lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  49                            struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Comment out osc_page_protected because it may sleep inside the
  57  * the client_obd_list_lock.
  58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  61  */
  62 #if 0
  63 static int osc_page_is_dlocked(const struct lu_env *env,
  64                                const struct osc_page *opg,
  65                                enum cl_lock_mode mode, int pending, int unref)
  66 {
  67         struct cl_page         *page;
  68         struct osc_object      *obj;
  69         struct osc_thread_info *info;
  70         struct ldlm_res_id     *resname;
  71         struct lustre_handle   *lockh;
  72         ldlm_policy_data_t     *policy;
  73         ldlm_mode_t             dlmmode;
  74         __u64                   flags;
  75
  76         might_sleep();
  77
  78         info = osc_env_info(env);
  79         resname = &info->oti_resname;
  80         policy = &info->oti_policy;
  81         lockh = &info->oti_handle;
  82         page = opg->ops_cl.cpl_page;
  83         obj = cl2osc(opg->ops_cl.cpl_obj);
  84
  85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  86         if (pending)
  87                 flags |= LDLM_FL_CBPENDING;
  88
  89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  90         osc_lock_build_res(env, obj, resname);
  91         osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
  92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  93                               dlmmode, &flags, NULL, lockh, unref);
  94 }
  95
  96 /**
  97  * Checks an invariant that a page in the cache is covered by a lock, as
  98  * needed.
  99  */
 100 static int osc_page_protected(const struct lu_env *env,
 101                               const struct osc_page *opg,
 102                               enum cl_lock_mode mode, int unref)
 103 {
 104         struct cl_object_header *hdr;
 105         struct cl_lock          *scan;
 106         struct cl_page          *page;
 107         struct cl_lock_descr    *descr;
 108         int result;
 109
 110         LINVRNT(!opg->ops_temp);
 111
 112         page = opg->ops_cl.cpl_page;
 113         if (page->cp_owner != NULL &&
 114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 115                 /*
 116                  * If IO is done without locks (liblustre, or lloop), lock is
 117                  * not required.
 118                  */
 119                 result = 1;
 120         else
 121                 /* otherwise check for a DLM lock */
 122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 123         if (result == 0) {
 124                 /* maybe this page is a part of a lockless io? */
 125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 126                 descr = &osc_env_info(env)->oti_descr;
 127                 descr->cld_mode = mode;
 128                 descr->cld_start = page->cp_index;
 129                 descr->cld_end   = page->cp_index;
 130                 spin_lock(&hdr->coh_lock_guard);
 131                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 132                         /*
 133                          * Lock-less sub-lock has to be either in HELD state
 134                          * (when io is actively going on), or in CACHED state,
 135                          * when top-lock is being unlocked:
 136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 137                          */
 138                         if ((scan->cll_state == CLS_HELD ||
 139                              scan->cll_state == CLS_CACHED) &&
 140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 141                                 struct osc_lock *olck;
 142
 143                                 olck = osc_lock_at(scan);
 144                                 result = osc_lock_is_lockless(olck);
 145                                 break;
 146                         }
 147                 }
 148                 spin_unlock(&hdr->coh_lock_guard);
 149         }
 150         return result;
 151 }
 152 #else
 153 static int osc_page_protected(const struct lu_env *env,
 154                               const struct osc_page *opg,
 155                               enum cl_lock_mode mode, int unref)
 156 {
 157         return 1;
 158 }
 159 #endif
 160
 161 /*****************************************************************************
 162  *
 163  * Page operations.
 164  *
 165  */
 166 static void osc_page_fini(const struct lu_env *env,
 167                           struct cl_page_slice *slice)
 168 {
 169 }
 170
 171 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 172 {
 173         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 174
 175         LASSERT(!opg->ops_transfer_pinned);
 176         cl_page_get(page);
 177         lu_ref_add_atomic(&page->cp_reference, label, page);
 178         opg->ops_transfer_pinned = 1;
 179 }
 180
 181 static void osc_page_transfer_put(const struct lu_env *env,
 182                                   struct osc_page *opg)
 183 {
 184         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 185
 186         if (opg->ops_transfer_pinned) {
 187                 opg->ops_transfer_pinned = 0;
 188                 lu_ref_del(&page->cp_reference, "transfer", page);
 189                 cl_page_put(env, page);
 190         }
 191 }
 192
 193 /**
 194  * This is called once for every page when it is submitted for a transfer
 195  * either opportunistic (osc_page_cache_add()), or immediate
 196  * (osc_page_submit()).
