lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * Implementation of cl_page for OSC layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38 #include <lustre_osc.h>
  39
  40 #include "osc_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 = osc_page_object(opg);
  85
  86         osc_lru_use(osc_cli(obj), opg);
  87 }
  88
  89 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  90                        struct cl_io *io, cl_commit_cbt cb)
  91 {
  92         int result;
  93         ENTRY;
  94
  95         osc_page_transfer_get(opg, "transfer\0cache");
  96         result = osc_queue_async_io(env, io, opg, cb);
  97         if (result != 0)
  98                 osc_page_transfer_put(env, opg);
  99         else
 100                 osc_page_transfer_add(env, opg, CRT_WRITE);
 101
 102         RETURN(result);
 103 }
 104
 105 void osc_index2policy(union ldlm_policy_data *policy,
 106                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 107 {
 108         memset(policy, 0, sizeof *policy);
 109         policy->l_extent.start = cl_offset(obj, start);
 110         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 111 }
 112
 113 static int osc_page_print(const struct lu_env *env,
 114                           const struct cl_page_slice *slice,
 115                           void *cookie, lu_printer_t printer)
 116 {
 117         struct osc_page *opg = cl2osc_page(slice);
 118         struct osc_async_page *oap = &opg->ops_oap;
 119         struct osc_object *obj = osc_page_object(opg);
 120         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 121
 122         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 123                           "1< %d %c %c > "
 124                           "2< %lld %u %u %#x %#x | %p %p %p > "
 125                           "3< %d %d > "
 126                           "4< %d %d %d %lu %c | %c %c %c %c > "
 127                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 128                           opg, osc_index(opg),
 129                           /* 1 */
 130                           oap->oap_cmd,
 131                           list_empty_marker(&oap->oap_pending_item),
 132                           list_empty_marker(&oap->oap_rpc_item),
 133                           /* 2 */
 134                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 135                           oap->oap_async_flags, oap->oap_brw_flags,
 136                           oap->oap_request, cli, obj,
 137                           /* 3 */
 138                           opg->ops_transfer_pinned,
 139                           opg->ops_srvlock,
 140                           /* 4 */
 141                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 142                           cli->cl_max_rpcs_in_flight,
 143                           cli->cl_avail_grant,
 144                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 145                           list_empty_marker(&cli->cl_loi_ready_list),
 146                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 147                           list_empty_marker(&cli->cl_loi_write_list),
 148                           list_empty_marker(&cli->cl_loi_read_list),
 149                           /* 5 */
 150                           list_empty_marker(&obj->oo_ready_item),
 151                           list_empty_marker(&obj->oo_hp_ready_item),
 152                           list_empty_marker(&obj->oo_write_item),
 153                           list_empty_marker(&obj->oo_read_item),
 154                           atomic_read(&obj->oo_nr_reads),
 155                           list_empty_marker(&obj->oo_reading_exts),
 156                           atomic_read(&obj->oo_nr_writes),
 157                           list_empty_marker(&obj->oo_hp_exts),
 158                           list_empty_marker(&obj->oo_urgent_exts));
 159 }
 160
 161 static void osc_page_delete(const struct lu_env *env,
 162                             const struct cl_page_slice *slice)
 163 {
 164         struct osc_page   *opg = cl2osc_page(slice);
 165         struct osc_object *obj = osc_page_object(opg);
 166         int rc;
 167
 168         ENTRY;
 169         CDEBUG(D_TRACE, "%p\n", opg);
 170         osc_page_transfer_put(env, opg);
 171         rc = osc_teardown_async_page(env, obj, opg);
 172         if (rc) {
 173                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 174                               "Trying to teardown failed: %d\n", rc);
 175                 LASSERT(0);
 176         }
 177
 178         osc_lru_del(osc_cli(obj), opg);
 179
 180         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 181                 void *value = NULL;
 182
 183                 spin_lock(&obj->oo_tree_lock);
 184                 if (opg->ops_intree) {
 185                         value = radix_tree_delete(&obj->oo_tree,
 186                                                   osc_index(opg));
 187                         if (value != NULL) {
 188                                 --obj->oo_npages;
