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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Page operations.
  54  */
  55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  56 {
  57         struct cl_page *page = opg->ops_cl.cpl_page;
  58
  59         LASSERT(!opg->ops_transfer_pinned);
  60         cl_page_get(page);
  61         lu_ref_add_atomic(&page->cp_reference, label, page);
  62         opg->ops_transfer_pinned = 1;
  63 }
  64
  65 static void osc_page_transfer_put(const struct lu_env *env,
  66                                   struct osc_page *opg)
  67 {
  68         struct cl_page *page = opg->ops_cl.cpl_page;
  69
  70         if (opg->ops_transfer_pinned) {
  71                 opg->ops_transfer_pinned = 0;
  72                 lu_ref_del(&page->cp_reference, "transfer", page);
  73                 cl_page_put(env, page);
  74         }
  75 }
  76
  77 /**
  78  * This is called once for every page when it is submitted for a transfer
  79  * either opportunistic (osc_page_cache_add()), or immediate
  80  * (osc_page_submit()).
  81  */
  82 static void osc_page_transfer_add(const struct lu_env *env,
  83                                   struct osc_page *opg, enum cl_req_type crt)
  84 {
  85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  86
  87         /* ops_lru and ops_inflight share the same field, so take it from LRU
  88          * first and then use it as inflight. */
  89         osc_lru_use(osc_cli(obj), opg);
  90 }
  91
  92 int osc_page_cache_add(const struct lu_env *env,
  93                         const struct cl_page_slice *slice, struct cl_io *io)
  94 {
  95         struct osc_page *opg = cl2osc_page(slice);
  96         int result;
  97         ENTRY;
  98
  99         osc_page_transfer_get(opg, "transfer\0cache");
 100         result = osc_queue_async_io(env, io, opg);
 101         if (result != 0)
 102                 osc_page_transfer_put(env, opg);
 103         else
 104                 osc_page_transfer_add(env, opg, CRT_WRITE);
 105
 106         RETURN(result);
 107 }
 108
 109 void osc_index2policy(union ldlm_policy_data *policy,
 110                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 111 {
 112         memset(policy, 0, sizeof *policy);
 113         policy->l_extent.start = cl_offset(obj, start);
 114         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 115 }
 116
 117 static const char *osc_list(struct list_head *head)
 118 {
 119         return list_empty(head) ? "-" : "+";
 120 }
 121
 122 static inline s64 osc_submit_duration(struct osc_page *opg)
 123 {
 124         if (ktime_to_ns(opg->ops_submit_time) == 0)
 125                 return 0;
 126
 127         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 128 }
 129
 130 static int osc_page_print(const struct lu_env *env,
 131                           const struct cl_page_slice *slice,
 132                           void *cookie, lu_printer_t printer)
 133 {
 134         struct osc_page       *opg = cl2osc_page(slice);
 135         struct osc_async_page *oap = &opg->ops_oap;
 136         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 137         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 138
 139         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 140                           "1< %#x %d %u %s %s > "
 141                           "2< %lld %u %u %#x %#x | %p %p %p > "
 142                           "3< %d %lld %d > "
 143                           "4< %d %d %d %lu %s | %s %s %s %s > "
 144                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 145                           opg, osc_index(opg),
 146                           /* 1 */
 147                           oap->oap_magic, oap->oap_cmd,
 148                           oap->oap_interrupted,
 149                           osc_list(&oap->oap_pending_item),
 150                           osc_list(&oap->oap_rpc_item),
 151                           /* 2 */
 152                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 153                           oap->oap_async_flags, oap->oap_brw_flags,
 154                           oap->oap_request, oap->oap_cli, obj,
 155                           /* 3 */
 156                           opg->ops_transfer_pinned,
 157                           osc_submit_duration(opg), opg->ops_srvlock,
 158                           /* 4 */
 159                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 160                           cli->cl_max_rpcs_in_flight,
 161                           cli->cl_avail_grant,
 162                           osc_list(&cli->cl_cache_waiters),
 163                           osc_list(&cli->cl_loi_ready_list),
 164                           osc_list(&cli->cl_loi_hp_ready_list),
 165                           osc_list(&cli->cl_loi_write_list),
 166                           osc_list(&cli->cl_loi_read_list),
 167                           /* 5 */
 168                           osc_list(&obj->oo_ready_item),
 169                           osc_list(&obj->oo_hp_ready_item),
 170                           osc_list(&obj->oo_write_item),
 171                           osc_list(&obj->oo_read_item),
 172                           atomic_read(&obj->oo_nr_reads),
 173                           osc_list(&obj->oo_reading_exts),
 174                           atomic_read(&obj->oo_nr_writes),
 175                           osc_list(&obj->oo_hp_exts),
 176                           osc_list(&obj->oo_urgent_exts));
 177 }
 178
 179 static void osc_page_delete(const struct lu_env *env,
