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  */
  40
  41 #define DEBUG_SUBSYSTEM S_OSC
  42
  43 #include "osc_cl_internal.h"
  44
  45 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  46 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  47 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  48                            struct osc_page *opg);
  49
  50 /** \addtogroup osc
  51  *  @{
  52  */
  53
  54 /*
  55  * Comment out osc_page_protected because it may sleep inside the
  56  * the client_obd_list_lock.
  57  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  58  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  59  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  60  */
  61 #if 0
  62 static int osc_page_is_dlocked(const struct lu_env *env,
  63                                const struct osc_page *opg,
  64                                enum cl_lock_mode mode, int pending, int unref)
  65 {
  66         struct cl_page         *page;
  67         struct osc_object      *obj;
  68         struct osc_thread_info *info;
  69         struct ldlm_res_id     *resname;
  70         struct lustre_handle   *lockh;
  71         ldlm_policy_data_t     *policy;
  72         ldlm_mode_t             dlmmode;
  73         __u64                   flags;
  74
  75         might_sleep();
  76
  77         info = osc_env_info(env);
  78         resname = &info->oti_resname;
  79         policy = &info->oti_policy;
  80         lockh = &info->oti_handle;
  81         page = opg->ops_cl.cpl_page;
  82         obj = cl2osc(opg->ops_cl.cpl_obj);
  83
  84         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  85         if (pending)
  86                 flags |= LDLM_FL_CBPENDING;
  87
  88         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  89         osc_lock_build_res(env, obj, resname);
  90         osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
  91         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  92                               dlmmode, &flags, NULL, lockh, unref);
  93 }
  94
  95 /**
  96  * Checks an invariant that a page in the cache is covered by a lock, as
  97  * needed.
  98  */
  99 static int osc_page_protected(const struct lu_env *env,
 100                               const struct osc_page *opg,
 101                               enum cl_lock_mode mode, int unref)
 102 {
 103         struct cl_object_header *hdr;
 104         struct cl_lock          *scan;
 105         struct cl_page          *page;
 106         struct cl_lock_descr    *descr;
 107         int result;
 108
 109         LINVRNT(!opg->ops_temp);
 110
 111         page = opg->ops_cl.cpl_page;
 112         if (page->cp_owner != NULL &&
 113             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 114                 /*
 115                  * If IO is done without locks (liblustre, or lloop), lock is
 116                  * not required.
 117                  */
 118                 result = 1;
 119         else
 120                 /* otherwise check for a DLM lock */
 121         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 122         if (result == 0) {
 123                 /* maybe this page is a part of a lockless io? */
 124                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 125                 descr = &osc_env_info(env)->oti_descr;
 126                 descr->cld_mode = mode;
 127                 descr->cld_start = page->cp_index;
 128                 descr->cld_end   = page->cp_index;
 129                 spin_lock(&hdr->coh_lock_guard);
 130                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 131                         /*
 132                          * Lock-less sub-lock has to be either in HELD state
 133                          * (when io is actively going on), or in CACHED state,
 134                          * when top-lock is being unlocked:
 135                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 136                          */
 137                         if ((scan->cll_state == CLS_HELD ||
 138                              scan->cll_state == CLS_CACHED) &&
 139                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 140                                 struct osc_lock *olck;
 141
 142                                 olck = osc_lock_at(scan);
 143                                 result = osc_lock_is_lockless(olck);
 144                                 break;
 145                         }
 146                 }
 147                 spin_unlock(&hdr->coh_lock_guard);
 148         }
 149         return result;
 150 }
 151 #else
 152 static int osc_page_protected(const struct lu_env *env,
 153                               const struct osc_page *opg,
 154                               enum cl_lock_mode mode, int unref)
 155 {
 156         return 1;
 157 }
 158 #endif
 159
 160 /*****************************************************************************
 161  *
 162  * Page operations.
