lustre/osc/osc_page.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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  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  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  49                            struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Comment out osc_page_protected because it may sleep inside the
  57  * the client_obd_list_lock.
  58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  61  */
  62 #if 0
  63 static int osc_page_is_dlocked(const struct lu_env *env,
  64                                const struct osc_page *opg,
  65                                enum cl_lock_mode mode, int pending, int unref)
  66 {
  67         struct cl_page         *page;
  68         struct osc_object      *obj;
  69         struct osc_thread_info *info;
  70         struct ldlm_res_id     *resname;
  71         struct lustre_handle   *lockh;
  72         ldlm_policy_data_t     *policy;
  73         ldlm_mode_t             dlmmode;
  74         __u64                   flags;
  75
  76         might_sleep();
  77
  78         info = osc_env_info(env);
  79         resname = &info->oti_resname;
  80         policy = &info->oti_policy;
  81         lockh = &info->oti_handle;
  82         page = opg->ops_cl.cpl_page;
  83         obj = cl2osc(opg->ops_cl.cpl_obj);
  84
  85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  86         if (pending)
  87                 flags |= LDLM_FL_CBPENDING;
  88
  89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  90         osc_lock_build_res(env, obj, resname);
  91         osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
  92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  93                               dlmmode, &flags, NULL, lockh, unref);
  94 }
  95
  96 /**
  97  * Checks an invariant that a page in the cache is covered by a lock, as
  98  * needed.
  99  */
 100 static int osc_page_protected(const struct lu_env *env,
 101                               const struct osc_page *opg,
 102                               enum cl_lock_mode mode, int unref)
 103 {
 104         struct cl_object_header *hdr;
 105         struct cl_lock          *scan;
 106         struct cl_page          *page;
 107         struct cl_lock_descr    *descr;
 108         int result;
 109
 110         LINVRNT(!opg->ops_temp);
 111
 112         page = opg->ops_cl.cpl_page;
 113         if (page->cp_owner != NULL &&
 114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 115                 /*
 116                  * If IO is done without locks (liblustre, or lloop), lock is
 117                  * not required.
 118                  */
 119                 result = 1;
 120         else
 121                 /* otherwise check for a DLM lock */
 122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 123         if (result == 0) {
 124                 /* maybe this page is a part of a lockless io? */
 125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 126                 descr = &osc_env_info(env)->oti_descr;
 127                 descr->cld_mode = mode;
 128                 descr->cld_start = page->cp_index;
 129                 descr->cld_end   = page->cp_index;
 130                 spin_lock(&hdr->coh_lock_guard);
 131                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 132                         /*
 133                          * Lock-less sub-lock has to be either in HELD state
 134                          * (when io is actively going on), or in CACHED state,
 135                          * when top-lock is being unlocked:
 136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 137                          */
 138                         if ((scan->cll_state == CLS_HELD ||
 139                              scan->cll_state == CLS_CACHED) &&
 140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 141                                 struct osc_lock *olck;
 142
 143                                 olck = osc_lock_at(scan);
 144                                 result = osc_lock_is_lockless(olck);
 145                                 break;
 146                         }
 147                 }
 148                 spin_unlock(&hdr->coh_lock_guard);
 149         }
 150         return result;
 151 }
 152 #else
 153 static int osc_page_protected(const struct lu_env *env,
 154                               const struct osc_page *opg,
 155                               enum cl_lock_mode mode, int unref)
 156 {
 157         return 1;
 158 }
 159 #endif
 160
 161 /*****************************************************************************
 162  *
 163  * Page operations.
