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