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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Page operations.
  54  */
  55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  56 {
  57         struct cl_page *page = opg->ops_cl.cpl_page;
  58
  59         LASSERT(!opg->ops_transfer_pinned);
  60         cl_page_get(page);
  61         lu_ref_add_atomic(&page->cp_reference, label, page);
  62         opg->ops_transfer_pinned = 1;
  63 }
  64
  65 static void osc_page_transfer_put(const struct lu_env *env,
  66                                   struct osc_page *opg)
  67 {
  68         struct cl_page *page = opg->ops_cl.cpl_page;
  69
  70         if (opg->ops_transfer_pinned) {
  71                 opg->ops_transfer_pinned = 0;
  72                 lu_ref_del(&page->cp_reference, "transfer", page);
  73                 cl_page_put(env, page);
  74         }
  75 }
  76
  77 /**
  78  * This is called once for every page when it is submitted for a transfer
  79  * either opportunistic (osc_page_cache_add()), or immediate
  80  * (osc_page_submit()).
  81  */
  82 static void osc_page_transfer_add(const struct lu_env *env,
  83                                   struct osc_page *opg, enum cl_req_type crt)
  84 {
  85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  86
  87         osc_lru_use(osc_cli(obj), opg);
  88 }
  89
  90 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  91                        struct cl_io *io, cl_commit_cbt cb)
  92 {
  93         int result;
  94         ENTRY;
  95
  96         osc_page_transfer_get(opg, "transfer\0cache");
  97         result = osc_queue_async_io(env, io, opg, cb);
  98         if (result != 0)
  99                 osc_page_transfer_put(env, opg);
 100         else
 101                 osc_page_transfer_add(env, opg, CRT_WRITE);
 102
 103         RETURN(result);
 104 }
 105
 106 void osc_index2policy(union ldlm_policy_data *policy,
 107                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 108 {
 109         memset(policy, 0, sizeof *policy);
 110         policy->l_extent.start = cl_offset(obj, start);
 111         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 112 }
 113
 114 static inline s64 osc_submit_duration(struct osc_page *opg)
 115 {
 116         if (ktime_to_ns(opg->ops_submit_time) == 0)
 117                 return 0;
 118
 119         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 120 }
 121
 122 static int osc_page_print(const struct lu_env *env,
 123                           const struct cl_page_slice *slice,
 124                           void *cookie, lu_printer_t printer)
 125 {
 126         struct osc_page *opg = cl2osc_page(slice);
 127         struct osc_async_page *oap = &opg->ops_oap;
 128         struct osc_object *obj = cl2osc(slice->cpl_obj);
 129         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 130
 131         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 132                           "1< %#x %d %c %c > "
 133                           "2< %lld %u %u %#x %#x | %p %p %p > "
 134                           "3< %d %lld %d > "
 135                           "4< %d %d %d %lu %c | %c %c %c %c > "
 136                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 137                           opg, osc_index(opg),
 138                           /* 1 */
 139                           oap->oap_magic, oap->oap_cmd,
 140                           list_empty_marker(&oap->oap_pending_item),
 141                           list_empty_marker(&oap->oap_rpc_item),
 142                           /* 2 */
 143                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 144                           oap->oap_async_flags, oap->oap_brw_flags,
 145                           oap->oap_request, oap->oap_cli, obj,
 146                           /* 3 */
 147                           opg->ops_transfer_pinned,
 148                           osc_submit_duration(opg), opg->ops_srvlock,
 149                           /* 4 */
 150                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 151                           cli->cl_max_rpcs_in_flight,
 152                           cli->cl_avail_grant,
 153                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 154                           list_empty_marker(&cli->cl_loi_ready_list),
 155                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 156                           list_empty_marker(&cli->cl_loi_write_list),
 157                           list_empty_marker(&cli->cl_loi_read_list),
 158                           /* 5 */
 159                           list_empty_marker(&obj->oo_ready_item),
 160                           list_empty_marker(&obj->oo_hp_ready_item),
 161                           list_empty_marker(&obj->oo_write_item),
 162                           list_empty_marker(&obj->oo_read_item),
 163                           atomic_read(&obj->oo_nr_reads),
 164                           list_empty_marker(&obj->oo_reading_exts),
 165                           atomic_read(&obj->oo_nr_writes),
 166                           list_empty_marker(&obj->oo_hp_exts),