 164  *
 165  */
 166 static void osc_page_fini(const struct lu_env *env,
 167                           struct cl_page_slice *slice)
 168 {
 169         struct osc_page *opg = cl2osc_page(slice);
 170         CDEBUG(D_TRACE, "%p\n", opg);
 171         LASSERT(opg->ops_lock == NULL);
 172 }
 173
 174 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 175 {
 176         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 177
 178         LASSERT(!opg->ops_transfer_pinned);
 179         cl_page_get(page);
 180         lu_ref_add_atomic(&page->cp_reference, label, page);
 181         opg->ops_transfer_pinned = 1;
 182 }
 183
 184 static void osc_page_transfer_put(const struct lu_env *env,
 185                                   struct osc_page *opg)
 186 {
 187         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 188
 189         if (opg->ops_transfer_pinned) {
 190                 lu_ref_del(&page->cp_reference, "transfer", page);
 191                 opg->ops_transfer_pinned = 0;
 192                 cl_page_put(env, page);
 193         }
 194 }
 195
 196 /**
 197  * This is called once for every page when it is submitted for a transfer
 198  * either opportunistic (osc_page_cache_add()), or immediate
 199  * (osc_page_submit()).
 200  */
 201 static void osc_page_transfer_add(const struct lu_env *env,
 202                                   struct osc_page *opg, enum cl_req_type crt)
 203 {
 204         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 205
 206         /* ops_lru and ops_inflight share the same field, so take it from LRU
 207          * first and then use it as inflight. */
 208         osc_lru_use(osc_cli(obj), opg);
 209
 210         spin_lock(&obj->oo_seatbelt);
 211         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 212         opg->ops_submitter = current;
 213         spin_unlock(&obj->oo_seatbelt);
 214 }
 215
 216 static int osc_page_cache_add(const struct lu_env *env,
 217                               const struct cl_page_slice *slice,
 218                               struct cl_io *io)
 219 {
 220         struct osc_io   *oio = osc_env_io(env);
 221         struct osc_page *opg = cl2osc_page(slice);
 222         int result;
 223         ENTRY;
 224
 225         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 226
 227         osc_page_transfer_get(opg, "transfer\0cache");
 228         result = osc_queue_async_io(env, io, opg);
 229         if (result != 0)
 230                 osc_page_transfer_put(env, opg);
 231         else
 232                 osc_page_transfer_add(env, opg, CRT_WRITE);
 233
 234         /* for sync write, kernel will wait for this page to be flushed before
 235          * osc_io_end() is called, so release it earlier.
 236          * for mkwrite(), it's known there is no further pages. */
 237         if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
 238                 if (oio->oi_active != NULL) {
 239                         osc_extent_release(env, oio->oi_active);
 240                         oio->oi_active = NULL;
 241                 }
 242         }
 243
 244         RETURN(result);
 245 }
 246
 247 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 248                       pgoff_t start, pgoff_t end)
 249 {
 250         memset(policy, 0, sizeof *policy);
 251         policy->l_extent.start = cl_offset(obj, start);
 252         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 253 }
 254
 255 static int osc_page_addref_lock(const struct lu_env *env,
 256                                 struct osc_page *opg,
 257                                 struct cl_lock *lock)
 258 {
 259         struct osc_lock *olock;
 260         int              rc;
 261
 262         LASSERT(opg->ops_lock == NULL);
 263
 264         olock = osc_lock_at(lock);
 265         if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
 266                 cfs_atomic_dec(&olock->ols_pageref);
 267                 rc = -ENODATA;
 268         } else {
 269                 cl_lock_get(lock);
 270                 opg->ops_lock = lock;
 271                 rc = 0;
 272         }
 273         return rc;
 274 }
 275
 276 static void osc_page_putref_lock(const struct lu_env *env,
 277                                  struct osc_page *opg)
 278 {
 279         struct cl_lock  *lock = opg->ops_lock;
 280         struct osc_lock *olock;
 281
 282         LASSERT(lock != NULL);
 283         olock = osc_lock_at(lock);
 284
 285         cfs_atomic_dec(&olock->ols_pageref);
 286         opg->ops_lock = NULL;
 287
