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         /* This isn't really necessary for transient pages, but we also don't
 219          * call clip on transient pages often, so it's OK.
 220          */
 221         spin_lock(&oap->oap_lock);
 222         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 223         spin_unlock(&oap->oap_lock);
 224 }
 225
 226 static int osc_page_flush(const struct lu_env *env,
 227                           const struct cl_page_slice *slice,
 228                           struct cl_io *io)
 229 {
 230         struct osc_page *opg = cl2osc_page(slice);
 231         int rc = 0;
 232         ENTRY;
 233         rc = osc_flush_async_page(env, io, opg);
 234         RETURN(rc);
 235 }
 236
 237 static void osc_page_touch(const struct lu_env *env,
 238                           const struct cl_page_slice *slice, size_t to)
 239 {
 240         struct osc_page *opg = cl2osc_page(slice);
 241         struct cl_object *obj = opg->ops_cl.cpl_obj;
 242
 243         osc_page_touch_at(env, obj, osc_index(opg), to);
 244 }
 245
 246 static const struct cl_page_operations osc_page_ops = {
 247         .cpo_print         = osc_page_print,
 248         .cpo_delete        = osc_page_delete,
 249         .cpo_clip           = osc_page_clip,
 250         .cpo_flush          = osc_page_flush,
 251         .cpo_page_touch    = osc_page_touch,
 252 };
 253
 254 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 255                   struct cl_page *cl_page, pgoff_t index)
 256 {
 257         struct osc_object *osc = cl2osc(obj);
 258         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 259         struct osc_io *oio = osc_env_io(env);
 260         int result;
 261
 262         opg->ops_from = 0;
 263         opg->ops_to = PAGE_SIZE - 1;
 264
 265         INIT_LIST_HEAD(&opg->ops_lru);
 266
 267         result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
 268         if (result != 0)
 269                 return result;
 270
 271         opg->ops_srvlock = osc_io_srvlock(oio);
 272         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 273         cl_page->cp_osc_index = index;
 274
 275         /* reserve an LRU space for this page */
 276         if (cl_page->cp_type == CPT_CACHEABLE) {
 277                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 278                 if (result == 0) {
 279                         result = radix_tree_preload(GFP_NOFS);
 280                         if (result == 0) {
 281                                 spin_lock(&osc->oo_tree_lock);
 282                                 result = radix_tree_insert(&osc->oo_tree,
 283                                                            index, opg);
 284                                 if (result == 0) {
 285                                         ++osc->oo_npages;
 286                                         opg->ops_intree = 1;
 287                                 }
 288                                 spin_unlock(&osc->oo_tree_lock);
 289
 290                                 radix_tree_preload_end();
 291                         }
 292                 }
 293         }
 294
 295         return result;
 296 }
 297 EXPORT_SYMBOL(osc_page_init);
 298
 299 /**
 300  * Helper function called by osc_io_submit() for every page in an immediate
 301  * transfer (i.e., transferred synchronously).
 302  */
 303 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 304                      enum cl_req_type crt, int brw_flags, ktime_t submit_time)
 305 {
 306         struct osc_io *oio = osc_env_io(env);
 307         struct osc_async_page *oap = &opg->ops_oap;
 308
 309         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 310                  "magic 0x%x\n", oap, oap->oap_magic);
 311         LASSERT(oap->oap_async_flags & ASYNC_READY);
 312         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 313
 314         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 315         oap->oap_page_off = opg->ops_from;
 316         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 317         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 318
 319         if (oio->oi_cap_sys_resource) {
 320                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 321                 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
 322         }
 323
 324         opg->ops_submit_time = submit_time;
 325         osc_page_transfer_get(opg, "transfer\0imm");
 326         osc_page_transfer_add(env, opg, crt);
 327 }
 328
 329 /* --------------- LRU page management ------------------ */
 330
 331 /* OSC is a natural place to manage LRU pages as applications are specialized
 332  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 333  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 334  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 335  * for free LRU slots - this will be very bad so the algorithm requires each
 336  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 337  * at any time.
 338  */
 339
 340 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 341
 342 /**
 343  * LRU pages are freed in batch mode. OSC should at least free this
 344  * number of pages to avoid running out of LRU slots.
