kernel/drivers/staging/lustre/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, 2012, 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, bool del);
  46 static void osc_lru_add(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                 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_del(osc_cli(obj), opg, false);
 208
 209         spin_lock(&obj->oo_seatbelt);
 210         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
 223         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 224
 225         osc_page_transfer_get(opg, "transfer\0cache");
 226         result = osc_queue_async_io(env, io, opg);
 227         if (result != 0)
 228                 osc_page_transfer_put(env, opg);
 229         else
 230                 osc_page_transfer_add(env, opg, CRT_WRITE);
 231
 232         /* for sync write, kernel will wait for this page to be flushed before
 233          * osc_io_end() is called, so release it earlier.
 234          * for mkwrite(), it's known there is no further pages. */
 235         if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
 236                 if (oio->oi_active != NULL) {
 237                         osc_extent_release(env, oio->oi_active);
 238                         oio->oi_active = NULL;
 239                 }
 240         }
 241
 242         return result;
 243 }
 244
 245 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 246                       pgoff_t start, pgoff_t end)
 247 {
 248         memset(policy, 0, sizeof(*policy));
 249         policy->l_extent.start = cl_offset(obj, start);
 250         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 251 }
 252
 253 static int osc_page_addref_lock(const struct lu_env *env,
 254                                 struct osc_page *opg,
 255                                 struct cl_lock *lock)
 256 {
 257         struct osc_lock *olock;
 258         int           rc;
 259
 260         LASSERT(opg->ops_lock == NULL);
 261
 262         olock = osc_lock_at(lock);
 263         if (atomic_inc_return(&olock->ols_pageref) <= 0) {
 264                 atomic_dec(&olock->ols_pageref);
 265                 rc = -ENODATA;
 266         } else {
 267                 cl_lock_get(lock);
 268                 opg->ops_lock = lock;
 269                 rc = 0;
 270         }
 271         return rc;
 272 }
 273
 274 static void osc_page_putref_lock(const struct lu_env *env,
 275                                  struct osc_page *opg)
 276 {
 277         struct cl_lock  *lock = opg->ops_lock;
 278         struct osc_lock *olock;
 279
 280         LASSERT(lock != NULL);
 281         olock = osc_lock_at(lock);
 282
 283         atomic_dec(&olock->ols_pageref);
 284         opg->ops_lock = NULL;
 285
 286         cl_lock_put(env, lock);
 287 }
 288
 289 static int osc_page_is_under_lock(const struct lu_env *env,
 290                                   const struct cl_page_slice *slice,
 291                                   struct cl_io *unused)
 292 {
 293         struct cl_lock *lock;
 294         int          result = -ENODATA;
 295
 296         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 297                                NULL, 1, 0);
 298         if (lock != NULL) {
 299                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 300                         result = -EBUSY;
 301                 cl_lock_put(env, lock);
 302         }
 303         return result;
 304 }
 305
 306 static void osc_page_disown(const struct lu_env *env,
 307                             const struct cl_page_slice *slice,
 308                             struct cl_io *io)
 309 {
 310         struct osc_page *opg = cl2osc_page(slice);
 311
 312         if (unlikely(opg->ops_lock))
 313                 osc_page_putref_lock(env, opg);
 314 }
 315
 316 static void osc_page_completion_read(const struct lu_env *env,
 317                                      const struct cl_page_slice *slice,
 318                                      int ioret)
 319 {
 320         struct osc_page   *opg = cl2osc_page(slice);
 321         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 322
 323         if (likely(opg->ops_lock))
 324                 osc_page_putref_lock(env, opg);
 325         osc_lru_add(osc_cli(obj), 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         struct osc_page   *opg = cl2osc_page(slice);
 333         struct osc_object *obj = cl2osc(slice->cpl_obj);
 334
 335         osc_lru_add(osc_cli(obj), opg);
 336 }
 337
 338 static int osc_page_fail(const struct lu_env *env,
 339                          const struct cl_page_slice *slice,
 340                          struct cl_io *unused)
 341 {
 342         /*
 343          * Cached read?
