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, 2016, 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
  40 #include "osc_cl_internal.h"
  41
  42 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  43 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  44 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  45                          struct osc_page *opg);
  46
  47 /** \addtogroup osc
  48  *  @{
  49  */
  50
  51 /*
  52  * Page operations.
  53  */
  54 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  55 {
  56         struct cl_page *page = opg->ops_cl.cpl_page;
  57
  58         LASSERT(!opg->ops_transfer_pinned);
  59         cl_page_get(page);
  60         lu_ref_add_atomic(&page->cp_reference, label, page);
  61         opg->ops_transfer_pinned = 1;
  62 }
  63
  64 static void osc_page_transfer_put(const struct lu_env *env,
  65                                   struct osc_page *opg)
  66 {
  67         struct cl_page *page = opg->ops_cl.cpl_page;
  68
  69         if (opg->ops_transfer_pinned) {
  70                 opg->ops_transfer_pinned = 0;
  71                 lu_ref_del(&page->cp_reference, "transfer", page);
  72                 cl_page_put(env, page);
  73         }
  74 }
  75
  76 /**
  77  * This is called once for every page when it is submitted for a transfer
  78  * either opportunistic (osc_page_cache_add()), or immediate
  79  * (osc_page_submit()).
  80  */
  81 static void osc_page_transfer_add(const struct lu_env *env,
  82                                   struct osc_page *opg, enum cl_req_type crt)
  83 {
  84         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  85
  86         /* ops_lru and ops_inflight share the same field, so take it from LRU
  87          * first and then use it as inflight. */
  88         osc_lru_use(osc_cli(obj), opg);
  89 }
  90
  91 int osc_page_cache_add(const struct lu_env *env,
  92                         const struct cl_page_slice *slice, struct cl_io *io)
  93 {
  94         struct osc_page *opg = cl2osc_page(slice);
  95         int result;
  96         ENTRY;
  97
  98         osc_page_transfer_get(opg, "transfer\0cache");
  99         result = osc_queue_async_io(env, io, opg);
 100         if (result != 0)
 101                 osc_page_transfer_put(env, opg);
 102         else
 103                 osc_page_transfer_add(env, opg, CRT_WRITE);
 104
 105         RETURN(result);
 106 }
 107
 108 void osc_index2policy(union ldlm_policy_data *policy,
 109                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 110 {
 111         memset(policy, 0, sizeof *policy);
 112         policy->l_extent.start = cl_offset(obj, start);
 113         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 114 }
 115
 116 static const char *osc_list(struct list_head *head)
 117 {
 118         return list_empty(head) ? "-" : "+";
 119 }
 120
 121 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 122 {
 123         if (opg->ops_submit_time == 0)
 124                 return 0;
 125
 126         return (cfs_time_current() - opg->ops_submit_time);
 127 }
 128
 129 static int osc_page_print(const struct lu_env *env,
 130                           const struct cl_page_slice *slice,
 131                           void *cookie, lu_printer_t printer)
 132 {
 133         struct osc_page       *opg = cl2osc_page(slice);
 134         struct osc_async_page *oap = &opg->ops_oap;
 135         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 136         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 137
 138         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 139                           "1< %#x %d %u %s %s > "
 140                           "2< %lld %u %u %#x %#x | %p %p %p > "
 141                           "3< %d %lu %d > "
 142                           "4< %d %d %d %lu %s | %s %s %s %s > "
 143                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 144                           opg, osc_index(opg),
 145                           /* 1 */
 146                           oap->oap_magic, oap->oap_cmd,
 147                           oap->oap_interrupted,
 148                           osc_list(&oap->oap_pending_item),
 149                           osc_list(&oap->oap_rpc_item),
 150                           /* 2 */
 151                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 152                           oap->oap_async_flags, oap->oap_brw_flags,
 153                           oap->oap_request, oap->oap_cli, obj,
 154                           /* 3 */
 155                           opg->ops_transfer_pinned,
 156                           osc_submit_duration(opg), opg->ops_srvlock,
 157                           /* 4 */
 158                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 159                           cli->cl_max_rpcs_in_flight,
 160                           cli->cl_avail_grant,
 161                           osc_list(&cli->cl_cache_waiters),
 162                           osc_list(&cli->cl_loi_ready_list),
 163                           osc_list(&cli->cl_loi_hp_ready_list),
 164                           osc_list(&cli->cl_loi_write_list),
 165                           osc_list(&cli->cl_loi_read_list),
 166                           /* 5 */
 167                           osc_list(&obj->oo_ready_item),
 168                           osc_list(&obj->oo_hp_ready_item),
 169                           osc_list(&obj->oo_write_item),
 170                           osc_list(&obj->oo_read_item),
 171                           atomic_read(&obj->oo_nr_reads),
