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
 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_del(osc_cli(obj), opg, false);
 209
 210         spin_lock(&obj->oo_seatbelt);
 211         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
 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 (atomic_inc_return(&olock->ols_pageref) <= 0) {
 265                 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         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         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 298                                NULL, 1, 0);
 299         if (lock != NULL) {
 300                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 301                         result = -EBUSY;
 302                 cl_lock_put(env, lock);
 303         }
 304         return result;
 305 }
 306
 307 static void osc_page_disown(const struct lu_env *env,
 308                             const struct cl_page_slice *slice,
 309                             struct cl_io *io)
 310 {
 311         struct osc_page *opg = cl2osc_page(slice);
 312
 313         if (unlikely(opg->ops_lock))
 314                 osc_page_putref_lock(env, opg);
 315 }
 316
 317 static void osc_page_completion_read(const struct lu_env *env,
 318                                      const struct cl_page_slice *slice,
 319                                      int ioret)
 320 {
 321         struct osc_page *opg = cl2osc_page(slice);
 322         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 323
 324         if (likely(opg->ops_lock))
 325                 osc_page_putref_lock(env, opg);
 326         osc_lru_add(osc_cli(obj), 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         struct osc_page *opg = cl2osc_page(slice);
 334         struct osc_object *obj = cl2osc(slice->cpl_obj);
 335
 336         osc_lru_add(osc_cli(obj), opg);
 337 }
 338
 339 static int osc_page_fail(const struct lu_env *env,
 340                          const struct cl_page_slice *slice,
 341                          struct cl_io *unused)
 342 {
 343         /*
 344          * Cached read?
 345          */
 346         LBUG();
 347         return 0;
 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;
 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
 516                 opg->ops_srvlock = osc_io_srvlock(oio);
 517                 cl_page_slice_add(page, &opg->ops_cl, obj,
 518                                 &osc_page_ops);
 519         }
 520         /*
 521          * Cannot assert osc_page_protected() here as read-ahead
 522          * creates temporary pages outside of a lock.
 523          */
 524         /* ops_inflight and ops_lru are the same field, but it doesn't
 525          * hurt to initialize it twice :-) */
 526         INIT_LIST_HEAD(&opg->ops_inflight);
 527         INIT_LIST_HEAD(&opg->ops_lru);
 528
 529         /* reserve an LRU space for this page */
 530         if (page->cp_type == CPT_CACHEABLE && result == 0)
 531                 result = osc_lru_reserve(env, osc, opg);
 532
 533         return result;
 534 }
 535
 536 /**
 537  * Helper function called by osc_io_submit() for every page in an immediate
 538  * transfer (i.e., transferred synchronously).
 539  */
 540 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 541                      enum cl_req_type crt, int brw_flags)
 542 {
 543         struct osc_async_page *oap = &opg->ops_oap;
 544         struct osc_object *obj = oap->oap_obj;
 545
 546         LINVRNT(osc_page_protected(env, opg,
 547                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 548
 549         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, magic 0x%x\n",
 550                  oap, oap->oap_magic);
 551         LASSERT(oap->oap_async_flags & ASYNC_READY);
 552         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 553
 554         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 555         oap->oap_page_off = opg->ops_from;
 556         oap->oap_count = opg->ops_to - opg->ops_from;
 557         oap->oap_brw_flags = brw_flags | OBD_BRW_SYNC;
 558
 559         if (!client_is_remote(osc_export(obj)) &&
 560                         capable(CFS_CAP_SYS_RESOURCE)) {
 561                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 562                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 563         }
 564
 565         opg->ops_submit_time = cfs_time_current();
 566         osc_page_transfer_get(opg, "transfer\0imm");
 567         osc_page_transfer_add(env, opg, crt);
 568 }
 569
 570 /* --------------- LRU page management ------------------ */
 571
 572 /* OSC is a natural place to manage LRU pages as applications are specialized
 573  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 574  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 575  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 576  * for free LRU slots - this will be very bad so the algorithm requires each
 577  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 578  * at any time.