 288         cl_lock_put(env, lock);
 289 }
 290
 291 static int osc_page_is_under_lock(const struct lu_env *env,
 292                                   const struct cl_page_slice *slice,
 293                                   struct cl_io *unused)
 294 {
 295         struct cl_lock *lock;
 296         int             result = -ENODATA;
 297
 298         ENTRY;
 299         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 300                                NULL, 1, 0);
 301         if (lock != NULL) {
 302                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 303                         result = -EBUSY;
 304                 cl_lock_put(env, lock);
 305         }
 306         RETURN(result);
 307 }
 308
 309 static void osc_page_disown(const struct lu_env *env,
 310                             const struct cl_page_slice *slice,
 311                             struct cl_io *io)
 312 {
 313         struct osc_page *opg = cl2osc_page(slice);
 314
 315         if (unlikely(opg->ops_lock))
 316                 osc_page_putref_lock(env, opg);
 317 }
 318
 319 static void osc_page_completion_read(const struct lu_env *env,
 320                                      const struct cl_page_slice *slice,
 321                                      int ioret)
 322 {
 323         struct osc_page   *opg = cl2osc_page(slice);
 324
 325         if (likely(opg->ops_lock))
 326                 osc_page_putref_lock(env, opg);
 327 }
 328
 329 static void osc_page_completion_write(const struct lu_env *env,
 330                                       const struct cl_page_slice *slice,
 331                                       int ioret)
 332 {
 333 }
 334
 335 static int osc_page_fail(const struct lu_env *env,
 336                          const struct cl_page_slice *slice,
 337                          struct cl_io *unused)
 338 {
 339         /*
 340          * Cached read?
 341          */
 342         LBUG();
 343         return 0;
 344 }
 345
 346
 347 static const char *osc_list(cfs_list_t *head)
 348 {
 349         return cfs_list_empty(head) ? "-" : "+";
 350 }
 351
 352 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 353 {
 354         if (opg->ops_submit_time == 0)
 355                 return 0;
 356
 357         return (cfs_time_current() - opg->ops_submit_time);
 358 }
 359
 360 static int osc_page_print(const struct lu_env *env,
 361                           const struct cl_page_slice *slice,
 362                           void *cookie, lu_printer_t printer)
 363 {
 364         struct osc_page       *opg = cl2osc_page(slice);
 365         struct osc_async_page *oap = &opg->ops_oap;
 366         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 367         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 368
 369         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 370                           "1< %#x %d %u %s %s > "
 371                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 372                           "3< %s %p %d %lu %d > "
 373                           "4< %d %d %d %lu %s | %s %s %s %s > "
 374                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 375                           opg,
 376                           /* 1 */
 377                           oap->oap_magic, oap->oap_cmd,
 378                           oap->oap_interrupted,
 379                           osc_list(&oap->oap_pending_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, oap->oap_cli, obj,
 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(&obj->oo_ready_item),
 400                           osc_list(&obj->oo_hp_ready_item),
 401                           osc_list(&obj->oo_write_item),
 402                           osc_list(&obj->oo_read_item),
 403                           cfs_atomic_read(&obj->oo_nr_reads),
 404                           osc_list(&obj->oo_reading_exts),
 405                           cfs_atomic_read(&obj->oo_nr_writes),
 406                           osc_list(&obj->oo_hp_exts),
 407                           osc_list(&obj->oo_urgent_exts));
 408 }
 409
 410 static void osc_page_delete(const struct lu_env *env,
 411                             const struct cl_page_slice *slice)
 412 {
 413         struct osc_page   *opg = cl2osc_page(slice);
 414         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 415         int rc;
 416
 417         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 418
 419         ENTRY;
 420         CDEBUG(D_TRACE, "%p\n", opg);
 421         osc_page_transfer_put(env, opg);