 172                           osc_list(&obj->oo_reading_exts),
 173                           atomic_read(&obj->oo_nr_writes),
 174                           osc_list(&obj->oo_hp_exts),
 175                           osc_list(&obj->oo_urgent_exts));
 176 }
 177
 178 static void osc_page_delete(const struct lu_env *env,
 179                             const struct cl_page_slice *slice)
 180 {
 181         struct osc_page   *opg = cl2osc_page(slice);
 182         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 183         int rc;
 184
 185         ENTRY;
 186         CDEBUG(D_TRACE, "%p\n", opg);
 187         osc_page_transfer_put(env, opg);
 188         rc = osc_teardown_async_page(env, obj, opg);
 189         if (rc) {
 190                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 191                               "Trying to teardown failed: %d\n", rc);
 192                 LASSERT(0);
 193         }
 194
 195         osc_lru_del(osc_cli(obj), opg);
 196
 197         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 198                 void *value = NULL;
 199
 200                 spin_lock(&obj->oo_tree_lock);
 201                 if (opg->ops_intree) {
 202                         value = radix_tree_delete(&obj->oo_tree,
 203                                                   osc_index(opg));
 204                         if (value != NULL) {
 205                                 --obj->oo_npages;
 206                                 opg->ops_intree = 0;
 207                         }
 208                 }
 209                 spin_unlock(&obj->oo_tree_lock);
 210
 211                 LASSERT(ergo(value != NULL, value == opg));
 212         }
 213
 214         EXIT;
 215 }
 216
 217 static void osc_page_clip(const struct lu_env *env,
 218                           const struct cl_page_slice *slice,
 219                           int from, int to)
 220 {
 221         struct osc_page       *opg = cl2osc_page(slice);
 222         struct osc_async_page *oap = &opg->ops_oap;
 223
 224         opg->ops_from = from;
 225         opg->ops_to   = to;
 226         spin_lock(&oap->oap_lock);
 227         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 228         spin_unlock(&oap->oap_lock);
 229 }
 230
 231 static int osc_page_cancel(const struct lu_env *env,
 232                            const struct cl_page_slice *slice)
 233 {
 234         struct osc_page *opg = cl2osc_page(slice);
 235         int rc = 0;
 236
 237         /* Check if the transferring against this page
 238          * is completed, or not even queued. */
 239         if (opg->ops_transfer_pinned)
 240                 /* FIXME: may not be interrupted.. */
 241                 rc = osc_cancel_async_page(env, opg);
 242         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 243         return rc;
 244 }
 245
 246 static int osc_page_flush(const struct lu_env *env,
 247                           const struct cl_page_slice *slice,
 248                           struct cl_io *io)
 249 {
 250         struct osc_page *opg = cl2osc_page(slice);
 251         int rc = 0;
 252         ENTRY;
 253         rc = osc_flush_async_page(env, io, opg);
 254         RETURN(rc);
 255 }
 256
 257 static const struct cl_page_operations osc_page_ops = {
 258         .cpo_print         = osc_page_print,
 259         .cpo_delete        = osc_page_delete,
 260         .cpo_clip           = osc_page_clip,
 261         .cpo_cancel         = osc_page_cancel,
 262         .cpo_flush          = osc_page_flush
 263 };
 264
 265 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 266                   struct cl_page *page, pgoff_t index)
 267 {
 268         struct osc_object *osc = cl2osc(obj);
 269         struct osc_page   *opg = cl_object_page_slice(obj, page);
 270         struct osc_io *oio = osc_env_io(env);
 271         int result;
 272
 273         opg->ops_from = 0;
 274         opg->ops_to   = PAGE_SIZE;
 275
 276         INIT_LIST_HEAD(&opg->ops_lru);
 277
 278         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 279                                      cl_offset(obj, index));
 280         if (result != 0)
 281                 return result;
 282
 283         opg->ops_srvlock = osc_io_srvlock(oio);
 284         cl_page_slice_add(page, &opg->ops_cl, obj, index,
 285                           &osc_page_ops);
 286
 287
 288         /* reserve an LRU space for this page */
 289         if (page->cp_type == CPT_CACHEABLE) {
 290                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 291                 if (result == 0) {
 292                         result = radix_tree_preload(GFP_NOFS);
 293                         if (result == 0) {
 294                                 spin_lock(&osc->oo_tree_lock);
 295                                 result = radix_tree_insert(&osc->oo_tree,
 296                                                            index, opg);
 297                                 if (result == 0) {
 298                                         ++osc->oo_npages;
 299                                         opg->ops_intree = 1;
 300                                 }
 301                                 spin_unlock(&osc->oo_tree_lock);
 302
 303                                 radix_tree_preload_end();
 304                         }
 305                 }
 306         }
 307
 308         return result;
 309 }
 310
 311 /**
 312  * Helper function called by osc_io_submit() for every page in an immediate
 313  * transfer (i.e., transferred synchronously).