 579  */
 580
 581 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 582 static atomic_t osc_lru_waiters = ATOMIC_INIT(0);
 583 /* LRU pages are freed in batch mode. OSC should at least free this
 584  * number of pages to avoid running out of LRU budget, and.. */
 585 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT);  /* 2M */
 586 /* free this number at most otherwise it will take too long time to finish. */
 587 static const int lru_shrink_max = 32 << (20 - PAGE_CACHE_SHIFT); /* 32M */
 588
 589 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 590  * we should free slots aggressively. In this way, slots are freed in a steady
 591  * step to maintain fairness among OSCs.
 592  *
 593  * Return how many LRU pages should be freed. */
 594 static int osc_cache_too_much(struct client_obd *cli)
 595 {
 596         struct cl_client_cache *cache = cli->cl_cache;
 597         int pages = atomic_read(&cli->cl_lru_in_list) >> 1;
 598
 599         if (atomic_read(&osc_lru_waiters) > 0 &&
 600             atomic_read(cli->cl_lru_left) < lru_shrink_max)
 601                 /* drop lru pages aggressively */
 602                 return min(pages, lru_shrink_max);
 603
 604         /* if it's going to run out LRU slots, we should free some, but not
 605          * too much to maintain fairness among OSCs. */
 606         if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 607                 unsigned long tmp;
 608
 609                 tmp = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 610                 if (pages > tmp)
 611                         return min(pages, lru_shrink_max);
 612
 613                 return pages > lru_shrink_min ? lru_shrink_min : 0;
 614         }
 615
 616         return 0;
 617 }
 618
 619 /* Return how many pages are not discarded in @pvec. */
 620 static int discard_pagevec(const struct lu_env *env, struct cl_io *io,
 621                            struct cl_page **pvec, int max_index)
 622 {
 623         int count;
 624         int i;
 625
 626         for (count = 0, i = 0; i < max_index; i++) {
 627                 struct cl_page *page = pvec[i];
 628
 629                 if (cl_page_own_try(env, io, page) == 0) {
 630                         /* free LRU page only if nobody is using it.
 631                          * This check is necessary to avoid freeing the pages
 632                          * having already been removed from LRU and pinned
 633                          * for IO. */
 634                         if (!cl_page_in_use(page)) {
 635                                 cl_page_unmap(env, io, page);
 636                                 cl_page_discard(env, io, page);
 637                                 ++count;
 638                         }
 639                         cl_page_disown(env, io, page);
 640                 }
 641                 cl_page_put(env, page);
 642                 pvec[i] = NULL;
 643         }
 644         return max_index - count;
 645 }
 646
 647 /**
 648  * Drop @target of pages from LRU at most.
 649  */
 650 int osc_lru_shrink(struct client_obd *cli, int target)
 651 {
 652         struct cl_env_nest nest;
 653         struct lu_env *env;
 654         struct cl_io *io;
 655         struct cl_object *clobj = NULL;
 656         struct cl_page **pvec;
 657         struct osc_page *opg;
 658         int maxscan = 0;
 659         int count = 0;
 660         int index = 0;
 661         int rc = 0;
 662
 663         LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0);
 664         if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 665                 return 0;
 666
 667         env = cl_env_nested_get(&nest);
 668         if (IS_ERR(env))
 669                 return PTR_ERR(env);
 670
 671         pvec = osc_env_info(env)->oti_pvec;
 672         io = &osc_env_info(env)->oti_io;
 673
 674         client_obd_list_lock(&cli->cl_lru_list_lock);
 675         atomic_inc(&cli->cl_lru_shrinkers);
 676         maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
 677         while (!list_empty(&cli->cl_lru_list)) {
 678                 struct cl_page *page;
 679
 680                 if (--maxscan < 0)
 681                         break;
 682
 683                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 684                                      ops_lru);
 685                 page = cl_page_top(opg->ops_cl.cpl_page);
 686                 if (cl_page_in_use_noref(page)) {
 687                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 688                         continue;
 689                 }
 690
 691                 LASSERT(page->cp_obj != NULL);