 345  */
 346 static inline int lru_shrink_min(struct client_obd *cli)
 347 {
 348         return cli->cl_max_pages_per_rpc * 2;
 349 }
 350
 351 /**
 352  * free this number at most otherwise it will take too long time to finsih.
 353  */
 354 static inline int lru_shrink_max(struct client_obd *cli)
 355 {
 356         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 357 }
 358
 359 /**
 360  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 361  * we should free slots aggressively. In this way, slots are freed in a steady
 362  * step to maintain fairness among OSCs.
 363  *
 364  * Return how many LRU pages should be freed.
 365  */
 366 static int osc_cache_too_much(struct client_obd *cli)
 367 {
 368         struct cl_client_cache *cache = cli->cl_cache;
 369         long pages = atomic_long_read(&cli->cl_lru_in_list);
 370         unsigned long budget;
 371
 372         LASSERT(cache != NULL);
 373         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 374
 375         /* if it's going to run out LRU slots, we should free some, but not
 376          * too much to maintain faireness among OSCs. */
 377         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 378                 if (pages >= budget)
 379                         return lru_shrink_max(cli);
 380                 else if (pages >= budget / 2)
 381                         return lru_shrink_min(cli);
 382         } else {
 383                 time64_t duration = ktime_get_real_seconds();
 384                 long timediff;
 385
 386                 /* knock out pages by duration of no IO activity */
 387                 duration -= cli->cl_lru_last_used;
 388                 /*
 389                  * The difference shouldn't be more than 70 years
 390                  * so we can safely case to a long. Round to
 391                  * approximately 1 minute.
 392                  */
 393                 timediff = (long)(duration >> 6);
 394                 if (timediff > 0 && pages >= budget / timediff)
 395                         return lru_shrink_min(cli);
 396         }
 397         return 0;
 398 }
 399
 400 int lru_queue_work(const struct lu_env *env, void *data)
 401 {
 402         struct client_obd *cli = data;
 403         int count;
 404
 405         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 406         count = osc_cache_too_much(cli);
 407         if (count > 0) {
 408                 int rc = osc_lru_shrink(env, cli, count, false);
 409
 410                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 411                        cli_name(cli), rc, count);
 412                 if (rc >= count) {
 413                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 414                         ptlrpcd_queue_work(cli->cl_lru_work);
 415                 }
 416         }
 417
 418         RETURN(0);
 419 }
 420
 421 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 422 {
 423         LIST_HEAD(lru);
 424         struct osc_async_page *oap;
 425         long npages = 0;
 426
 427         list_for_each_entry(oap, plist, oap_pending_item) {
 428                 struct osc_page *opg = oap2osc_page(oap);
 429
 430                 if (!opg->ops_in_lru)
 431                         continue;
 432
 433                 ++npages;
 434                 LASSERT(list_empty(&opg->ops_lru));
 435                 list_add(&opg->ops_lru, &lru);
 436         }
 437
 438         if (npages > 0) {
 439                 spin_lock(&cli->cl_lru_list_lock);
 440                 list_splice_tail(&lru, &cli->cl_lru_list);
 441                 atomic_long_sub(npages, &cli->cl_lru_busy);
 442                 atomic_long_add(npages, &cli->cl_lru_in_list);
 443                 cli->cl_lru_last_used = ktime_get_real_seconds();
 444                 spin_unlock(&cli->cl_lru_list_lock);
 445
 446                 if (waitqueue_active(&osc_lru_waitq))
 447                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 448         }
 449 }
 450
 451 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 452 {
 453         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 454         list_del_init(&opg->ops_lru);
 455         atomic_long_dec(&cli->cl_lru_in_list);
 456 }
 457
 458 /**
 459  * Page is being destroyed. The page may be not in LRU list, if the transfer
 460  * has never finished(error occurred).