 314  */
 315 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 316                      enum cl_req_type crt, int brw_flags)
 317 {
 318         struct osc_async_page *oap = &opg->ops_oap;
 319
 320         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 321                  "magic 0x%x\n", oap, oap->oap_magic);
 322         LASSERT(oap->oap_async_flags & ASYNC_READY);
 323         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 324
 325         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 326         oap->oap_page_off  = opg->ops_from;
 327         oap->oap_count     = opg->ops_to - opg->ops_from;
 328         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 329
 330         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 331                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 332                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 333         }
 334
 335         opg->ops_submit_time = cfs_time_current();
 336         osc_page_transfer_get(opg, "transfer\0imm");
 337         osc_page_transfer_add(env, opg, crt);
 338 }
 339
 340 /* --------------- LRU page management ------------------ */
 341
 342 /* OSC is a natural place to manage LRU pages as applications are specialized
 343  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 344  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 345  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 346  * for free LRU slots - this will be very bad so the algorithm requires each
 347  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 348  * at any time.
 349  */
 350
 351 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 352
 353 /**
 354  * LRU pages are freed in batch mode. OSC should at least free this
 355  * number of pages to avoid running out of LRU slots.
 356  */
 357 static inline int lru_shrink_min(struct client_obd *cli)
 358 {
 359         return cli->cl_max_pages_per_rpc * 2;
 360 }
 361
 362 /**
 363  * free this number at most otherwise it will take too long time to finsih.
 364  */
 365 static inline int lru_shrink_max(struct client_obd *cli)
 366 {
 367         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 368 }
 369
 370 /**
 371  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 372  * we should free slots aggressively. In this way, slots are freed in a steady
 373  * step to maintain fairness among OSCs.
 374  *
 375  * Return how many LRU pages should be freed.
 376  */
 377 static int osc_cache_too_much(struct client_obd *cli)
 378 {
 379         struct cl_client_cache *cache = cli->cl_cache;
 380         long pages = atomic_long_read(&cli->cl_lru_in_list);
 381         unsigned long budget;
 382
 383         LASSERT(cache != NULL);
 384         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 385
 386         /* if it's going to run out LRU slots, we should free some, but not
 387          * too much to maintain faireness among OSCs. */
 388         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 389                 if (pages >= budget)
 390                         return lru_shrink_max(cli);
 391                 else if (pages >= budget / 2)
 392                         return lru_shrink_min(cli);
 393         } else {
 394                 time64_t duration = ktime_get_real_seconds();
 395                 long timediff;
 396
 397                 /* knock out pages by duration of no IO activity */
 398                 duration -= cli->cl_lru_last_used;
 399                 /*
 400                  * The difference shouldn't be more than 70 years
 401                  * so we can safely case to a long. Round to
 402                  * approximately 1 minute.