 167                           list_empty_marker(&obj->oo_urgent_exts));
 168 }
 169
 170 static void osc_page_delete(const struct lu_env *env,
 171                             const struct cl_page_slice *slice)
 172 {
 173         struct osc_page   *opg = cl2osc_page(slice);
 174         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 175         int rc;
 176
 177         ENTRY;
 178         CDEBUG(D_TRACE, "%p\n", opg);
 179         osc_page_transfer_put(env, opg);
 180         rc = osc_teardown_async_page(env, obj, opg);
 181         if (rc) {
 182                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 183                               "Trying to teardown failed: %d\n", rc);
 184                 LASSERT(0);
 185         }
 186
 187         osc_lru_del(osc_cli(obj), opg);
 188
 189         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 190                 void *value = NULL;
 191
 192                 spin_lock(&obj->oo_tree_lock);
 193                 if (opg->ops_intree) {
 194                         value = radix_tree_delete(&obj->oo_tree,
 195                                                   osc_index(opg));
 196                         if (value != NULL) {
 197                                 --obj->oo_npages;
 198                                 opg->ops_intree = 0;
 199                         }
 200                 }
 201                 spin_unlock(&obj->oo_tree_lock);
 202
 203                 LASSERT(ergo(value != NULL, value == opg));
 204         }
 205
 206         EXIT;
 207 }
 208
 209 static void osc_page_clip(const struct lu_env *env,
 210                           const struct cl_page_slice *slice,
 211                           int from, int to)
 212 {
 213         struct osc_page       *opg = cl2osc_page(slice);
 214         struct osc_async_page *oap = &opg->ops_oap;
 215
 216         opg->ops_from = from;
 217         /* argument @to is exclusive, but @ops_to is inclusive */
 218         opg->ops_to   = to - 1;
 219         spin_lock(&oap->oap_lock);
 220         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 221         spin_unlock(&oap->oap_lock);
 222 }
 223
 224 static int osc_page_flush(const struct lu_env *env,
 225                           const struct cl_page_slice *slice,
 226                           struct cl_io *io)
 227 {
 228         struct osc_page *opg = cl2osc_page(slice);
 229         int rc = 0;
 230         ENTRY;
 231         rc = osc_flush_async_page(env, io, opg);
 232         RETURN(rc);
 233 }
 234
 235 static void osc_page_touch(const struct lu_env *env,
 236                           const struct cl_page_slice *slice, size_t to)
 237 {
 238         struct osc_page *opg = cl2osc_page(slice);
 239         struct cl_object *obj = opg->ops_cl.cpl_obj;
 240
 241         osc_page_touch_at(env, obj, osc_index(opg), to);
 242 }
 243
 244 static const struct cl_page_operations osc_page_ops = {
 245         .cpo_print         = osc_page_print,
 246         .cpo_delete        = osc_page_delete,
 247         .cpo_clip           = osc_page_clip,
 248         .cpo_flush          = osc_page_flush,
 249         .cpo_page_touch    = osc_page_touch,
 250 };
 251
 252 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 253                   struct cl_page *cl_page, pgoff_t index)
 254 {
 255         struct osc_object *osc = cl2osc(obj);
 256         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 257         struct osc_io *oio = osc_env_io(env);
 258         int result;
 259
 260         opg->ops_from = 0;
 261         opg->ops_to = PAGE_SIZE - 1;
 262
 263         INIT_LIST_HEAD(&opg->ops_lru);
 264
 265         result = osc_prep_async_page(osc, opg, cl_page->cp_vmpage,
 266                                      cl_offset(obj, index));
 267         if (result != 0)
 268                 return result;
 269
 270         opg->ops_srvlock = osc_io_srvlock(oio);
 271         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 272         cl_page->cp_osc_index = index;
 273
 274         /* reserve an LRU space for this page */
 275         if (cl_page->cp_type == CPT_CACHEABLE) {
 276                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 277                 if (result == 0) {
 278                         result = radix_tree_preload(GFP_NOFS);
 279                         if (result == 0) {
 280                                 spin_lock(&osc->oo_tree_lock);
 281                                 result = radix_tree_insert(&osc->oo_tree,
 282                                                            index, opg);
 283                                 if (result == 0) {
 284                                         ++osc->oo_npages;
 285                                         opg->ops_intree = 1;
 286                                 }
 287                                 spin_unlock(&osc->oo_tree_lock);
 288
 289                                 radix_tree_preload_end();
 290                         }
 291                 }
 292         }
 293
 294         return result;
 295 }
 296 EXPORT_SYMBOL(osc_page_init);
 297
 298 /**
 299  * Helper function called by osc_io_submit() for every page in an immediate
 300  * transfer (i.e., transferred synchronously).