 180                             const struct cl_page_slice *slice)
 181 {
 182         struct osc_page   *opg = cl2osc_page(slice);
 183         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 184         int rc;
 185
 186         ENTRY;
 187         CDEBUG(D_TRACE, "%p\n", opg);
 188         osc_page_transfer_put(env, opg);
 189         rc = osc_teardown_async_page(env, obj, opg);
 190         if (rc) {
 191                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 192                               "Trying to teardown failed: %d\n", rc);
 193                 LASSERT(0);
 194         }
 195
 196         osc_lru_del(osc_cli(obj), opg);
 197
 198         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 199                 void *value = NULL;
 200
 201                 spin_lock(&obj->oo_tree_lock);
 202                 if (opg->ops_intree) {
 203                         value = radix_tree_delete(&obj->oo_tree,
 204                                                   osc_index(opg));
 205                         if (value != NULL) {
 206                                 --obj->oo_npages;
 207                                 opg->ops_intree = 0;
 208                         }
 209                 }
 210                 spin_unlock(&obj->oo_tree_lock);
 211
 212                 LASSERT(ergo(value != NULL, value == opg));
 213         }
 214
 215         EXIT;
 216 }
 217
 218 static void osc_page_clip(const struct lu_env *env,
 219                           const struct cl_page_slice *slice,
 220                           int from, int to)
 221 {
 222         struct osc_page       *opg = cl2osc_page(slice);
 223         struct osc_async_page *oap = &opg->ops_oap;
 224
 225         opg->ops_from = from;
 226         opg->ops_to   = to;
 227         spin_lock(&oap->oap_lock);
 228         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 229         spin_unlock(&oap->oap_lock);
 230 }
 231
 232 static int osc_page_cancel(const struct lu_env *env,
 233                            const struct cl_page_slice *slice)
 234 {
 235         struct osc_page *opg = cl2osc_page(slice);
 236         int rc = 0;
 237
 238         /* Check if the transferring against this page
 239          * is completed, or not even queued. */
 240         if (opg->ops_transfer_pinned)
 241                 /* FIXME: may not be interrupted.. */
 242                 rc = osc_cancel_async_page(env, opg);
 243         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 244         return rc;
 245 }
 246
 247 static int osc_page_flush(const struct lu_env *env,
 248                           const struct cl_page_slice *slice,
 249                           struct cl_io *io)
 250 {
 251         struct osc_page *opg = cl2osc_page(slice);
 252         int rc = 0;
 253         ENTRY;
 254         rc = osc_flush_async_page(env, io, opg);
 255         RETURN(rc);
 256 }
 257
 258 static void osc_page_touch(const struct lu_env *env,
 259                           const struct cl_page_slice *slice, size_t to)
 260 {
 261         struct osc_page *opg = cl2osc_page(slice);
 262         struct cl_object *obj = opg->ops_cl.cpl_obj;
 263
 264         osc_page_touch_at(env, obj, osc_index(opg), to);
 265 }
 266
 267 static const struct cl_page_operations osc_page_ops = {
 268         .cpo_print         = osc_page_print,
 269         .cpo_delete        = osc_page_delete,
 270         .cpo_clip           = osc_page_clip,
 271         .cpo_cancel         = osc_page_cancel,
 272         .cpo_flush          = osc_page_flush,
 273         .cpo_page_touch    = osc_page_touch,
 274 };
 275
 276 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 277                   struct cl_page *page, pgoff_t index)
 278 {
 279         struct osc_object *osc = cl2osc(obj);
 280         struct osc_page   *opg = cl_object_page_slice(obj, page);
 281         struct osc_io *oio = osc_env_io(env);
 282         int result;
 283
 284         opg->ops_from = 0;
 285         opg->ops_to   = PAGE_SIZE;
 286
 287         INIT_LIST_HEAD(&opg->ops_lru);
 288
 289         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 290                                      cl_offset(obj, index));
 291         if (result != 0)
 292                 return result;
 293
 294         opg->ops_srvlock = osc_io_srvlock(oio);
 295         cl_page_slice_add(page, &opg->ops_cl, obj, index,
 296                           &osc_page_ops);
 297
 298
 299         /* reserve an LRU space for this page */
 300         if (page->cp_type == CPT_CACHEABLE) {
 301                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 302                 if (result == 0) {
 303                         result = radix_tree_preload(GFP_NOFS);
 304                         if (result == 0) {
 305                                 spin_lock(&osc->oo_tree_lock);
 306                                 result = radix_tree_insert(&osc->oo_tree,
 307                                                            index, opg);
 308                                 if (result == 0) {
 309                                         ++osc->oo_npages;
 310                                         opg->ops_intree = 1;
 311                                 }
 312                                 spin_unlock(&osc->oo_tree_lock);
 313
 314                                 radix_tree_preload_end();
 315                         }
 316                 }
 317         }
 318
 319         return result;
 320 }
 321 EXPORT_SYMBOL(osc_page_init);
 322
 323 /**
 324  * Helper function called by osc_io_submit() for every page in an immediate
 325  * transfer (i.e., transferred synchronously).