 461  */
 462 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 463 {
 464         if (opg->ops_in_lru) {
 465                 spin_lock(&cli->cl_lru_list_lock);
 466                 if (!list_empty(&opg->ops_lru)) {
 467                         __osc_lru_del(cli, opg);
 468                 } else {
 469                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 470                         atomic_long_dec(&cli->cl_lru_busy);
 471                 }
 472                 spin_unlock(&cli->cl_lru_list_lock);
 473
 474                 atomic_long_inc(cli->cl_lru_left);
 475                 /* this is a great place to release more LRU pages if
 476                  * this osc occupies too many LRU pages and kernel is
 477                  * stealing one of them. */
 478                 if (osc_cache_too_much(cli)) {
 479                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 480                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 481                 }
 482                 wake_up(&osc_lru_waitq);
 483         } else {
 484                 LASSERT(list_empty(&opg->ops_lru));
 485         }
 486 }
 487
 488 /**
 489  * Delete page from LRU list for redirty.
 490  */
 491 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 492 {
 493         /* If page is being transferred for the first time,
 494          * ops_lru should be empty */
 495         if (opg->ops_in_lru) {
 496                 if (list_empty(&opg->ops_lru))
 497                         return;
 498                 spin_lock(&cli->cl_lru_list_lock);
 499                 if (!list_empty(&opg->ops_lru)) {
 500                         __osc_lru_del(cli, opg);
 501                         atomic_long_inc(&cli->cl_lru_busy);
 502                 }
 503                 spin_unlock(&cli->cl_lru_list_lock);
 504         }
 505 }
 506
 507 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 508                                 struct cl_page **pvec, int max_index)
 509 {
 510         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 511         int i;
 512
 513         ll_pagevec_init(pagevec, 0);
 514         for (i = 0; i < max_index; i++) {
 515                 struct cl_page *page = pvec[i];
 516
 517                 LASSERT(cl_page_is_owned(page, io));
 518                 cl_page_delete(env, page);
 519                 cl_page_discard(env, io, page);
 520                 cl_page_disown(env, io, page);
 521                 cl_pagevec_put(env, page, pagevec);
 522
 523                 pvec[i] = NULL;
 524         }
 525         pagevec_release(pagevec);
 526 }
 527
 528 /**
 529  * Check if a cl_page can be released, i.e, it's not being used.
 530  *
 531  * If unstable account is turned on, bulk transfer may hold one refcount
 532  * for recovery so we need to check vmpage refcount as well; otherwise,
 533  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 534  */
 535 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 536 {
 537         if (cl_page_in_use_noref(page))
 538                 return true;
 539
 540         if (cli->cl_cache->ccc_unstable_check) {
 541                 struct page *vmpage = cl_page_vmpage(page);
 542
 543                 /* vmpage have two known users: cl_page and VM page cache */
 544                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 545                         return true;
 546         }
 547         return false;
 548 }
 549
 550 /**
 551  * Drop @target of pages from LRU at most.
 552  */
 553 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 554                    long target, bool force)
 555 {
 556         struct cl_io *io;
 557         struct cl_object *clobj = NULL;
 558         struct cl_page **pvec;
 559         struct osc_page *opg;
 560         long count = 0;
 561         int maxscan = 0;
 562         int index = 0;
 563         int rc = 0;
 564         ENTRY;
 565
 566         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 567         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 568                 RETURN(0);
 569
 570         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 571                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 572         if (!force) {
 573                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 574                         RETURN(-EBUSY);
 575
 576                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 577                         atomic_dec(&cli->cl_lru_shrinkers);
 578                         RETURN(-EBUSY);
 579                 }
 580         } else {
 581                 atomic_inc(&cli->cl_lru_shrinkers);
 582         }
 583
 584         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 585         io = osc_env_thread_io(env);
 586
 587         spin_lock(&cli->cl_lru_list_lock);
 588         if (force)
 589                 cli->cl_lru_reclaim++;
 590         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 591         while (!list_empty(&cli->cl_lru_list)) {
 592                 struct cl_page *page;
 593                 bool will_free = false;
 594
 595                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 596                         break;
 597
 598                 if (--maxscan < 0)
 599                         break;
 600
 601                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 602                                        ops_lru);
 603                 page = opg->ops_cl.cpl_page;
 604                 if (lru_page_busy(cli, page)) {
 605                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 606                         continue;
 607                 }
 608
 609                 LASSERT(page->cp_obj != NULL);
 610                 if (clobj != page->cp_obj) {
 611                         struct cl_object *tmp = page->cp_obj;
 612
 613                         cl_object_get(tmp);
 614                         spin_unlock(&cli->cl_lru_list_lock);
 615
 616                         if (clobj != NULL) {
 617                                 discard_pagevec(env, io, pvec, index);
 618                                 index = 0;
 619
 620                                 cl_io_fini(env, io);
