351 {
 352         if (opg->ops_submit_time == 0)
 353                 return 0;
 354
 355         return (cfs_time_current() - opg->ops_submit_time);
 356 }
 357
 358 static int osc_page_print(const struct lu_env *env,
 359                           const struct cl_page_slice *slice,
 360                           void *cookie, lu_printer_t printer)
 361 {
 362         struct osc_page       *opg = cl2osc_page(slice);
 363         struct osc_async_page *oap = &opg->ops_oap;
 364         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 365         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 366         struct lov_oinfo      *loi = obj->oo_oinfo;
 367
 368         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 369                           "1< %#x %d %u %s %s %s > "
 370                           "2< "LPU64" %u %u %#x %#x | %p %p %p %p %p > "
 371                           "3< %s %p %d %lu %d > "
 372                           "4< %d %d %d %lu %s | %s %s %s %s > "
 373                           "5< %s %s %s %s | %d %s %s | %d %s %s>\n",
 374                           opg,
 375                           /* 1 */
 376                           oap->oap_magic, oap->oap_cmd,
 377                           oap->oap_interrupted,
 378                           osc_list(&oap->oap_pending_item),
 379                           osc_list(&oap->oap_urgent_item),
 380                           osc_list(&oap->oap_rpc_item),
 381                           /* 2 */
 382                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 383                           oap->oap_async_flags, oap->oap_brw_flags,
 384                           oap->oap_request,
 385                           oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
 386                           oap->oap_caller_data,
 387                           /* 3 */
 388                           osc_list(&opg->ops_inflight),
 389                           opg->ops_submitter, opg->ops_transfer_pinned,
 390                           osc_submit_duration(opg), opg->ops_srvlock,
 391                           /* 4 */
 392                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 393                           cli->cl_max_rpcs_in_flight,
 394                           cli->cl_avail_grant,
 395                           osc_list(&cli->cl_cache_waiters),
 396                           osc_list(&cli->cl_loi_ready_list),
 397                           osc_list(&cli->cl_loi_hp_ready_list),
 398                           osc_list(&cli->cl_loi_write_list),
 399                           osc_list(&cli->cl_loi_read_list),
 400                           /* 5 */
 401                           osc_list(&loi->loi_ready_item),
 402                           osc_list(&loi->loi_hp_ready_item),
 403                           osc_list(&loi->loi_write_item),
 404                           osc_list(&loi->loi_read_item),
 405                           loi->loi_read_lop.lop_num_pending,
 406                           osc_list(&loi->loi_read_lop.lop_pending),
 407                           osc_list(&loi->loi_read_lop.lop_urgent),
 408                           loi->loi_write_lop.lop_num_pending,
 409                           osc_list(&loi->loi_write_lop.lop_pending),
 410                           osc_list(&loi->loi_write_lop.lop_urgent));
 411 }
 412
 413 static void osc_page_delete(const struct lu_env *env,
 414                             const struct cl_page_slice *slice)
 415 {
 416         struct osc_page       *opg = cl2osc_page(slice);
 417         struct osc_object     *obj = cl2osc(opg->ops_cl.cpl_obj);
 418         struct osc_async_page *oap = &opg->ops_oap;
 419         int rc;
 420
 421         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 422
 423         ENTRY;
 424         CDEBUG(D_TRACE, "%p\n", opg);
 425         osc_page_transfer_put(env, opg);
 426         rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
 427         if (rc) {
 428                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 429                               "Trying to teardown failed: %d\n", rc);
 430                 LASSERT(0);
 431         }
 432         cfs_spin_lock(&obj->oo_seatbelt);
 433         cfs_list_del_init(&opg->ops_inflight);
 434         cfs_spin_unlock(&obj->oo_seatbelt);
 435         EXIT;
 436 }
 437
 438 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 439                    int from, int to)
 440 {
 441         struct osc_page       *opg = cl2osc_page(slice);
 442         struct osc_async_page *oap = &opg->ops_oap;
 443
 444         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 445