 163  *
 164  */
 165 static void osc_page_fini(const struct lu_env *env,
 166                           struct cl_page_slice *slice)
 167 {
 168         struct osc_page *opg = cl2osc_page(slice);
 169         CDEBUG(D_TRACE, "%p\n", opg);
 170         LASSERT(opg->ops_lock == NULL);
 171 }
 172
 173 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 174 {
 175         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 176
 177         LASSERT(!opg->ops_transfer_pinned);
 178         cl_page_get(page);
 179         lu_ref_add_atomic(&page->cp_reference, label, page);
 180         opg->ops_transfer_pinned = 1;
 181 }
 182
 183 static void osc_page_transfer_put(const struct lu_env *env,
 184                                   struct osc_page *opg)
 185 {
 186         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 187
 188         if (opg->ops_transfer_pinned) {
 189                 lu_ref_del(&page->cp_reference, "transfer", page);
 190                 opg->ops_transfer_pinned = 0;
 191                 cl_page_put(env, page);
 192         }
 193 }
 194
 195 /**
 196  * This is called once for every page when it is submitted for a transfer
 197  * either opportunistic (osc_page_cache_add()), or immediate
 198  * (osc_page_submit()).
 199  */
 200 static void osc_page_transfer_add(const struct lu_env *env,
 201                                   struct osc_page *opg, enum cl_req_type crt)
 202 {
 203         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 204
 205         /* ops_lru and ops_inflight share the same field, so take it from LRU
 206          * first and then use it as inflight. */
 207         osc_lru_use(osc_cli(obj), opg);
 208
 209         spin_lock(&obj->oo_seatbelt);
 210         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 211         opg->ops_submitter = current;
 212         spin_unlock(&obj->oo_seatbelt);
 213 }
 214
 215 static int osc_page_cache_add(const struct lu_env *env,
 216                               const struct cl_page_slice *slice,
 217                               struct cl_io *io)
 218 {
 219         struct osc_io   *oio = osc_env_io(env);
 220         struct osc_page *opg = cl2osc_page(slice);
 221         int result;
 222         ENTRY;
 223
 224         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 225
 226         osc_page_transfer_get(opg, "transfer\0cache");
 227         result = osc_queue_async_io(env, io, opg);
 228         if (result != 0)
 229                 osc_page_transfer_put(env, opg);
 230         else
 231                 osc_page_transfer_add(env, opg, CRT_WRITE);
 232
 233         /* for sync write, kernel will wait for this page to be flushed before
 234          * osc_io_end() is called, so release it earlier.
 235          * for mkwrite(), it's known there is no further pages. */
 236         if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
 237                 if (oio->oi_active != NULL) {
 238                         osc_extent_release(env, oio->oi_active);
 239                         oio->oi_active = NULL;
 240                 }
 241         }
 242
 243         RETURN(result);
 244 }
 245
 246 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 247                       pgoff_t start, pgoff_t end)
 248 {
 249         memset(policy, 0, sizeof *policy);
 250         policy->l_extent.start = cl_offset(obj, start);
 251         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 252 }
 253
 254 static int osc_page_addref_lock(const struct lu_env *env,
 255                                 struct osc_page *opg,
 256                                 struct cl_lock *lock)
 257 {
 258         struct osc_lock *olock;
 259         int              rc;
 260
 261         LASSERT(opg->ops_lock == NULL);
 262
 263         olock = osc_lock_at(lock);
 264         if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
 265                 cfs_atomic_dec(&olock->ols_pageref);
 266                 rc = -ENODATA;
 267         } else {
 268                 cl_lock_get(lock);
 269                 opg->ops_lock = lock;
 270                 rc = 0;
 271         }
 272         return rc;
 273 }
 274
 275 static void osc_page_putref_lock(const struct lu_env *env,
 276                                  struct osc_page *opg)
 277 {
 278         struct cl_lock  *lock = opg->ops_lock;
 279         struct osc_lock *olock;
 280
 281         LASSERT(lock != NULL);
 282         olock = osc_lock_at(lock);
 283
 284         cfs_atomic_dec(&olock->ols_pageref);
 285         opg->ops_lock = NULL;
 286
 287         cl_lock_put(env, lock);
 288 }
 289
 290 static int osc_page_is_under_lock(const struct lu_env *env,