 189                                 opg->ops_intree = 0;
 190                         }
 191                 }
 192                 spin_unlock(&obj->oo_tree_lock);
 193
 194                 LASSERT(ergo(value != NULL, value == opg));
 195         }
 196
 197         EXIT;
 198 }
 199
 200 static void osc_page_clip(const struct lu_env *env,
 201                           const struct cl_page_slice *slice,
 202                           int from, int to)
 203 {
 204         struct osc_page       *opg = cl2osc_page(slice);
 205         struct osc_async_page *oap = &opg->ops_oap;
 206
 207         opg->ops_from = from;
 208         /* argument @to is exclusive, but @ops_to is inclusive */
 209         opg->ops_to   = to - 1;
 210         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 211 }
 212
 213 static int osc_page_flush(const struct lu_env *env,
 214                           const struct cl_page_slice *slice,
 215                           struct cl_io *io)
 216 {
 217         struct osc_page *opg = cl2osc_page(slice);
 218         int rc = 0;
 219         ENTRY;
 220         rc = osc_flush_async_page(env, io, opg);
 221         RETURN(rc);
 222 }
 223
 224 static void osc_page_touch(const struct lu_env *env,
 225                           const struct cl_page_slice *slice, size_t to)
 226 {
 227         struct osc_page *opg = cl2osc_page(slice);
 228         struct cl_object *obj = osc2cl(osc_page_object(opg));
 229
 230         osc_page_touch_at(env, obj, osc_index(opg), to);
 231 }
 232
 233 static const struct cl_page_operations osc_page_ops = {
 234         .cpo_print         = osc_page_print,
 235         .cpo_delete        = osc_page_delete,
 236         .cpo_clip           = osc_page_clip,
 237         .cpo_flush          = osc_page_flush,
 238         .cpo_page_touch    = osc_page_touch,
 239 };
 240
 241 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 242                   struct cl_page *cl_page, pgoff_t index)
 243 {
 244         struct osc_object *osc = cl2osc(obj);
 245         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 246         struct osc_io *oio = osc_env_io(env);
 247         int result;
 248
 249         opg->ops_from = 0;
 250         opg->ops_to = PAGE_SIZE - 1;
 251
 252         INIT_LIST_HEAD(&opg->ops_lru);
 253
 254         result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
 255         if (result != 0)
 256                 return result;
 257
 258         opg->ops_srvlock = osc_io_srvlock(oio);
 259         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 260
 261         /* reserve an LRU space for this page */
 262         if (cl_page->cp_type == CPT_CACHEABLE) {
 263                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 264                 if (result == 0) {
 265                         result = radix_tree_preload(GFP_NOFS);
 266                         if (result == 0) {
 267                                 spin_lock(&osc->oo_tree_lock);
 268                                 result = radix_tree_insert(&osc->oo_tree,
 269                                                            index, opg);
 270                                 if (result == 0) {
 271                                         ++osc->oo_npages;
 272                                         opg->ops_intree = 1;
 273                                 }
 274                                 spin_unlock(&osc->oo_tree_lock);
 275
 276                                 radix_tree_preload_end();
 277                         }
 278                 }
 279         }
 280
 281         return result;
 282 }
 283 EXPORT_SYMBOL(osc_page_init);
 284
 285 /**
 286  * Helper function called by osc_io_submit() for every page in an immediate
 287  * transfer (i.e., transferred synchronously).
 288  */
 289 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 290                      enum cl_req_type crt, int brw_flags)
 291 {
 292         struct osc_io *oio = osc_env_io(env);
 293         struct osc_async_page *oap = &opg->ops_oap;
 294
 295         LASSERT(oap->oap_async_flags & ASYNC_READY);
 296         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 297
 298         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 299         oap->oap_page_off = opg->ops_from;
 300         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 301         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 302
 303         if (oio->oi_cap_sys_resource)
 304                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 305
 306         osc_page_transfer_get(opg, "transfer\0imm");
 307         osc_page_transfer_add(env, opg, crt);
 308 }
 309
 310 /* --------------- LRU page management ------------------ */
 311
 312 /* OSC is a natural place to manage LRU pages as applications are specialized
 313  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 314  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 315  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 316  * for free LRU slots - this will be very bad so the algorithm requires each
 317  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 318  * at any time.
