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 LRUlist 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 && !list_empty(&opg->ops_lru)) {
 492                 spin_lock(&cli->cl_lru_list_lock);
 493                 __osc_lru_del(cli, opg);
 494                 spin_unlock(&cli->cl_lru_list_lock);
 495                 atomic_long_inc(&cli->cl_lru_busy);
 496         }
 497 }
 498
 499 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 500                                 struct cl_page **pvec, int max_index)
 501 {
 502         int i;
 503
 504         for (i = 0; i < max_index; i++) {
 505                 struct cl_page *page = pvec[i];
 506
 507                 LASSERT(cl_page_is_owned(page, io));
 508                 cl_page_delete(env, page);
 509                 cl_page_discard(env, io, page);
 510                 cl_page_disown(env, io, page);
 511                 cl_page_put(env, page);
 512
 513                 pvec[i] = NULL;
 514         }
 515 }
 516
 517 /**
 518  * Check if a cl_page can be released, i.e, it's not being used.
 519  *
 520  * If unstable account is turned on, bulk transfer may hold one refcount
 521  * for recovery so we need to check vmpage refcount as well; otherwise,
 522  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 523  */
 524 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 525 {
 526         if (cl_page_in_use_noref(page))
 527                 return true;
 528
 529         if (cli->cl_cache->ccc_unstable_check) {
 530                 struct page *vmpage = cl_page_vmpage(page);
 531
 532                 /* vmpage have two known users: cl_page and VM page cache */
 533                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 534                         return true;
 535         }
 536         return false;
 537 }
 538
 539 /**
 540  * Drop @target of pages from LRU at most.
 541  */
 542 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 543                    long target, bool force)
 544 {
 545         struct cl_io *io;
 546         struct cl_object *clobj = NULL;
 547         struct cl_page **pvec;
 548         struct osc_page *opg;
 549         long count = 0;
 550         int maxscan = 0;
 551         int index = 0;
 552         int rc = 0;
 553         ENTRY;
 554
 555         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 556         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 557                 RETURN(0);
 558
 559         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 560                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 561         if (!force) {
 562                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 563                         RETURN(-EBUSY);
 564
 565                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 566                         atomic_dec(&cli->cl_lru_shrinkers);
 567                         RETURN(-EBUSY);
 568                 }
 569         } else {
 570                 atomic_inc(&cli->cl_lru_shrinkers);
 571         }
 572
 573         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 574         io = &osc_env_info(env)->oti_io;
 575
 576         spin_lock(&cli->cl_lru_list_lock);
 577         if (force)
 578                 cli->cl_lru_reclaim++;
 579         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 580         while (!list_empty(&cli->cl_lru_list)) {
 581                 struct cl_page *page;
 582                 bool will_free = false;
 583
 584                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 585                         break;
 586
 587                 if (--maxscan < 0)
 588                         break;
 589
 590                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 591                                  ops_lru);
 592                 page = opg->ops_cl.cpl_page;
 593                 if (lru_page_busy(cli, page)) {
 594                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 595                         continue;
 596                 }
 597
 598                 LASSERT(page->cp_obj != NULL);
 599                 if (clobj != page->cp_obj) {
 600                         struct cl_object *tmp = page->cp_obj;
 601
 602                         cl_object_get(tmp);
 603                         spin_unlock(&cli->cl_lru_list_lock);
 604
 605                         if (clobj != NULL) {
 606                                 discard_pagevec(env, io, pvec, index);
 607                                 index = 0;
 608
 609                                 cl_io_fini(env, io);
 610                                 cl_object_put(env, clobj);
 611                                 clobj = NULL;
 612                         }
 613
 614                         clobj = tmp;
 615                         io->ci_obj = clobj;
 616                         io->ci_ignore_layout = 1;
 617                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 618
 619                         spin_lock(&cli->cl_lru_list_lock);
 620
 621                         if (rc != 0)
 622                                 break;
 623
 624                         ++maxscan;
 625                         continue;
 626                 }
 627
 628                 if (cl_page_own_try(env, io, page) == 0) {
 629                         if (!lru_page_busy(cli, page)) {