 403                  */
 404                 timediff = (long)(duration >> 6);
 405                 if (timediff > 0 && pages >= budget / timediff)
 406                         return lru_shrink_min(cli);
 407         }
 408         return 0;
 409 }
 410
 411 int lru_queue_work(const struct lu_env *env, void *data)
 412 {
 413         struct client_obd *cli = data;
 414         int count;
 415
 416         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 417         count = osc_cache_too_much(cli);
 418         if (count > 0) {
 419                 int rc = osc_lru_shrink(env, cli, count, false);
 420
 421                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 422                        cli_name(cli), rc, count);
 423                 if (rc >= count) {
 424                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 425                         ptlrpcd_queue_work(cli->cl_lru_work);
 426                 }
 427         }
 428
 429         RETURN(0);
 430 }
 431
 432 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 433 {
 434         struct list_head lru = LIST_HEAD_INIT(lru);
 435         struct osc_async_page *oap;
 436         long npages = 0;
 437
 438         list_for_each_entry(oap, plist, oap_pending_item) {
 439                 struct osc_page *opg = oap2osc_page(oap);
 440
 441                 if (!opg->ops_in_lru)
 442                         continue;
 443
 444                 ++npages;
 445                 LASSERT(list_empty(&opg->ops_lru));
 446                 list_add(&opg->ops_lru, &lru);
 447         }
 448
 449         if (npages > 0) {
 450                 spin_lock(&cli->cl_lru_list_lock);
 451                 list_splice_tail(&lru, &cli->cl_lru_list);
 452                 atomic_long_sub(npages, &cli->cl_lru_busy);
 453                 atomic_long_add(npages, &cli->cl_lru_in_list);
 454                 cli->cl_lru_last_used = ktime_get_real_seconds();
 455                 spin_unlock(&cli->cl_lru_list_lock);
 456
 457                 if (waitqueue_active(&osc_lru_waitq))
 458                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 459         }
 460 }
 461
 462 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 463 {
 464         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 465         list_del_init(&opg->ops_lru);
 466         atomic_long_dec(&cli->cl_lru_in_list);
 467 }
 468
 469 /**
 470  * Page is being destroyed. The page may be not in LRU list, if the transfer
 471  * has never finished(error occurred).
 472  */
 473 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 474 {
 475         if (opg->ops_in_lru) {
 476                 spin_lock(&cli->cl_lru_list_lock);
 477                 if (!list_empty(&opg->ops_lru)) {
 478                         __osc_lru_del(cli, opg);
 479                 } else {
 480                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 481                         atomic_long_dec(&cli->cl_lru_busy);
 482                 }
 483                 spin_unlock(&cli->cl_lru_list_lock);
 484
 485                 atomic_long_inc(cli->cl_lru_left);
 486                 /* this is a great place to release more LRU pages if
 487                  * this osc occupies too many LRU pages and kernel is
 488                  * stealing one of them. */
 489                 if (osc_cache_too_much(cli)) {
 490                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 491                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 492                 }
 493                 wake_up(&osc_lru_waitq);
 494         } else {
 495                 LASSERT(list_empty(&opg->ops_lru));
 496         }
 497 }
 498
 499 /**
 500  * Delete page from LRU list for redirty.
 501  */
 502 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 503 {
 504         /* If page is being transferred for the first time,
 505          * ops_lru should be empty */
 506         if (opg->ops_in_lru) {
 507                 spin_lock(&cli->cl_lru_list_lock);
 508                 if (!list_empty(&opg->ops_lru)) {
 509                         __osc_lru_del(cli, opg);
 510                         atomic_long_inc(&cli->cl_lru_busy);
 511                 }
 512                 spin_unlock(&cli->cl_lru_list_lock);
 513         }
 514 }
 515
 516 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 517                                 struct cl_page **pvec, int max_index)
 518 {
 519         int i;
 520
 521         for (i = 0; i < max_index; i++) {
 522                 struct cl_page *page = pvec[i];
 523
 524                 LASSERT(cl_page_is_owned(page, io));
 525                 cl_page_delete(env, page);
 526                 cl_page_discard(env, io, page);
 527                 cl_page_disown(env, io, page);
 528                 cl_page_put(env, page);
 529
 530                 pvec[i] = NULL;
 531         }
 532 }
 533
 534 /**
 535  * Check if a cl_page can be released, i.e, it's not being used.