 344          */
 345         LBUG();
 346         return 0;
 347 }
 348
 349
 350 static const char *osc_list(struct list_head *head)
 351 {
 352         return list_empty(head) ? "-" : "+";
 353 }
 354
 355 static inline unsigned long osc_submit_duration(struct osc_page *opg)
 356 {
 357         if (opg->ops_submit_time == 0)
 358                 return 0;
 359
 360         return (cfs_time_current() - opg->ops_submit_time);
 361 }
 362
 363 static int osc_page_print(const struct lu_env *env,
 364                           const struct cl_page_slice *slice,
 365                           void *cookie, lu_printer_t printer)
 366 {
 367         struct osc_page       *opg = cl2osc_page(slice);
 368         struct osc_async_page *oap = &opg->ops_oap;
 369         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 370         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 371
 372         return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
 373                           opg,
 374                           /* 1 */
 375                           oap->oap_magic, oap->oap_cmd,
 376                           oap->oap_interrupted,
 377                           osc_list(&oap->oap_pending_item),
 378                           osc_list(&oap->oap_rpc_item),
 379                           /* 2 */
 380                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 381                           oap->oap_async_flags, oap->oap_brw_flags,
 382                           oap->oap_request, oap->oap_cli, obj,
 383                           /* 3 */
 384                           osc_list(&opg->ops_inflight),
 385                           opg->ops_submitter, opg->ops_transfer_pinned,
 386                           osc_submit_duration(opg), opg->ops_srvlock,
 387                           /* 4 */
 388                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 389                           cli->cl_max_rpcs_in_flight,
 390                           cli->cl_avail_grant,
 391                           osc_list(&cli->cl_cache_waiters),
 392                           osc_list(&cli->cl_loi_ready_list),
 393                           osc_list(&cli->cl_loi_hp_ready_list),
 394                           osc_list(&cli->cl_loi_write_list),
 395                           osc_list(&cli->cl_loi_read_list),
 396                           /* 5 */
 397                           osc_list(&obj->oo_ready_item),
 398                           osc_list(&obj->oo_hp_ready_item),
 399                           osc_list(&obj->oo_write_item),
 400                           osc_list(&obj->oo_read_item),
 401                           atomic_read(&obj->oo_nr_reads),
 402                           osc_list(&obj->oo_reading_exts),
 403                           atomic_read(&obj->oo_nr_writes),
 404                           osc_list(&obj->oo_hp_exts),
 405                           osc_list(&obj->oo_urgent_exts));
 406 }
 407
 408 static void osc_page_delete(const struct lu_env *env,
 409                             const struct cl_page_slice *slice)
 410 {
 411         struct osc_page   *opg = cl2osc_page(slice);
 412         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 413         int rc;
 414
 415         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 416
 417         CDEBUG(D_TRACE, "%p\n", opg);
 418         osc_page_transfer_put(env, opg);
 419         rc = osc_teardown_async_page(env, obj, opg);
 420         if (rc) {
 421                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 422                               "Trying to teardown failed: %d\n", rc);
 423                 LASSERT(0);
 424         }
 425
 426         spin_lock(&obj->oo_seatbelt);
 427         if (opg->ops_submitter != NULL) {
 428                 LASSERT(!list_empty(&opg->ops_inflight));
 429                 list_del_init(&opg->ops_inflight);
 430                 opg->ops_submitter = NULL;
 431         }
 432         spin_unlock(&obj->oo_seatbelt);
 433
 434         osc_lru_del(osc_cli(obj), opg, true);
 435 }
 436
 437 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 438                    int from, int to)
 439 {
 440         struct osc_page       *opg = cl2osc_page(slice);
 441         struct osc_async_page *oap = &opg->ops_oap;
 442
 443         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 444
 445         opg->ops_from = from;
 446         opg->ops_to   = to;
 447         spin_lock(&oap->oap_lock);
 448         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 449         spin_unlock(&oap->oap_lock);
 450 }
 451
 452 static int osc_page_cancel(const struct lu_env *env,
 453                            const struct cl_page_slice *slice)
 454 {
 455         struct osc_page *opg = cl2osc_page(slice);
 456         int rc = 0;
 457
 458         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 459
 460         /* Check if the transferring against this page