 197  */
 198 static void osc_page_transfer_add(const struct lu_env *env,
 199                                   struct osc_page *opg, enum cl_req_type crt)
 200 {
 201         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 202
 203         /* ops_lru and ops_inflight share the same field, so take it from LRU
 204          * first and then use it as inflight. */
 205         osc_lru_use(osc_cli(obj), opg);
 206
 207         spin_lock(&obj->oo_seatbelt);
 208         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 209         opg->ops_submitter = current;
 210         spin_unlock(&obj->oo_seatbelt);
 211 }
 212
 213 int osc_page_cache_add(const struct lu_env *env,
 214                         const struct cl_page_slice *slice, struct cl_io *io)
 215 {
 216         struct osc_page *opg = cl2osc_page(slice);
 217         int result;
 218         ENTRY;
 219
 220         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 221
 222         osc_page_transfer_get(opg, "transfer\0cache");
 223         result = osc_queue_async_io(env, io, opg);
 224         if (result != 0)
 225                 osc_page_transfer_put(env, opg);
 226         else
 227                 osc_page_transfer_add(env, opg, CRT_WRITE);
 228
 229         RETURN(result);
 230 }
 231
 232 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 233                       pgoff_t start, pgoff_t end)
 234 {
 235         memset(policy, 0, sizeof *policy);
 236         policy->l_extent.start = cl_offset(obj, start);
 237         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 238 }
 239
 240 static int osc_page_is_under_lock(const struct lu_env *env,
 241                                   const struct cl_page_slice *slice,
 242                                   struct cl_io *unused)
 243 {
 244         struct cl_lock *lock;
 245         int             result = -ENODATA;
 246
 247         ENTRY;
 248         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 249                                NULL, 1, 0);
 250         if (lock != NULL) {
 251                 cl_lock_put(env, lock);
 252                 result = -EBUSY;
 253         }
 254         RETURN(result);
 255 }
 256
 257 static const char *osc_list(cfs_list_t *head)
 258 {
 259         return cfs_list_empty(head) ? "-" : "+";
 260 }
 261
 262 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 263 {
 264         if (opg->ops_submit_time == 0)
 265                 return 0;
 266
 267         return (cfs_time_current() - opg->ops_submit_time);
 268 }
 269
 270 static int osc_page_print(const struct lu_env *env,
 271                           const struct cl_page_slice *slice,
 272                           void *cookie, lu_printer_t printer)
 273 {
 274         struct osc_page       *opg = cl2osc_page(slice);
 275         struct osc_async_page *oap = &opg->ops_oap;
 276         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 277         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 278
 279         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 280                           "1< %#x %d %u %s %s > "
 281                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 282                           "3< %s %p %d %lu %d > "
 283                           "4< %d %d %d %lu %s | %s %s %s %s > "
 284                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 285                           opg,
 286                           /* 1 */
 287                           oap->oap_magic, oap->oap_cmd,
 288                           oap->oap_interrupted,
 289                           osc_list(&oap->oap_pending_item),
 290                           osc_list(&oap->oap_rpc_item),
 291                           /* 2 */
 292                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 293                           oap->oap_async_flags, oap->oap_brw_flags,
 294                           oap->oap_request, oap->oap_cli, obj,
 295                           /* 3 */
 296                           osc_list(&opg->ops_inflight),
 297                           opg->ops_submitter, opg->ops_transfer_pinned,
 298                           osc_submit_duration(opg), opg->ops_srvlock,
 299                           /* 4 */
 300                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 301                           cli->cl_max_rpcs_in_flight,
 302                           cli->cl_avail_grant,
 303                           osc_list(&cli->cl_cache_waiters),
 304                           osc_list(&cli->cl_loi_ready_list),
 305                           osc_list(&cli->cl_loi_hp_ready_list),
 306                           osc_list(&cli->cl_loi_write_list),
 307                           osc_list(&cli->cl_loi_read_list),
 308                           /* 5 */
 309                           osc_list(&obj->oo_ready_item),
 310                           osc_list(&obj->oo_hp_ready_item),
 311                           osc_list(&obj->oo_write_item),