 422         rc = osc_teardown_async_page(env, obj, opg);
 423         if (rc) {
 424                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 425                               "Trying to teardown failed: %d\n", rc);
 426                 LASSERT(0);
 427         }
 428
 429         spin_lock(&obj->oo_seatbelt);
 430         if (opg->ops_submitter != NULL) {
 431                 LASSERT(!cfs_list_empty(&opg->ops_inflight));
 432                 cfs_list_del_init(&opg->ops_inflight);
 433                 opg->ops_submitter = NULL;
 434         }
 435         spin_unlock(&obj->oo_seatbelt);
 436
 437         osc_lru_del(osc_cli(obj), opg);
 438
 439         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 440                 void *value;
 441
 442                 spin_lock(&obj->oo_tree_lock);
 443                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 444                 if (value != NULL)
 445                         --obj->oo_npages;
 446                 spin_unlock(&obj->oo_tree_lock);
 447
 448                 LASSERT(ergo(value != NULL, value == opg));
 449         }
 450
 451         EXIT;
 452 }
 453
 454 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 455                    int from, int to)
 456 {
 457         struct osc_page       *opg = cl2osc_page(slice);
 458         struct osc_async_page *oap = &opg->ops_oap;
 459
 460         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 461
 462         opg->ops_from = from;
 463         opg->ops_to   = to;
 464         spin_lock(&oap->oap_lock);
 465         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 466         spin_unlock(&oap->oap_lock);
 467 }
 468
 469 static int osc_page_cancel(const struct lu_env *env,
 470                            const struct cl_page_slice *slice)
 471 {
 472         struct osc_page *opg = cl2osc_page(slice);
 473         int rc = 0;
 474
 475         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 476
 477         /* Check if the transferring against this page
 478          * is completed, or not even queued. */
 479         if (opg->ops_transfer_pinned)
 480                 /* FIXME: may not be interrupted.. */
 481                 rc = osc_cancel_async_page(env, opg);
 482         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 483         return rc;
 484 }
 485
 486 static int osc_page_flush(const struct lu_env *env,
 487                           const struct cl_page_slice *slice,
 488                           struct cl_io *io)
 489 {
 490         struct osc_page *opg = cl2osc_page(slice);
 491         int rc = 0;
 492         ENTRY;
 493         rc = osc_flush_async_page(env, io, opg);
 494         RETURN(rc);
 495 }
 496
 497 static const struct cl_page_operations osc_page_ops = {
 498         .cpo_fini          = osc_page_fini,
 499         .cpo_print         = osc_page_print,
 500         .cpo_delete        = osc_page_delete,
 501         .cpo_is_under_lock = osc_page_is_under_lock,
 502         .cpo_disown        = osc_page_disown,
 503         .io = {
 504                 [CRT_READ] = {
 505                         .cpo_cache_add  = osc_page_fail,
 506                         .cpo_completion = osc_page_completion_read
 507                 },
 508                 [CRT_WRITE] = {
 509                         .cpo_cache_add  = osc_page_cache_add,
 510                         .cpo_completion = osc_page_completion_write
 511                 }
 512         },
 513         .cpo_clip           = osc_page_clip,
 514         .cpo_cancel         = osc_page_cancel,
 515         .cpo_flush          = osc_page_flush
 516 };
 517
 518 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 519                   struct cl_page *page, struct page *vmpage)
 520 {
 521         struct osc_object *osc = cl2osc(obj);
 522         struct osc_page   *opg = cl_object_page_slice(obj, page);
 523         int result;
 524
 525         opg->ops_from = 0;
 526         opg->ops_to   = PAGE_CACHE_SIZE;
 527
 528         result = osc_prep_async_page(osc, opg, vmpage,
 529                                         cl_offset(obj, page->cp_index));
 530         if (result == 0) {
 531                 struct osc_io *oio = osc_env_io(env);
 532                 opg->ops_srvlock = osc_io_srvlock(oio);
 533                 cl_page_slice_add(page, &opg->ops_cl, obj,
 534                                 &osc_page_ops);
 535         }
 536         /*
 537          * Cannot assert osc_page_protected() here as read-ahead
 538          * creates temporary pages outside of a lock.