 630                                 /* remove it from lru list earlier to avoid
 631                                  * lock contention */
 632                                 __osc_lru_del(cli, opg);
 633                                 opg->ops_in_lru = 0; /* will be discarded */
 634
 635                                 cl_page_get(page);
 636                                 will_free = true;
 637                         } else {
 638                                 cl_page_disown(env, io, page);
 639                         }
 640                 }
 641
 642                 if (!will_free) {
 643                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 644                         continue;
 645                 }
 646
 647                 /* Don't discard and free the page with cl_lru_list held */
 648                 pvec[index++] = page;
 649                 if (unlikely(index == OTI_PVEC_SIZE)) {
 650                         spin_unlock(&cli->cl_lru_list_lock);
 651                         discard_pagevec(env, io, pvec, index);
 652                         index = 0;
 653
 654                         spin_lock(&cli->cl_lru_list_lock);
 655                 }
 656
 657                 if (++count >= target)
 658                         break;
 659         }
 660         spin_unlock(&cli->cl_lru_list_lock);
 661
 662         if (clobj != NULL) {
 663                 discard_pagevec(env, io, pvec, index);
 664
 665                 cl_io_fini(env, io);
 666                 cl_object_put(env, clobj);
 667         }
 668
 669         atomic_dec(&cli->cl_lru_shrinkers);
 670         if (count > 0) {
 671                 atomic_long_add(count, cli->cl_lru_left);
 672                 wake_up_all(&osc_lru_waitq);
 673         }
 674         RETURN(count > 0 ? count : rc);
 675 }
 676
 677 /**
 678  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 679  * \@npages of LRU slots. For performance consideration, it's better to drop
 680  * LRU pages in batch. Therefore, the actual number is adjusted at least
 681  * max_pages_per_rpc.
 682  */
 683 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 684 {
 685         struct lu_env *env;
 686         struct cl_client_cache *cache = cli->cl_cache;
 687         int max_scans;
 688         __u16 refcheck;
 689         long rc = 0;
 690         ENTRY;
 691
 692         LASSERT(cache != NULL);
 693
 694         env = cl_env_get(&refcheck);
 695         if (IS_ERR(env))
 696                 RETURN(rc);
 697
 698         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 699         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 700                cli_name(cli), npages);
 701         rc = osc_lru_shrink(env, cli, npages, true);
 702         if (rc >= npages) {
 703                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 704                        cli_name(cli), rc, npages);
 705                 if (osc_cache_too_much(cli) > 0)
 706                         ptlrpcd_queue_work(cli->cl_lru_work);
 707                 GOTO(out, rc);
 708         } else if (rc > 0) {
 709                 npages -= rc;
 710         }
 711
 712         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 713                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 714                 atomic_long_read(&cli->cl_lru_busy), npages);
 715
 716         /* Reclaim LRU slots from other client_obd as it can't free enough
 717          * from its own. This should rarely happen. */
 718         spin_lock(&cache->ccc_lru_lock);
 719         LASSERT(!list_empty(&cache->ccc_lru));
 720
 721         cache->ccc_lru_shrinkers++;
 722         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 723
 724         max_scans = atomic_read(&cache->ccc_users) - 2;
 725         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 726                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 727                                  cl_lru_osc);
 728
 729                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 730                         cli_name(cli), cli,
 731                         atomic_long_read(&cli->cl_lru_in_list),
 732                         atomic_long_read(&cli->cl_lru_busy));
 733
 734                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 735                 if (osc_cache_too_much(cli) > 0) {
 736                         spin_unlock(&cache->ccc_lru_lock);
 737
 738                         rc = osc_lru_shrink(env, cli, npages, true);
 739                         spin_lock(&cache->ccc_lru_lock);
 740                         if (rc >= npages)
 741                                 break;
 742                         if (rc > 0)
 743                                 npages -= rc;
 744                 }
 745         }
 746         spin_unlock(&cache->ccc_lru_lock);
 747
 748 out:
 749         cl_env_put(env, &refcheck);
 750         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 751                 cli_name(cli), cli, rc);
 752         return rc;
 753 }
 754
 755 /**
 756  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 757  *
 758  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 759  * Only in the case that the LRU slots are in extreme shortage, it should
 760  * have reserved enough slots for an IO.