 312                           osc_list(&obj->oo_read_item),
 313                           cfs_atomic_read(&obj->oo_nr_reads),
 314                           osc_list(&obj->oo_reading_exts),
 315                           cfs_atomic_read(&obj->oo_nr_writes),
 316                           osc_list(&obj->oo_hp_exts),
 317                           osc_list(&obj->oo_urgent_exts));
 318 }
 319
 320 static void osc_page_delete(const struct lu_env *env,
 321                             const struct cl_page_slice *slice)
 322 {
 323         struct osc_page   *opg = cl2osc_page(slice);
 324         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 325         int rc;
 326
 327         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 328
 329         ENTRY;
 330         CDEBUG(D_TRACE, "%p\n", opg);
 331         osc_page_transfer_put(env, opg);
 332         rc = osc_teardown_async_page(env, obj, opg);
 333         if (rc) {
 334                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 335                               "Trying to teardown failed: %d\n", rc);
 336                 LASSERT(0);
 337         }
 338
 339         spin_lock(&obj->oo_seatbelt);
 340         if (opg->ops_submitter != NULL) {
 341                 LASSERT(!cfs_list_empty(&opg->ops_inflight));
 342                 cfs_list_del_init(&opg->ops_inflight);
 343                 opg->ops_submitter = NULL;
 344         }
 345         spin_unlock(&obj->oo_seatbelt);
 346
 347         osc_lru_del(osc_cli(obj), opg);
 348
 349         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 350                 void *value;
 351
 352                 spin_lock(&obj->oo_tree_lock);
 353                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 354                 if (value != NULL)
 355                         --obj->oo_npages;
 356                 spin_unlock(&obj->oo_tree_lock);
 357
 358                 LASSERT(ergo(value != NULL, value == opg));
 359         }
 360
 361         EXIT;
 362 }
 363
 364 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 365                    int from, int to)
 366 {
 367         struct osc_page       *opg = cl2osc_page(slice);
 368         struct osc_async_page *oap = &opg->ops_oap;
 369
 370         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 371
 372         opg->ops_from = from;
 373         opg->ops_to   = to;
 374         spin_lock(&oap->oap_lock);
 375         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 376         spin_unlock(&oap->oap_lock);
 377 }
 378
 379 static int osc_page_cancel(const struct lu_env *env,
 380                            const struct cl_page_slice *slice)
 381 {
 382         struct osc_page *opg = cl2osc_page(slice);
 383         int rc = 0;
 384
 385         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 386
 387         /* Check if the transferring against this page
 388          * is completed, or not even queued. */
 389         if (opg->ops_transfer_pinned)
 390                 /* FIXME: may not be interrupted.. */
 391                 rc = osc_cancel_async_page(env, opg);
 392         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 393         return rc;
 394 }
 395
 396 static int osc_page_flush(const struct lu_env *env,
 397                           const struct cl_page_slice *slice,
 398                           struct cl_io *io)
 399 {
 400         struct osc_page *opg = cl2osc_page(slice);
 401         int rc = 0;
 402         ENTRY;
 403         rc = osc_flush_async_page(env, io, opg);
 404         RETURN(rc);
 405 }
 406
 407 static const struct cl_page_operations osc_page_ops = {
 408         .cpo_fini          = osc_page_fini,
 409         .cpo_print         = osc_page_print,
 410         .cpo_delete        = osc_page_delete,
 411         .cpo_is_under_lock = osc_page_is_under_lock,
 412         .cpo_clip           = osc_page_clip,
 413         .cpo_cancel         = osc_page_cancel,
 414         .cpo_flush          = osc_page_flush
 415 };
 416
 417 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 418                   struct cl_page *page, struct page *vmpage)
 419 {
 420         struct osc_object *osc = cl2osc(obj);
 421         struct osc_page   *opg = cl_object_page_slice(obj, page);
 422         int result;
 423
 424         opg->ops_from = 0;
 425         opg->ops_to   = PAGE_CACHE_SIZE;
 426
 427         result = osc_prep_async_page(osc, opg, vmpage,
 428                                         cl_offset(obj, page->cp_index));
 429         if (result == 0) {
 430                 struct osc_io *oio = osc_env_io(env);
 431                 opg->ops_srvlock = osc_io_srvlock(oio);
 432                 cl_page_slice_add(page, &opg->ops_cl, obj,
 433                                 &osc_page_ops);
 434         }
 435         /*
 436          * Cannot assert osc_page_protected() here as read-ahead
 437          * creates temporary pages outside of a lock.