 539          */
 540 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 541         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 542 #endif
 543         /* ops_inflight and ops_lru are the same field, but it doesn't
 544          * hurt to initialize it twice :-) */
 545         CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 546         CFS_INIT_LIST_HEAD(&opg->ops_lru);
 547
 548         /* reserve an LRU space for this page */
 549         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 550                 result = osc_lru_reserve(env, osc, opg);
 551                 if (result == 0) {
 552                         spin_lock(&osc->oo_tree_lock);
 553                         result = radix_tree_insert(&osc->oo_tree,
 554                                                    page->cp_index, opg);
 555                         if (result == 0)
 556                                 ++osc->oo_npages;
 557                         spin_unlock(&osc->oo_tree_lock);
 558                         LASSERT(result == 0);
 559                 }
 560         }
 561
 562         return result;
 563 }
 564
 565 int osc_over_unstable_soft_limit(struct client_obd *cli)
 566 {
 567         long obd_upages, obd_dpages, osc_upages;
 568
 569         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 570         if (cli == NULL)
 571                 return 0;
 572
 573         obd_upages = cfs_atomic_read(&obd_unstable_pages);
 574         obd_dpages = cfs_atomic_read(&obd_dirty_pages);
 575
 576         osc_upages = cfs_atomic_read(&cli->cl_unstable_count);
 577
 578         /* obd_max_dirty_pages is the max number of (dirty + unstable)
 579          * pages allowed at any given time. To simulate an unstable page
 580          * only limit, we subtract the current number of dirty pages
 581          * from this max. This difference is roughly the amount of pages
 582          * currently available for unstable pages. Thus, the soft limit
 583          * is half of that difference. Check osc_upages to ensure we don't
 584          * set SOFT_SYNC for OSCs without any outstanding unstable pages. */
 585         return osc_upages != 0 &&
 586                obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
 587 }
 588
 589 /**
 590  * Helper function called by osc_io_submit() for every page in an immediate
 591  * transfer (i.e., transferred synchronously).
 592  */
 593 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 594                      enum cl_req_type crt, int brw_flags)
 595 {
 596         struct osc_async_page *oap = &opg->ops_oap;
 597         struct osc_object     *obj = oap->oap_obj;
 598
 599         LINVRNT(osc_page_protected(env, opg,
 600                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 601
 602         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 603                  "magic 0x%x\n", oap, oap->oap_magic);
 604         LASSERT(oap->oap_async_flags & ASYNC_READY);
 605         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 606
 607         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 608         oap->oap_page_off  = opg->ops_from;
 609         oap->oap_count     = opg->ops_to - opg->ops_from;
 610         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 611
 612         if (osc_over_unstable_soft_limit(oap->oap_cli))
 613                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
 614
 615         if (!client_is_remote(osc_export(obj)) &&
 616                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 617                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 618                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 619         }
 620
 621         opg->ops_submit_time = cfs_time_current();
 622         osc_page_transfer_get(opg, "transfer\0imm");
 623         osc_page_transfer_add(env, opg, crt);
 624 }
 625
 626 /* --------------- LRU page management ------------------ */
 627
 628 /* OSC is a natural place to manage LRU pages as applications are specialized
 629  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 630  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 631  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 632  * for free LRU slots - this will be very bad so the algorithm requires each
 633  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 634  * at any time.
 635  */
 636
 637 static CFS_DECL_WAITQ(osc_lru_waitq);
 638 /* LRU pages are freed in batch mode. OSC should at least free this
 639  * number of pages to avoid running out of LRU budget, and.. */
 640 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 641 /* free this number at most otherwise it will take too long time to finsih. */
 642 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 643
 644 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 645  * we should free slots aggressively. In this way, slots are freed in a steady
 646  * step to maintain fairness among OSCs.