 446         opg->ops_from = from;
 447         opg->ops_to   = to;
 448         cfs_spin_lock(&oap->oap_lock);
 449         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 450         cfs_spin_unlock(&oap->oap_lock);
 451 }
 452
 453 static int osc_page_cancel(const struct lu_env *env,
 454                            const struct cl_page_slice *slice)
 455 {
 456         struct osc_page *opg       = cl2osc_page(slice);
 457         struct osc_async_page *oap = &opg->ops_oap;
 458         int rc = 0;
 459
 460         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 461
 462         client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
 463         /* Check if the transferring against this page
 464          * is completed, or not even queued. */
 465         if (opg->ops_transfer_pinned)
 466                 /* FIXME: may not be interrupted.. */
 467                 rc = osc_oap_interrupted(env, oap);
 468         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 469         client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
 470         return rc;
 471 }
 472
 473 static const struct cl_page_operations osc_page_ops = {
 474         .cpo_fini          = osc_page_fini,
 475         .cpo_print         = osc_page_print,
 476         .cpo_delete        = osc_page_delete,
 477         .cpo_is_under_lock = osc_page_is_under_lock,
 478         .cpo_disown        = osc_page_disown,
 479         .io = {
 480                 [CRT_READ] = {
 481                         .cpo_cache_add  = osc_page_fail,
 482                         .cpo_completion = osc_page_completion_read
 483                 },
 484                 [CRT_WRITE] = {
 485                         .cpo_cache_add  = osc_page_cache_add
 486                 }
 487         },
 488         .cpo_clip           = osc_page_clip,
 489         .cpo_cancel         = osc_page_cancel
 490 };
 491
 492 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
 493 {
 494         struct osc_page *opg  = data;
 495         struct cl_page  *page = cl_page_top(opg->ops_cl.cpl_page);
 496         int result;
 497
 498         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
 499         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
 500
 501         ENTRY;
 502         result = cl_page_make_ready(env, page, CRT_WRITE);
 503         if (result == 0)
 504                 opg->ops_submit_time = cfs_time_current();
 505         RETURN(result);
 506 }
 507
 508 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
 509 {
 510         struct cl_page   *page;
 511         struct osc_page  *osc = data;
 512         struct cl_object *obj;
 513         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
 514
 515         int result;
 516         loff_t kms;
 517
 518         LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
 519
 520         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 521         LASSERT(!(cmd & OBD_BRW_READ));
 522         LASSERT(osc != NULL);
 523         page = osc->ops_cl.cpl_page;
 524         obj = osc->ops_cl.cpl_obj;
 525
 526         cl_object_attr_lock(obj);
 527         result = cl_object_attr_get(env, obj, attr);
 528         cl_object_attr_unlock(obj);
 529         if (result < 0)
 530                 return result;
 531         kms = attr->cat_kms;
 532         if (cl_offset(obj, page->cp_index) >= kms)
 533                 /* catch race with truncate */
 534                 return 0;
 535         else if (cl_offset(obj, page->cp_index + 1) > kms)
 536                 /* catch sub-page write at end of file */
 537                 return kms % CFS_PAGE_SIZE;
 538         else
 539                 return CFS_PAGE_SIZE;
 540 }
 541
 542 static int osc_completion(const struct lu_env *env,
 543                           void *data, int cmd, struct obdo *oa, int rc)
 544 {
 545         struct osc_page       *opg  = data;
 546         struct osc_async_page *oap  = &opg->ops_oap;
 547         struct cl_page        *page = cl_page_top(opg->ops_cl.cpl_page);
 548         struct osc_object     *obj  = cl2osc(opg->ops_cl.cpl_obj);
 549         enum cl_req_type crt;
 550         int srvlock;
 551
 552         LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
 553
 554         ENTRY;
 555
 556         cmd &= ~OBD_BRW_NOQUOTA;
 557         LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
 558         LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
 559         LASSERT(opg->ops_transfer_pinned);
 560
 561         /*
 562          * page->cp_req can be NULL if io submission failed before
 563          * cl_req was allocated.