 438          */
 439 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 440         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 441 #endif
 442         /* ops_inflight and ops_lru are the same field, but it doesn't
 443          * hurt to initialize it twice :-) */
 444         CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 445         CFS_INIT_LIST_HEAD(&opg->ops_lru);
 446
 447         /* reserve an LRU space for this page */
 448         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 449                 result = osc_lru_reserve(env, osc, opg);
 450                 if (result == 0) {
 451                         spin_lock(&osc->oo_tree_lock);
 452                         result = radix_tree_insert(&osc->oo_tree,
 453                                                    page->cp_index, opg);
 454                         if (result == 0)
 455                                 ++osc->oo_npages;
 456                         spin_unlock(&osc->oo_tree_lock);
 457                         LASSERT(result == 0);
 458                 }
 459         }
 460
 461         return result;
 462 }
 463
 464 int osc_over_unstable_soft_limit(struct client_obd *cli)
 465 {
 466         long obd_upages, obd_dpages, osc_upages;
 467
 468         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 469         if (cli == NULL)
 470                 return 0;
 471
 472         obd_upages = cfs_atomic_read(&obd_unstable_pages);
 473         obd_dpages = cfs_atomic_read(&obd_dirty_pages);
 474
 475         osc_upages = cfs_atomic_read(&cli->cl_unstable_count);
 476
 477         /* obd_max_dirty_pages is the max number of (dirty + unstable)
 478          * pages allowed at any given time. To simulate an unstable page
 479          * only limit, we subtract the current number of dirty pages
 480          * from this max. This difference is roughly the amount of pages
 481          * currently available for unstable pages. Thus, the soft limit
 482          * is half of that difference. Check osc_upages to ensure we don't
 483          * set SOFT_SYNC for OSCs without any outstanding unstable pages. */
 484         return osc_upages != 0 &&
 485                obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
 486 }
 487
 488 /**
 489  * Helper function called by osc_io_submit() for every page in an immediate
 490  * transfer (i.e., transferred synchronously).
 491  */
 492 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 493                      enum cl_req_type crt, int brw_flags)
 494 {
 495         struct osc_async_page *oap = &opg->ops_oap;
 496         struct osc_object     *obj = oap->oap_obj;
 497
 498         LINVRNT(osc_page_protected(env, opg,
 499                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 500
 501         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 502                  "magic 0x%x\n", oap, oap->oap_magic);
 503         LASSERT(oap->oap_async_flags & ASYNC_READY);
 504         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 505
 506         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 507         oap->oap_page_off  = opg->ops_from;
 508         oap->oap_count     = opg->ops_to - opg->ops_from;
 509         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 510
 511         if (osc_over_unstable_soft_limit(oap->oap_cli))
 512                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
 513
 514         if (!client_is_remote(osc_export(obj)) &&
 515                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 516                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 517                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 518         }
 519
 520         opg->ops_submit_time = cfs_time_current();
 521         osc_page_transfer_get(opg, "transfer\0imm");
 522         osc_page_transfer_add(env, opg, crt);
 523 }
 524
 525 /* --------------- LRU page management ------------------ */
 526
 527 /* OSC is a natural place to manage LRU pages as applications are specialized
 528  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 529  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 530  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 531  * for free LRU slots - this will be very bad so the algorithm requires each
 532  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 533  * at any time.
 534  */
 535
 536 static CFS_DECL_WAITQ(osc_lru_waitq);
 537 /* LRU pages are freed in batch mode. OSC should at least free this
 538  * number of pages to avoid running out of LRU budget, and.. */
 539 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 540 /* free this number at most otherwise it will take too long time to finsih. */
 541 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 542
 543 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 544  * we should free slots aggressively. In this way, slots are freed in a steady
 545  * step to maintain fairness among OSCs.