 536  *
 537  * If unstable account is turned on, bulk transfer may hold one refcount
 538  * for recovery so we need to check vmpage refcount as well; otherwise,
 539  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 540  */
 541 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 542 {
 543         if (cl_page_in_use_noref(page))
 544                 return true;
 545
 546         if (cli->cl_cache->ccc_unstable_check) {
 547                 struct page *vmpage = cl_page_vmpage(page);
 548
 549                 /* vmpage have two known users: cl_page and VM page cache */
 550                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 551                         return true;
 552         }
 553         return false;
 554 }
 555
 556 /**
 557  * Drop @target of pages from LRU at most.
 558  */
 559 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 560                    long target, bool force)
 561 {
 562         struct cl_io *io;
 563         struct cl_object *clobj = NULL;
 564         struct cl_page **pvec;
 565         struct osc_page *opg;
 566         long count = 0;
 567         int maxscan = 0;
 568         int index = 0;
 569         int rc = 0;
 570         ENTRY;
 571
 572         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 573         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 574                 RETURN(0);
 575
 576         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 577                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 578         if (!force) {
 579                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 580                         RETURN(-EBUSY);
 581
 582                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 583                         atomic_dec(&cli->cl_lru_shrinkers);
 584                         RETURN(-EBUSY);
 585                 }
 586         } else {
 587                 atomic_inc(&cli->cl_lru_shrinkers);
 588         }
 589
 590         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 591         io = &osc_env_info(env)->oti_io;
 592
 593         spin_lock(&cli->cl_lru_list_lock);
 594         if (force)
 595                 cli->cl_lru_reclaim++;
 596         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 597         while (!list_empty(&cli->cl_lru_list)) {
 598                 struct cl_page *page;
 599                 bool will_free = false;
 600
 601                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 602                         break;
 603
 604                 if (--maxscan < 0)
 605                         break;
 606
 607                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 608                                  ops_lru);
 609                 page = opg->ops_cl.cpl_page;
 610                 if (lru_page_busy(cli, page)) {
 611                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 612                         continue;
 613                 }
 614
 615                 LASSERT(page->cp_obj != NULL);
 616                 if (clobj != page->cp_obj) {
 617                         struct cl_object *tmp = page->cp_obj;
 618
 619                         cl_object_get(tmp);
 620                         spin_unlock(&cli->cl_lru_list_lock);
 621
 622                         if (clobj != NULL) {
 623                                 discard_pagevec(env, io, pvec, index);
 624                                 index = 0;
 625
 626                                 cl_io_fini(env, io);
 627                                 cl_object_put(env, clobj);
 628                                 clobj = NULL;
 629                         }
 630
 631                         clobj = tmp;
 632                         io->ci_obj = clobj;
 633                         io->ci_ignore_layout = 1;
 634                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 635
 636                         spin_lock(&cli->cl_lru_list_lock);
 637
 638                         if (rc != 0)
 639                                 break;
 640
 641                         ++maxscan;
 642                         continue;
 643                 }
 644
 645                 if (cl_page_own_try(env, io, page) == 0) {
 646                         if (!lru_page_busy(cli, page)) {
 647                                 /* remove it from lru list earlier to avoid
 648                                  * lock contention */
 649                                 __osc_lru_del(cli, opg);
 650                                 opg->ops_in_lru = 0; /* will be discarded */
 651
 652                                 cl_page_get(page);
 653                                 will_free = true;
 654                         } else {
 655                                 cl_page_disown(env, io, page);
 656                         }
 657                 }
 658
 659                 if (!will_free) {
 660                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 661                         continue;
 662                 }
 663
 664                 /* Don't discard and free the page with cl_lru_list held */
 665                 pvec[index++] = page;
 666                 if (unlikely(index == OTI_PVEC_SIZE)) {
 667                         spin_unlock(&cli->cl_lru_list_lock);
 668                         discard_pagevec(env, io, pvec, index);
 669                         index = 0;
 670
 671                         spin_lock(&cli->cl_lru_list_lock);
 672                 }
 673
 674                 if (++count >= target)
 675                         break;
 676         }
 677         spin_unlock(&cli->cl_lru_list_lock);
 678
 679         if (clobj != NULL) {
 680                 discard_pagevec(env, io, pvec, index);
 681
 682                 cl_io_fini(env, io);
 683                 cl_object_put(env, clobj);
 684         }
 685
 686         atomic_dec(&cli->cl_lru_shrinkers);
 687         if (count > 0) {
 688                 atomic_long_add(count, cli->cl_lru_left);
 689                 wake_up_all(&osc_lru_waitq);
 690         }
 691         RETURN(count > 0 ? count : rc);
 692 }
 693
 694 /**
 695  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 696  * \@npages of LRU slots. For performance consideration, it's better to drop
 697  * LRU pages in batch. Therefore, the actual number is adjusted at least
 698  * max_pages_per_rpc.