 291                                   const struct cl_page_slice *slice,
 292                                   struct cl_io *unused)
 293 {
 294         struct cl_lock *lock;
 295         int             result = -ENODATA;
 296
 297         ENTRY;
 298         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 299                                NULL, 1, 0);
 300         if (lock != NULL) {
 301                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 302                         result = -EBUSY;
 303                 cl_lock_put(env, lock);
 304         }
 305         RETURN(result);
 306 }
 307
 308 static void osc_page_disown(const struct lu_env *env,
 309                             const struct cl_page_slice *slice,
 310                             struct cl_io *io)
 311 {
 312         struct osc_page *opg = cl2osc_page(slice);
 313
 314         if (unlikely(opg->ops_lock))
 315                 osc_page_putref_lock(env, opg);
 316 }
 317
 318 static void osc_page_completion_read(const struct lu_env *env,
 319                                      const struct cl_page_slice *slice,
 320                                      int ioret)
 321 {
 322         struct osc_page   *opg = cl2osc_page(slice);
 323
 324         if (likely(opg->ops_lock))
 325                 osc_page_putref_lock(env, opg);
 326 }
 327
 328 static void osc_page_completion_write(const struct lu_env *env,
 329                                       const struct cl_page_slice *slice,
 330                                       int ioret)
 331 {
 332 }
 333
 334 static int osc_page_fail(const struct lu_env *env,
 335                          const struct cl_page_slice *slice,
 336                          struct cl_io *unused)
 337 {
 338         /*
 339          * Cached read?
 340          */
 341         LBUG();
 342         return 0;
 343 }
 344
 345
 346 static const char *osc_list(cfs_list_t *head)
 347 {
 348         return cfs_list_empty(head) ? "-" : "+";
 349 }
 350
 351 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 352 {
 353         if (opg->ops_submit_time == 0)
 354                 return 0;
 355
 356         return (cfs_time_current() - opg->ops_submit_time);
 357 }
 358
 359 static int osc_page_print(const struct lu_env *env,
 360                           const struct cl_page_slice *slice,
 361                           void *cookie, lu_printer_t printer)
 362 {
 363         struct osc_page       *opg = cl2osc_page(slice);
 364         struct osc_async_page *oap = &opg->ops_oap;
 365         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 366         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 367
 368         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 369                           "1< %#x %d %u %s %s > "
 370                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 371                           "3< %s %p %d %lu %d > "
 372                           "4< %d %d %d %lu %s | %s %s %s %s > "
 373                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 374                           opg,
 375                           /* 1 */
 376                           oap->oap_magic, oap->oap_cmd,
 377                           oap->oap_interrupted,
 378                           osc_list(&oap->oap_pending_item),
 379                           osc_list(&oap->oap_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, oap->oap_cli, obj,
 384                           /* 3 */
 385                           osc_list(&opg->ops_inflight),
 386                           opg->ops_submitter, opg->ops_transfer_pinned,
 387                           osc_submit_duration(opg), opg->ops_srvlock,
 388                           /* 4 */
 389                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 390                           cli->cl_max_rpcs_in_flight,
 391                           cli->cl_avail_grant,
 392                           osc_list(&cli->cl_cache_waiters),
 393                           osc_list(&cli->cl_loi_ready_list),
 394                           osc_list(&cli->cl_loi_hp_ready_list),
 395                           osc_list(&cli->cl_loi_write_list),
 396                           osc_list(&cli->cl_loi_read_list),
 397                           /* 5 */
 398                           osc_list(&obj->oo_ready_item),
 399                           osc_list(&obj->oo_hp_ready_item),
 400                           osc_list(&obj->oo_write_item),
 401                           osc_list(&obj->oo_read_item),
 402                           cfs_atomic_read(&obj->oo_nr_reads),
 403                           osc_list(&obj->oo_reading_exts),
 404                           cfs_atomic_read(&obj->oo_nr_writes),