 319  */
 320
 321 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 322
 323 /**
 324  * LRU pages are freed in batch mode. OSC should at least free this
 325  * number of pages to avoid running out of LRU slots.
 326  */
 327 static inline int lru_shrink_min(struct client_obd *cli)
 328 {
 329         return cli->cl_max_pages_per_rpc * 2;
 330 }
 331
 332 /**
 333  * free this number at most otherwise it will take too long time to finsih.
 334  */
 335 static inline int lru_shrink_max(struct client_obd *cli)
 336 {
 337         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 338 }
 339
 340 /**
 341  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 342  * we should free slots aggressively. In this way, slots are freed in a steady
 343  * step to maintain fairness among OSCs.
 344  *
 345  * Return how many LRU pages should be freed.
 346  */
 347 static int osc_cache_too_much(struct client_obd *cli)
 348 {
 349         struct cl_client_cache *cache = cli->cl_cache;
 350         long pages = atomic_long_read(&cli->cl_lru_in_list);
 351         unsigned long budget;
 352
 353         LASSERT(cache != NULL);
 354         budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
 355
 356         /* if it's going to run out LRU slots, we should free some, but not
 357          * too much to maintain faireness among OSCs. */
 358         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 359                 if (pages >= budget)
 360                         return lru_shrink_max(cli);
 361                 else if (pages >= budget / 2)
 362                         return lru_shrink_min(cli);
 363         } else {
 364                 time64_t duration = ktime_get_real_seconds();
 365                 long timediff;
 366
 367                 /* knock out pages by duration of no IO activity */
 368                 duration -= cli->cl_lru_last_used;
 369                 /*
 370                  * The difference shouldn't be more than 70 years
 371                  * so we can safely case to a long. Round to
 372                  * approximately 1 minute.
 373                  */
 374                 timediff = (long)(duration >> 6);
 375                 if (timediff > 0 && pages >= budget / timediff)
 376                         return lru_shrink_min(cli);
 377         }
 378         return 0;
 379 }
 380
 381 int lru_queue_work(const struct lu_env *env, void *data)
 382 {
 383         struct client_obd *cli = data;
 384         int count;
 385
 386         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 387         count = osc_cache_too_much(cli);
 388         if (count > 0) {
 389                 int rc = osc_lru_shrink(env, cli, count, false);
 390
 391                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 392                        cli_name(cli), rc, count);
 393                 if (rc >= count) {
 394                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 395                         ptlrpcd_queue_work(cli->cl_lru_work);
 396                 }
 397         }
 398
 399         RETURN(0);
 400 }
 401
 402 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 403 {
 404         LIST_HEAD(lru);
 405         struct osc_async_page *oap;
 406         long npages = 0;
 407
 408         list_for_each_entry(oap, plist, oap_pending_item) {
 409                 struct osc_page *opg = oap2osc_page(oap);
 410
 411                 if (!opg->ops_in_lru)
 412                         continue;
 413
 414                 ++npages;
 415                 LASSERT(list_empty(&opg->ops_lru));
 416                 list_add(&opg->ops_lru, &lru);
 417         }
 418
 419         if (npages > 0) {
 420                 spin_lock(&cli->cl_lru_list_lock);
 421                 list_splice_tail(&lru, &cli->cl_lru_list);
 422                 atomic_long_sub(npages, &cli->cl_lru_busy);
 423                 atomic_long_add(npages, &cli->cl_lru_in_list);
 424                 cli->cl_lru_last_used = ktime_get_real_seconds();
 425                 spin_unlock(&cli->cl_lru_list_lock);
 426
 427                 if (waitqueue_active(&osc_lru_waitq))
 428                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 429         }
 430 }
 431
 432 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 433 {
 434         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 435         list_del_init(&opg->ops_lru);
 436         atomic_long_dec(&cli->cl_lru_in_list);
 437 }
 438
 439 /**
 440  * Page is being destroyed. The page may be not in LRU list, if the transfer
 441  * has never finished(error occurred).
