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 = cl2osc(opg->ops_cl.cpl_obj);
  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 inline s64 osc_submit_duration(struct osc_page *opg)
 114 {
 115         if (ktime_to_ns(opg->ops_submit_time) == 0)
 116                 return 0;
 117
 118         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 119 }
 120
 121 static int osc_page_print(const struct lu_env *env,
 122                           const struct cl_page_slice *slice,
 123                           void *cookie, lu_printer_t printer)
 124 {
 125         struct osc_page *opg = cl2osc_page(slice);
 126         struct osc_async_page *oap = &opg->ops_oap;
 127         struct osc_object *obj = cl2osc(slice->cpl_obj);
 128         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 129
 130         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 131                           "1< %#x %d %c %c > "
 132                           "2< %lld %u %u %#x %#x | %p %p %p > "
 133                           "3< %d %lld %d > "
 134                           "4< %d %d %d %lu %c | %c %c %c %c > "
 135                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 136                           opg, osc_index(opg),
 137                           /* 1 */
 138                           oap->oap_magic, oap->oap_cmd,
 139                           list_empty_marker(&oap->oap_pending_item),
 140                           list_empty_marker(&oap->oap_rpc_item),
 141                           /* 2 */
 142                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 143                           oap->oap_async_flags, oap->oap_brw_flags,
 144                           oap->oap_request, oap->oap_cli, obj,
 145                           /* 3 */
 146                           opg->ops_transfer_pinned,
 147                           osc_submit_duration(opg), opg->ops_srvlock,
 148                           /* 4 */
 149                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 150                           cli->cl_max_rpcs_in_flight,
 151                           cli->cl_avail_grant,
 152                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 153                           list_empty_marker(&cli->cl_loi_ready_list),
 154                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 155                           list_empty_marker(&cli->cl_loi_write_list),
 156                           list_empty_marker(&cli->cl_loi_read_list),
 157                           /* 5 */
 158                           list_empty_marker(&obj->oo_ready_item),
 159                           list_empty_marker(&obj->oo_hp_ready_item),
 160                           list_empty_marker(&obj->oo_write_item),
 161                           list_empty_marker(&obj->oo_read_item),
 162                           atomic_read(&obj->oo_nr_reads),
 163                           list_empty_marker(&obj->oo_reading_exts),
 164                           atomic_read(&obj->oo_nr_writes),
 165                           list_empty_marker(&obj->oo_hp_exts),
 166                           list_empty_marker(&obj->oo_urgent_exts));
 167 }
 168
 169 static void osc_page_delete(const struct lu_env *env,
 170                             const struct cl_page_slice *slice)
 171 {
 172         struct osc_page   *opg = cl2osc_page(slice);
 173         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 174         int rc;
 175
 176         ENTRY;
 177         CDEBUG(D_TRACE, "%p\n", opg);
 178         osc_page_transfer_put(env, opg);
 179         rc = osc_teardown_async_page(env, obj, opg);
 180         if (rc) {
 181                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 182                               "Trying to teardown failed: %d\n", rc);
 183                 LASSERT(0);
 184         }
 185
 186         osc_lru_del(osc_cli(obj), opg);
 187
 188         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 189                 void *value = NULL;
 190
 191                 spin_lock(&obj->oo_tree_lock);
 192                 if (opg->ops_intree) {
 193                         value = radix_tree_delete(&obj->oo_tree,
 194                                                   osc_index(opg));
 195                         if (value != NULL) {
 196                                 --obj->oo_npages;
 197                                 opg->ops_intree = 0;
 198                         }
 199                 }
 200                 spin_unlock(&obj->oo_tree_lock);
 201
 202                 LASSERT(ergo(value != NULL, value == opg));
 203         }
 204
 205         EXIT;
 206 }
 207
 208 static void osc_page_clip(const struct lu_env *env,
 209                           const struct cl_page_slice *slice,
 210                           int from, int to)
 211 {
 212         struct osc_page       *opg = cl2osc_page(slice);
 213         struct osc_async_page *oap = &opg->ops_oap;
 214
 215         opg->ops_from = from;
 216         /* argument @to is exclusive, but @ops_to is inclusive */
 217         opg->ops_to   = to - 1;