 326  */
 327 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 328                      enum cl_req_type crt, int brw_flags)
 329 {
 330         struct osc_async_page *oap = &opg->ops_oap;
 331
 332         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 333                  "magic 0x%x\n", oap, oap->oap_magic);
 334         LASSERT(oap->oap_async_flags & ASYNC_READY);
 335         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 336
 337         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 338         oap->oap_page_off  = opg->ops_from;
 339         oap->oap_count     = opg->ops_to - opg->ops_from;
 340         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 341
 342         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 343                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 344                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 345         }
 346
 347         opg->ops_submit_time = ktime_get();
 348         osc_page_transfer_get(opg, "transfer\0imm");
 349         osc_page_transfer_add(env, opg, crt);
 350 }
 351
 352 /* --------------- LRU page management ------------------ */
 353
 354 /* OSC is a natural place to manage LRU pages as applications are specialized
 355  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 356  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 357  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 358  * for free LRU slots - this will be very bad so the algorithm requires each
 359  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 360  * at any time.
 361  */
 362
 363 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 364
 365 /**
 366  * LRU pages are freed in batch mode. OSC should at least free this
 367  * number of pages to avoid running out of LRU slots.
 368  */
 369 static inline int lru_shrink_min(struct client_obd *cli)
 370 {
 371         return cli->cl_max_pages_per_rpc * 2;
 372 }
 373
 374 /**
 375  * free this number at most otherwise it will take too long time to finsih.
 376  */
 377 static inline int lru_shrink_max(struct client_obd *cli)
 378 {
 379         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 380 }
 381
 382 /**
 383  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 384  * we should free slots aggressively. In this way, slots are freed in a steady
 385  * step to maintain fairness among OSCs.
 386  *
 387  * Return how many LRU pages should be freed.
 388  */
 389 static int osc_cache_too_much(struct client_obd *cli)
 390 {
 391         struct cl_client_cache *cache = cli->cl_cache;
 392         long pages = atomic_long_read(&cli->cl_lru_in_list);
 393         unsigned long budget;
 394
 395         LASSERT(cache != NULL);
 396         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 397
 398         /* if it's going to run out LRU slots, we should free some, but not
 399          * too much to maintain faireness among OSCs. */
 400         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 401                 if (pages >= budget)
 402                         return lru_shrink_max(cli);
 403                 else if (pages >= budget / 2)
 404                         return lru_shrink_min(cli);
 405         } else {
 406                 time64_t duration = ktime_get_real_seconds();
 407                 long timediff;
 408
 409                 /* knock out pages by duration of no IO activity */
 410                 duration -= cli->cl_lru_last_used;
 411                 /*
 412                  * The difference shouldn't be more than 70 years
 413                  * so we can safely case to a long. Round to
 414                  * approximately 1 minute.