 442  */
 443 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 444 {
 445         if (opg->ops_in_lru) {
 446                 spin_lock(&cli->cl_lru_list_lock);
 447                 if (!list_empty(&opg->ops_lru)) {
 448                         __osc_lru_del(cli, opg);
 449                 } else {
 450                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 451                         atomic_long_dec(&cli->cl_lru_busy);
 452                 }
 453                 spin_unlock(&cli->cl_lru_list_lock);
 454
 455                 atomic_long_inc(cli->cl_lru_left);
 456                 /* this is a great place to release more LRU pages if
 457                  * this osc occupies too many LRU pages and kernel is
 458                  * stealing one of them. */
 459                 if (osc_cache_too_much(cli)) {
 460                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 461                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 462                 }
 463                 wake_up(&osc_lru_waitq);
 464         } else {
 465                 LASSERT(list_empty(&opg->ops_lru));
 466         }
 467 }
 468
 469 /**
 470  * Delete page from LRU list for redirty.
 471  */
 472 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 473 {
 474         /* If page is being transferred for the first time,
 475          * ops_lru should be empty */
 476         if (opg->ops_in_lru) {
 477                 if (list_empty(&opg->ops_lru))
 478                         return;
 479                 spin_lock(&cli->cl_lru_list_lock);
 480                 if (!list_empty(&opg->ops_lru)) {
 481                         __osc_lru_del(cli, opg);
 482                         atomic_long_inc(&cli->cl_lru_busy);
 483                 }
 484                 spin_unlock(&cli->cl_lru_list_lock);
 485         }
 486 }
 487
 488 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 489                                 struct cl_page **pvec, int max_index)
 490 {
 491         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 492         int i;
 493
 494         ll_pagevec_init(pagevec, 0);
 495         for (i = 0; i < max_index; i++) {
 496                 struct cl_page *page = pvec[i];
 497
 498                 LASSERT(cl_page_is_owned(page, io));
 499                 cl_page_delete(env, page);
 500                 cl_page_discard(env, io, page);
 501                 cl_page_disown(env, io, page);
 502                 cl_pagevec_put(env, page, pagevec);
 503
 504                 pvec[i] = NULL;
 505         }
 506         pagevec_release(pagevec);
 507 }
 508
 509 /**
 510  * Check if a cl_page can be released, i.e, it's not being used.
 511  *
 512  * If unstable account is turned on, bulk transfer may hold one refcount
 513  * for recovery so we need to check vmpage refcount as well; otherwise,
 514  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 515  */
 516 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 517 {
 518         if (cl_page_in_use_noref(page))
 519                 return true;
 520
 521         if (cli->cl_cache->ccc_unstable_check) {
 522                 struct page *vmpage = cl_page_vmpage(page);
 523
 524                 /* this check is racy because the vmpage is not locked, but
 525                  * that's OK - the code which does the actual page release
 526                  * checks this again before releasing
 527                  *
 528                  * vmpage have two known users: cl_page and VM page cache
 529                  */
 530                 if (vmpage_in_use(vmpage, 0))
 531                         return true;
 532         }
 533         return false;
 534 }
 535
 536 /**
 537  * Drop @target of pages from LRU at most.
 538  */
 539 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 540                    long target, bool force)
 541 {
 542         struct cl_io *io;
 543         struct cl_object *clobj = NULL;
 544         struct cl_page **pvec;
 545         struct osc_page *opg;
 546         long count = 0;
 547         int maxscan = 0;
 548         int index = 0;
 549         int rc = 0;
 550         ENTRY;
 551
 552         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 553         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 554                 RETURN(0);
 555
 556         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 557                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 558         if (!force) {
 559                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 560                         RETURN(-EBUSY);
 561
 562                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 563                         atomic_dec(&cli->cl_lru_shrinkers);
 564                         RETURN(-EBUSY);
 565                 }
 566         } else {
 567                 atomic_inc(&cli->cl_lru_shrinkers);
 568         }
 569
 570         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 571         io = osc_env_thread_io(env);
 572
 573         spin_lock(&cli->cl_lru_list_lock);
 574         if (force)
 575                 cli->cl_lru_reclaim++;
 576         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 577         while (!list_empty(&cli->cl_lru_list)) {
 578                 struct cl_page *page;
 579                 bool will_free = false;
 580
 581                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 582                         break;
 583
 584                 if (--maxscan < 0)
 585                         break;
 586