 218         spin_lock(&oap->oap_lock);
 219         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 220         spin_unlock(&oap->oap_lock);
 221 }
 222
 223 static int osc_page_flush(const struct lu_env *env,
 224                           const struct cl_page_slice *slice,
 225                           struct cl_io *io)
 226 {
 227         struct osc_page *opg = cl2osc_page(slice);
 228         int rc = 0;
 229         ENTRY;
 230         rc = osc_flush_async_page(env, io, opg);
 231         RETURN(rc);
 232 }
 233
 234 static void osc_page_touch(const struct lu_env *env,
 235                           const struct cl_page_slice *slice, size_t to)
 236 {
 237         struct osc_page *opg = cl2osc_page(slice);
 238         struct cl_object *obj = opg->ops_cl.cpl_obj;
 239
 240         osc_page_touch_at(env, obj, osc_index(opg), to);
 241 }
 242
 243 static const struct cl_page_operations osc_page_ops = {
 244         .cpo_print         = osc_page_print,
 245         .cpo_delete        = osc_page_delete,
 246         .cpo_clip           = osc_page_clip,
 247         .cpo_flush          = osc_page_flush,
 248         .cpo_page_touch    = osc_page_touch,
 249 };
 250
 251 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 252                   struct cl_page *cl_page, pgoff_t index)
 253 {
 254         struct osc_object *osc = cl2osc(obj);
 255         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 256         struct osc_io *oio = osc_env_io(env);
 257         int result;
 258
 259         opg->ops_from = 0;
 260         opg->ops_to = PAGE_SIZE - 1;
 261
 262         INIT_LIST_HEAD(&opg->ops_lru);
 263
 264         result = osc_prep_async_page(osc, opg, cl_page->cp_vmpage,
 265                                      cl_offset(obj, index));
 266         if (result != 0)
 267                 return result;
 268
 269         opg->ops_srvlock = osc_io_srvlock(oio);
 270         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 271         cl_page->cp_osc_index = index;
 272
 273         /* reserve an LRU space for this page */
 274         if (cl_page->cp_type == CPT_CACHEABLE) {
 275                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 276                 if (result == 0) {
 277                         result = radix_tree_preload(GFP_NOFS);
 278                         if (result == 0) {
 279                                 spin_lock(&osc->oo_tree_lock);
 280                                 result = radix_tree_insert(&osc->oo_tree,
 281                                                            index, opg);
 282                                 if (result == 0) {
 283                                         ++osc->oo_npages;
 284                                         opg->ops_intree = 1;
 285                                 }
 286                                 spin_unlock(&osc->oo_tree_lock);
 287
 288                                 radix_tree_preload_end();
 289                         }
 290                 }
 291         }
 292
 293         return result;
 294 }
 295 EXPORT_SYMBOL(osc_page_init);
 296
 297 /**
 298  * Helper function called by osc_io_submit() for every page in an immediate
 299  * transfer (i.e., transferred synchronously).
 300  */
 301 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 302                      enum cl_req_type crt, int brw_flags, ktime_t submit_time)
 303 {
 304         struct osc_io *oio = osc_env_io(env);
 305         struct osc_async_page *oap = &opg->ops_oap;
 306
 307         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 308                  "magic 0x%x\n", oap, oap->oap_magic);
 309         LASSERT(oap->oap_async_flags & ASYNC_READY);
 310         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 311
 312         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 313         oap->oap_page_off = opg->ops_from;
 314         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 315         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 316
 317         if (oio->oi_cap_sys_resource) {
 318                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 319                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 320         }
 321
 322         opg->ops_submit_time = submit_time;
 323         osc_page_transfer_get(opg, "transfer\0imm");
 324         osc_page_transfer_add(env, opg, crt);
 325 }
 326
 327 /* --------------- LRU page management ------------------ */
 328
 329 /* OSC is a natural place to manage LRU pages as applications are specialized
 330  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 331  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 332  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 333  * for free LRU slots - this will be very bad so the algorithm requires each
 334  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 335  * at any time.
 336  */
 337
 338 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 339
 340 /**
 341  * LRU pages are freed in batch mode. OSC should at least free this
 342  * number of pages to avoid running out of LRU slots.