 647  *
 648  * Return how many LRU pages should be freed. */
 649 static int osc_cache_too_much(struct client_obd *cli)
 650 {
 651         struct cl_client_cache *cache = cli->cl_cache;
 652         int pages = cfs_atomic_read(&cli->cl_lru_in_list);
 653         unsigned long budget;
 654
 655         budget = cache->ccc_lru_max / cfs_atomic_read(&cache->ccc_users);
 656
 657         /* if it's going to run out LRU slots, we should free some, but not
 658          * too much to maintain faireness among OSCs. */
 659         if (cfs_atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 660                 if (pages >= budget)
 661                         return lru_shrink_max;
 662                 else if (pages >= budget / 2)
 663                         return lru_shrink_min;
 664         } else if (pages >= budget * 2)
 665                 return lru_shrink_min;
 666         return 0;
 667 }
 668
 669 int lru_queue_work(const struct lu_env *env, void *data)
 670 {
 671         struct client_obd *cli = data;
 672
 673         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 674
 675         if (osc_cache_too_much(cli))
 676                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 677
 678         RETURN(0);
 679 }
 680
 681 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *plist)
 682 {
 683         CFS_LIST_HEAD(lru);
 684         struct osc_async_page *oap;
 685         int npages = 0;
 686
 687         cfs_list_for_each_entry(oap, plist, oap_pending_item) {
 688                 struct osc_page *opg = oap2osc_page(oap);
 689
 690                 if (!opg->ops_in_lru)
 691                         continue;
 692
 693                 ++npages;
 694                 LASSERT(cfs_list_empty(&opg->ops_lru));
 695                 cfs_list_add(&opg->ops_lru, &lru);
 696         }
 697
 698         if (npages > 0) {
 699                 client_obd_list_lock(&cli->cl_lru_list_lock);
 700                 cfs_list_splice_tail(&lru, &cli->cl_lru_list);
 701                 cfs_atomic_sub(npages, &cli->cl_lru_busy);
 702                 cfs_atomic_add(npages, &cli->cl_lru_in_list);
 703                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 704
 705                 /* XXX: May set force to be true for better performance */
 706                 if (osc_cache_too_much(cli))
 707                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 708         }
 709 }
 710
 711 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 712 {
 713         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) > 0);
 714         cfs_list_del_init(&opg->ops_lru);
 715         cfs_atomic_dec(&cli->cl_lru_in_list);
 716 }
 717
 718 /**
 719  * Page is being destroyed. The page may be not in LRU list, if the transfer
 720  * has never finished(error occurred).
 721  */
 722 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 723 {
 724         if (opg->ops_in_lru) {
 725                 client_obd_list_lock(&cli->cl_lru_list_lock);
 726                 if (!cfs_list_empty(&opg->ops_lru)) {
 727                         __osc_lru_del(cli, opg);
 728                 } else {
 729                         LASSERT(cfs_atomic_read(&cli->cl_lru_busy) > 0);
 730                         cfs_atomic_dec(&cli->cl_lru_busy);
 731                 }
 732                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 733
 734                 cfs_atomic_inc(cli->cl_lru_left);
 735                 /* this is a great place to release more LRU pages if
 736                  * this osc occupies too many LRU pages and kernel is
 737                  * stealing one of them. */
 738                 if (!memory_pressure_get())
 739                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 740                 wake_up(&osc_lru_waitq);
 741         } else {
 742                 LASSERT(cfs_list_empty(&opg->ops_lru));
 743         }
 744 }
 745
 746 /**
 747  * Delete page from LRUlist for redirty.
 748  */
 749 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 750 {
 751         /* If page is being transfered for the first time,
 752          * ops_lru should be empty */
 753         if (opg->ops_in_lru && !cfs_list_empty(&opg->ops_lru)) {
 754                 client_obd_list_lock(&cli->cl_lru_list_lock);
 755                 __osc_lru_del(cli, opg);
 756                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 757                 cfs_atomic_inc(&cli->cl_lru_busy);
 758         }
 759 }
 760
 761 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 762                                 struct cl_page **pvec, int max_index)
 763 {
 764         int i;
 765
 766         for (i = 0; i < max_index; i++) {
 767                 struct cl_page *page = pvec[i];
 768
 769                 LASSERT(cl_page_is_owned(page, io));
 770                 cl_page_discard(env, io, page);
 771                 cl_page_disown(env, io, page);
 772                 cl_page_put(env, page);
 773
 774                 pvec[i] = NULL;
 775         }
 776 }
 777
 778 /**
 779  * Drop @target of pages from LRU at most.