 699  */
 700 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 701 {
 702         struct lu_env *env;
 703         struct cl_client_cache *cache = cli->cl_cache;
 704         int max_scans;
 705         __u16 refcheck;
 706         long rc = 0;
 707         ENTRY;
 708
 709         LASSERT(cache != NULL);
 710
 711         env = cl_env_get(&refcheck);
 712         if (IS_ERR(env))
 713                 RETURN(rc);
 714
 715         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 716         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 717                cli_name(cli), npages);
 718         rc = osc_lru_shrink(env, cli, npages, true);
 719         if (rc >= npages) {
 720                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 721                        cli_name(cli), rc, npages);
 722                 if (osc_cache_too_much(cli) > 0)
 723                         ptlrpcd_queue_work(cli->cl_lru_work);
 724                 GOTO(out, rc);
 725         } else if (rc > 0) {
 726                 npages -= rc;
 727         }
 728
 729         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 730                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 731                 atomic_long_read(&cli->cl_lru_busy), npages);
 732
 733         /* Reclaim LRU slots from other client_obd as it can't free enough
 734          * from its own. This should rarely happen. */
 735         spin_lock(&cache->ccc_lru_lock);
 736         LASSERT(!list_empty(&cache->ccc_lru));
 737
 738         cache->ccc_lru_shrinkers++;
 739         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 740
 741         max_scans = atomic_read(&cache->ccc_users) - 2;
 742         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 743                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 744                                  cl_lru_osc);
 745
 746                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 747                         cli_name(cli), cli,
 748                         atomic_long_read(&cli->cl_lru_in_list),
 749                         atomic_long_read(&cli->cl_lru_busy));
 750
 751                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 752                 if (osc_cache_too_much(cli) > 0) {
 753                         spin_unlock(&cache->ccc_lru_lock);
 754
 755                         rc = osc_lru_shrink(env, cli, npages, true);
 756                         spin_lock(&cache->ccc_lru_lock);
 757                         if (rc >= npages)
 758                                 break;
 759                         if (rc > 0)
 760                                 npages -= rc;
 761                 }
 762         }
 763         spin_unlock(&cache->ccc_lru_lock);
 764
 765 out:
 766         cl_env_put(env, &refcheck);
 767         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 768                 cli_name(cli), cli, rc);
 769         return rc;
 770 }
 771
 772 /**
 773  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 774  *
 775  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 776  * Only in the case that the LRU slots are in extreme shortage, it should
 777  * have reserved enough slots for an IO.
 778  */
 779 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 780                          struct osc_page *opg)
 781 {
 782         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 783         struct osc_io *oio = osc_env_io(env);
 784         int rc = 0;
 785         ENTRY;
 786
 787         if (cli->cl_cache == NULL) /* shall not be in LRU */
 788                 RETURN(0);
 789
 790         if (oio->oi_lru_reserved > 0) {
 791                 --oio->oi_lru_reserved;
 792                 goto out;
 793         }
 794
 795         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 796         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 797                 /* run out of LRU spaces, try to drop some by itself */
 798                 rc = osc_lru_reclaim(cli, 1);
 799                 if (rc < 0)
 800                         break;
 801                 if (rc > 0)
 802                         continue;
 803
 804                 cond_resched();
 805                 rc = l_wait_event(osc_lru_waitq,
 806                                 atomic_long_read(cli->cl_lru_left) > 0,
 807                                 &lwi);
 808                 if (rc < 0)
 809                         break;
 810         }
 811
 812 out:
 813         if (rc >= 0) {
 814                 atomic_long_inc(&cli->cl_lru_busy);
 815                 opg->ops_in_lru = 1;
 816                 rc = 0;
 817         }
 818
 819         RETURN(rc);
 820 }
 821
 822 /**
 823  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 824  *
 825  * The benefit of doing this is to reduce contention against atomic counter
 826  * cl_lru_left by changing it from per-page access to per-IO access.