 415                  */
 416                 timediff = (long)(duration >> 6);
 417                 if (timediff > 0 && pages >= budget / timediff)
 418                         return lru_shrink_min(cli);
 419         }
 420         return 0;
 421 }
 422
 423 int lru_queue_work(const struct lu_env *env, void *data)
 424 {
 425         struct client_obd *cli = data;
 426         int count;
 427
 428         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 429         count = osc_cache_too_much(cli);
 430         if (count > 0) {
 431                 int rc = osc_lru_shrink(env, cli, count, false);
 432
 433                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 434                        cli_name(cli), rc, count);
 435                 if (rc >= count) {
 436                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 437                         ptlrpcd_queue_work(cli->cl_lru_work);
 438                 }
 439         }
 440
 441         RETURN(0);
 442 }
 443
 444 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 445 {
 446         struct list_head lru = LIST_HEAD_INIT(lru);
 447         struct osc_async_page *oap;
 448         long npages = 0;
 449
 450         list_for_each_entry(oap, plist, oap_pending_item) {
 451                 struct osc_page *opg = oap2osc_page(oap);
 452
 453                 if (!opg->ops_in_lru)
 454                         continue;
 455
 456                 ++npages;
 457                 LASSERT(list_empty(&opg->ops_lru));
 458                 list_add(&opg->ops_lru, &lru);
 459         }
 460
 461         if (npages > 0) {
 462                 spin_lock(&cli->cl_lru_list_lock);
 463                 list_splice_tail(&lru, &cli->cl_lru_list);
 464                 atomic_long_sub(npages, &cli->cl_lru_busy);
 465                 atomic_long_add(npages, &cli->cl_lru_in_list);
 466                 cli->cl_lru_last_used = ktime_get_real_seconds();
 467                 spin_unlock(&cli->cl_lru_list_lock);
 468
 469                 if (waitqueue_active(&osc_lru_waitq))
 470                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 471         }
 472 }
 473
 474 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 475 {
 476         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 477         list_del_init(&opg->ops_lru);
 478         atomic_long_dec(&cli->cl_lru_in_list);
 479 }
 480
 481 /**
 482  * Page is being destroyed. The page may be not in LRU list, if the transfer
 483  * has never finished(error occurred).
 484  */
 485 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 486 {
 487         if (opg->ops_in_lru) {
 488                 spin_lock(&cli->cl_lru_list_lock);
 489                 if (!list_empty(&opg->ops_lru)) {
 490                         __osc_lru_del(cli, opg);
 491                 } else {
 492                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 493                         atomic_long_dec(&cli->cl_lru_busy);
 494                 }
 495                 spin_unlock(&cli->cl_lru_list_lock);
 496
 497                 atomic_long_inc(cli->cl_lru_left);
 498                 /* this is a great place to release more LRU pages if
 499                  * this osc occupies too many LRU pages and kernel is
 500                  * stealing one of them. */
 501                 if (osc_cache_too_much(cli)) {
 502                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 503                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 504                 }
 505                 wake_up(&osc_lru_waitq);
 506         } else {
 507                 LASSERT(list_empty(&opg->ops_lru));
 508         }
 509 }
 510
 511 /**
 512  * Delete page from LRU list for redirty.
 513  */
 514 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 515 {
 516         /* If page is being transferred for the first time,
 517          * ops_lru should be empty */
 518         if (opg->ops_in_lru) {
 519                 spin_lock(&cli->cl_lru_list_lock);
 520                 if (!list_empty(&opg->ops_lru)) {
 521                         __osc_lru_del(cli, opg);
 522                         atomic_long_inc(&cli->cl_lru_busy);
 523                 }
 524                 spin_unlock(&cli->cl_lru_list_lock);
 525         }
 526 }
 527
 528 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 529                                 struct cl_page **pvec, int max_index)
 530 {
 531         int i;
 532
 533         for (i = 0; i < max_index; i++) {
 534                 struct cl_page *page = pvec[i];
 535
 536                 LASSERT(cl_page_is_owned(page, io));
 537                 cl_page_delete(env, page);
 538                 cl_page_discard(env, io, page);
 539                 cl_page_disown(env, io, page);
 540                 cl_page_put(env, page);
 541
 542                 pvec[i] = NULL;
 543         }
 544 }
 545
 546 /**
 547  * Check if a cl_page can be released, i.e, it's not being used.
 548  *
 549  * If unstable account is turned on, bulk transfer may hold one refcount
 550  * for recovery so we need to check vmpage refcount as well; otherwise,
 551  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 552  */
 553 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 554 {
 555         if (cl_page_in_use_noref(page))
 556                 return true;
 557
 558         if (cli->cl_cache->ccc_unstable_check) {
 559                 struct page *vmpage = cl_page_vmpage(page);
 560
 561                 /* vmpage have two known users: cl_page and VM page cache */
 562                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 563                         return true;
 564         }
 565         return false;
 566 }
 567
 568 /**
 569  * Drop @target of pages from LRU at most.