 780  */
 781 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 782                    int target, bool force)
 783 {
 784         struct cl_io *io;
 785         struct cl_object *clobj = NULL;
 786         struct cl_page **pvec;
 787         struct osc_page *opg;
 788         int maxscan = 0;
 789         int count = 0;
 790         int index = 0;
 791         int rc = 0;
 792         ENTRY;
 793
 794         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) >= 0);
 795         if (cfs_atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 796                 RETURN(0);
 797
 798         if (!force) {
 799                 if (cfs_atomic_read(&cli->cl_lru_shrinkers) > 0)
 800                         RETURN(-EBUSY);
 801
 802                 if (cfs_atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 803                         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 804                         RETURN(-EBUSY);
 805                 }
 806         } else {
 807                 cfs_atomic_inc(&cli->cl_lru_shrinkers);
 808         }
 809
 810         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 811         io = &osc_env_info(env)->oti_io;
 812
 813         client_obd_list_lock(&cli->cl_lru_list_lock);
 814         maxscan = min(target << 1, cfs_atomic_read(&cli->cl_lru_in_list));
 815         while (!cfs_list_empty(&cli->cl_lru_list)) {
 816                 struct cl_page *page;
 817                 bool will_free = false;
 818
 819                 if (--maxscan < 0)
 820                         break;
 821
 822                 opg = cfs_list_entry(cli->cl_lru_list.next, struct osc_page,
 823                                      ops_lru);
 824                 page = cl_page_top(opg->ops_cl.cpl_page);
 825                 if (cl_page_in_use_noref(page)) {
 826                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 827                         continue;
 828                 }
 829
 830                 LASSERT(page->cp_obj != NULL);
 831                 if (clobj != page->cp_obj) {
 832                         struct cl_object *tmp = page->cp_obj;
 833
 834                         cl_object_get(tmp);
 835                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 836
 837                         if (clobj != NULL) {
 838                                 discard_pagevec(env, io, pvec, index);
 839                                 index = 0;
 840
 841                                 cl_io_fini(env, io);
 842                                 cl_object_put(env, clobj);
 843                                 clobj = NULL;
 844                         }
 845
 846                         clobj = tmp;
 847                         io->ci_obj = clobj;
 848                         io->ci_ignore_layout = 1;
 849                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 850
 851                         client_obd_list_lock(&cli->cl_lru_list_lock);
 852
 853                         if (rc != 0)
 854                                 break;
 855
 856                         ++maxscan;
 857                         continue;
 858                 }
 859
 860                 if (cl_page_own_try(env, io, page) == 0) {
 861                         if (!cl_page_in_use_noref(page)) {
 862                                 /* remove it from lru list earlier to avoid
 863                                  * lock contention */
 864                                 __osc_lru_del(cli, opg);
 865                                 opg->ops_in_lru = 0; /* will be discarded */
 866
 867                                 cl_page_get(page);
 868                                 will_free = true;
 869                         } else {
 870                                 cl_page_disown(env, io, page);
 871                         }
 872                 }
 873
 874                 if (!will_free) {
 875                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 876                         continue;
 877                 }
 878
 879                 /* Don't discard and free the page with cl_lru_list held */
 880                 pvec[index++] = page;
 881                 if (unlikely(index == OTI_PVEC_SIZE)) {
 882                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 883                         discard_pagevec(env, io, pvec, index);
 884                         index = 0;
 885
 886                         client_obd_list_lock(&cli->cl_lru_list_lock);
 887                 }
 888
 889                 if (++count >= target)
 890                         break;
 891         }
 892         client_obd_list_unlock(&cli->cl_lru_list_lock);
 893
 894         if (clobj != NULL) {
 895                 discard_pagevec(env, io, pvec, index);
 896
 897                 cl_io_fini(env, io);
