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