 621                                 cl_object_put(env, clobj);
 622                                 clobj = NULL;
 623                         }
 624
 625                         clobj = tmp;
 626                         io->ci_obj = clobj;
 627                         io->ci_ignore_layout = 1;
 628                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 629
 630                         spin_lock(&cli->cl_lru_list_lock);
 631
 632                         if (rc != 0)
 633                                 break;
 634
 635                         ++maxscan;
 636                         continue;
 637                 }
 638
 639                 if (cl_page_own_try(env, io, page) == 0) {
 640                         if (!lru_page_busy(cli, page)) {
 641                                 /* remove it from lru list earlier to avoid
 642                                  * lock contention */
 643                                 __osc_lru_del(cli, opg);
 644                                 opg->ops_in_lru = 0; /* will be discarded */
 645
 646                                 cl_page_get(page);
 647                                 will_free = true;
 648                         } else {
 649                                 cl_page_disown(env, io, page);
 650                         }
 651                 }
 652
 653                 if (!will_free) {
 654                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 655                         continue;
 656                 }
 657
 658                 /* Don't discard and free the page with cl_lru_list held */
 659                 pvec[index++] = page;
 660                 if (unlikely(index == OTI_PVEC_SIZE)) {
 661                         spin_unlock(&cli->cl_lru_list_lock);
 662                         discard_pagevec(env, io, pvec, index);
 663                         index = 0;
 664
 665                         spin_lock(&cli->cl_lru_list_lock);
 666                 }
 667
 668                 if (++count >= target)
 669                         break;
 670         }
 671         spin_unlock(&cli->cl_lru_list_lock);
 672
 673         if (clobj != NULL) {
 674                 discard_pagevec(env, io, pvec, index);
 675
 676                 cl_io_fini(env, io);
 677                 cl_object_put(env, clobj);
 678         }
 679
 680         atomic_dec(&cli->cl_lru_shrinkers);
 681         if (count > 0) {
 682                 atomic_long_add(count, cli->cl_lru_left);
 683                 wake_up(&osc_lru_waitq);
 684         }
 685         RETURN(count > 0 ? count : rc);
 686 }
 687 EXPORT_SYMBOL(osc_lru_shrink);
 688
 689 /**
 690  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 691  * \@npages of LRU slots. For performance consideration, it's better to drop
 692  * LRU pages in batch. Therefore, the actual number is adjusted at least
 693  * max_pages_per_rpc.
 694  */
 695 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 696 {
 697         struct lu_env *env;
 698         struct cl_client_cache *cache = cli->cl_cache;
 699         int max_scans;
 700         __u16 refcheck;
 701         long rc = 0;
 702         ENTRY;
 703
 704         LASSERT(cache != NULL);
 705
 706         env = cl_env_get(&refcheck);
 707         if (IS_ERR(env))
 708                 RETURN(rc);
 709
 710         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 711         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 712                cli_name(cli), npages);
 713         rc = osc_lru_shrink(env, cli, npages, true);
 714         if (rc >= npages) {
 715                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 716                        cli_name(cli), rc, npages);
 717                 if (osc_cache_too_much(cli) > 0)
 718                         ptlrpcd_queue_work(cli->cl_lru_work);
 719                 GOTO(out, rc);
 720         } else if (rc > 0) {
 721                 npages -= rc;
 722         }
 723
 724         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 725                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 726                 atomic_long_read(&cli->cl_lru_busy), npages);
 727
 728         /* Reclaim LRU slots from other client_obd as it can't free enough
 729          * from its own. This should rarely happen. */
 730         spin_lock(&cache->ccc_lru_lock);
 731         LASSERT(!list_empty(&cache->ccc_lru));
 732
 733         cache->ccc_lru_shrinkers++;
 734         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 735
 736         max_scans = atomic_read(&cache->ccc_users) - 2;
 737         while (--max_scans > 0 &&
 738                (cli = list_first_entry_or_null(&cache->ccc_lru,
 739                                                struct client_obd,
 740                                                cl_lru_osc)) != NULL) {
 741                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 742                         cli_name(cli), cli,
 743                         atomic_long_read(&cli->cl_lru_in_list),
 744                         atomic_long_read(&cli->cl_lru_busy));
 745
 746                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 747                 if (osc_cache_too_much(cli) > 0) {
 748                         spin_unlock(&cache->ccc_lru_lock);
 749
 750                         rc = osc_lru_shrink(env, cli, npages, true);
 751                         spin_lock(&cache->ccc_lru_lock);
 752                         if (rc >= npages)
 753                                 break;
 754                         if (rc > 0)
 755                                 npages -= rc;
 756                 }
 757         }
 758         spin_unlock(&cache->ccc_lru_lock);
 759
 760 out:
 761         cl_env_put(env, &refcheck);
 762         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 763                 cli_name(cli), cli, rc);
 764         return rc;
 765 }
 766
 767 /**
 768  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 769  *
 770  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 771  * Only in the case that the LRU slots are in extreme shortage, it should
 772  * have reserved enough slots for an IO.
