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