 301  */
 302 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 303                      enum cl_req_type crt, int brw_flags)
 304 {
 305         struct osc_io *oio = osc_env_io(env);
 306         struct osc_async_page *oap = &opg->ops_oap;
 307
 308         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 309                  "magic 0x%x\n", oap, oap->oap_magic);
 310         LASSERT(oap->oap_async_flags & ASYNC_READY);
 311         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 312
 313         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 314         oap->oap_page_off = opg->ops_from;
 315         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 316         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 317
 318         if (oio->oi_cap_sys_resource) {
 319                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 320                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 321         }
 322
 323         opg->ops_submit_time = ktime_get();
 324         osc_page_transfer_get(opg, "transfer\0imm");
 325         osc_page_transfer_add(env, opg, crt);
 326 }
 327
 328 /* --------------- LRU page management ------------------ */
 329
 330 /* OSC is a natural place to manage LRU pages as applications are specialized
 331  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 332  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 333  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 334  * for free LRU slots - this will be very bad so the algorithm requires each
 335  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 336  * at any time.
 337  */
 338
 339 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 340
 341 /**
 342  * LRU pages are freed in batch mode. OSC should at least free this
 343  * number of pages to avoid running out of LRU slots.
 344  */
 345 static inline int lru_shrink_min(struct client_obd *cli)
 346 {
 347         return cli->cl_max_pages_per_rpc * 2;
 348 }
 349
 350 /**
 351  * free this number at most otherwise it will take too long time to finsih.
 352  */
 353 static inline int lru_shrink_max(struct client_obd *cli)
 354 {
 355         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 356 }
 357
 358 /**
 359  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 360  * we should free slots aggressively. In this way, slots are freed in a steady
 361  * step to maintain fairness among OSCs.
 362  *
 363  * Return how many LRU pages should be freed.
 364  */
 365 static int osc_cache_too_much(struct client_obd *cli)
 366 {
 367         struct cl_client_cache *cache = cli->cl_cache;
 368         long pages = atomic_long_read(&cli->cl_lru_in_list);
 369         unsigned long budget;
 370
 371         LASSERT(cache != NULL);
 372         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 373
 374         /* if it's going to run out LRU slots, we should free some, but not
 375          * too much to maintain faireness among OSCs. */
 376         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 377                 if (pages >= budget)
 378                         return lru_shrink_max(cli);
 379                 else if (pages >= budget / 2)
 380                         return lru_shrink_min(cli);
 381         } else {
 382                 time64_t duration = ktime_get_real_seconds();
 383                 long timediff;
 384
 385                 /* knock out pages by duration of no IO activity */
 386                 duration -= cli->cl_lru_last_used;
 387                 /*
 388                  * The difference shouldn't be more than 70 years
 389                  * so we can safely case to a long. Round to
 390                  * approximately 1 minute.