 570  */
 571 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 572                    long target, bool force)
 573 {
 574         struct cl_io *io;
 575         struct cl_object *clobj = NULL;
 576         struct cl_page **pvec;
 577         struct osc_page *opg;
 578         long count = 0;
 579         int maxscan = 0;
 580         int index = 0;
 581         int rc = 0;
 582         ENTRY;
 583
 584         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 585         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 586                 RETURN(0);
 587
 588         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 589                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 590         if (!force) {
 591                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 592                         RETURN(-EBUSY);
 593
 594                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 595                         atomic_dec(&cli->cl_lru_shrinkers);
 596                         RETURN(-EBUSY);
 597                 }
 598         } else {
 599                 atomic_inc(&cli->cl_lru_shrinkers);
 600         }
 601
 602         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 603         io = osc_env_thread_io(env);
 604
 605         spin_lock(&cli->cl_lru_list_lock);
 606         if (force)
 607                 cli->cl_lru_reclaim++;
 608         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 609         while (!list_empty(&cli->cl_lru_list)) {
 610                 struct cl_page *page;
 611                 bool will_free = false;
 612
 613                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 614                         break;
 615
 616                 if (--maxscan < 0)
 617                         break;
 618
 619                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 620                                  ops_lru);
 621                 page = opg->ops_cl.cpl_page;
 622                 if (lru_page_busy(cli, page)) {
 623                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 624                         continue;
 625                 }
 626
 627                 LASSERT(page->cp_obj != NULL);
 628                 if (clobj != page->cp_obj) {
 629                         struct cl_object *tmp = page->cp_obj;
 630
 631                         cl_object_get(tmp);
 632                         spin_unlock(&cli->cl_lru_list_lock);
 633
 634                         if (clobj != NULL) {
 635                                 discard_pagevec(env, io, pvec, index);
 636                                 index = 0;
 637
 638                                 cl_io_fini(env, io);
 639                                 cl_object_put(env, clobj);
 640                                 clobj = NULL;
 641                         }
 642
 643                         clobj = tmp;
 644                         io->ci_obj = clobj;
 645                         io->ci_ignore_layout = 1;
 646                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 647
 648                         spin_lock(&cli->cl_lru_list_lock);
 649
 650                         if (rc != 0)
 651                                 break;
 652
 653                         ++maxscan;
 654                         continue;
 655                 }
 656
 657                 if (cl_page_own_try(env, io, page) == 0) {
 658                         if (!lru_page_busy(cli, page)) {
 659                                 /* remove it from lru list earlier to avoid
 660                                  * lock contention */
 661                                 __osc_lru_del(cli, opg);
 662                                 opg->ops_in_lru = 0; /* will be discarded */
 663
 664                                 cl_page_get(page);
 665                                 will_free = true;
 666                         } else {
 667                                 cl_page_disown(env, io, page);
 668                         }
 669                 }
 670
 671                 if (!will_free) {
 672                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 673                         continue;
 674                 }
 675
 676                 /* Don't discard and free the page with cl_lru_list held */
 677                 pvec[index++] = page;
 678                 if (unlikely(index == OTI_PVEC_SIZE)) {
 679                         spin_unlock(&cli->cl_lru_list_lock);
 680                         discard_pagevec(env, io, pvec, index);
 681                         index = 0;
 682
 683                         spin_lock(&cli->cl_lru_list_lock);
 684                 }
 685
 686                 if (++count >= target)
 687                         break;
 688         }
 689         spin_unlock(&cli->cl_lru_list_lock);
 690
 691         if (clobj != NULL) {
 692                 discard_pagevec(env, io, pvec, index);
 693
 694                 cl_io_fini(env, io);
 695                 cl_object_put(env, clobj);
 696         }
 697
 698         atomic_dec(&cli->cl_lru_shrinkers);
 699         if (count > 0) {
 700                 atomic_long_add(count, cli->cl_lru_left);
 701                 wake_up_all(&osc_lru_waitq);
 702         }
 703         RETURN(count > 0 ? count : rc);
 704 }
 705 EXPORT_SYMBOL(osc_lru_shrink);
 706
 707 /**
 708  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 709  * \@npages of LRU slots. For performance consideration, it's better to drop
 710  * LRU pages in batch. Therefore, the actual number is adjusted at least
 711  * max_pages_per_rpc.