 546  *
 547  * Return how many LRU pages should be freed. */
 548 static int osc_cache_too_much(struct client_obd *cli)
 549 {
 550         struct cl_client_cache *cache = cli->cl_cache;
 551         int pages = cfs_atomic_read(&cli->cl_lru_in_list);
 552         unsigned long budget;
 553
 554         budget = cache->ccc_lru_max / cfs_atomic_read(&cache->ccc_users);
 555
 556         /* if it's going to run out LRU slots, we should free some, but not
 557          * too much to maintain faireness among OSCs. */
 558         if (cfs_atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 559                 if (pages >= budget)
 560                         return lru_shrink_max;
 561                 else if (pages >= budget / 2)
 562                         return lru_shrink_min;
 563         } else if (pages >= budget * 2)
 564                 return lru_shrink_min;
 565         return 0;
 566 }
 567
 568 int lru_queue_work(const struct lu_env *env, void *data)
 569 {
 570         struct client_obd *cli = data;
 571
 572         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 573
 574         if (osc_cache_too_much(cli))
 575                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 576
 577         RETURN(0);
 578 }
 579
 580 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *plist)
 581 {
 582         CFS_LIST_HEAD(lru);
 583         struct osc_async_page *oap;
 584         int npages = 0;
 585
 586         cfs_list_for_each_entry(oap, plist, oap_pending_item) {
 587                 struct osc_page *opg = oap2osc_page(oap);
 588
 589                 if (!opg->ops_in_lru)
 590                         continue;
 591
 592                 ++npages;
 593                 LASSERT(cfs_list_empty(&opg->ops_lru));
 594                 cfs_list_add(&opg->ops_lru, &lru);
 595         }
 596
 597         if (npages > 0) {
 598                 client_obd_list_lock(&cli->cl_lru_list_lock);
 599                 cfs_list_splice_tail(&lru, &cli->cl_lru_list);
 600                 cfs_atomic_sub(npages, &cli->cl_lru_busy);
 601                 cfs_atomic_add(npages, &cli->cl_lru_in_list);
 602                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 603
 604                 /* XXX: May set force to be true for better performance */
 605                 if (osc_cache_too_much(cli))
 606                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 607         }
 608 }
 609
 610 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 611 {
 612         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) > 0);
 613         cfs_list_del_init(&opg->ops_lru);
 614         cfs_atomic_dec(&cli->cl_lru_in_list);
 615 }
 616
 617 /**
 618  * Page is being destroyed. The page may be not in LRU list, if the transfer
 619  * has never finished(error occurred).
 620  */
 621 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 622 {
 623         if (opg->ops_in_lru) {
 624                 client_obd_list_lock(&cli->cl_lru_list_lock);
 625                 if (!cfs_list_empty(&opg->ops_lru)) {
 626                         __osc_lru_del(cli, opg);
 627                 } else {
 628                         LASSERT(cfs_atomic_read(&cli->cl_lru_busy) > 0);
 629                         cfs_atomic_dec(&cli->cl_lru_busy);
 630                 }
 631                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 632
 633                 cfs_atomic_inc(cli->cl_lru_left);
 634                 /* this is a great place to release more LRU pages if
 635                  * this osc occupies too many LRU pages and kernel is
 636                  * stealing one of them. */
 637                 if (!memory_pressure_get())
 638                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 639                 wake_up(&osc_lru_waitq);
 640         } else {
 641                 LASSERT(cfs_list_empty(&opg->ops_lru));
 642         }
 643 }
 644
 645 /**
 646  * Delete page from LRUlist for redirty.
 647  */
 648 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 649 {
 650         /* If page is being transfered for the first time,
 651          * ops_lru should be empty */
 652         if (opg->ops_in_lru && !cfs_list_empty(&opg->ops_lru)) {
 653                 client_obd_list_lock(&cli->cl_lru_list_lock);
 654                 __osc_lru_del(cli, opg);
 655                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 656                 cfs_atomic_inc(&cli->cl_lru_busy);
 657         }
 658 }
 659
 660 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 661                                 struct cl_page **pvec, int max_index)
 662 {
 663         int i;
 664
 665         for (i = 0; i < max_index; i++) {
 666                 struct cl_page *page = pvec[i];
 667
 668                 LASSERT(cl_page_is_owned(page, io));
 669                 cl_page_discard(env, io, page);
 670                 cl_page_disown(env, io, page);
 671                 cl_page_put(env, page);
 672
 673                 pvec[i] = NULL;
 674         }
 675 }
 676
 677 /**
 678  * Drop @target of pages from LRU at most.