 761  */
 762 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 763                          struct osc_page *opg)
 764 {
 765         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 766         struct osc_io *oio = osc_env_io(env);
 767         int rc = 0;
 768         ENTRY;
 769
 770         if (cli->cl_cache == NULL) /* shall not be in LRU */
 771                 RETURN(0);
 772
 773         if (oio->oi_lru_reserved > 0) {
 774                 --oio->oi_lru_reserved;
 775                 goto out;
 776         }
 777
 778         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 779         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 780                 /* run out of LRU spaces, try to drop some by itself */
 781                 rc = osc_lru_reclaim(cli, 1);
 782                 if (rc < 0)
 783                         break;
 784                 if (rc > 0)
 785                         continue;
 786
 787                 cond_resched();
 788                 rc = l_wait_event(osc_lru_waitq,
 789                                 atomic_long_read(cli->cl_lru_left) > 0,
 790                                 &lwi);
 791                 if (rc < 0)
 792                         break;
 793         }
 794
 795 out:
 796         if (rc >= 0) {
 797                 atomic_long_inc(&cli->cl_lru_busy);
 798                 opg->ops_in_lru = 1;
 799                 rc = 0;
 800         }
 801
 802         RETURN(rc);
 803 }
 804
 805 /**
 806  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 807  *
 808  * The benefit of doing this is to reduce contention against atomic counter
 809  * cl_lru_left by changing it from per-page access to per-IO access.
 810  */
 811 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 812 {
 813         unsigned long reserved = 0;
 814         unsigned long max_pages;
 815         unsigned long c;
 816
 817         /* reserve a full RPC window at most to avoid that a thread accidentally
 818          * consumes too many LRU slots */
 819         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 820         if (npages > max_pages)
 821                 npages = max_pages;
 822
 823         c = atomic_long_read(cli->cl_lru_left);
 824         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 825                 c = atomic_long_read(cli->cl_lru_left);
 826         while (c >= npages) {
 827                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 828                         reserved = npages;
 829                         break;
 830                 }
 831                 c = atomic_long_read(cli->cl_lru_left);
 832         }
 833         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 834                 /* If there aren't enough pages in the per-OSC LRU then
 835                  * wake up the LRU thread to try and clear out space, so
 836                  * we don't block if pages are being dirtied quickly. */
 837                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 838                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 839                        max_pages);
 840                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 841         }
 842
 843         return reserved;
 844 }
 845
 846 /**
 847  * osc_lru_unreserve() is called to unreserve LRU slots.
 848  *
 849  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 850  * reasons such as page already existing or I/O error. Those reserved slots
 851  * should be freed by calling this function.
 852  */
 853 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 854 {
 855         atomic_long_add(npages, cli->cl_lru_left);
 856         wake_up_all(&osc_lru_waitq);
 857 }
 858
 859 /**
 860  * Atomic operations are expensive. We accumulate the accounting for the
 861  * same page zone to get better performance.
 862  * In practice this can work pretty good because the pages in the same RPC
 863  * are likely from the same page zone.
 864  */
 865 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 866                                             int factor)
 867 {
 868         int page_count = desc->bd_iov_count;
 869         void *zone = NULL;
 870         int count = 0;
 871         int i;
 872
 873         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 874
 875         for (i = 0; i < page_count; i++) {
 876                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 877
 878                 if (likely(pz == zone)) {
 879                         ++count;
 880                         continue;
 881                 }
 882
 883                 if (count > 0) {
 884                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 885                                             factor * count);
 886                         count = 0;
 887                 }
 888                 zone = pz;
 889                 ++count;
 890         }
 891         if (count > 0)
 892                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 893 }
 894
 895 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 896 {
 897         unstable_page_accounting(desc, 1);
 898 }
 899
 900 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 901 {
 902         unstable_page_accounting(desc, -1);
 903 }
 904
 905 /**
 906  * Performs "unstable" page accounting. This function balances the
 907  * increment operations performed in osc_inc_unstable_pages. It is
 908  * registered as the RPC request callback, and is executed when the
 909  * bulk RPC is committed on the server. Thus at this point, the pages
 910  * involved in the bulk transfer are no longer considered unstable.
 911  *
 912  * If this function is called, the request should have been committed
 913  * or req:rq_unstable must have been set; it implies that the unstable
 914  * statistic have been added.
 915  */
 916 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 917 {
 918         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 919         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 920         int                      page_count = desc->bd_iov_count;
 921         long                     unstable_count;
 922
 923         LASSERT(page_count >= 0);
 924         dec_unstable_page_accounting(desc);
 925
 926         unstable_count = atomic_long_sub_return(page_count,
 927                                                 &cli->cl_unstable_count);
 928         LASSERT(unstable_count >= 0);
 929
 930         unstable_count = atomic_long_sub_return(page_count,
 931                                            &cli->cl_cache->ccc_unstable_nr);
 932         LASSERT(unstable_count >= 0);
 933         if (unstable_count == 0)
 934                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 935
 936         if (waitqueue_active(&osc_lru_waitq))
 937                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 938 }
 939
 940 /**
 941  * "unstable" page accounting. See: osc_dec_unstable_pages.