 343  */
 344 static inline int lru_shrink_min(struct client_obd *cli)
 345 {
 346         return cli->cl_max_pages_per_rpc * 2;
 347 }
 348
 349 /**
 350  * free this number at most otherwise it will take too long time to finsih.
 351  */
 352 static inline int lru_shrink_max(struct client_obd *cli)
 353 {
 354         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 355 }
 356
 357 /**
 358  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 359  * we should free slots aggressively. In this way, slots are freed in a steady
 360  * step to maintain fairness among OSCs.
 361  *
 362  * Return how many LRU pages should be freed.
 363  */
 364 static int osc_cache_too_much(struct client_obd *cli)
 365 {
 366         struct cl_client_cache *cache = cli->cl_cache;
 367         long pages = atomic_long_read(&cli->cl_lru_in_list);
 368         unsigned long budget;
 369
 370         LASSERT(cache != NULL);
 371         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 372
 373         /* if it's going to run out LRU slots, we should free some, but not
 374          * too much to maintain faireness among OSCs. */
 375         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 376                 if (pages >= budget)
 377                         return lru_shrink_max(cli);
 378                 else if (pages >= budget / 2)
 379                         return lru_shrink_min(cli);
 380         } else {
 381                 time64_t duration = ktime_get_real_seconds();
 382                 long timediff;
 383
 384                 /* knock out pages by duration of no IO activity */
 385                 duration -= cli->cl_lru_last_used;
 386                 /*
 387                  * The difference shouldn't be more than 70 years
 388                  * so we can safely case to a long. Round to
 389                  * approximately 1 minute.
 390                  */
 391                 timediff = (long)(duration >> 6);
 392                 if (timediff > 0 && pages >= budget / timediff)
 393                         return lru_shrink_min(cli);
 394         }
 395         return 0;
 396 }
 397
 398 int lru_queue_work(const struct lu_env *env, void *data)
 399 {
 400         struct client_obd *cli = data;
 401         int count;
 402
 403         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 404         count = osc_cache_too_much(cli);
 405         if (count > 0) {
 406                 int rc = osc_lru_shrink(env, cli, count, false);
 407
 408                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 409                        cli_name(cli), rc, count);
 410                 if (rc >= count) {
 411                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 412                         ptlrpcd_queue_work(cli->cl_lru_work);
 413                 }
 414         }
 415
 416         RETURN(0);
 417 }
 418
 419 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 420 {
 421         LIST_HEAD(lru);
 422         struct osc_async_page *oap;
 423         long npages = 0;
 424
 425         list_for_each_entry(oap, plist, oap_pending_item) {
 426                 struct osc_page *opg = oap2osc_page(oap);
 427
 428                 if (!opg->ops_in_lru)
 429                         continue;
 430
 431                 ++npages;
 432                 LASSERT(list_empty(&opg->ops_lru));
 433                 list_add(&opg->ops_lru, &lru);
 434         }
 435
 436         if (npages > 0) {
 437                 spin_lock(&cli->cl_lru_list_lock);
 438                 list_splice_tail(&lru, &cli->cl_lru_list);
 439                 atomic_long_sub(npages, &cli->cl_lru_busy);
 440                 atomic_long_add(npages, &cli->cl_lru_in_list);
 441                 cli->cl_lru_last_used = ktime_get_real_seconds();
 442                 spin_unlock(&cli->cl_lru_list_lock);
 443
 444                 if (waitqueue_active(&osc_lru_waitq))
 445                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 446         }
 447 }
 448
 449 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 450 {
 451         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 452         list_del_init(&opg->ops_lru);
 453         atomic_long_dec(&cli->cl_lru_in_list);
 454 }
 455
 456 /**
 457  * Page is being destroyed. The page may be not in LRU list, if the transfer
 458  * has never finished(error occurred).