 405                           osc_list(&obj->oo_hp_exts),
 406                           osc_list(&obj->oo_urgent_exts));
 407 }
 408
 409 static void osc_page_delete(const struct lu_env *env,
 410                             const struct cl_page_slice *slice)
 411 {
 412         struct osc_page   *opg = cl2osc_page(slice);
 413         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 414         int rc;
 415
 416         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 417
 418         ENTRY;
 419         CDEBUG(D_TRACE, "%p\n", opg);
 420         osc_page_transfer_put(env, opg);
 421         rc = osc_teardown_async_page(env, obj, opg);
 422         if (rc) {
 423                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 424                               "Trying to teardown failed: %d\n", rc);
 425                 LASSERT(0);
 426         }
 427
 428         spin_lock(&obj->oo_seatbelt);
 429         if (opg->ops_submitter != NULL) {
 430                 LASSERT(!cfs_list_empty(&opg->ops_inflight));
 431                 cfs_list_del_init(&opg->ops_inflight);
 432                 opg->ops_submitter = NULL;
 433         }
 434         spin_unlock(&obj->oo_seatbelt);
 435
 436         osc_lru_del(osc_cli(obj), opg);
 437
 438         EXIT;
 439 }
 440
 441 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 442                    int from, int to)
 443 {
 444         struct osc_page       *opg = cl2osc_page(slice);
 445         struct osc_async_page *oap = &opg->ops_oap;
 446
 447         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 448
 449         opg->ops_from = from;
 450         opg->ops_to   = to;
 451         spin_lock(&oap->oap_lock);
 452         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 453         spin_unlock(&oap->oap_lock);
 454 }
 455
 456 static int osc_page_cancel(const struct lu_env *env,
 457                            const struct cl_page_slice *slice)
 458 {
 459         struct osc_page *opg = cl2osc_page(slice);
 460         int rc = 0;
 461
 462         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 463
 464         /* Check if the transferring against this page
 465          * is completed, or not even queued. */
 466         if (opg->ops_transfer_pinned)
 467                 /* FIXME: may not be interrupted.. */
 468                 rc = osc_cancel_async_page(env, opg);
 469         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 470         return rc;
 471 }
 472
 473 static int osc_page_flush(const struct lu_env *env,
 474                           const struct cl_page_slice *slice,
 475                           struct cl_io *io)
 476 {
 477         struct osc_page *opg = cl2osc_page(slice);
 478         int rc = 0;
 479         ENTRY;
 480         rc = osc_flush_async_page(env, io, opg);
 481         RETURN(rc);
 482 }
 483
 484 static const struct cl_page_operations osc_page_ops = {
 485         .cpo_fini          = osc_page_fini,
 486         .cpo_print         = osc_page_print,
 487         .cpo_delete        = osc_page_delete,
 488         .cpo_is_under_lock = osc_page_is_under_lock,
 489         .cpo_disown        = osc_page_disown,
 490         .io = {
 491                 [CRT_READ] = {
 492                         .cpo_cache_add  = osc_page_fail,
 493                         .cpo_completion = osc_page_completion_read
 494                 },
 495                 [CRT_WRITE] = {
 496                         .cpo_cache_add  = osc_page_cache_add,
 497                         .cpo_completion = osc_page_completion_write
 498                 }
 499         },
 500         .cpo_clip           = osc_page_clip,
 501         .cpo_cancel         = osc_page_cancel,
 502         .cpo_flush          = osc_page_flush
 503 };
 504
 505 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 506                 struct cl_page *page, struct page *vmpage)
 507 {
 508         struct osc_object *osc = cl2osc(obj);
 509         struct osc_page   *opg = cl_object_page_slice(obj, page);
 510         int result;
 511
 512         opg->ops_from = 0;
 513         opg->ops_to   = PAGE_CACHE_SIZE;
 514
 515         result = osc_prep_async_page(osc, opg, vmpage,
 516                                         cl_offset(obj, page->cp_index));
 517         if (result == 0) {
 518                 struct osc_io *oio = osc_env_io(env);
 519                 opg->ops_srvlock = osc_io_srvlock(oio);
 520                 cl_page_slice_add(page, &opg->ops_cl, obj,
 521                                 &osc_page_ops);
 522         }
 523         /*
 524          * Cannot assert osc_page_protected() here as read-ahead
 525          * creates temporary pages outside of a lock.