 898                 cl_object_put(env, clobj);
 899         }
 900
 901         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 902         if (count > 0) {
 903                 cfs_atomic_add(count, cli->cl_lru_left);
 904                 wake_up_all(&osc_lru_waitq);
 905         }
 906         RETURN(count > 0 ? count : rc);
 907 }
 908
 909 static inline int max_to_shrink(struct client_obd *cli)
 910 {
 911         return min(cfs_atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 912 }
 913
 914 int osc_lru_reclaim(struct client_obd *cli)
 915 {
 916         struct cl_env_nest nest;
 917         struct lu_env *env;
 918         struct cl_client_cache *cache = cli->cl_cache;
 919         int max_scans;
 920         int rc = 0;
 921         ENTRY;
 922
 923         LASSERT(cache != NULL);
 924         LASSERT(!cfs_list_empty(&cache->ccc_lru));
 925
 926         env = cl_env_nested_get(&nest);
 927         if (IS_ERR(env))
 928                 RETURN(rc);
 929
 930         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 931         if (rc != 0) {
 932                 if (rc == -EBUSY)
 933                         rc = 0;
 934
 935                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 936                         cli->cl_import->imp_obd->obd_name, rc, cli);
 937                 GOTO(out, rc);
 938         }
 939
 940         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 941                 cli->cl_import->imp_obd->obd_name, cli,
 942                 cfs_atomic_read(&cli->cl_lru_in_list),
 943                 cfs_atomic_read(&cli->cl_lru_busy));
 944
 945         /* Reclaim LRU slots from other client_obd as it can't free enough
 946          * from its own. This should rarely happen. */
 947         spin_lock(&cache->ccc_lru_lock);
 948         cache->ccc_lru_shrinkers++;
 949         cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 950
 951         max_scans = cfs_atomic_read(&cache->ccc_users);
 952         while (--max_scans > 0 && !cfs_list_empty(&cache->ccc_lru)) {
 953                 cli = cfs_list_entry(cache->ccc_lru.next, struct client_obd,
 954                                         cl_lru_osc);
 955
 956                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 957                         cli->cl_import->imp_obd->obd_name, cli,
 958                         cfs_atomic_read(&cli->cl_lru_in_list),
 959                         cfs_atomic_read(&cli->cl_lru_busy));
 960
 961                 cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 962                 if (osc_cache_too_much(cli) > 0) {
 963                         spin_unlock(&cache->ccc_lru_lock);
 964
 965                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 966                                             true);
 967                         spin_lock(&cache->ccc_lru_lock);
 968                         if (rc != 0)
 969                                 break;
 970                 }
 971         }
 972         spin_unlock(&cache->ccc_lru_lock);
 973
 974 out:
 975         cl_env_nested_put(&nest, env);
 976         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 977                 cli->cl_import->imp_obd->obd_name, cli, rc);
 978         return rc;
 979 }
 980
 981 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 982                            struct osc_page *opg)
 983 {
 984         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 985         struct osc_io *oio = osc_env_io(env);
 986         struct client_obd *cli = osc_cli(obj);
 987         int rc = 0;
 988         ENTRY;
 989
 990         if (cli->cl_cache == NULL) /* shall not be in LRU */
 991                 RETURN(0);
 992
 993         if (oio->oi_lru_reserved > 0) {
 994                 --oio->oi_lru_reserved;
 995                 goto out;
 996         }
 997
 998         LASSERT(cfs_atomic_read(cli->cl_lru_left) >= 0);
 999         while (!cfs_atomic_add_unless(cli->cl_lru_left, -1, 0)) {
1000
1001                 /* run out of LRU spaces, try to drop some by itself */
1002                 rc = osc_lru_reclaim(cli);
1003                 if (rc < 0)
1004                         break;
1005                 if (rc > 0)
1006                         continue;
1007
1008                 cond_resched();
1009                 rc = l_wait_event(osc_lru_waitq,
1010                                 cfs_atomic_read(cli->cl_lru_left) > 0,
1011                                 &lwi);
1012                 if (rc < 0)
1013                         break;
1014         }
1015
1016 out:
1017         if (rc >= 0) {
1018                 cfs_atomic_inc(&cli->cl_lru_busy);
1019                 opg->ops_in_lru = 1;
1020                 rc = 0;
1021         }
1022
1023         RETURN(rc);
1024 }
1025
1026 /** @} osc */