 692                 if (clobj != page->cp_obj) {
 693                         struct cl_object *tmp = page->cp_obj;
 694
 695                         cl_object_get(tmp);
 696                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 697
 698                         if (clobj != NULL) {
 699                                 count -= discard_pagevec(env, io, pvec, index);
 700                                 index = 0;
 701
 702                                 cl_io_fini(env, io);
 703                                 cl_object_put(env, clobj);
 704                                 clobj = NULL;
 705                         }
 706
 707                         clobj = tmp;
 708                         io->ci_obj = clobj;
 709                         io->ci_ignore_layout = 1;
 710                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 711
 712                         client_obd_list_lock(&cli->cl_lru_list_lock);
 713
 714                         if (rc != 0)
 715                                 break;
 716
 717                         ++maxscan;
 718                         continue;
 719                 }
 720
 721                 /* move this page to the end of list as it will be discarded
 722                  * soon. The page will be finally removed from LRU list in
 723                  * osc_page_delete().  */
 724                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 725
 726                 /* it's okay to grab a refcount here w/o holding lock because
 727                  * it has to grab cl_lru_list_lock to delete the page. */
 728                 cl_page_get(page);
 729                 pvec[index++] = page;
 730                 if (++count >= target)
 731                         break;
 732
 733                 if (unlikely(index == OTI_PVEC_SIZE)) {
 734                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 735                         count -= discard_pagevec(env, io, pvec, index);
 736                         index = 0;
 737
 738                         client_obd_list_lock(&cli->cl_lru_list_lock);
 739                 }
 740         }
 741         client_obd_list_unlock(&cli->cl_lru_list_lock);
 742
 743         if (clobj != NULL) {
 744                 count -= discard_pagevec(env, io, pvec, index);
 745
 746                 cl_io_fini(env, io);
 747                 cl_object_put(env, clobj);
 748         }
 749         cl_env_nested_put(&nest, env);
 750
 751         atomic_dec(&cli->cl_lru_shrinkers);
 752         return count > 0 ? count : rc;
 753 }
 754
 755 static void osc_lru_add(struct client_obd *cli, struct osc_page *opg)
 756 {
 757         bool wakeup = false;
 758
 759         if (!opg->ops_in_lru)
 760                 return;
 761
 762         atomic_dec(&cli->cl_lru_busy);
 763         client_obd_list_lock(&cli->cl_lru_list_lock);
 764         if (list_empty(&opg->ops_lru)) {
 765                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 766                 atomic_inc_return(&cli->cl_lru_in_list);
 767                 wakeup = atomic_read(&osc_lru_waiters) > 0;
 768         }
 769         client_obd_list_unlock(&cli->cl_lru_list_lock);
 770
 771         if (wakeup) {
 772                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 773                 wake_up_all(&osc_lru_waitq);
 774         }
 775 }
 776
 777 /* delete page from LRUlist. The page can be deleted from LRUlist for two
 778  * reasons: redirtied or deleted from page cache. */
 779 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del)
 780 {
 781         if (opg->ops_in_lru) {
 782                 client_obd_list_lock(&cli->cl_lru_list_lock);
 783                 if (!list_empty(&opg->ops_lru)) {
 784                         LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
 785                         list_del_init(&opg->ops_lru);
 786                         atomic_dec(&cli->cl_lru_in_list);
 787                         if (!del)
 788                                 atomic_inc(&cli->cl_lru_busy);
 789                 } else if (del) {
 790                         LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
 791                         atomic_dec(&cli->cl_lru_busy);
 792                 }
 793                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 794                 if (del) {
 795                         atomic_inc(cli->cl_lru_left);
 796                         /* this is a great place to release more LRU pages if
 797                          * this osc occupies too many LRU pages and kernel is
 798                          * stealing one of them.