 459  */
 460 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 461 {
 462         if (opg->ops_in_lru) {
 463                 spin_lock(&cli->cl_lru_list_lock);
 464                 if (!list_empty(&opg->ops_lru)) {
 465                         __osc_lru_del(cli, opg);
 466                 } else {
 467                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 468                         atomic_long_dec(&cli->cl_lru_busy);
 469                 }
 470                 spin_unlock(&cli->cl_lru_list_lock);
 471
 472                 atomic_long_inc(cli->cl_lru_left);
 473                 /* this is a great place to release more LRU pages if
 474                  * this osc occupies too many LRU pages and kernel is
 475                  * stealing one of them. */
 476                 if (osc_cache_too_much(cli)) {
 477                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 478                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 479                 }
 480                 wake_up(&osc_lru_waitq);
 481         } else {
 482                 LASSERT(list_empty(&opg->ops_lru));
 483         }
 484 }
 485
 486 /**
 487  * Delete page from LRU list for redirty.
 488  */
 489 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 490 {
 491         /* If page is being transferred for the first time,
 492          * ops_lru should be empty */
 493         if (opg->ops_in_lru) {
 494                 if (list_empty(&opg->ops_lru))
 495                         return;
 496                 spin_lock(&cli->cl_lru_list_lock);
 497                 if (!list_empty(&opg->ops_lru)) {
 498                         __osc_lru_del(cli, opg);
 499                         atomic_long_inc(&cli->cl_lru_busy);
 500                 }
 501                 spin_unlock(&cli->cl_lru_list_lock);
 502         }
 503 }
 504
 505 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 506                                 struct cl_page **pvec, int max_index)
 507 {
 508         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 509         int i;
 510
 511         ll_pagevec_init(pagevec, 0);
 512         for (i = 0; i < max_index; i++) {
 513                 struct cl_page *page = pvec[i];
 514
 515                 LASSERT(cl_page_is_owned(page, io));
 516                 cl_page_delete(env, page);
 517                 cl_page_discard(env, io, page);
 518                 cl_page_disown(env, io, page);
 519                 cl_pagevec_put(env, page, pagevec);
 520
 521                 pvec[i] = NULL;
 522         }
 523         pagevec_release(pagevec);
 524 }
 525
 526 /**
 527  * Check if a cl_page can be released, i.e, it's not being used.
 528  *
 529  * If unstable account is turned on, bulk transfer may hold one refcount
 530  * for recovery so we need to check vmpage refcount as well; otherwise,
 531  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 532  */
 533 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 534 {
 535         if (cl_page_in_use_noref(page))
 536                 return true;
 537
 538         if (cli->cl_cache->ccc_unstable_check) {
 539                 struct page *vmpage = cl_page_vmpage(page);
 540
 541                 /* vmpage have two known users: cl_page and VM page cache */
 542                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 543                         return true;
 544         }
 545         return false;
 546 }
 547
 548 /**
 549  * Drop @target of pages from LRU at most.
 550  */
 551 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 552                    long target, bool force)
 553 {
 554         struct cl_io *io;
 555         struct cl_object *clobj = NULL;
 556         struct cl_page **pvec;
 557         struct osc_page *opg;
 558         long count = 0;
 559         int maxscan = 0;
 560         int index = 0;
 561         int rc = 0;
 562         ENTRY;
 563
 564         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 565         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 566                 RETURN(0);
 567
 568         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 569                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 570         if (!force) {
 571                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 572                         RETURN(-EBUSY);
 573
 574                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 575                         atomic_dec(&cli->cl_lru_shrinkers);
 576                         RETURN(-EBUSY);
 577                 }
 578         } else {
 579                 atomic_inc(&cli->cl_lru_shrinkers);
 580         }
 581
 582         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 583         io = osc_env_thread_io(env);
 584
 585         spin_lock(&cli->cl_lru_list_lock);
 586         if (force)
 587                 cli->cl_lru_reclaim++;
 588         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 589         while (!list_empty(&cli->cl_lru_list)) {
 590                 struct cl_page *page;
 591                 bool will_free = false;
 592
 593                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 594                         break;
 595
 596                 if (--maxscan < 0)
 597                         break;
 598
 599                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 600                                  ops_lru);
 601                 page = opg->ops_cl.cpl_page;
 602                 if (lru_page_busy(cli, page)) {
 603                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 604                         continue;
 605                 }
 606
 607                 LASSERT(page->cp_obj != NULL);
 608                 if (clobj != page->cp_obj) {
 609                         struct cl_object *tmp = page->cp_obj;
 610
 611                         cl_object_get(tmp);
 612                         spin_unlock(&cli->cl_lru_list_lock);
 613
 614                         if (clobj != NULL) {
 615                                 discard_pagevec(env, io, pvec, index);
 616                                 index = 0;
 617