 712  */
 713 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 714 {
 715         struct lu_env *env;
 716         struct cl_client_cache *cache = cli->cl_cache;
 717         int max_scans;
 718         __u16 refcheck;
 719         long rc = 0;
 720         ENTRY;
 721
 722         LASSERT(cache != NULL);
 723
 724         env = cl_env_get(&refcheck);
 725         if (IS_ERR(env))
 726                 RETURN(rc);
 727
 728         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 729         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 730                cli_name(cli), npages);
 731         rc = osc_lru_shrink(env, cli, npages, true);
 732         if (rc >= npages) {
 733                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 734                        cli_name(cli), rc, npages);
 735                 if (osc_cache_too_much(cli) > 0)
 736                         ptlrpcd_queue_work(cli->cl_lru_work);
 737                 GOTO(out, rc);
 738         } else if (rc > 0) {
 739                 npages -= rc;
 740         }
 741
 742         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 743                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 744                 atomic_long_read(&cli->cl_lru_busy), npages);
 745
 746         /* Reclaim LRU slots from other client_obd as it can't free enough
 747          * from its own. This should rarely happen. */
 748         spin_lock(&cache->ccc_lru_lock);
 749         LASSERT(!list_empty(&cache->ccc_lru));
 750
 751         cache->ccc_lru_shrinkers++;
 752         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 753
 754         max_scans = atomic_read(&cache->ccc_users) - 2;
 755         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 756                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 757                                  cl_lru_osc);
 758
 759                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 760                         cli_name(cli), cli,
 761                         atomic_long_read(&cli->cl_lru_in_list),
 762                         atomic_long_read(&cli->cl_lru_busy));
 763
 764                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 765                 if (osc_cache_too_much(cli) > 0) {
 766                         spin_unlock(&cache->ccc_lru_lock);
 767
 768                         rc = osc_lru_shrink(env, cli, npages, true);
 769                         spin_lock(&cache->ccc_lru_lock);
 770                         if (rc >= npages)
 771                                 break;
 772                         if (rc > 0)
 773                                 npages -= rc;
 774                 }
 775         }
 776         spin_unlock(&cache->ccc_lru_lock);
 777
 778 out:
 779         cl_env_put(env, &refcheck);
 780         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 781                 cli_name(cli), cli, rc);
 782         return rc;
 783 }
 784
 785 /**
 786  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 787  *
 788  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 789  * Only in the case that the LRU slots are in extreme shortage, it should
 790  * have reserved enough slots for an IO.
 791  */
 792 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 793                          struct osc_page *opg)
 794 {
 795         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 796         struct osc_io *oio = osc_env_io(env);
 797         int rc = 0;
 798
 799         ENTRY;
 800
 801         if (cli->cl_cache == NULL) /* shall not be in LRU */
 802                 RETURN(0);
 803
 804         if (oio->oi_lru_reserved > 0) {
 805                 --oio->oi_lru_reserved;
 806                 goto out;
 807         }
 808
 809         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 810         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 811                 /* run out of LRU spaces, try to drop some by itself */
 812                 rc = osc_lru_reclaim(cli, 1);
 813                 if (rc < 0)
 814                         break;
 815                 if (rc > 0)
 816                         continue;
 817
 818                 cond_resched();
 819                 rc = l_wait_event(osc_lru_waitq,
 820                                 atomic_long_read(cli->cl_lru_left) > 0,
 821                                 &lwi);
 822                 if (rc < 0)
 823                         break;
 824         }
 825
 826 out:
 827         if (rc >= 0) {
 828                 atomic_long_inc(&cli->cl_lru_busy);
 829                 opg->ops_in_lru = 1;
 830                 rc = 0;
 831         }
 832
 833         RETURN(rc);
 834 }
 835
 836 /**
 837  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 838  *
 839  * The benefit of doing this is to reduce contention against atomic counter
 840  * cl_lru_left by changing it from per-page access to per-IO access.