 587                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 588                                        ops_lru);
 589                 page = opg->ops_cl.cpl_page;
 590                 if (lru_page_busy(cli, page)) {
 591                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 592                         continue;
 593                 }
 594
 595                 LASSERT(page->cp_obj != NULL);
 596                 if (clobj != page->cp_obj) {
 597                         struct cl_object *tmp = page->cp_obj;
 598
 599                         cl_object_get(tmp);
 600                         spin_unlock(&cli->cl_lru_list_lock);
 601
 602                         if (clobj != NULL) {
 603                                 discard_pagevec(env, io, pvec, index);
 604                                 index = 0;
 605
 606                                 cl_io_fini(env, io);
 607                                 cl_object_put(env, clobj);
 608                                 clobj = NULL;
 609                         }
 610
 611                         clobj = tmp;
 612                         io->ci_obj = clobj;
 613                         io->ci_ignore_layout = 1;
 614                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 615
 616                         spin_lock(&cli->cl_lru_list_lock);
 617
 618                         if (rc != 0)
 619                                 break;
 620
 621                         ++maxscan;
 622                         continue;
 623                 }
 624
 625                 if (cl_page_own_try(env, io, page) == 0) {
 626                         if (!lru_page_busy(cli, page)) {
 627                                 /* remove it from lru list earlier to avoid
 628                                  * lock contention */
 629                                 __osc_lru_del(cli, opg);
 630                                 opg->ops_in_lru = 0; /* will be discarded */
 631
 632                                 cl_page_get(page);
 633                                 will_free = true;
 634                         } else {
 635                                 cl_page_disown(env, io, page);
 636                         }
 637                 }
 638
 639                 if (!will_free) {
 640                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 641                         continue;
 642                 }
 643
 644                 /* Don't discard and free the page with cl_lru_list held */
 645                 pvec[index++] = page;
 646                 if (unlikely(index == OTI_PVEC_SIZE)) {
 647                         spin_unlock(&cli->cl_lru_list_lock);
 648                         discard_pagevec(env, io, pvec, index);
 649                         index = 0;
 650
 651                         spin_lock(&cli->cl_lru_list_lock);
 652                 }
 653
 654                 if (++count >= target)
 655                         break;
 656         }
 657         spin_unlock(&cli->cl_lru_list_lock);
 658
 659         if (clobj != NULL) {
 660                 discard_pagevec(env, io, pvec, index);
 661
 662                 cl_io_fini(env, io);
 663                 cl_object_put(env, clobj);
 664         }
 665
 666         atomic_dec(&cli->cl_lru_shrinkers);
 667         if (count > 0) {
 668                 atomic_long_add(count, cli->cl_lru_left);
 669                 wake_up(&osc_lru_waitq);
 670         }
 671         RETURN(count > 0 ? count : rc);
 672 }
 673 EXPORT_SYMBOL(osc_lru_shrink);
 674
 675 /**
 676  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 677  * \@npages of LRU slots. For performance consideration, it's better to drop
 678  * LRU pages in batch. Therefore, the actual number is adjusted at least
 679  * max_pages_per_rpc.
 680  */
 681 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 682 {
 683         struct lu_env *env;
 684         struct cl_client_cache *cache = cli->cl_cache;
 685         struct client_obd *scan;
 686         int max_scans;
 687         __u16 refcheck;
 688         long rc = 0;
 689         ENTRY;
 690
 691         LASSERT(cache != NULL);
 692
 693         env = cl_env_get(&refcheck);
 694         if (IS_ERR(env))
 695                 RETURN(rc);
 696
 697         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 698         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 699                cli_name(cli), npages);
 700         rc = osc_lru_shrink(env, cli, npages, true);
 701         if (rc >= npages) {
 702                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 703                        cli_name(cli), rc, npages);
 704                 if (osc_cache_too_much(cli) > 0)
 705                         ptlrpcd_queue_work(cli->cl_lru_work);
 706                 GOTO(out, rc);
 707         } else if (rc > 0) {
 708                 npages -= rc;
 709         }
 710
 711         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 712                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 713                 atomic_long_read(&cli->cl_lru_busy), npages);
 714
 715         /* Reclaim LRU slots from other client_obd as it can't free enough
 716          * from its own. This should rarely happen. */
 717         spin_lock(&cache->ccc_lru_lock);
 718         LASSERT(!list_empty(&cache->ccc_lru));
 719
 720         cache->ccc_lru_shrinkers++;
 721         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 722
 723         max_scans = refcount_read(&cache->ccc_users) - 2;
 724         while (--max_scans > 0 &&
 725                (scan = list_first_entry_or_null(&cache->ccc_lru,