 618                                 cl_io_fini(env, io);
 619                                 cl_object_put(env, clobj);
 620                                 clobj = NULL;
 621                         }
 622
 623                         clobj = tmp;
 624                         io->ci_obj = clobj;
 625                         io->ci_ignore_layout = 1;
 626                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 627
 628                         spin_lock(&cli->cl_lru_list_lock);
 629
 630                         if (rc != 0)
 631                                 break;
 632
 633                         ++maxscan;
 634                         continue;
 635                 }
 636
 637                 if (cl_page_own_try(env, io, page) == 0) {
 638                         if (!lru_page_busy(cli, page)) {
 639                                 /* remove it from lru list earlier to avoid
 640                                  * lock contention */
 641                                 __osc_lru_del(cli, opg);
 642                                 opg->ops_in_lru = 0; /* will be discarded */
 643
 644                                 cl_page_get(page);
 645                                 will_free = true;
 646                         } else {
 647                                 cl_page_disown(env, io, page);
 648                         }
 649                 }
 650
 651                 if (!will_free) {
 652                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 653                         continue;
 654                 }
 655
 656                 /* Don't discard and free the page with cl_lru_list held */
 657                 pvec[index++] = page;
 658                 if (unlikely(index == OTI_PVEC_SIZE)) {
 659                         spin_unlock(&cli->cl_lru_list_lock);
 660                         discard_pagevec(env, io, pvec, index);
 661                         index = 0;
 662
 663                         spin_lock(&cli->cl_lru_list_lock);
 664                 }
 665
 666                 if (++count >= target)
 667                         break;
 668         }
 669         spin_unlock(&cli->cl_lru_list_lock);
 670
 671         if (clobj != NULL) {
 672                 discard_pagevec(env, io, pvec, index);
 673
 674                 cl_io_fini(env, io);
 675                 cl_object_put(env, clobj);
 676         }
 677
 678         atomic_dec(&cli->cl_lru_shrinkers);
 679         if (count > 0) {
 680                 atomic_long_add(count, cli->cl_lru_left);
 681                 wake_up(&osc_lru_waitq);
 682         }
 683         RETURN(count > 0 ? count : rc);
 684 }
 685 EXPORT_SYMBOL(osc_lru_shrink);
 686
 687 /**
 688  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 689  * \@npages of LRU slots. For performance consideration, it's better to drop
 690  * LRU pages in batch. Therefore, the actual number is adjusted at least
 691  * max_pages_per_rpc.
 692  */
 693 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 694 {
 695         struct lu_env *env;
 696         struct cl_client_cache *cache = cli->cl_cache;
 697         int max_scans;
 698         __u16 refcheck;
 699         long rc = 0;
 700         ENTRY;
 701
 702         LASSERT(cache != NULL);
 703
 704         env = cl_env_get(&refcheck);
 705         if (IS_ERR(env))
 706                 RETURN(rc);
 707
 708         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 709         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 710                cli_name(cli), npages);
 711         rc = osc_lru_shrink(env, cli, npages, true);
 712         if (rc >= npages) {
 713                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 714                        cli_name(cli), rc, npages);
 715                 if (osc_cache_too_much(cli) > 0)
 716                         ptlrpcd_queue_work(cli->cl_lru_work);
 717                 GOTO(out, rc);
 718         } else if (rc > 0) {
 719                 npages -= rc;
 720         }
 721
 722         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 723                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 724                 atomic_long_read(&cli->cl_lru_busy), npages);
 725
 726         /* Reclaim LRU slots from other client_obd as it can't free enough
 727          * from its own. This should rarely happen. */
 728         spin_lock(&cache->ccc_lru_lock);
 729         LASSERT(!list_empty(&cache->ccc_lru));
 730
 731         cache->ccc_lru_shrinkers++;
 732         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 733
 734         max_scans = atomic_read(&cache->ccc_users) - 2;
 735         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 736                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 737                                  cl_lru_osc);
 738
 739                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 740                         cli_name(cli), cli,
 741                         atomic_long_read(&cli->cl_lru_in_list),
 742                         atomic_long_read(&cli->cl_lru_busy));
 743
 744                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 745                 if (osc_cache_too_much(cli) > 0) {
 746                         spin_unlock(&cache->ccc_lru_lock);
 747
 748                         rc = osc_lru_shrink(env, cli, npages, true);
 749                         spin_lock(&cache->ccc_lru_lock);
 750                         if (rc >= npages)
 751                                 break;
 752                         if (rc > 0)
 753                                 npages -= rc;
 754                 }
 755         }
 756         spin_unlock(&cache->ccc_lru_lock);
 757
 758 out:
 759         cl_env_put(env, &refcheck);
 760         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 761                 cli_name(cli), cli, rc);
 762         return rc;
 763 }
 764
 765 /**
 766  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 767  *
 768  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 769  * Only in the case that the LRU slots are in extreme shortage, it should
 770  * have reserved enough slots for an IO.