 841  */
 842 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 843 {
 844         unsigned long reserved = 0;
 845         unsigned long max_pages;
 846         unsigned long c;
 847
 848         /* reserve a full RPC window at most to avoid that a thread accidentally
 849          * consumes too many LRU slots */
 850         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 851         if (npages > max_pages)
 852                 npages = max_pages;
 853
 854         c = atomic_long_read(cli->cl_lru_left);
 855         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 856                 c = atomic_long_read(cli->cl_lru_left);
 857         while (c >= npages) {
 858                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 859                         reserved = npages;
 860                         break;
 861                 }
 862                 c = atomic_long_read(cli->cl_lru_left);
 863         }
 864         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 865                 /* If there aren't enough pages in the per-OSC LRU then
 866                  * wake up the LRU thread to try and clear out space, so
 867                  * we don't block if pages are being dirtied quickly. */
 868                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 869                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 870                        max_pages);
 871                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 872         }
 873
 874         return reserved;
 875 }
 876
 877 /**
 878  * osc_lru_unreserve() is called to unreserve LRU slots.
 879  *
 880  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 881  * reasons such as page already existing or I/O error. Those reserved slots
 882  * should be freed by calling this function.
 883  */
 884 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 885 {
 886         atomic_long_add(npages, cli->cl_lru_left);
 887         wake_up_all(&osc_lru_waitq);
 888 }
 889
 890 /**
 891  * Atomic operations are expensive. We accumulate the accounting for the
 892  * same page zone to get better performance.
 893  * In practice this can work pretty good because the pages in the same RPC
 894  * are likely from the same page zone.
 895  */
 896 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 897                                             struct osc_brw_async_args *aa,
 898                                             int factor)
 899 {
 900         int page_count;
 901         void *zone = NULL;
 902         int count = 0;
 903         int i;
 904
 905         if (desc != NULL) {
 906                 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 907                 page_count = desc->bd_iov_count;
 908         } else {
 909                 page_count = aa->aa_page_count;
 910         }
 911
 912         for (i = 0; i < page_count; i++) {
 913                 void *pz;
 914                 if (desc)
 915                         pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 916                 else
 917                         pz = page_zone(aa->aa_ppga[i]->pg);
 918
 919                 if (likely(pz == zone)) {
 920                         ++count;
 921                         continue;
 922                 }
 923
 924                 if (count > 0) {
 925                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 926                                             factor * count);
 927                         count = 0;
 928                 }
 929                 zone = pz;
 930                 ++count;
 931         }
 932         if (count > 0)
 933                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 934 }
 935
 936 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 937                                                 struct osc_brw_async_args *aa)
 938 {
 939         unstable_page_accounting(desc, aa, 1);
 940 }
 941
 942 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 943                                                 struct osc_brw_async_args *aa)
 944 {
 945         unstable_page_accounting(desc, aa, -1);
 946 }
 947
 948 /**
 949  * Performs "unstable" page accounting. This function balances the
 950  * increment operations performed in osc_inc_unstable_pages. It is
 951  * registered as the RPC request callback, and is executed when the
 952  * bulk RPC is committed on the server. Thus at this point, the pages
 953  * involved in the bulk transfer are no longer considered unstable.
 954  *
 955  * If this function is called, the request should have been committed
 956  * or req:rq_unstable must have been set; it implies that the unstable
 957  * statistic have been added.
 958  */
 959 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 960 {
 961         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 962         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 963         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 964         int                      page_count;
 965         long                     unstable_count;
 966
 967         if (desc)
 968                 page_count = desc->bd_iov_count;
 969         else
 970                 page_count = aa->aa_page_count;
 971
 972         LASSERT(page_count >= 0);
 973
 974         dec_unstable_page_accounting(desc, aa);
 975
 976         unstable_count = atomic_long_sub_return(page_count,
 977                                                 &cli->cl_unstable_count);
 978         LASSERT(unstable_count >= 0);
 979
 980         unstable_count = atomic_long_sub_return(page_count,
 981                                            &cli->cl_cache->ccc_unstable_nr);
 982         LASSERT(unstable_count >= 0);
 983         if (unstable_count == 0)
 984                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 985
 986         if (waitqueue_active(&osc_lru_waitq))
 987                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 988 }
 989
 990 /**
 991  * "unstable" page accounting. See: osc_dec_unstable_pages.