 827  */
 828 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 829 {
 830         unsigned long reserved = 0;
 831         unsigned long max_pages;
 832         unsigned long c;
 833
 834         /* reserve a full RPC window at most to avoid that a thread accidentally
 835          * consumes too many LRU slots */
 836         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 837         if (npages > max_pages)
 838                 npages = max_pages;
 839
 840         c = atomic_long_read(cli->cl_lru_left);
 841         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 842                 c = atomic_long_read(cli->cl_lru_left);
 843         while (c >= npages) {
 844                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 845                         reserved = npages;
 846                         break;
 847                 }
 848                 c = atomic_long_read(cli->cl_lru_left);
 849         }
 850         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 851                 /* If there aren't enough pages in the per-OSC LRU then
 852                  * wake up the LRU thread to try and clear out space, so
 853                  * we don't block if pages are being dirtied quickly. */
 854                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 855                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 856                        max_pages);
 857                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 858         }
 859
 860         return reserved;
 861 }
 862
 863 /**
 864  * osc_lru_unreserve() is called to unreserve LRU slots.
 865  *
 866  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 867  * reasons such as page already existing or I/O error. Those reserved slots
 868  * should be freed by calling this function.
 869  */
 870 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 871 {
 872         atomic_long_add(npages, cli->cl_lru_left);
 873         wake_up_all(&osc_lru_waitq);
 874 }
 875
 876 /**
 877  * Atomic operations are expensive. We accumulate the accounting for the
 878  * same page zone to get better performance.
 879  * In practice this can work pretty good because the pages in the same RPC
 880  * are likely from the same page zone.
 881  */
 882 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 883                                             int factor)
 884 {
 885         int page_count = desc->bd_iov_count;
 886         void *zone = NULL;
 887         int count = 0;
 888         int i;
 889
 890         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 891
 892         for (i = 0; i < page_count; i++) {
 893                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 894
 895                 if (likely(pz == zone)) {
 896                         ++count;
 897                         continue;
 898                 }
 899
 900                 if (count > 0) {
 901                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 902                                             factor * count);
 903                         count = 0;
 904                 }
 905                 zone = pz;
 906                 ++count;
 907         }
 908         if (count > 0)
 909                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 910 }
 911
 912 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 913 {
 914         unstable_page_accounting(desc, 1);
 915 }
 916
 917 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 918 {
 919         unstable_page_accounting(desc, -1);
 920 }
 921
 922 /**
 923  * Performs "unstable" page accounting. This function balances the
 924  * increment operations performed in osc_inc_unstable_pages. It is
 925  * registered as the RPC request callback, and is executed when the
 926  * bulk RPC is committed on the server. Thus at this point, the pages
 927  * involved in the bulk transfer are no longer considered unstable.
 928  *
 929  * If this function is called, the request should have been committed
 930  * or req:rq_unstable must have been set; it implies that the unstable
 931  * statistic have been added.
 932  */
 933 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 934 {
 935         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 936         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 937         int                      page_count = desc->bd_iov_count;
 938         long                     unstable_count;
 939
 940         LASSERT(page_count >= 0);
 941         dec_unstable_page_accounting(desc);
 942
 943         unstable_count = atomic_long_sub_return(page_count,
 944                                                 &cli->cl_unstable_count);
 945         LASSERT(unstable_count >= 0);
 946
 947         unstable_count = atomic_long_sub_return(page_count,
 948                                            &cli->cl_cache->ccc_unstable_nr);
 949         LASSERT(unstable_count >= 0);
 950         if (unstable_count == 0)
 951                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 952
 953         if (waitqueue_active(&osc_lru_waitq))
 954                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 955 }
 956
 957 /**
 958  * "unstable" page accounting. See: osc_dec_unstable_pages.