 771  */
 772 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 773                          struct osc_page *opg)
 774 {
 775         struct osc_io *oio = osc_env_io(env);
 776         int rc = 0;
 777
 778         ENTRY;
 779
 780         if (cli->cl_cache == NULL) /* shall not be in LRU */
 781                 RETURN(0);
 782
 783         if (oio->oi_lru_reserved > 0) {
 784                 --oio->oi_lru_reserved;
 785                 goto out;
 786         }
 787
 788         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 789         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 790                 /* run out of LRU spaces, try to drop some by itself */
 791                 rc = osc_lru_reclaim(cli, 1);
 792                 if (rc < 0)
 793                         break;
 794                 if (rc > 0)
 795                         continue;
 796                 /* IO issued by readahead, don't try hard */
 797                 if (oio->oi_is_readahead) {
 798                         if (atomic_long_read(cli->cl_lru_left) > 0)
 799                                 continue;
 800                         rc = -EBUSY;
 801                         break;
 802                 }
 803
 804                 cond_resched();
 805                 rc = l_wait_event_abortable(
 806                         osc_lru_waitq,
 807                         atomic_long_read(cli->cl_lru_left) > 0);
 808                 if (rc < 0) {
 809                         rc = -EINTR;
 810                         break;
 811                 }
 812         }
 813
 814 out:
 815         if (rc >= 0) {
 816                 atomic_long_inc(&cli->cl_lru_busy);
 817                 opg->ops_in_lru = 1;
 818                 rc = 0;
 819         }
 820
 821         RETURN(rc);
 822 }
 823
 824 /**
 825  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 826  *
 827  * The benefit of doing this is to reduce contention against atomic counter
 828  * cl_lru_left by changing it from per-page access to per-IO access.
 829  */
 830 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 831 {
 832         unsigned long reserved = 0;
 833         unsigned long max_pages;
 834         unsigned long c;
 835         int rc;
 836
 837 again:
 838         c = atomic_long_read(cli->cl_lru_left);
 839         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 840                 c = atomic_long_read(cli->cl_lru_left);
 841
 842         if (c < npages) {
 843                 /*
 844                  * Trigger writeback in the hope some LRU slot could
 845                  * be freed.
 846                  */
 847                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 848                 if (rc)
 849                         return 0;
 850         }
 851
 852         while (c >= npages) {
 853                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 854                         reserved = npages;
 855                         break;
 856                 }
 857                 c = atomic_long_read(cli->cl_lru_left);
 858         }
 859
 860         if (reserved != npages) {
 861                 cond_resched();
 862                 rc = l_wait_event_abortable(
 863                         osc_lru_waitq,
 864                         atomic_long_read(cli->cl_lru_left) > 0);
 865                 goto again;
 866         }
 867
 868         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 869         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 870                 /* If there aren't enough pages in the per-OSC LRU then
 871                  * wake up the LRU thread to try and clear out space, so
 872                  * we don't block if pages are being dirtied quickly. */
 873                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 874                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 875                        max_pages);
 876                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 877         }
 878
 879         return reserved;
 880 }
 881
 882 /**
 883  * osc_lru_unreserve() is called to unreserve LRU slots.