 461          * is completed, or not even queued. */
 462         if (opg->ops_transfer_pinned)
 463                 /* FIXME: may not be interrupted.. */
 464                 rc = osc_cancel_async_page(env, opg);
 465         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 466         return rc;
 467 }
 468
 469 static int osc_page_flush(const struct lu_env *env,
 470                           const struct cl_page_slice *slice,
 471                           struct cl_io *io)
 472 {
 473         struct osc_page *opg = cl2osc_page(slice);
 474         int rc = 0;
 475
 476         rc = osc_flush_async_page(env, io, opg);
 477         return rc;
 478 }
 479
 480 static const struct cl_page_operations osc_page_ops = {
 481         .cpo_fini         = osc_page_fini,
 482         .cpo_print       = osc_page_print,
 483         .cpo_delete     = osc_page_delete,
 484         .cpo_is_under_lock = osc_page_is_under_lock,
 485         .cpo_disown     = osc_page_disown,
 486         .io = {
 487                 [CRT_READ] = {
 488                         .cpo_cache_add  = osc_page_fail,
 489                         .cpo_completion = osc_page_completion_read
 490                 },
 491                 [CRT_WRITE] = {
 492                         .cpo_cache_add  = osc_page_cache_add,
 493                         .cpo_completion = osc_page_completion_write
 494                 }
 495         },
 496         .cpo_clip          = osc_page_clip,
 497         .cpo_cancel      = osc_page_cancel,
 498         .cpo_flush        = osc_page_flush
 499 };
 500
 501 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 502                 struct cl_page *page, struct page *vmpage)
 503 {
 504         struct osc_object *osc = cl2osc(obj);
 505         struct osc_page   *opg = cl_object_page_slice(obj, page);
 506         int result;
 507
 508         opg->ops_from = 0;
 509         opg->ops_to   = PAGE_CACHE_SIZE;
 510
 511         result = osc_prep_async_page(osc, opg, vmpage,
 512                                         cl_offset(obj, page->cp_index));
 513         if (result == 0) {
 514                 struct osc_io *oio = osc_env_io(env);
 515                 opg->ops_srvlock = osc_io_srvlock(oio);
 516                 cl_page_slice_add(page, &opg->ops_cl, obj,
 517                                 &osc_page_ops);
 518         }
 519         /*
 520          * Cannot assert osc_page_protected() here as read-ahead
 521          * creates temporary pages outside of a lock.
 522          */
 523         /* ops_inflight and ops_lru are the same field, but it doesn't
 524          * hurt to initialize it twice :-) */
 525         INIT_LIST_HEAD(&opg->ops_inflight);
 526         INIT_LIST_HEAD(&opg->ops_lru);
 527
 528         /* reserve an LRU space for this page */
 529         if (page->cp_type == CPT_CACHEABLE && result == 0)
 530                 result = osc_lru_reserve(env, osc, opg);
 531
 532         return result;
 533 }
 534
 535 /**
 536  * Helper function called by osc_io_submit() for every page in an immediate
 537  * transfer (i.e., transferred synchronously).
 538  */
 539 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 540                      enum cl_req_type crt, int brw_flags)
 541 {
 542         struct osc_async_page *oap = &opg->ops_oap;
 543         struct osc_object     *obj = oap->oap_obj;
 544
 545         LINVRNT(osc_page_protected(env, opg,
 546                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 547
 548         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, magic 0x%x\n",
 549                  oap, oap->oap_magic);
 550         LASSERT(oap->oap_async_flags & ASYNC_READY);
 551         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 552
 553         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 554         oap->oap_page_off  = opg->ops_from;
 555         oap->oap_count     = opg->ops_to - opg->ops_from;
 556         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 557
 558         if (!client_is_remote(osc_export(obj)) &&
 559                         capable(CFS_CAP_SYS_RESOURCE)) {
 560                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 561                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 562         }
 563
 564         opg->ops_submit_time = cfs_time_current();
 565         osc_page_transfer_get(opg, "transfer\0imm");
 566         osc_page_transfer_add(env, opg, crt);
 567 }
 568
 569 /* --------------- LRU page management ------------------ */
 570
 571 /* OSC is a natural place to manage LRU pages as applications are specialized
 572  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 573  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 574  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 575  * for free LRU slots - this will be very bad so the algorithm requires each
 576  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 577  * at any time.