 526          */
 527 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 528         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 529 #endif
 530         /* ops_inflight and ops_lru are the same field, but it doesn't
 531          * hurt to initialize it twice :-) */
 532         CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 533         CFS_INIT_LIST_HEAD(&opg->ops_lru);
 534
 535         /* reserve an LRU space for this page */
 536         if (page->cp_type == CPT_CACHEABLE && result == 0)
 537                 result = osc_lru_reserve(env, osc, opg);
 538
 539         return result;
 540 }
 541
 542 int osc_over_unstable_soft_limit(struct client_obd *cli)
 543 {
 544         long obd_upages, obd_dpages, osc_upages;
 545
 546         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 547         if (cli == NULL)
 548                 return 0;
 549
 550         obd_upages = cfs_atomic_read(&obd_unstable_pages);
 551         obd_dpages = cfs_atomic_read(&obd_dirty_pages);
 552
 553         osc_upages = cfs_atomic_read(&cli->cl_unstable_count);
 554
 555         /* obd_max_dirty_pages is the max number of (dirty + unstable)
 556          * pages allowed at any given time. To simulate an unstable page
 557          * only limit, we subtract the current number of dirty pages
 558          * from this max. This difference is roughly the amount of pages
 559          * currently available for unstable pages. Thus, the soft limit
 560          * is half of that difference. Check osc_upages to ensure we don't
 561          * set SOFT_SYNC for OSCs without any outstanding unstable pages. */
 562         return osc_upages != 0 &&
 563                obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
 564 }
 565
 566 /**
 567  * Helper function called by osc_io_submit() for every page in an immediate
 568  * transfer (i.e., transferred synchronously).
 569  */
 570 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 571                      enum cl_req_type crt, int brw_flags)
 572 {
 573         struct osc_async_page *oap = &opg->ops_oap;
 574         struct osc_object     *obj = oap->oap_obj;
 575
 576         LINVRNT(osc_page_protected(env, opg,
 577                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 578
 579         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 580                  "magic 0x%x\n", oap, oap->oap_magic);
 581         LASSERT(oap->oap_async_flags & ASYNC_READY);
 582         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 583
 584         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 585         oap->oap_page_off  = opg->ops_from;
 586         oap->oap_count     = opg->ops_to - opg->ops_from;
 587         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 588
 589         if (osc_over_unstable_soft_limit(oap->oap_cli))
 590                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
 591
 592         if (!client_is_remote(osc_export(obj)) &&
 593                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 594                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 595                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 596         }
 597
 598         opg->ops_submit_time = cfs_time_current();
 599         osc_page_transfer_get(opg, "transfer\0imm");
 600         osc_page_transfer_add(env, opg, crt);
 601 }
 602
 603 /* --------------- LRU page management ------------------ */
 604
 605 /* OSC is a natural place to manage LRU pages as applications are specialized
 606  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 607  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 608  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 609  * for free LRU slots - this will be very bad so the algorithm requires each
 610  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 611  * at any time.
 612  */
 613
 614 static CFS_DECL_WAITQ(osc_lru_waitq);
 615 /* LRU pages are freed in batch mode. OSC should at least free this
 616  * number of pages to avoid running out of LRU budget, and.. */
 617 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 618 /* free this number at most otherwise it will take too long time to finsih. */
 619 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 620
 621 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 622  * we should free slots aggressively. In this way, slots are freed in a steady
 623  * step to maintain fairness among OSCs.