 884  *
 885  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 886  * reasons such as page already existing or I/O error. Those reserved slots
 887  * should be freed by calling this function.
 888  */
 889 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 890 {
 891         atomic_long_add(npages, cli->cl_lru_left);
 892         wake_up(&osc_lru_waitq);
 893 }
 894
 895 /**
 896  * Atomic operations are expensive. We accumulate the accounting for the
 897  * same page zone to get better performance.
 898  * In practice this can work pretty good because the pages in the same RPC
 899  * are likely from the same page zone.
 900  */
 901 #ifdef HAVE_NR_UNSTABLE_NFS
 902 /* Old kernels use a separate counter for unstable pages,
 903  * newer kernels treat them like any other writeback.
 904  */
 905 #define NR_WRITEBACK NR_UNSTABLE_NFS
 906 #endif
 907
 908 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 909                                             struct osc_brw_async_args *aa,
 910                                             int factor)
 911 {
 912         int page_count;
 913         void *zone = NULL;
 914         int count = 0;
 915         int i;
 916
 917         if (desc != NULL) {
 918                 page_count = desc->bd_iov_count;
 919         } else {
 920                 page_count = aa->aa_page_count;
 921         }
 922
 923         for (i = 0; i < page_count; i++) {
 924                 void *pz;
 925                 if (desc)
 926                         pz = page_zone(desc->bd_vec[i].bv_page);
 927                 else
 928                         pz = page_zone(aa->aa_ppga[i]->pg);
 929
 930                 if (likely(pz == zone)) {
 931                         ++count;
 932                         continue;
 933                 }
 934
 935                 if (count > 0) {
 936                         mod_zone_page_state(zone, NR_WRITEBACK,
 937                                             factor * count);
 938                         count = 0;
 939                 }
 940                 zone = pz;
 941                 ++count;
 942         }
 943         if (count > 0)
 944                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 945 }
 946
 947 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 948                                                 struct osc_brw_async_args *aa)
 949 {
 950         unstable_page_accounting(desc, aa, 1);
 951 }
 952
 953 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 954                                                 struct osc_brw_async_args *aa)
 955 {
 956         unstable_page_accounting(desc, aa, -1);
 957 }
 958
 959 /**
 960  * Performs "unstable" page accounting. This function balances the
 961  * increment operations performed in osc_inc_unstable_pages. It is
 962  * registered as the RPC request callback, and is executed when the
 963  * bulk RPC is committed on the server. Thus at this point, the pages
 964  * involved in the bulk transfer are no longer considered unstable.
 965  *
 966  * If this function is called, the request should have been committed
 967  * or req:rq_unstable must have been set; it implies that the unstable
 968  * statistic have been added.
 969  */
 970 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 971 {
 972         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 973         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 974         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 975         int                      page_count;
 976         long                     unstable_count;
 977
 978         if (desc)
 979                 page_count = desc->bd_iov_count;
 980         else
 981                 page_count = aa->aa_page_count;
 982
 983         LASSERT(page_count >= 0);
 984
 985         dec_unstable_page_accounting(desc, aa);
 986
 987         unstable_count = atomic_long_sub_return(page_count,
 988                                                 &cli->cl_unstable_count);
 989         LASSERT(unstable_count >= 0);
 990
 991         unstable_count = atomic_long_sub_return(page_count,
 992                                            &cli->cl_cache->ccc_unstable_nr);
 993         LASSERT(unstable_count >= 0);
 994         if (unstable_count == 0)
 995                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 996
 997         if (waitqueue_active(&osc_lru_waitq))
 998                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 999 }
1000
1001 /**
1002  * "unstable" page accounting. See: osc_dec_unstable_pages.