 578  */
 579
 580 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 581 static atomic_t osc_lru_waiters = ATOMIC_INIT(0);
 582 /* LRU pages are freed in batch mode. OSC should at least free this
 583  * number of pages to avoid running out of LRU budget, and.. */
 584 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT);  /* 2M */
 585 /* free this number at most otherwise it will take too long time to finish. */
 586 static const int lru_shrink_max = 32 << (20 - PAGE_CACHE_SHIFT); /* 32M */
 587
 588 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 589  * we should free slots aggressively. In this way, slots are freed in a steady
 590  * step to maintain fairness among OSCs.
 591  *
 592  * Return how many LRU pages should be freed. */
 593 static int osc_cache_too_much(struct client_obd *cli)
 594 {
 595         struct cl_client_cache *cache = cli->cl_cache;
 596         int pages = atomic_read(&cli->cl_lru_in_list) >> 1;
 597
 598         if (atomic_read(&osc_lru_waiters) > 0 &&
 599             atomic_read(cli->cl_lru_left) < lru_shrink_max)
 600                 /* drop lru pages aggressively */
 601                 return min(pages, lru_shrink_max);
 602
 603         /* if it's going to run out LRU slots, we should free some, but not
 604          * too much to maintain fairness among OSCs. */
 605         if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 606                 unsigned long tmp;
 607
 608                 tmp = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 609                 if (pages > tmp)
 610                         return min(pages, lru_shrink_max);
 611
 612                 return pages > lru_shrink_min ? lru_shrink_min : 0;
 613         }
 614
 615         return 0;
 616 }
 617
 618 /* Return how many pages are not discarded in @pvec. */
 619 static int discard_pagevec(const struct lu_env *env, struct cl_io *io,
 620                            struct cl_page **pvec, int max_index)
 621 {
 622         int count;
 623         int i;
 624
 625         for (count = 0, i = 0; i < max_index; i++) {
 626                 struct cl_page *page = pvec[i];
 627                 if (cl_page_own_try(env, io, page) == 0) {
 628                         /* free LRU page only if nobody is using it.
 629                          * This check is necessary to avoid freeing the pages
 630                          * having already been removed from LRU and pinned
 631                          * for IO. */
 632                         if (!cl_page_in_use(page)) {
 633                                 cl_page_unmap(env, io, page);
 634                                 cl_page_discard(env, io, page);
 635                                 ++count;
 636                         }
 637                         cl_page_disown(env, io, page);
 638                 }
 639                 cl_page_put(env, page);
 640                 pvec[i] = NULL;
 641         }
 642         return max_index - count;
 643 }
 644
 645 /**
 646  * Drop @target of pages from LRU at most.