 624  *
 625  * Return how many LRU pages should be freed. */
 626 static int osc_cache_too_much(struct client_obd *cli)
 627 {
 628         struct cl_client_cache *cache = cli->cl_cache;
 629         int pages = cfs_atomic_read(&cli->cl_lru_in_list);
 630         unsigned long budget;
 631
 632         budget = cache->ccc_lru_max / cfs_atomic_read(&cache->ccc_users);
 633
 634         /* if it's going to run out LRU slots, we should free some, but not
 635          * too much to maintain faireness among OSCs. */
 636         if (cfs_atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 637                 if (pages >= budget)
 638                         return lru_shrink_max;
 639                 else if (pages >= budget / 2)
 640                         return lru_shrink_min;
 641         } else if (pages >= budget * 2)
 642                 return lru_shrink_min;
 643         return 0;
 644 }
 645
 646 int lru_queue_work(const struct lu_env *env, void *data)
 647 {
 648         struct client_obd *cli = data;
 649
 650         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 651
 652         if (osc_cache_too_much(cli))
 653                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 654
 655         RETURN(0);
 656 }
 657
 658 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *plist)
 659 {
 660         CFS_LIST_HEAD(lru);
 661         struct osc_async_page *oap;
 662         int npages = 0;
 663
 664         cfs_list_for_each_entry(oap, plist, oap_pending_item) {
 665                 struct osc_page *opg = oap2osc_page(oap);
 666
 667                 if (!opg->ops_in_lru)
 668                         continue;
 669
 670                 ++npages;
 671                 LASSERT(cfs_list_empty(&opg->ops_lru));
 672                 cfs_list_add(&opg->ops_lru, &lru);
 673         }
 674
 675         if (npages > 0) {
 676                 client_obd_list_lock(&cli->cl_lru_list_lock);
 677                 cfs_list_splice_tail(&lru, &cli->cl_lru_list);
 678                 cfs_atomic_sub(npages, &cli->cl_lru_busy);
 679                 cfs_atomic_add(npages, &cli->cl_lru_in_list);
 680                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 681
 682                 /* XXX: May set force to be true for better performance */
 683                 if (osc_cache_too_much(cli))
 684                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 685         }
 686 }
 687
 688 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 689 {
 690         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) > 0);
 691         cfs_list_del_init(&opg->ops_lru);
 692         cfs_atomic_dec(&cli->cl_lru_in_list);
 693 }
 694
 695 /**
 696  * Page is being destroyed. The page may be not in LRU list, if the transfer
 697  * has never finished(error occurred).
 698  */
 699 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 700 {
 701         if (opg->ops_in_lru) {
 702                 client_obd_list_lock(&cli->cl_lru_list_lock);
 703                 if (!cfs_list_empty(&opg->ops_lru)) {
 704                         __osc_lru_del(cli, opg);
 705                 } else {
 706                         LASSERT(cfs_atomic_read(&cli->cl_lru_busy) > 0);
 707                         cfs_atomic_dec(&cli->cl_lru_busy);
 708                 }
 709                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 710
 711                 cfs_atomic_inc(cli->cl_lru_left);
 712                 /* this is a great place to release more LRU pages if
 713                  * this osc occupies too many LRU pages and kernel is
 714                  * stealing one of them. */
 715                 if (!memory_pressure_get())
 716                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 717                 wake_up(&osc_lru_waitq);
 718         } else {
 719                 LASSERT(cfs_list_empty(&opg->ops_lru));
 720         }
 721 }
 722
 723 /**
 724  * Delete page from LRUlist for redirty.
 725  */
 726 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 727 {
 728         /* If page is being transfered for the first time,
 729          * ops_lru should be empty */
 730         if (opg->ops_in_lru && !cfs_list_empty(&opg->ops_lru)) {
 731                 client_obd_list_lock(&cli->cl_lru_list_lock);
 732                 __osc_lru_del(cli, opg);
 733                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 734                 cfs_atomic_inc(&cli->cl_lru_busy);
 735         }
 736 }
 737
 738 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 739                                 struct cl_page **pvec, int max_index)
 740 {
 741         int i;
 742
 743         for (i = 0; i < max_index; i++) {
 744                 struct cl_page *page = pvec[i];
 745
 746                 LASSERT(cl_page_is_owned(page, io));
 747                 cl_page_unmap(env, io, page);
 748                 cl_page_discard(env, io, page);
 749                 cl_page_disown(env, io, page);
 750                 cl_page_put(env, page);
 751
 752                 pvec[i] = NULL;
 753         }
 754 }
 755
 756 /**
 757  * Drop @target of pages from LRU at most.