 959  */
 960 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 961 {
 962         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 963         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 964         long                     page_count = desc->bd_iov_count;
 965
 966         /* No unstable page tracking */
 967         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 968                 return;
 969
 970         add_unstable_page_accounting(desc);
 971         atomic_long_add(page_count, &cli->cl_unstable_count);
 972         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 973
 974         /* If the request has already been committed (i.e. brw_commit
 975          * called via rq_commit_cb), we need to undo the unstable page
 976          * increments we just performed because rq_commit_cb wont be
 977          * called again. */
 978         spin_lock(&req->rq_lock);
 979         if (unlikely(req->rq_committed)) {
 980                 spin_unlock(&req->rq_lock);
 981
 982                 osc_dec_unstable_pages(req);
 983         } else {
 984                 req->rq_unstable = 1;
 985                 spin_unlock(&req->rq_lock);
 986         }
 987 }
 988
 989 /**
 990  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 991  * This function will be called by every BRW RPC so it's critical
 992  * to make this function fast.
 993  */
 994 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 995 {
 996         long unstable_nr, osc_unstable_count;
 997
 998         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 999         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1000                 return false;
1001
1002         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1003         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1004
1005         CDEBUG(D_CACHE,
1006                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1007                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1008
1009         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1010          * has one full RPC window of unstable pages, it's a good chance
1011          * to piggyback a SOFT_SYNC flag.
1012          * Please notice that the OST won't take immediate response for the
1013          * SOFT_SYNC request so active OSCs will have more chance to carry
1014          * the flag, this is reasonable. */
1015         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1016                osc_unstable_count > cli->cl_max_pages_per_rpc *
1017                                     cli->cl_max_rpcs_in_flight;
1018 }
1019
1020 /**
1021  * Return how many LRU pages in the cache of all OSC devices
1022  *
1023  * \retval      return # of cached LRU pages times reclaimation tendency
1024  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1025  */
1026 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1027                                      struct shrink_control *sc)
1028 {
1029         struct client_obd *cli;
1030         unsigned long cached = 0;
1031
1032         spin_lock(&osc_shrink_lock);
1033         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1034                 cached += atomic_long_read(&cli->cl_lru_in_list);
1035         spin_unlock(&osc_shrink_lock);
1036
1037         return (cached  * sysctl_vfs_cache_pressure) / 100;
1038 }
1039
1040 /**
1041  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1042  *
1043  * \retval      number of cached LRU pages reclaimed
1044  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1045  *
1046  * Linux kernel will loop calling this shrinker scan routine with
1047  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1048  *
1049  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1050  * to scan and try to reclaim LRU pages, just return 0 and
1051  * osc_cache_shrink_count() will report the LRU page number.
1052  */
1053 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1054                                     struct shrink_control *sc)
1055 {
1056         struct client_obd *cli;
1057         struct client_obd *stop_anchor = NULL;
1058         struct lu_env *env;
1059         long shrank = 0;
1060         int rc;
1061         __u16 refcheck;
1062
1063         if (sc->nr_to_scan == 0)
1064                 return 0;
1065
1066         if (!(sc->gfp_mask & __GFP_FS))
1067                 return SHRINK_STOP;
1068
1069         env = cl_env_get(&refcheck);
1070         if (IS_ERR(env))
1071                 return SHRINK_STOP;
1072
1073         spin_lock(&osc_shrink_lock);
1074         while (!list_empty(&osc_shrink_list)) {
1075                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1076                                  cl_shrink_list);
1077
1078                 if (stop_anchor == NULL)
1079                         stop_anchor = cli;
1080                 else if (cli == stop_anchor)
1081                         break;
1082
1083                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1084                 spin_unlock(&osc_shrink_lock);
1085
1086                 /* shrink no more than max_pages_per_rpc for an OSC */
1087                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1088                                     cli->cl_max_pages_per_rpc ?
1089                                     cli->cl_max_pages_per_rpc :
1090                                     sc->nr_to_scan - shrank, true);
1091                 if (rc > 0)
1092                         shrank += rc;
1093
1094                 if (shrank >= sc->nr_to_scan)
1095                         goto out;
1096
1097                 spin_lock(&osc_shrink_lock);
1098         }
1099         spin_unlock(&osc_shrink_lock);
1100
1101 out:
1102         cl_env_put(env, &refcheck);
1103
1104         return shrank;
1105 }
1106
1107 /** @} osc */