 647  */
 648 int osc_lru_shrink(struct client_obd *cli, int target)
 649 {
 650         struct cl_env_nest nest;
 651         struct lu_env *env;
 652         struct cl_io *io;
 653         struct cl_object *clobj = NULL;
 654         struct cl_page **pvec;
 655         struct osc_page *opg;
 656         int maxscan = 0;
 657         int count = 0;
 658         int index = 0;
 659         int rc = 0;
 660
 661         LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0);
 662         if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 663                 return 0;
 664
 665         env = cl_env_nested_get(&nest);
 666         if (IS_ERR(env))
 667                 return PTR_ERR(env);
 668
 669         pvec = osc_env_info(env)->oti_pvec;
 670         io = &osc_env_info(env)->oti_io;
 671
 672         client_obd_list_lock(&cli->cl_lru_list_lock);
 673         atomic_inc(&cli->cl_lru_shrinkers);
 674         maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
 675         while (!list_empty(&cli->cl_lru_list)) {
 676                 struct cl_page *page;
 677
 678                 if (--maxscan < 0)
 679                         break;
 680
 681                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 682                                      ops_lru);
 683                 page = cl_page_top(opg->ops_cl.cpl_page);
 684                 if (cl_page_in_use_noref(page)) {
 685                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 686                         continue;
 687                 }
 688
 689                 LASSERT(page->cp_obj != NULL);
 690                 if (clobj != page->cp_obj) {
 691                         struct cl_object *tmp = page->cp_obj;
 692
 693                         cl_object_get(tmp);
 694                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 695
 696                         if (clobj != NULL) {
 697                                 count -= discard_pagevec(env, io, pvec, index);
 698                                 index = 0;
 699
 700                                 cl_io_fini(env, io);
 701                                 cl_object_put(env, clobj);
 702                                 clobj = NULL;
 703                         }
 704
 705                         clobj = tmp;
 706                         io->ci_obj = clobj;
 707                         io->ci_ignore_layout = 1;
 708                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 709
 710                         client_obd_list_lock(&cli->cl_lru_list_lock);
 711
 712                         if (rc != 0)
 713                                 break;
 714
 715                         ++maxscan;
 716                         continue;
 717                 }
 718
 719                 /* move this page to the end of list as it will be discarded
 720                  * soon. The page will be finally removed from LRU list in
 721                  * osc_page_delete().  */
 722                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 723
 724                 /* it's okay to grab a refcount here w/o holding lock because
 725                  * it has to grab cl_lru_list_lock to delete the page. */
 726                 cl_page_get(page);
 727                 pvec[index++] = page;
 728                 if (++count >= target)
 729                         break;
 730
 731                 if (unlikely(index == OTI_PVEC_SIZE)) {
 732                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 733                         count -= discard_pagevec(env, io, pvec, index);
 734                         index = 0;
 735
 736                         client_obd_list_lock(&cli->cl_lru_list_lock);
 737                 }
 738         }
 739         client_obd_list_unlock(&cli->cl_lru_list_lock);
 740
 741         if (clobj != NULL) {
 742                 count -= discard_pagevec(env, io, pvec, index);
 743
 744                 cl_io_fini(env, io);
 745                 cl_object_put(env, clobj);
 746         }
 747         cl_env_nested_put(&nest, env);
 748
 749         atomic_dec(&cli->cl_lru_shrinkers);
 750         return count > 0 ? count : rc;
 751 }
 752
 753 static void osc_lru_add(struct client_obd *cli, struct osc_page *opg)
 754 {
 755         bool wakeup = false;
 756
 757         if (!opg->ops_in_lru)
 758                 return;
 759
 760         atomic_dec(&cli->cl_lru_busy);
 761         client_obd_list_lock(&cli->cl_lru_list_lock);
 762         if (list_empty(&opg->ops_lru)) {
 763                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 764                 atomic_inc_return(&cli->cl_lru_in_list);
 765                 wakeup = atomic_read(&osc_lru_waiters) > 0;
 766         }
 767         client_obd_list_unlock(&cli->cl_lru_list_lock);
 768
 769         if (wakeup) {
 770                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 771                 wake_up_all(&osc_lru_waitq);
 772         }
 773 }
 774
 775 /* delete page from LRUlist. The page can be deleted from LRUlist for two
 776  * reasons: redirtied or deleted from page cache. */
 777 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del)
 778 {
 779         if (opg->ops_in_lru) {
 780                 client_obd_list_lock(&cli->cl_lru_list_lock);
 781                 if (!list_empty(&opg->ops_lru)) {
 782                         LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
 783                         list_del_init(&opg->ops_lru);
 784                         atomic_dec(&cli->cl_lru_in_list);
 785                         if (!del)
 786                                 atomic_inc(&cli->cl_lru_busy);
 787                 } else if (del) {
 788                         LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
 789                         atomic_dec(&cli->cl_lru_busy);
 790                 }
 791                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 792                 if (del) {
 793                         atomic_inc(cli->cl_lru_left);
 794                         /* this is a great place to release more LRU pages if
 795                          * this osc occupies too many LRU pages and kernel is
 796                          * stealing one of them.