 564          */
 565         if (page->cp_req != NULL)
 566                 cl_req_page_done(env, page);
 567         LASSERT(page->cp_req == NULL);
 568
 569         /* As the transfer for this page is being done, clear the flags */
 570         cfs_spin_lock(&oap->oap_lock);
 571         oap->oap_async_flags = 0;
 572         cfs_spin_unlock(&oap->oap_lock);
 573
 574         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
 575         /* Clear opg->ops_transfer_pinned before VM lock is released. */
 576         opg->ops_transfer_pinned = 0;
 577
 578         cfs_spin_lock(&obj->oo_seatbelt);
 579         LASSERT(opg->ops_submitter != NULL);
 580         LASSERT(!cfs_list_empty(&opg->ops_inflight));
 581         cfs_list_del_init(&opg->ops_inflight);
 582         cfs_spin_unlock(&obj->oo_seatbelt);
 583
 584         opg->ops_submit_time = 0;
 585         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
 586
 587         cl_page_completion(env, page, crt, rc);
 588
 589         /* statistic */
 590         if (rc == 0 && srvlock) {
 591                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
 592                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
 593                 int bytes = oap->oap_count;
 594
 595                 if (crt == CRT_READ)
 596                         stats->os_lockless_reads += bytes;
 597                 else
 598                         stats->os_lockless_writes += bytes;
 599         }
 600
 601         /*
 602          * This has to be the last operation with the page, as locks are
 603          * released in cl_page_completion() and nothing except for the
 604          * reference counter protects page from concurrent reclaim.
 605          */
 606         lu_ref_del(&page->cp_reference, "transfer", page);
 607         /*
 608          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 609          * safe to call cl_page_put() without risking object destruction in a
 610          * non-blocking context.
 611          */
 612         cl_page_put(env, page);
 613         RETURN(0);
 614 }
 615
 616 const static struct obd_async_page_ops osc_async_page_ops = {
 617         .ap_make_ready    = osc_make_ready,
 618         .ap_refresh_count = osc_refresh_count,
 619         .ap_completion    = osc_completion
 620 };
 621
 622 struct cl_page *osc_page_init(const struct lu_env *env,
 623                               struct cl_object *obj,
 624                               struct cl_page *page, cfs_page_t *vmpage)
 625 {
 626         struct osc_object *osc = cl2osc(obj);
 627         struct osc_page   *opg;
 628         int result;
 629
 630         OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
 631         if (opg != NULL) {
 632                 void *oap = &opg->ops_oap;
 633
 634                 opg->ops_from = 0;
 635                 opg->ops_to   = CFS_PAGE_SIZE;
 636
 637                 result = osc_prep_async_page(osc_export(osc),
 638                                              NULL, osc->oo_oinfo, vmpage,
 639                                              cl_offset(obj, page->cp_index),
 640                                              &osc_async_page_ops,
 641                                              opg, (void **)&oap, 1, NULL);
 642                 if (result == 0) {
 643                         struct osc_io *oio = osc_env_io(env);
 644                         opg->ops_srvlock = osc_io_srvlock(oio);
 645                         cl_page_slice_add(page, &opg->ops_cl, obj,
 646                                           &osc_page_ops);
 647                 }
 648                 /*
 649                  * Cannot assert osc_page_protected() here as read-ahead
 650                  * creates temporary pages outside of a lock.
 651                  */
 652 #ifdef INVARIANT_CHECK
 653                 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 654 #endif
 655                 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 656         } else
 657                 result = -ENOMEM;
 658         return ERR_PTR(result);
 659 }
 660
 661 /**
 662  * Helper function called by osc_io_submit() for every page in an immediate
 663  * transfer (i.e., transferred synchronously).
 664  */
 665 void osc_io_submit_page(const struct lu_env *env,
 666                         struct osc_io *oio, struct osc_page *opg,
 667                         enum cl_req_type crt)
 668 {
 669         struct osc_async_page *oap = &opg->ops_oap;
 670         struct client_obd     *cli = oap->oap_cli;
 671         int flags = 0;
 672
 673         LINVRNT(osc_page_protected(env, opg,
 674                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 675
 676         oap->oap_page_off   = opg->ops_from;
 677         oap->oap_count      = opg->ops_to - opg->ops_from;
 678         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
 679         if (cfs_memory_pressure_get())
 680                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
 681         oap->oap_brw_flags |= OBD_BRW_SYNC;
 682         if (osc_io_srvlock(oio))
 683                 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
 684
 685         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 686         if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
 687             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 688                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 689                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 690         }
 691
 692         if (oap->oap_cmd & OBD_BRW_READ)
 693                 flags = ASYNC_COUNT_STABLE;
 694         else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
 695                 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
 696
 697         cfs_spin_lock(&oap->oap_lock);
 698         oap->oap_async_flags |= OSC_FLAGS | flags;
 699         cfs_spin_unlock(&oap->oap_lock);
 700
 701         osc_oap_to_pending(oap);
 702         osc_page_transfer_get(opg, "transfer\0imm");
 703         osc_page_transfer_add(env, opg, crt);
 704 }
 705
 706 /** @} osc */