 391                  */
 392                 timediff = (long)(duration >> 6);
 393                 if (timediff > 0 && pages >= budget / timediff)
 394                         return lru_shrink_min(cli);
 395         }
 396         return 0;
 397 }
 398
 399 int lru_queue_work(const struct lu_env *env, void *data)
 400 {
 401         struct client_obd *cli = data;
 402         int count;
 403
 404         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 405         count = osc_cache_too_much(cli);
 406         if (count > 0) {
 407                 int rc = osc_lru_shrink(env, cli, count, false);
 408
 409                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 410                        cli_name(cli), rc, count);
 411                 if (rc >= count) {
 412                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 413                         ptlrpcd_queue_work(cli->cl_lru_work);
 414                 }
 415         }
 416
 417         RETURN(0);
 418 }
 419
 420 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 421 {
 422         LIST_HEAD(lru);
 423         struct osc_async_page *oap;
 424         long npages = 0;
 425
 426         list_for_each_entry(oap, plist, oap_pending_item) {
 427                 struct osc_page *opg = oap2osc_page(oap);
 428
 429                 if (!opg->ops_in_lru)
 430                         continue;
 431
 432                 ++npages;
 433                 LASSERT(list_empty(&opg->ops_lru));
 434                 list_add(&opg->ops_lru, &lru);
 435         }
 436
 437         if (npages > 0) {
 438                 spin_lock(&cli->cl_lru_list_lock);
 439                 list_splice_tail(&lru, &cli->cl_lru_list);
 440                 atomic_long_sub(npages, &cli->cl_lru_busy);
 441                 atomic_long_add(npages, &cli->cl_lru_in_list);
 442                 cli->cl_lru_last_used = ktime_get_real_seconds();
 443                 spin_unlock(&cli->cl_lru_list_lock);
 444
 445                 if (waitqueue_active(&osc_lru_waitq))
 446                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 447         }
 448 }
 449
 450 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 451 {
 452         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 453         list_del_init(&opg->ops_lru);
 454         atomic_long_dec(&cli->cl_lru_in_list);
 455 }
 456
 457 /**
 458  * Page is being destroyed. The page may be not in LRU list, if the transfer
 459  * has never finished(error occurred).
 460  */
 461 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 462 {
 463         if (opg->ops_in_lru) {
 464                 spin_lock(&cli->cl_lru_list_lock);
 465                 if (!list_empty(&opg->ops_lru)) {
 466                         __osc_lru_del(cli, opg);
 467                 } else {
 468                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 469                         atomic_long_dec(&cli->cl_lru_busy);
 470                 }
 471                 spin_unlock(&cli->cl_lru_list_lock);
 472
 473                 atomic_long_inc(cli->cl_lru_left);
 474                 /* this is a great place to release more LRU pages if
 475                  * this osc occupies too many LRU pages and kernel is
 476                  * stealing one of them. */
 477                 if (osc_cache_too_much(cli)) {
 478                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 479                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 480                 }
 481                 wake_up(&osc_lru_waitq);
 482         } else {
 483                 LASSERT(list_empty(&opg->ops_lru));
 484         }
 485 }
 486
 487 /**
 488  * Delete page from LRU list for redirty.
 489  */
 490 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 491 {
 492         /* If page is being transferred for the first time,
 493          * ops_lru should be empty */
 494         if (opg->ops_in_lru) {
 495                 if (list_empty(&opg->ops_lru))
 496                         return;
 497                 spin_lock(&cli->cl_lru_list_lock);
 498                 if (!list_empty(&opg->ops_lru)) {
 499                         __osc_lru_del(cli, opg);
 500                         atomic_long_inc(&cli->cl_lru_busy);
 501                 }
 502                 spin_unlock(&cli->cl_lru_list_lock);
 503         }
 504 }
 505
 506 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 507                                 struct cl_page **pvec, int max_index)
 508 {
 509         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 510         int i;
 511
 512         ll_pagevec_init(pagevec, 0);
 513         for (i = 0; i < max_index; i++) {
 514                 struct cl_page *page = pvec[i];
 515
 516                 LASSERT(cl_page_is_owned(page, io));
 517                 cl_page_delete(env, page);
 518                 cl_page_discard(env, io, page);
 519                 cl_page_disown(env, io, page);
 520                 cl_pagevec_put(env, page, pagevec);
 521
 522                 pvec[i] = NULL;
 523         }
 524         pagevec_release(pagevec);
 525 }
 526
 527 /**
 528  * Check if a cl_page can be released, i.e, it's not being used.
 529  *
 530  * If unstable account is turned on, bulk transfer may hold one refcount
 531  * for recovery so we need to check vmpage refcount as well; otherwise,
 532  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 533  */
 534 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 535 {
 536         if (cl_page_in_use_noref(page))
 537                 return true;
 538
 539         if (cli->cl_cache->ccc_unstable_check) {
 540                 struct page *vmpage = cl_page_vmpage(page);
 541
 542                 /* vmpage have two known users: cl_page and VM page cache */
 543                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 544                         return true;
 545         }
 546         return false;
 547 }
 548
 549 /**
 550  * Drop @target of pages from LRU at most.