 773  */
 774 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 775                          struct osc_page *opg)
 776 {
 777         struct osc_io *oio = osc_env_io(env);
 778         int rc = 0;
 779
 780         ENTRY;
 781
 782         if (cli->cl_cache == NULL) /* shall not be in LRU */
 783                 RETURN(0);
 784
 785         if (oio->oi_lru_reserved > 0) {
 786                 --oio->oi_lru_reserved;
 787                 goto out;
 788         }
 789
 790         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 791         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 792                 /* run out of LRU spaces, try to drop some by itself */
 793                 rc = osc_lru_reclaim(cli, 1);
 794                 if (rc < 0)
 795                         break;
 796                 if (rc > 0)
 797                         continue;
 798                 /* IO issued by readahead, don't try hard */
 799                 if (oio->oi_is_readahead) {
 800                         if (atomic_long_read(cli->cl_lru_left) > 0)
 801                                 continue;
 802                         rc = -EBUSY;
 803                         break;
 804                 }
 805
 806                 cond_resched();
 807                 rc = l_wait_event_abortable(
 808                         osc_lru_waitq,
 809                         atomic_long_read(cli->cl_lru_left) > 0);
 810                 if (rc < 0) {
 811                         rc = -EINTR;
 812                         break;
 813                 }
 814         }
 815
 816 out:
 817         if (rc >= 0) {
 818                 atomic_long_inc(&cli->cl_lru_busy);
 819                 opg->ops_in_lru = 1;
 820                 rc = 0;
 821         }
 822
 823         RETURN(rc);
 824 }
 825
 826 /**
 827  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 828  *
 829  * The benefit of doing this is to reduce contention against atomic counter
 830  * cl_lru_left by changing it from per-page access to per-IO access.
 831  */
 832 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 833 {
 834         unsigned long reserved = 0;
 835         unsigned long max_pages;
 836         unsigned long c;
 837         int rc;
 838
 839 again:
 840         c = atomic_long_read(cli->cl_lru_left);
 841         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 842                 c = atomic_long_read(cli->cl_lru_left);
 843
 844         if (c < npages) {
 845                 /*
 846                  * Trigger writeback in the hope some LRU slot could
 847                  * be freed.
 848                  */
 849                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 850                 if (rc)
 851                         return 0;
 852         }
 853
 854         while (c >= npages) {
 855                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 856                         reserved = npages;
 857                         break;
 858                 }
 859                 c = atomic_long_read(cli->cl_lru_left);
 860         }
 861
 862         if (reserved != npages) {
 863                 cond_resched();
 864                 rc = l_wait_event_abortable(
 865                         osc_lru_waitq,
 866                         atomic_long_read(cli->cl_lru_left) > 0);
 867                 goto again;
 868         }
 869
 870         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 871         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 872                 /* If there aren't enough pages in the per-OSC LRU then
 873                  * wake up the LRU thread to try and clear out space, so
 874                  * we don't block if pages are being dirtied quickly. */
 875                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 876                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 877                        max_pages);
 878                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 879         }
 880
 881         return reserved;
 882 }
 883
 884 /**
 885  * osc_lru_unreserve() is called to unreserve LRU slots.