 799                          * cl_lru_shrinkers is to avoid recursive call in case
 800                          * we're already in the context of osc_lru_shrink(). */
 801                         if (atomic_read(&cli->cl_lru_shrinkers) == 0 &&
 802                             !memory_pressure_get())
 803                                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 804                         wake_up(&osc_lru_waitq);
 805                 }
 806         } else {
 807                 LASSERT(list_empty(&opg->ops_lru));
 808         }
 809 }
 810
 811 static inline int max_to_shrink(struct client_obd *cli)
 812 {
 813         return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 814 }
 815
 816 static int osc_lru_reclaim(struct client_obd *cli)
 817 {
 818         struct cl_client_cache *cache = cli->cl_cache;
 819         int max_scans;
 820         int rc;
 821
 822         LASSERT(cache != NULL);
 823
 824         rc = osc_lru_shrink(cli, lru_shrink_min);
 825         if (rc != 0) {
 826                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 827                         cli->cl_import->imp_obd->obd_name, rc, cli);
 828                 return rc;
 829         }
 830
 831         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 832                 cli->cl_import->imp_obd->obd_name, cli,
 833                 atomic_read(&cli->cl_lru_in_list),
 834                 atomic_read(&cli->cl_lru_busy));
 835
 836         /* Reclaim LRU slots from other client_obd as it can't free enough
 837          * from its own. This should rarely happen. */
 838         spin_lock(&cache->ccc_lru_lock);
 839         LASSERT(!list_empty(&cache->ccc_lru));
 840
 841         cache->ccc_lru_shrinkers++;
 842         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 843
 844         max_scans = atomic_read(&cache->ccc_users);
 845         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 846                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 847                                         cl_lru_osc);
 848
 849                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 850                         cli->cl_import->imp_obd->obd_name, cli,
 851                         atomic_read(&cli->cl_lru_in_list),
 852                         atomic_read(&cli->cl_lru_busy));
 853
 854                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 855                 if (atomic_read(&cli->cl_lru_in_list) > 0) {
 856                         spin_unlock(&cache->ccc_lru_lock);
 857
 858                         rc = osc_lru_shrink(cli, max_to_shrink(cli));
 859                         spin_lock(&cache->ccc_lru_lock);
 860                         if (rc != 0)
 861                                 break;
 862                 }
 863         }
 864         spin_unlock(&cache->ccc_lru_lock);
 865
 866         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 867                 cli->cl_import->imp_obd->obd_name, cli, rc);
 868         return rc;
 869 }
 870
 871 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 872                            struct osc_page *opg)
 873 {
 874         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 875         struct client_obd *cli = osc_cli(obj);
 876         int rc = 0;
 877
 878         if (cli->cl_cache == NULL) /* shall not be in LRU */
 879                 return 0;
 880
 881         LASSERT(atomic_read(cli->cl_lru_left) >= 0);
 882         while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 883                 int gen;
 884
 885                 /* run out of LRU spaces, try to drop some by itself */
 886                 rc = osc_lru_reclaim(cli);
 887                 if (rc < 0)
 888                         break;
 889                 if (rc > 0)
 890                         continue;
 891
 892                 cond_resched();
 893
 894                 /* slowest case, all of caching pages are busy, notifying
 895                  * other OSCs that we're lack of LRU slots. */
 896                 atomic_inc(&osc_lru_waiters);
 897
 898                 gen = atomic_read(&cli->cl_lru_in_list);
 899                 rc = l_wait_event(osc_lru_waitq,
 900                                 atomic_read(cli->cl_lru_left) > 0 ||
 901                                 (atomic_read(&cli->cl_lru_in_list) > 0 &&
 902                                  gen != atomic_read(&cli->cl_lru_in_list)),
 903                                 &lwi);
 904
 905                 atomic_dec(&osc_lru_waiters);
 906                 if (rc < 0)
 907                         break;
 908         }
 909
 910         if (rc >= 0) {
 911                 atomic_inc(&cli->cl_lru_busy);
 912                 opg->ops_in_lru = 1;
 913                 rc = 0;
 914         }
 915
 916         return rc;
 917 }
 918
 919 /** @} osc */