 551  */
 552 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 553                    long target, bool force)
 554 {
 555         struct cl_io *io;
 556         struct cl_object *clobj = NULL;
 557         struct cl_page **pvec;
 558         struct osc_page *opg;
 559         long count = 0;
 560         int maxscan = 0;
 561         int index = 0;
 562         int rc = 0;
 563         ENTRY;
 564
 565         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 566         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 567                 RETURN(0);
 568
 569         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 570                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 571         if (!force) {
 572                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 573                         RETURN(-EBUSY);
 574
 575                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 576                         atomic_dec(&cli->cl_lru_shrinkers);
 577                         RETURN(-EBUSY);
 578                 }
 579         } else {
 580                 atomic_inc(&cli->cl_lru_shrinkers);
 581         }
 582
 583         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 584         io = osc_env_thread_io(env);
 585
 586         spin_lock(&cli->cl_lru_list_lock);
 587         if (force)
 588                 cli->cl_lru_reclaim++;
 589         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 590         while (!list_empty(&cli->cl_lru_list)) {
 591                 struct cl_page *page;
 592                 bool will_free = false;
 593
 594                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 595                         break;
 596
 597                 if (--maxscan < 0)
 598                         break;
 599
 600                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 601                                  ops_lru);
 602                 page = opg->ops_cl.cpl_page;
 603                 if (lru_page_busy(cli, page)) {
 604                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 605                         continue;
 606                 }
 607
 608                 LASSERT(page->cp_obj != NULL);
 609                 if (clobj != page->cp_obj) {
 610                         struct cl_object *tmp = page->cp_obj;
 611
 612                         cl_object_get(tmp);
 613                         spin_unlock(&cli->cl_lru_list_lock);
 614
 615                         if (clobj != NULL) {
 616                                 discard_pagevec(env, io, pvec, index);
 617                                 index = 0;
 618
 619                                 cl_io_fini(env, io);
 620                                 cl_object_put(env, clobj);
 621                                 clobj = NULL;
 622                         }
 623
 624                         clobj = tmp;
 625                         io->ci_obj = clobj;
 626                         io->ci_ignore_layout = 1;
 627                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 628
 629                         spin_lock(&cli->cl_lru_list_lock);
 630
 631                         if (rc != 0)
 632                                 break;
 633
 634                         ++maxscan;
 635                         continue;
 636                 }
 637
 638                 if (cl_page_own_try(env, io, page) == 0) {
 639                         if (!lru_page_busy(cli, page)) {
 640                                 /* remove it from lru list earlier to avoid
 641                                  * lock contention */
 642                                 __osc_lru_del(cli, opg);
 643                                 opg->ops_in_lru = 0; /* will be discarded */
 644
 645                                 cl_page_get(page);
 646                                 will_free = true;
 647                         } else {
 648                                 cl_page_disown(env, io, page);
 649                         }
 650                 }
 651
 652                 if (!will_free) {
 653                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 654                         continue;
 655                 }
 656
 657                 /* Don't discard and free the page with cl_lru_list held */
 658                 pvec[index++] = page;
 659                 if (unlikely(index == OTI_PVEC_SIZE)) {
 660                         spin_unlock(&cli->cl_lru_list_lock);
 661                         discard_pagevec(env, io, pvec, index);
 662                         index = 0;
 663
 664                         spin_lock(&cli->cl_lru_list_lock);
 665                 }
 666
 667                 if (++count >= target)
 668                         break;
 669         }
 670         spin_unlock(&cli->cl_lru_list_lock);
 671
 672         if (clobj != NULL) {
 673                 discard_pagevec(env, io, pvec, index);
 674
 675                 cl_io_fini(env, io);
 676                 cl_object_put(env, clobj);
 677         }
 678
 679         atomic_dec(&cli->cl_lru_shrinkers);
 680         if (count > 0) {
 681                 atomic_long_add(count, cli->cl_lru_left);
 682                 wake_up_all(&osc_lru_waitq);
 683         }
 684         RETURN(count > 0 ? count : rc);
 685 }
 686 EXPORT_SYMBOL(osc_lru_shrink);
 687
 688 /**
 689  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 690  * \@npages of LRU slots. For performance consideration, it's better to drop
 691  * LRU pages in batch. Therefore, the actual number is adjusted at least
 692  * max_pages_per_rpc.