 886  *
 887  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 888  * reasons such as page already existing or I/O error. Those reserved slots
 889  * should be freed by calling this function.
 890  */
 891 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 892 {
 893         atomic_long_add(npages, cli->cl_lru_left);
 894         wake_up(&osc_lru_waitq);
 895 }
 896
 897 /**
 898  * Atomic operations are expensive. We accumulate the accounting for the
 899  * same page zone to get better performance.
 900  * In practice this can work pretty good because the pages in the same RPC
 901  * are likely from the same page zone.
 902  */
 903 #ifdef HAVE_NR_UNSTABLE_NFS
 904 /* Old kernels use a separate counter for unstable pages,
 905  * newer kernels treat them like any other writeback.
 906  */
 907 #define NR_WRITEBACK NR_UNSTABLE_NFS
 908 #endif
 909
 910 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 911                                             struct osc_brw_async_args *aa,
 912                                             int factor)
 913 {
 914         int page_count;
 915         void *zone = NULL;
 916         int count = 0;
 917         int i;
 918
 919         if (desc != NULL) {
 920                 page_count = desc->bd_iov_count;
 921         } else {
 922                 page_count = aa->aa_page_count;
 923         }
 924
 925         for (i = 0; i < page_count; i++) {
 926                 void *pz;
 927                 if (desc)
 928                         pz = page_zone(desc->bd_vec[i].bv_page);
 929                 else
 930                         pz = page_zone(aa->aa_ppga[i]->pg);
 931
 932                 if (likely(pz == zone)) {
 933                         ++count;
 934                         continue;
 935                 }
 936
 937                 if (count > 0) {
 938                         mod_zone_page_state(zone, NR_WRITEBACK,
 939                                             factor * count);
 940                         count = 0;
 941                 }
 942                 zone = pz;
 943                 ++count;
 944         }
 945         if (count > 0)
 946                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 947 }
 948
 949 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 950                                                 struct osc_brw_async_args *aa)
 951 {
 952         unstable_page_accounting(desc, aa, 1);
 953 }
 954
 955 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 956                                                 struct osc_brw_async_args *aa)
 957 {
 958         unstable_page_accounting(desc, aa, -1);
 959 }
 960
 961 /**
 962  * Performs "unstable" page accounting. This function balances the
 963  * increment operations performed in osc_inc_unstable_pages. It is
 964  * registered as the RPC request callback, and is executed when the
 965  * bulk RPC is committed on the server. Thus at this point, the pages
 966  * involved in the bulk transfer are no longer considered unstable.
 967  *
 968  * If this function is called, the request should have been committed
 969  * or req:rq_unstable must have been set; it implies that the unstable
 970  * statistic have been added.
 971  */
 972 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 973 {
 974         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 975         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 976         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 977         int                      page_count;
 978         long                     unstable_count;
 979
 980         if (desc)
 981                 page_count = desc->bd_iov_count;
 982         else
 983                 page_count = aa->aa_page_count;
 984
 985         LASSERT(page_count >= 0);
 986
 987         dec_unstable_page_accounting(desc, aa);
 988
 989         unstable_count = atomic_long_sub_return(page_count,
 990                                                 &cli->cl_unstable_count);
 991         LASSERT(unstable_count >= 0);
 992
 993         unstable_count = atomic_long_sub_return(page_count,
 994                                            &cli->cl_cache->ccc_unstable_nr);
 995         LASSERT(unstable_count >= 0);
 996         if (unstable_count == 0)
 997                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 998
 999         if (waitqueue_active(&osc_lru_waitq))
1000                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1001 }
1002
1003 /**
1004  * "unstable" page accounting. See: osc_dec_unstable_pages.