1003  */
1004 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1005 {
1006         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1007         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1008         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1009         long                     page_count;
1010
1011         /* No unstable page tracking */
1012         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1013                 return;
1014
1015         if (desc)
1016                 page_count = desc->bd_iov_count;
1017         else
1018                 page_count = aa->aa_page_count;
1019
1020         add_unstable_page_accounting(desc, aa);
1021         atomic_long_add(page_count, &cli->cl_unstable_count);
1022         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1023
1024         /* If the request has already been committed (i.e. brw_commit
1025          * called via rq_commit_cb), we need to undo the unstable page
1026          * increments we just performed because rq_commit_cb wont be
1027          * called again. */
1028         spin_lock(&req->rq_lock);
1029         if (unlikely(req->rq_committed)) {
1030                 spin_unlock(&req->rq_lock);
1031
1032                 osc_dec_unstable_pages(req);
1033         } else {
1034                 req->rq_unstable = 1;
1035                 spin_unlock(&req->rq_lock);
1036         }
1037 }
1038
1039 /**
1040  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1041  * This function will be called by every BRW RPC so it's critical
1042  * to make this function fast.
1043  */
1044 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1045 {
1046         long unstable_nr, osc_unstable_count;
1047
1048         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1049         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1050                 return false;
1051
1052         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1053         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1054
1055         CDEBUG(D_CACHE,
1056                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1057                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1058
1059         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1060          * has one full RPC window of unstable pages, it's a good chance
1061          * to piggyback a SOFT_SYNC flag.
1062          * Please notice that the OST won't take immediate response for the
1063          * SOFT_SYNC request so active OSCs will have more chance to carry
1064          * the flag, this is reasonable. */
1065         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1066                osc_unstable_count > cli->cl_max_pages_per_rpc *
1067                                     cli->cl_max_rpcs_in_flight;
1068 }
1069
1070 /**
1071  * Return how many LRU pages in the cache of all OSC devices
1072  *
1073  * \retval      return # of cached LRU pages times reclaimation tendency
1074  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1075  */
1076 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1077                                      struct shrink_control *sc)
1078 {
1079         struct client_obd *cli;
1080         unsigned long cached = 0;
1081
1082         spin_lock(&osc_shrink_lock);
1083         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1084                 cached += atomic_long_read(&cli->cl_lru_in_list);
1085         spin_unlock(&osc_shrink_lock);
1086
1087         return (cached  * sysctl_vfs_cache_pressure) / 100;
1088 }
1089
1090 /**
1091  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1092  *
1093  * \retval      number of cached LRU pages reclaimed
1094  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1095  *
1096  * Linux kernel will loop calling this shrinker scan routine with
1097  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1098  *
1099  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1100  * to scan and try to reclaim LRU pages, just return 0 and
1101  * osc_cache_shrink_count() will report the LRU page number.
1102  */
1103 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1104                                     struct shrink_control *sc)
1105 {
1106         struct client_obd *cli;
1107         struct client_obd *stop_anchor = NULL;
1108         struct lu_env *env;
1109         long shrank = 0;
1110         int rc;
1111         __u16 refcheck;
1112
1113         if (sc->nr_to_scan == 0)
1114                 return 0;
1115
1116         if (!(sc->gfp_mask & __GFP_FS))
1117                 return SHRINK_STOP;
1118
1119         env = cl_env_get(&refcheck);
1120         if (IS_ERR(env))
1121                 return SHRINK_STOP;
1122
1123         spin_lock(&osc_shrink_lock);
1124         while (!list_empty(&osc_shrink_list)) {
1125                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1126                                  cl_shrink_list);
1127
1128                 if (stop_anchor == NULL)
1129                         stop_anchor = cli;
1130                 else if (cli == stop_anchor)
1131                         break;
1132
1133                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1134                 spin_unlock(&osc_shrink_lock);
1135
1136                 /* shrink no more than max_pages_per_rpc for an OSC */
1137                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1138                                     cli->cl_max_pages_per_rpc ?
1139                                     cli->cl_max_pages_per_rpc :
1140                                     sc->nr_to_scan - shrank, true);
1141                 if (rc > 0)
1142                         shrank += rc;
1143
1144                 if (shrank >= sc->nr_to_scan)
1145                         goto out;
1146
1147                 spin_lock(&osc_shrink_lock);
1148         }
1149         spin_unlock(&osc_shrink_lock);
1150
1151 out:
1152         cl_env_put(env, &refcheck);
1153
1154         return shrank;
1155 }
1156
1157 /** @} osc */