 693  */
 694 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 695 {
 696         struct lu_env *env;
 697         struct cl_client_cache *cache = cli->cl_cache;
 698         int max_scans;
 699         __u16 refcheck;
 700         long rc = 0;
 701         ENTRY;
 702
 703         LASSERT(cache != NULL);
 704
 705         env = cl_env_get(&refcheck);
 706         if (IS_ERR(env))
 707                 RETURN(rc);
 708
 709         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 710         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 711                cli_name(cli), npages);
 712         rc = osc_lru_shrink(env, cli, npages, true);
 713         if (rc >= npages) {
 714                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 715                        cli_name(cli), rc, npages);
 716                 if (osc_cache_too_much(cli) > 0)
 717                         ptlrpcd_queue_work(cli->cl_lru_work);
 718                 GOTO(out, rc);
 719         } else if (rc > 0) {
 720                 npages -= rc;
 721         }
 722
 723         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 724                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 725                 atomic_long_read(&cli->cl_lru_busy), npages);
 726
 727         /* Reclaim LRU slots from other client_obd as it can't free enough
 728          * from its own. This should rarely happen. */
 729         spin_lock(&cache->ccc_lru_lock);
 730         LASSERT(!list_empty(&cache->ccc_lru));
 731
 732         cache->ccc_lru_shrinkers++;
 733         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 734
 735         max_scans = atomic_read(&cache->ccc_users) - 2;
 736         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 737                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 738                                  cl_lru_osc);
 739
 740                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 741                         cli_name(cli), cli,
 742                         atomic_long_read(&cli->cl_lru_in_list),
 743                         atomic_long_read(&cli->cl_lru_busy));
 744
 745                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 746                 if (osc_cache_too_much(cli) > 0) {
 747                         spin_unlock(&cache->ccc_lru_lock);
 748
 749                         rc = osc_lru_shrink(env, cli, npages, true);
 750                         spin_lock(&cache->ccc_lru_lock);
 751                         if (rc >= npages)
 752                                 break;
 753                         if (rc > 0)
 754                                 npages -= rc;
 755                 }
 756         }
 757         spin_unlock(&cache->ccc_lru_lock);
 758
 759 out:
 760         cl_env_put(env, &refcheck);
 761         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 762                 cli_name(cli), cli, rc);
 763         return rc;
 764 }
 765
 766 /**
 767  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 768  *
 769  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 770  * Only in the case that the LRU slots are in extreme shortage, it should
 771  * have reserved enough slots for an IO.
 772  */
 773 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 774                          struct osc_page *opg)
 775 {
 776         struct osc_io *oio = osc_env_io(env);
 777         int rc = 0;
 778
 779         ENTRY;
 780
 781         if (cli->cl_cache == NULL) /* shall not be in LRU */
 782                 RETURN(0);
 783
 784         if (oio->oi_lru_reserved > 0) {
 785                 --oio->oi_lru_reserved;
 786                 goto out;
 787         }
 788
 789         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 790         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 791                 /* run out of LRU spaces, try to drop some by itself */
 792                 rc = osc_lru_reclaim(cli, 1);
 793                 if (rc < 0)
 794                         break;
 795                 if (rc > 0)
 796                         continue;
 797                 /* IO issued by readahead, don't try hard */
 798                 if (oio->oi_is_readahead) {
 799                         if (atomic_long_read(cli->cl_lru_left) > 0)
 800                                 continue;
 801                         rc = -EBUSY;
 802                         break;
 803                 }
 804
 805                 cond_resched();
 806                 rc = l_wait_event_abortable(
 807                         osc_lru_waitq,
 808                         atomic_long_read(cli->cl_lru_left) > 0);
 809                 if (rc < 0) {
 810                         rc = -EINTR;
 811                         break;
 812                 }
 813         }
 814
 815 out:
 816         if (rc >= 0) {
 817                 atomic_long_inc(&cli->cl_lru_busy);
 818                 opg->ops_in_lru = 1;
 819                 rc = 0;
 820         }
 821
 822         RETURN(rc);
 823 }
 824
 825 /**
 826  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 827  *
 828  * The benefit of doing this is to reduce contention against atomic counter
 829  * cl_lru_left by changing it from per-page access to per-IO access.