 679  */
 680 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 681                    int target, bool force)
 682 {
 683         struct cl_io *io;
 684         struct cl_object *clobj = NULL;
 685         struct cl_page **pvec;
 686         struct osc_page *opg;
 687         int maxscan = 0;
 688         int count = 0;
 689         int index = 0;
 690         int rc = 0;
 691         ENTRY;
 692
 693         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) >= 0);
 694         if (cfs_atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 695                 RETURN(0);
 696
 697         if (!force) {
 698                 if (cfs_atomic_read(&cli->cl_lru_shrinkers) > 0)
 699                         RETURN(-EBUSY);
 700
 701                 if (cfs_atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 702                         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 703                         RETURN(-EBUSY);
 704                 }
 705         } else {
 706                 cfs_atomic_inc(&cli->cl_lru_shrinkers);
 707         }
 708
 709         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 710         io = &osc_env_info(env)->oti_io;
 711
 712         client_obd_list_lock(&cli->cl_lru_list_lock);
 713         maxscan = min(target << 1, cfs_atomic_read(&cli->cl_lru_in_list));
 714         while (!cfs_list_empty(&cli->cl_lru_list)) {
 715                 struct cl_page *page;
 716                 bool will_free = false;
 717
 718                 if (--maxscan < 0)
 719                         break;
 720
 721                 opg = cfs_list_entry(cli->cl_lru_list.next, struct osc_page,
 722                                      ops_lru);
 723                 page = cl_page_top(opg->ops_cl.cpl_page);
 724                 if (cl_page_in_use_noref(page)) {
 725                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 726                         continue;
 727                 }
 728
 729                 LASSERT(page->cp_obj != NULL);
 730                 if (clobj != page->cp_obj) {
 731                         struct cl_object *tmp = page->cp_obj;
 732
 733                         cl_object_get(tmp);
 734                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 735
 736                         if (clobj != NULL) {
 737                                 discard_pagevec(env, io, pvec, index);
 738                                 index = 0;
 739
 740                                 cl_io_fini(env, io);
 741                                 cl_object_put(env, clobj);
 742                                 clobj = NULL;
 743                         }
 744
 745                         clobj = tmp;
 746                         io->ci_obj = clobj;
 747                         io->ci_ignore_layout = 1;
 748                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 749
 750                         client_obd_list_lock(&cli->cl_lru_list_lock);
 751
 752                         if (rc != 0)
 753                                 break;
 754
 755                         ++maxscan;
 756                         continue;
 757                 }
 758
 759                 if (cl_page_own_try(env, io, page) == 0) {
 760                         if (!cl_page_in_use_noref(page)) {
 761                                 /* remove it from lru list earlier to avoid
 762                                  * lock contention */
 763                                 __osc_lru_del(cli, opg);
 764                                 opg->ops_in_lru = 0; /* will be discarded */
 765
 766                                 cl_page_get(page);
 767                                 will_free = true;
 768                         } else {
 769                                 cl_page_disown(env, io, page);
 770                         }
 771                 }
 772
 773                 if (!will_free) {
 774                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 775                         continue;
 776                 }
 777
 778                 /* Don't discard and free the page with cl_lru_list held */
 779                 pvec[index++] = page;
 780                 if (unlikely(index == OTI_PVEC_SIZE)) {
 781                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 782                         discard_pagevec(env, io, pvec, index);
 783                         index = 0;
 784
 785                         client_obd_list_lock(&cli->cl_lru_list_lock);
 786                 }
 787
 788                 if (++count >= target)
 789                         break;
 790         }
 791         client_obd_list_unlock(&cli->cl_lru_list_lock);
 792
 793         if (clobj != NULL) {
 794                 discard_pagevec(env, io, pvec, index);
 795
 796                 cl_io_fini(env, io);