 992  */
 993 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 994 {
 995         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 996         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 997         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 998         long                     page_count;
 999
1000         /* No unstable page tracking */
1001         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1002                 return;
1003
1004         if (desc)
1005                 page_count = desc->bd_iov_count;
1006         else
1007                 page_count = aa->aa_page_count;
1008
1009         add_unstable_page_accounting(desc, aa);
1010         atomic_long_add(page_count, &cli->cl_unstable_count);
1011         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1012
1013         /* If the request has already been committed (i.e. brw_commit
1014          * called via rq_commit_cb), we need to undo the unstable page
1015          * increments we just performed because rq_commit_cb wont be
1016          * called again. */
1017         spin_lock(&req->rq_lock);
1018         if (unlikely(req->rq_committed)) {
1019                 spin_unlock(&req->rq_lock);
1020
1021                 osc_dec_unstable_pages(req);
1022         } else {
1023                 req->rq_unstable = 1;
1024                 spin_unlock(&req->rq_lock);
1025         }
1026 }
1027
1028 /**
1029  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1030  * This function will be called by every BRW RPC so it's critical
1031  * to make this function fast.
1032  */
1033 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1034 {
1035         long unstable_nr, osc_unstable_count;
1036
1037         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1038         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1039                 return false;
1040
1041         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1042         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1043
1044         CDEBUG(D_CACHE,
1045                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1046                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1047
1048         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1049          * has one full RPC window of unstable pages, it's a good chance
1050          * to piggyback a SOFT_SYNC flag.
1051          * Please notice that the OST won't take immediate response for the
1052          * SOFT_SYNC request so active OSCs will have more chance to carry
1053          * the flag, this is reasonable. */
1054         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1055                osc_unstable_count > cli->cl_max_pages_per_rpc *
1056                                     cli->cl_max_rpcs_in_flight;
1057 }
1058
1059 /**
1060  * Return how many LRU pages in the cache of all OSC devices
1061  *
1062  * \retval      return # of cached LRU pages times reclaimation tendency
1063  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1064  */
1065 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1066                                      struct shrink_control *sc)
1067 {
1068         struct client_obd *cli;
1069         unsigned long cached = 0;
1070
1071         spin_lock(&osc_shrink_lock);
1072         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1073                 cached += atomic_long_read(&cli->cl_lru_in_list);
1074         spin_unlock(&osc_shrink_lock);
1075
1076         return (cached  * sysctl_vfs_cache_pressure) / 100;
1077 }
1078
1079 /**
1080  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1081  *
1082  * \retval      number of cached LRU pages reclaimed
1083  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1084  *
1085  * Linux kernel will loop calling this shrinker scan routine with
1086  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1087  *
1088  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1089  * to scan and try to reclaim LRU pages, just return 0 and
1090  * osc_cache_shrink_count() will report the LRU page number.
1091  */
1092 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1093                                     struct shrink_control *sc)
1094 {
1095         struct client_obd *cli;
1096         struct client_obd *stop_anchor = NULL;
1097         struct lu_env *env;
1098         long shrank = 0;
1099         int rc;
1100         __u16 refcheck;
1101
1102         if (sc->nr_to_scan == 0)
1103                 return 0;
1104
1105         if (!(sc->gfp_mask & __GFP_FS))
1106                 return SHRINK_STOP;
1107
1108         env = cl_env_get(&refcheck);
1109         if (IS_ERR(env))
1110                 return SHRINK_STOP;
1111
1112         spin_lock(&osc_shrink_lock);
1113         while (!list_empty(&osc_shrink_list)) {
1114                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1115                                  cl_shrink_list);
1116
1117                 if (stop_anchor == NULL)
1118                         stop_anchor = cli;
1119                 else if (cli == stop_anchor)
1120                         break;
1121
1122                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1123                 spin_unlock(&osc_shrink_lock);
1124
1125                 /* shrink no more than max_pages_per_rpc for an OSC */
1126                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1127                                     cli->cl_max_pages_per_rpc ?
1128                                     cli->cl_max_pages_per_rpc :
1129                                     sc->nr_to_scan - shrank, true);
1130                 if (rc > 0)
1131                         shrank += rc;
1132
1133                 if (shrank >= sc->nr_to_scan)
1134                         goto out;
1135
1136                 spin_lock(&osc_shrink_lock);
1137         }
1138         spin_unlock(&osc_shrink_lock);
1139
1140 out:
1141         cl_env_put(env, &refcheck);
1142
1143         return shrank;
1144 }
1145
1146 /** @} osc */