 726                                                   struct client_obd,
 727                                                   cl_lru_osc)) != NULL) {
 728                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 729                        cli_name(scan), scan,
 730                        atomic_long_read(&scan->cl_lru_in_list),
 731                        atomic_long_read(&scan->cl_lru_busy));
 732
 733                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 734                 if (osc_cache_too_much(scan) > 0) {
 735                         spin_unlock(&cache->ccc_lru_lock);
 736
 737                         rc = osc_lru_shrink(env, scan, npages, true);
 738                         spin_lock(&cache->ccc_lru_lock);
 739                         if (rc >= npages)
 740                                 break;
 741                         if (rc > 0)
 742                                 npages -= rc;
 743                 }
 744         }
 745         spin_unlock(&cache->ccc_lru_lock);
 746
 747 out:
 748         cl_env_put(env, &refcheck);
 749         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 750                cli_name(cli), cli, rc);
 751         return rc;
 752 }
 753
 754 /**
 755  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 756  *
 757  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 758  * Only in the case that the LRU slots are in extreme shortage, it should
 759  * have reserved enough slots for an IO.
 760  */
 761 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 762                          struct osc_page *opg)
 763 {
 764         struct osc_io *oio = osc_env_io(env);
 765         int rc = 0;
 766
 767         ENTRY;
 768
 769         if (cli->cl_cache == NULL) /* shall not be in LRU */
 770                 RETURN(0);
 771
 772         if (oio->oi_lru_reserved > 0) {
 773                 --oio->oi_lru_reserved;
 774                 goto out;
 775         }
 776
 777         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 778         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 779                 /* run out of LRU spaces, try to drop some by itself */
 780                 rc = osc_lru_reclaim(cli, 1);
 781                 if (rc < 0)
 782                         break;
 783                 if (rc > 0)
 784                         continue;
 785                 /* IO issued by readahead, don't try hard */
 786                 if (oio->oi_is_readahead) {
 787                         if (atomic_long_read(cli->cl_lru_left) > 0)
 788                                 continue;
 789                         rc = -EBUSY;
 790                         break;
 791                 }
 792
 793                 cond_resched();
 794                 rc = l_wait_event_abortable(
 795                         osc_lru_waitq,
 796                         atomic_long_read(cli->cl_lru_left) > 0);
 797                 if (rc < 0) {
 798                         rc = -EINTR;
 799                         break;
 800                 }
 801         }
 802
 803 out:
 804         if (rc >= 0) {
 805                 atomic_long_inc(&cli->cl_lru_busy);
 806                 opg->ops_in_lru = 1;
 807                 rc = 0;
 808         }
 809
 810         RETURN(rc);
 811 }
 812
 813 /**
 814  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 815  *
 816  * The benefit of doing this is to reduce contention against atomic counter
 817  * cl_lru_left by changing it from per-page access to per-IO access.
 818  */
 819 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 820 {
 821         unsigned long reserved = 0;
 822         unsigned long max_pages;
 823         unsigned long c;
 824         int rc;
 825
 826 again:
 827         c = atomic_long_read(cli->cl_lru_left);
 828         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 829                 c = atomic_long_read(cli->cl_lru_left);
 830
 831         if (c < npages) {
 832                 /*
 833                  * Trigger writeback in the hope some LRU slot could
 834                  * be freed.
 835                  */
 836                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 837                 if (rc)
 838                         return 0;
 839         }
 840
 841         while (c >= npages) {
 842                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 843                         reserved = npages;
 844                         break;
 845                 }
 846                 c = atomic_long_read(cli->cl_lru_left);
 847         }
 848
 849         if (reserved != npages) {
 850                 cond_resched();
 851                 rc = l_wait_event_abortable(
 852                         osc_lru_waitq,
 853                         atomic_long_read(cli->cl_lru_left) > 0);
 854                 goto again;
 855         }
 856
 857         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 858         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 859                 /* If there aren't enough pages in the per-OSC LRU then
 860                  * wake up the LRU thread to try and clear out space, so
 861                  * we don't block if pages are being dirtied quickly. */
 862                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 863                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 864                        max_pages);
 865                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 866         }
 867
 868         return reserved;
 869 }
 870
 871 /**
 872  * osc_lru_unreserve() is called to unreserve LRU slots.