 758  */
 759 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 760                    int target, bool force)
 761 {
 762         struct cl_io *io;
 763         struct cl_object *clobj = NULL;
 764         struct cl_page **pvec;
 765         struct osc_page *opg;
 766         int maxscan = 0;
 767         int count = 0;
 768         int index = 0;
 769         int rc = 0;
 770         ENTRY;
 771
 772         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) >= 0);
 773         if (cfs_atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 774                 RETURN(0);
 775
 776         if (!force) {
 777                 if (cfs_atomic_read(&cli->cl_lru_shrinkers) > 0)
 778                         RETURN(-EBUSY);
 779
 780                 if (cfs_atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 781                         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 782                         RETURN(-EBUSY);
 783                 }
 784         } else {
 785                 cfs_atomic_inc(&cli->cl_lru_shrinkers);
 786         }
 787
 788         pvec = osc_env_info(env)->oti_pvec;
 789         io = &osc_env_info(env)->oti_io;
 790
 791         client_obd_list_lock(&cli->cl_lru_list_lock);
 792         maxscan = min(target << 1, cfs_atomic_read(&cli->cl_lru_in_list));
 793         while (!cfs_list_empty(&cli->cl_lru_list)) {
 794                 struct cl_page *page;
 795                 bool will_free = false;
 796
 797                 if (--maxscan < 0)
 798                         break;
 799
 800                 opg = cfs_list_entry(cli->cl_lru_list.next, struct osc_page,
 801                                      ops_lru);
 802                 page = cl_page_top(opg->ops_cl.cpl_page);
 803                 if (cl_page_in_use_noref(page)) {
 804                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 805                         continue;
 806                 }
 807
 808                 LASSERT(page->cp_obj != NULL);
 809                 if (clobj != page->cp_obj) {
 810                         struct cl_object *tmp = page->cp_obj;
 811
 812                         cl_object_get(tmp);
 813                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 814
 815                         if (clobj != NULL) {
 816                                 discard_pagevec(env, io, pvec, index);
 817                                 index = 0;
 818
 819                                 cl_io_fini(env, io);
 820                                 cl_object_put(env, clobj);
 821                                 clobj = NULL;
 822                         }
 823
 824                         clobj = tmp;
 825                         io->ci_obj = clobj;
 826                         io->ci_ignore_layout = 1;
 827                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 828
 829                         client_obd_list_lock(&cli->cl_lru_list_lock);
 830
 831                         if (rc != 0)
 832                                 break;
 833
 834                         ++maxscan;
 835                         continue;
 836                 }
 837
 838                 if (cl_page_own_try(env, io, page) == 0) {
 839                         if (!cl_page_in_use_noref(page)) {
 840                                 /* remove it from lru list earlier to avoid
 841                                  * lock contention */
 842                                 __osc_lru_del(cli, opg);
 843                                 opg->ops_in_lru = 0; /* will be discarded */
 844
 845                                 cl_page_get(page);
 846                                 will_free = true;
 847                         } else {
 848                                 cl_page_disown(env, io, page);
 849                         }
 850                 }
 851
 852                 if (!will_free) {
 853                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 854                         continue;
 855                 }
 856
 857                 /* Don't discard and free the page with cl_lru_list held */
 858                 pvec[index++] = page;
 859                 if (unlikely(index == OTI_PVEC_SIZE)) {
 860                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 861                         discard_pagevec(env, io, pvec, index);
 862                         index = 0;
 863
 864                         client_obd_list_lock(&cli->cl_lru_list_lock);
 865                 }
 866
 867                 if (++count >= target)
 868                         break;
 869         }
 870         client_obd_list_unlock(&cli->cl_lru_list_lock);
 871
 872         if (clobj != NULL) {
 873                 discard_pagevec(env, io, pvec, index);
 874
 875                 cl_io_fini(env, io);