 942  */
 943 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 944 {
 945         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 946         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 947         long                     page_count = desc->bd_iov_count;
 948
 949         /* No unstable page tracking */
 950         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 951                 return;
 952
 953         add_unstable_page_accounting(desc);
 954         atomic_long_add(page_count, &cli->cl_unstable_count);
 955         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 956
 957         /* If the request has already been committed (i.e. brw_commit
 958          * called via rq_commit_cb), we need to undo the unstable page
 959          * increments we just performed because rq_commit_cb wont be
 960          * called again. */
 961         spin_lock(&req->rq_lock);
 962         if (unlikely(req->rq_committed)) {
 963                 spin_unlock(&req->rq_lock);
 964
 965                 osc_dec_unstable_pages(req);
 966         } else {
 967                 req->rq_unstable = 1;
 968                 spin_unlock(&req->rq_lock);
 969         }
 970 }
 971
 972 /**
 973  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 974  * This function will be called by every BRW RPC so it's critical
 975  * to make this function fast.
 976  */
 977 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 978 {
 979         long unstable_nr, osc_unstable_count;
 980
 981         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 982         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 983                 return false;
 984
 985         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 986         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 987
 988         CDEBUG(D_CACHE,
 989                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
 990                cli_name(cli), cli, unstable_nr, osc_unstable_count);
 991
 992         /* If the LRU slots are in shortage - 25% remaining AND this OSC
 993          * has one full RPC window of unstable pages, it's a good chance
 994          * to piggyback a SOFT_SYNC flag.
 995          * Please notice that the OST won't take immediate response for the
 996          * SOFT_SYNC request so active OSCs will have more chance to carry
 997          * the flag, this is reasonable. */
 998         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
 999                osc_unstable_count > cli->cl_max_pages_per_rpc *
1000                                     cli->cl_max_rpcs_in_flight;
1001 }
1002
1003 /**
1004  * Return how many LRU pages in the cache of all OSC devices
1005  *
1006  * \retval      return # of cached LRU pages times reclaimation tendency
1007  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1008  */
1009 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1010                                      struct shrink_control *sc)
1011 {
1012         struct client_obd *cli;
1013         unsigned long cached = 0;
1014
1015         spin_lock(&osc_shrink_lock);
1016         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1017                 cached += atomic_long_read(&cli->cl_lru_in_list);
1018         spin_unlock(&osc_shrink_lock);
1019
1020         return (cached  * sysctl_vfs_cache_pressure) / 100;
1021 }
1022
1023 /**
1024  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1025  *
1026  * \retval      number of cached LRU pages reclaimed
1027  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1028  *
1029  * Linux kernel will loop calling this shrinker scan routine with
1030  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1031  *
1032  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1033  * to scan and try to reclaim LRU pages, just return 0 and
1034  * osc_cache_shrink_count() will report the LRU page number.
1035  */
1036 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1037                                     struct shrink_control *sc)
1038 {
1039         struct client_obd *cli;
1040         struct client_obd *stop_anchor = NULL;
1041         struct lu_env *env;
1042         long shrank = 0;
1043         int rc;
1044         __u16 refcheck;
1045
1046         if (sc->nr_to_scan == 0)
1047                 return 0;
1048
1049         if (!(sc->gfp_mask & __GFP_FS))
1050                 return SHRINK_STOP;
1051
1052         env = cl_env_get(&refcheck);
1053         if (IS_ERR(env))
1054                 return SHRINK_STOP;
1055
1056         spin_lock(&osc_shrink_lock);
1057         while (!list_empty(&osc_shrink_list)) {
1058                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1059                                  cl_shrink_list);
1060
1061                 if (stop_anchor == NULL)
1062                         stop_anchor = cli;
1063                 else if (cli == stop_anchor)
1064                         break;
1065
1066                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1067                 spin_unlock(&osc_shrink_lock);
1068
1069                 /* shrink no more than max_pages_per_rpc for an OSC */
1070                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1071                                     cli->cl_max_pages_per_rpc ?
1072                                     cli->cl_max_pages_per_rpc :
1073                                     sc->nr_to_scan - shrank, true);
1074                 if (rc > 0)
1075                         shrank += rc;
1076
1077                 if (shrank >= sc->nr_to_scan)
1078                         goto out;
1079
1080                 spin_lock(&osc_shrink_lock);
1081         }
1082         spin_unlock(&osc_shrink_lock);
1083
1084 out:
1085         cl_env_put(env, &refcheck);
1086
1087         return shrank;
1088 }
1089
1090 /** @} osc */