 797                          * cl_lru_shrinkers is to avoid recursive call in case
 798                          * we're already in the context of osc_lru_shrink(). */
 799                         if (atomic_read(&cli->cl_lru_shrinkers) == 0 &&
 800                             !memory_pressure_get())
 801                                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 802                         wake_up(&osc_lru_waitq);
 803                 }
 804         } else {
 805                 LASSERT(list_empty(&opg->ops_lru));
 806         }
 807 }
 808
 809 static inline int max_to_shrink(struct client_obd *cli)
 810 {
 811         return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 812 }
 813
 814 static int osc_lru_reclaim(struct client_obd *cli)
 815 {
 816         struct cl_client_cache *cache = cli->cl_cache;
 817         int max_scans;
 818         int rc;
 819
 820         LASSERT(cache != NULL);
 821         LASSERT(!list_empty(&cache->ccc_lru));
 822
 823         rc = osc_lru_shrink(cli, lru_shrink_min);
 824         if (rc != 0) {
 825                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 826                         cli->cl_import->imp_obd->obd_name, rc, cli);
 827                 return rc;
 828         }
 829
 830         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 831                 cli->cl_import->imp_obd->obd_name, cli,
 832                 atomic_read(&cli->cl_lru_in_list),
 833                 atomic_read(&cli->cl_lru_busy));
 834
 835         /* Reclaim LRU slots from other client_obd as it can't free enough
 836          * from its own. This should rarely happen. */
 837         spin_lock(&cache->ccc_lru_lock);
 838         cache->ccc_lru_shrinkers++;
 839         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 840
 841         max_scans = atomic_read(&cache->ccc_users);
 842         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 843                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 844                                         cl_lru_osc);
 845
 846                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 847                         cli->cl_import->imp_obd->obd_name, cli,
 848                         atomic_read(&cli->cl_lru_in_list),
 849                         atomic_read(&cli->cl_lru_busy));
 850
 851                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 852                 if (atomic_read(&cli->cl_lru_in_list) > 0) {
 853                         spin_unlock(&cache->ccc_lru_lock);
 854
 855                         rc = osc_lru_shrink(cli, max_to_shrink(cli));
 856                         spin_lock(&cache->ccc_lru_lock);
 857                         if (rc != 0)
 858                                 break;
 859                 }
 860         }
 861         spin_unlock(&cache->ccc_lru_lock);
 862
 863         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 864                 cli->cl_import->imp_obd->obd_name, cli, rc);
 865         return rc;
 866 }
 867
 868 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 869                            struct osc_page *opg)
 870 {
 871         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 872         struct client_obd *cli = osc_cli(obj);
 873         int rc = 0;
 874
 875         if (cli->cl_cache == NULL) /* shall not be in LRU */
 876                 return 0;
 877
 878         LASSERT(atomic_read(cli->cl_lru_left) >= 0);
 879         while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 880                 int gen;
 881
 882                 /* run out of LRU spaces, try to drop some by itself */
 883                 rc = osc_lru_reclaim(cli);
 884                 if (rc < 0)
 885                         break;
 886                 if (rc > 0)
 887                         continue;
 888
 889                 cond_resched();
 890
 891                 /* slowest case, all of caching pages are busy, notifying
 892                  * other OSCs that we're lack of LRU slots. */
 893                 atomic_inc(&osc_lru_waiters);
 894
 895                 gen = atomic_read(&cli->cl_lru_in_list);
 896                 rc = l_wait_event(osc_lru_waitq,
 897                                 atomic_read(cli->cl_lru_left) > 0 ||
 898                                 (atomic_read(&cli->cl_lru_in_list) > 0 &&
 899                                  gen != atomic_read(&cli->cl_lru_in_list)),
 900                                 &lwi);
 901
 902                 atomic_dec(&osc_lru_waiters);
 903                 if (rc < 0)
 904                         break;
 905         }
 906
 907         if (rc >= 0) {
 908                 atomic_inc(&cli->cl_lru_busy);
 909                 opg->ops_in_lru = 1;
 910                 rc = 0;
 911         }
 912
 913         return rc;
 914 }
 915
 916 /** @} osc */