 876                 cl_object_put(env, clobj);
 877         }
 878
 879         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 880         if (count > 0) {
 881                 cfs_atomic_add(count, cli->cl_lru_left);
 882                 wake_up_all(&osc_lru_waitq);
 883         }
 884         RETURN(count > 0 ? count : rc);
 885 }
 886
 887 static inline int max_to_shrink(struct client_obd *cli)
 888 {
 889         return min(cfs_atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 890 }
 891
 892 int osc_lru_reclaim(struct client_obd *cli)
 893 {
 894         struct cl_env_nest nest;
 895         struct lu_env *env;
 896         struct cl_client_cache *cache = cli->cl_cache;
 897         int max_scans;
 898         int rc = 0;
 899         ENTRY;
 900
 901         LASSERT(cache != NULL);
 902         LASSERT(!cfs_list_empty(&cache->ccc_lru));
 903
 904         env = cl_env_nested_get(&nest);
 905         if (IS_ERR(env))
 906                 RETURN(rc);
 907
 908         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 909         if (rc != 0) {
 910                 if (rc == -EBUSY)
 911                         rc = 0;
 912
 913                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 914                         cli->cl_import->imp_obd->obd_name, rc, cli);
 915                 GOTO(out, rc);
 916         }
 917
 918         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 919                 cli->cl_import->imp_obd->obd_name, cli,
 920                 cfs_atomic_read(&cli->cl_lru_in_list),
 921                 cfs_atomic_read(&cli->cl_lru_busy));
 922
 923         /* Reclaim LRU slots from other client_obd as it can't free enough
 924          * from its own. This should rarely happen. */
 925         spin_lock(&cache->ccc_lru_lock);
 926         cache->ccc_lru_shrinkers++;
 927         cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 928
 929         max_scans = cfs_atomic_read(&cache->ccc_users);
 930         while (--max_scans > 0 && !cfs_list_empty(&cache->ccc_lru)) {
 931                 cli = cfs_list_entry(cache->ccc_lru.next, struct client_obd,
 932                                         cl_lru_osc);
 933
 934                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 935                         cli->cl_import->imp_obd->obd_name, cli,
 936                         cfs_atomic_read(&cli->cl_lru_in_list),
 937                         cfs_atomic_read(&cli->cl_lru_busy));
 938
 939                 cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 940                 if (osc_cache_too_much(cli) > 0) {
 941                         spin_unlock(&cache->ccc_lru_lock);
 942
 943                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 944                                             true);
 945                         spin_lock(&cache->ccc_lru_lock);
 946                         if (rc != 0)
 947                                 break;
 948                 }
 949         }
 950         spin_unlock(&cache->ccc_lru_lock);
 951
 952 out:
 953         cl_env_nested_put(&nest, env);
 954         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 955                 cli->cl_import->imp_obd->obd_name, cli, rc);
 956         return rc;
 957 }
 958
 959 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 960                            struct osc_page *opg)
 961 {
 962         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 963         struct osc_io *oio = osc_env_io(env);
 964         struct client_obd *cli = osc_cli(obj);
 965         int rc = 0;
 966         ENTRY;
 967
 968         if (cli->cl_cache == NULL) /* shall not be in LRU */
 969                 RETURN(0);
 970
 971         if (oio->oi_lru_reserved > 0) {
 972                 --oio->oi_lru_reserved;
 973                 goto out;
 974         }
 975
 976         LASSERT(cfs_atomic_read(cli->cl_lru_left) >= 0);
 977         while (!cfs_atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 978
 979                 /* run out of LRU spaces, try to drop some by itself */
 980                 rc = osc_lru_reclaim(cli);
 981                 if (rc < 0)
 982                         break;
 983                 if (rc > 0)
 984                         continue;
 985
 986                 cond_resched();
 987                 rc = l_wait_event(osc_lru_waitq,
 988                                 cfs_atomic_read(cli->cl_lru_left) > 0,
 989                                 &lwi);
 990                 if (rc < 0)
 991                         break;
 992         }
 993
 994 out:
 995         if (rc >= 0) {
 996                 cfs_atomic_inc(&cli->cl_lru_busy);
 997                 opg->ops_in_lru = 1;
 998                 rc = 0;
 999         }
1000
1001         RETURN(rc);
1002 }
1003
1004 /** @} osc */