1005  */
1006 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1007 {
1008         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1009         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1010         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1011         long                     page_count;
1012
1013         /* No unstable page tracking */
1014         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1015                 return;
1016
1017         if (desc)
1018                 page_count = desc->bd_iov_count;
1019         else
1020                 page_count = aa->aa_page_count;
1021
1022         add_unstable_page_accounting(desc, aa);
1023         atomic_long_add(page_count, &cli->cl_unstable_count);
1024         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1025
1026         /* If the request has already been committed (i.e. brw_commit
1027          * called via rq_commit_cb), we need to undo the unstable page
1028          * increments we just performed because rq_commit_cb wont be
1029          * called again. */
1030         spin_lock(&req->rq_lock);
1031         if (unlikely(req->rq_committed)) {
1032                 spin_unlock(&req->rq_lock);
1033
1034                 osc_dec_unstable_pages(req);
1035         } else {
1036                 req->rq_unstable = 1;
1037                 spin_unlock(&req->rq_lock);
1038         }
1039 }
1040
1041 /**
1042  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1043  * This function will be called by every BRW RPC so it's critical
1044  * to make this function fast.
1045  */
1046 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1047 {
1048         long unstable_nr, osc_unstable_count;
1049
1050         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1051         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1052                 return false;
1053
1054         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1055         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1056
1057         CDEBUG(D_CACHE,
1058                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1059                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1060
1061         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1062          * has one full RPC window of unstable pages, it's a good chance
1063          * to piggyback a SOFT_SYNC flag.
1064          * Please notice that the OST won't take immediate response for the
1065          * SOFT_SYNC request so active OSCs will have more chance to carry
1066          * the flag, this is reasonable. */
1067         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1068                osc_unstable_count > cli->cl_max_pages_per_rpc *
1069                                     cli->cl_max_rpcs_in_flight;
1070 }
1071
1072 /**
1073  * Return how many LRU pages in the cache of all OSC devices
1074  *
1075  * \retval      return # of cached LRU pages times reclaimation tendency
1076  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1077  */
1078 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1079                                      struct shrink_control *sc)
1080 {
1081         struct client_obd *cli;
1082         unsigned long cached = 0;
1083
1084         spin_lock(&osc_shrink_lock);
1085         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1086                 cached += atomic_long_read(&cli->cl_lru_in_list);
1087         spin_unlock(&osc_shrink_lock);
1088
1089         return (cached  * sysctl_vfs_cache_pressure) / 100;
1090 }
1091
1092 /**
1093  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1094  *
1095  * \retval      number of cached LRU pages reclaimed
1096  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1097  *
1098  * Linux kernel will loop calling this shrinker scan routine with
1099  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1100  *
1101  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1102  * to scan and try to reclaim LRU pages, just return 0 and
1103  * osc_cache_shrink_count() will report the LRU page number.
1104  */
1105 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1106                                     struct shrink_control *sc)
1107 {
1108         struct client_obd *cli;
1109         struct client_obd *stop_anchor = NULL;
1110         struct lu_env *env;
1111         long shrank = 0;
1112         int rc;
1113         __u16 refcheck;
1114
1115         if (sc->nr_to_scan == 0)
1116                 return 0;
1117
1118         if (!(sc->gfp_mask & __GFP_FS))
1119                 return SHRINK_STOP;
1120
1121         env = cl_env_get(&refcheck);
1122         if (IS_ERR(env))
1123                 return SHRINK_STOP;
1124
1125         spin_lock(&osc_shrink_lock);
1126         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1127                                                struct client_obd,
1128                                                cl_shrink_list)) != NULL) {
1129                 if (stop_anchor == NULL)
1130                         stop_anchor = cli;
1131                 else if (cli == stop_anchor)
1132                         break;
1133
1134                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1135                 spin_unlock(&osc_shrink_lock);
1136
1137                 /* shrink no more than max_pages_per_rpc for an OSC */
1138                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1139                                     cli->cl_max_pages_per_rpc ?
1140                                     cli->cl_max_pages_per_rpc :
1141                                     sc->nr_to_scan - shrank, true);
1142                 if (rc > 0)
1143                         shrank += rc;
1144
1145                 if (shrank >= sc->nr_to_scan)
1146                         goto out;
1147
1148                 spin_lock(&osc_shrink_lock);
1149         }
1150         spin_unlock(&osc_shrink_lock);
1151
1152 out:
1153         cl_env_put(env, &refcheck);
1154
1155         return shrank;
1156 }
1157
1158 /** @} osc */