 873  *
 874  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 875  * reasons such as page already existing or I/O error. Those reserved slots
 876  * should be freed by calling this function.
 877  */
 878 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 879 {
 880         atomic_long_add(npages, cli->cl_lru_left);
 881         wake_up(&osc_lru_waitq);
 882 }
 883
 884 /**
 885  * Atomic operations are expensive. We accumulate the accounting for the
 886  * same page zone to get better performance.
 887  * In practice this can work pretty good because the pages in the same RPC
 888  * are likely from the same page zone.
 889  */
 890 #ifdef HAVE_NR_UNSTABLE_NFS
 891 /* Old kernels use a separate counter for unstable pages,
 892  * newer kernels treat them like any other writeback.
 893  */
 894 #define NR_WRITEBACK NR_UNSTABLE_NFS
 895 #endif
 896
 897 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 898                                             struct osc_brw_async_args *aa,
 899                                             int factor)
 900 {
 901         int page_count;
 902         void *zone = NULL;
 903         int count = 0;
 904         int i;
 905
 906         if (desc != NULL) {
 907                 page_count = desc->bd_iov_count;
 908         } else {
 909                 page_count = aa->aa_page_count;
 910         }
 911
 912         for (i = 0; i < page_count; i++) {
 913                 void *pz;
 914                 if (desc)
 915                         pz = page_zone(desc->bd_vec[i].bv_page);
 916                 else
 917                         pz = page_zone(aa->aa_ppga[i]->pg);
 918
 919                 if (likely(pz == zone)) {
 920                         ++count;
 921                         continue;
 922                 }
 923
 924                 if (count > 0) {
 925                         mod_zone_page_state(zone, NR_WRITEBACK,
 926                                             factor * count);
 927                         count = 0;
 928                 }
 929                 zone = pz;
 930                 ++count;
 931         }
 932         if (count > 0)
 933                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 934 }
 935
 936 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 937                                                 struct osc_brw_async_args *aa)
 938 {
 939         unstable_page_accounting(desc, aa, 1);
 940 }
 941
 942 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 943                                                 struct osc_brw_async_args *aa)
 944 {
 945         unstable_page_accounting(desc, aa, -1);
 946 }
 947
 948 /**
 949  * Performs "unstable" page accounting. This function balances the
 950  * increment operations performed in osc_inc_unstable_pages. It is
 951  * registered as the RPC request callback, and is executed when the
 952  * bulk RPC is committed on the server. Thus at this point, the pages
 953  * involved in the bulk transfer are no longer considered unstable.
 954  *
 955  * If this function is called, the request should have been committed
 956  * or req:rq_unstable must have been set; it implies that the unstable
 957  * statistic have been added.
 958  */
 959 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 960 {
 961         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 962         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 963         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 964         int                      page_count;
 965         long                     unstable_count;
 966
 967         if (desc)
 968                 page_count = desc->bd_iov_count;
 969         else
 970                 page_count = aa->aa_page_count;
 971
 972         LASSERT(page_count >= 0);
 973
 974         dec_unstable_page_accounting(desc, aa);
 975
 976         unstable_count = atomic_long_sub_return(page_count,
 977                                                 &cli->cl_unstable_count);
 978         LASSERT(unstable_count >= 0);
 979
 980         unstable_count = atomic_long_sub_return(page_count,
 981                                            &cli->cl_cache->ccc_unstable_nr);
 982         LASSERT(unstable_count >= 0);
 983         if (unstable_count == 0)
 984                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 985
 986         if (waitqueue_active(&osc_lru_waitq))
 987                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 988 }
 989
 990 /**
 991  * "unstable" page accounting. See: osc_dec_unstable_pages.