 797                 cl_object_put(env, clobj);
 798         }
 799
 800         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 801         if (count > 0) {
 802                 cfs_atomic_add(count, cli->cl_lru_left);
 803                 wake_up_all(&osc_lru_waitq);
 804         }
 805         RETURN(count > 0 ? count : rc);
 806 }
 807
 808 static inline int max_to_shrink(struct client_obd *cli)
 809 {
 810         return min(cfs_atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 811 }
 812
 813 int osc_lru_reclaim(struct client_obd *cli)
 814 {
 815         struct cl_env_nest nest;
 816         struct lu_env *env;
 817         struct cl_client_cache *cache = cli->cl_cache;
 818         int max_scans;
 819         int rc = 0;
 820         ENTRY;
 821
 822         LASSERT(cache != NULL);
 823         LASSERT(!cfs_list_empty(&cache->ccc_lru));
 824
 825         env = cl_env_nested_get(&nest);
 826         if (IS_ERR(env))
 827                 RETURN(rc);
 828
 829         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 830         if (rc != 0) {
 831                 if (rc == -EBUSY)
 832                         rc = 0;
 833
 834                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 835                         cli->cl_import->imp_obd->obd_name, rc, cli);
 836                 GOTO(out, rc);
 837         }
 838
 839         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 840                 cli->cl_import->imp_obd->obd_name, cli,
 841                 cfs_atomic_read(&cli->cl_lru_in_list),
 842                 cfs_atomic_read(&cli->cl_lru_busy));
 843
 844         /* Reclaim LRU slots from other client_obd as it can't free enough
 845          * from its own. This should rarely happen. */
 846         spin_lock(&cache->ccc_lru_lock);
 847         cache->ccc_lru_shrinkers++;
 848         cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 849
 850         max_scans = cfs_atomic_read(&cache->ccc_users);
 851         while (--max_scans > 0 && !cfs_list_empty(&cache->ccc_lru)) {
 852                 cli = cfs_list_entry(cache->ccc_lru.next, struct client_obd,
 853                                         cl_lru_osc);
 854
 855                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 856                         cli->cl_import->imp_obd->obd_name, cli,
 857                         cfs_atomic_read(&cli->cl_lru_in_list),
 858                         cfs_atomic_read(&cli->cl_lru_busy));
 859
 860                 cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 861                 if (osc_cache_too_much(cli) > 0) {
 862                         spin_unlock(&cache->ccc_lru_lock);
 863
 864                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 865                                             true);
 866                         spin_lock(&cache->ccc_lru_lock);
 867                         if (rc != 0)
 868                                 break;
 869                 }
 870         }
 871         spin_unlock(&cache->ccc_lru_lock);
 872
 873 out:
 874         cl_env_nested_put(&nest, env);
 875         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 876                 cli->cl_import->imp_obd->obd_name, cli, rc);
 877         return rc;
 878 }
 879
 880 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 881                            struct osc_page *opg)
 882 {
 883         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 884         struct osc_io *oio = osc_env_io(env);
 885         struct client_obd *cli = osc_cli(obj);
 886         int rc = 0;
 887         ENTRY;
 888
 889         if (cli->cl_cache == NULL) /* shall not be in LRU */
 890                 RETURN(0);
 891
 892         if (oio->oi_lru_reserved > 0) {
 893                 --oio->oi_lru_reserved;
 894                 goto out;
 895         }
 896
 897         LASSERT(cfs_atomic_read(cli->cl_lru_left) >= 0);
 898         while (!cfs_atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 899
 900                 /* run out of LRU spaces, try to drop some by itself */
 901                 rc = osc_lru_reclaim(cli);
 902                 if (rc < 0)
 903                         break;
 904                 if (rc > 0)
 905                         continue;
 906
 907                 cond_resched();
 908                 rc = l_wait_event(osc_lru_waitq,
 909                                 cfs_atomic_read(cli->cl_lru_left) > 0,
 910                                 &lwi);
 911                 if (rc < 0)
 912                         break;
 913         }
 914
 915 out:
 916         if (rc >= 0) {
 917                 cfs_atomic_inc(&cli->cl_lru_busy);
 918                 opg->ops_in_lru = 1;
 919                 rc = 0;
 920         }
 921
 922         RETURN(rc);
 923 }
 924
 925 /** @} osc */