 992  */
 993 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 994 {
 995         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 996         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 997         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 998         long                     page_count;
 999
1000         /* No unstable page tracking */
1001         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1002                 return;
1003
1004         if (desc)
1005                 page_count = desc->bd_iov_count;
1006         else
1007                 page_count = aa->aa_page_count;
1008
1009         add_unstable_page_accounting(desc, aa);
1010         atomic_long_add(page_count, &cli->cl_unstable_count);
1011         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1012
1013         /* If the request has already been committed (i.e. brw_commit
1014          * called via rq_commit_cb), we need to undo the unstable page
1015          * increments we just performed because rq_commit_cb wont be
1016          * called again. */
1017         spin_lock(&req->rq_lock);
1018         if (unlikely(req->rq_committed)) {
1019                 spin_unlock(&req->rq_lock);
1020
1021                 osc_dec_unstable_pages(req);
1022         } else {
1023                 req->rq_unstable = 1;
1024                 spin_unlock(&req->rq_lock);
1025         }
1026 }
1027
1028 /**
1029  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1030  * This function will be called by every BRW RPC so it's critical
1031  * to make this function fast.
1032  */
1033 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1034 {
1035         long unstable_nr, osc_unstable_count;
1036
1037         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1038         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1039                 return false;
1040
1041         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1042         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1043
1044         CDEBUG(D_CACHE,
1045                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1046                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1047
1048         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1049          * has one full RPC window of unstable pages, it's a good chance
1050          * to piggyback a SOFT_SYNC flag.
1051          * Please notice that the OST won't take immediate response for the
1052          * SOFT_SYNC request so active OSCs will have more chance to carry
1053          * the flag, this is reasonable. */
1054         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1055                osc_unstable_count > cli->cl_max_pages_per_rpc *
1056                                     cli->cl_max_rpcs_in_flight;
1057 }
1058
1059 /**
1060  * Return how many LRU pages in the cache of all OSC devices
1061  *
1062  * \retval      return # of cached LRU pages times reclaimation tendency
1063  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1064  */
1065 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1066                                      struct shrink_control *sc)
1067 {
1068         struct client_obd *cli;
1069         unsigned long cached = 0;
1070
1071         spin_lock(&osc_shrink_lock);
1072         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1073                 cached += atomic_long_read(&cli->cl_lru_in_list);
1074         spin_unlock(&osc_shrink_lock);
1075
1076         return (cached  * sysctl_vfs_cache_pressure) / 100;
1077 }
1078
1079 /**
1080  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1081  *
1082  * \retval      number of cached LRU pages reclaimed
1083  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1084  *
1085  * Linux kernel will loop calling this shrinker scan routine with
1086  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1087  *
1088  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1089  * to scan and try to reclaim LRU pages, just return 0 and
1090  * osc_cache_shrink_count() will report the LRU page number.
1091  */
1092 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1093                                     struct shrink_control *sc)
1094 {
1095         struct client_obd *cli;
1096         struct client_obd *stop_anchor = NULL;
1097         struct lu_env *env;
1098         long shrank = 0;
1099         int rc;
1100         __u16 refcheck;
1101
1102         if (sc->nr_to_scan == 0)
1103                 return 0;
1104
1105         if (!(sc->gfp_mask & __GFP_FS))
1106                 return SHRINK_STOP;
1107
1108         env = cl_env_get(&refcheck);
1109         if (IS_ERR(env))
1110                 return SHRINK_STOP;
1111
1112         spin_lock(&osc_shrink_lock);
1113         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1114                                                struct client_obd,
1115                                                cl_shrink_list)) != NULL) {
1116                 if (stop_anchor == NULL)
1117                         stop_anchor = cli;
1118                 else if (cli == stop_anchor)
1119                         break;
1120
1121                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1122                 spin_unlock(&osc_shrink_lock);
1123
1124                 /* shrink no more than max_pages_per_rpc for an OSC */
1125                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1126                                     cli->cl_max_pages_per_rpc ?
1127                                     cli->cl_max_pages_per_rpc :
1128                                     sc->nr_to_scan - shrank, true);
1129                 if (rc > 0)
1130                         shrank += rc;
1131
1132                 if (shrank >= sc->nr_to_scan)
1133                         goto out;
1134
1135                 spin_lock(&osc_shrink_lock);
1136         }
1137         spin_unlock(&osc_shrink_lock);
1138
1139 out:
1140         cl_env_put(env, &refcheck);
1141
1142         return shrank;
1143 }
1144
1145 /** @} osc */