 830  */
 831 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 832 {
 833         unsigned long reserved = 0;
 834         unsigned long max_pages;
 835         unsigned long c;
 836         int rc;
 837
 838 again:
 839         c = atomic_long_read(cli->cl_lru_left);
 840         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 841                 c = atomic_long_read(cli->cl_lru_left);
 842
 843         if (c < npages) {
 844                 /*
 845                  * Trigger writeback in the hope some LRU slot could
 846                  * be freed.
 847                  */
 848                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 849                 if (rc)
 850                         return 0;
 851         }
 852
 853         while (c >= npages) {
 854                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 855                         reserved = npages;
 856                         break;
 857                 }
 858                 c = atomic_long_read(cli->cl_lru_left);
 859         }
 860
 861         if (reserved != npages) {
 862                 cond_resched();
 863                 rc = l_wait_event_abortable(
 864                         osc_lru_waitq,
 865                         atomic_long_read(cli->cl_lru_left) > 0);
 866                 goto again;
 867         }
 868
 869         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 870         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 871                 /* If there aren't enough pages in the per-OSC LRU then
 872                  * wake up the LRU thread to try and clear out space, so
 873                  * we don't block if pages are being dirtied quickly. */
 874                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 875                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 876                        max_pages);
 877                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 878         }
 879
 880         return reserved;
 881 }
 882
 883 /**
 884  * osc_lru_unreserve() is called to unreserve LRU slots.
 885  *
 886  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 887  * reasons such as page already existing or I/O error. Those reserved slots
 888  * should be freed by calling this function.
 889  */
 890 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 891 {
 892         atomic_long_add(npages, cli->cl_lru_left);
 893         wake_up_all(&osc_lru_waitq);
 894 }
 895
 896 /**
 897  * Atomic operations are expensive. We accumulate the accounting for the
 898  * same page zone to get better performance.
 899  * In practice this can work pretty good because the pages in the same RPC
 900  * are likely from the same page zone.
 901  */
 902 #ifdef HAVE_NR_UNSTABLE_NFS
 903 /* Old kernels use a separate counter for unstable pages,
 904  * newer kernels treat them like any other writeback.
 905  */
 906 #define NR_WRITEBACK NR_UNSTABLE_NFS
 907 #endif
 908
 909 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 910                                             struct osc_brw_async_args *aa,
 911                                             int factor)
 912 {
 913         int page_count;
 914         void *zone = NULL;
 915         int count = 0;
 916         int i;
 917
 918         if (desc != NULL) {
 919                 page_count = desc->bd_iov_count;
 920         } else {
 921                 page_count = aa->aa_page_count;
 922         }
 923
 924         for (i = 0; i < page_count; i++) {
 925                 void *pz;
 926                 if (desc)
 927                         pz = page_zone(desc->bd_vec[i].bv_page);
 928                 else
 929                         pz = page_zone(aa->aa_ppga[i]->pg);
 930
 931                 if (likely(pz == zone)) {
 932                         ++count;
 933                         continue;
 934                 }
 935
 936                 if (count > 0) {
 937                         mod_zone_page_state(zone, NR_WRITEBACK,
 938                                             factor * count);
 939                         count = 0;
 940                 }
 941                 zone = pz;
 942                 ++count;
 943         }
 944         if (count > 0)
 945                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 946 }
 947
 948 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 949                                                 struct osc_brw_async_args *aa)
 950 {
 951         unstable_page_accounting(desc, aa, 1);
 952 }
 953
 954 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 955                                                 struct osc_brw_async_args *aa)
 956 {
 957         unstable_page_accounting(desc, aa, -1);
 958 }
 959
 960 /**
 961  * Performs "unstable" page accounting. This function balances the
 962  * increment operations performed in osc_inc_unstable_pages. It is
 963  * registered as the RPC request callback, and is executed when the
 964  * bulk RPC is committed on the server. Thus at this point, the pages
 965  * involved in the bulk transfer are no longer considered unstable.
 966  *
 967  * If this function is called, the request should have been committed
 968  * or req:rq_unstable must have been set; it implies that the unstable
 969  * statistic have been added.
 970  */
 971 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 972 {
 973         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 974         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 975         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 976         int                      page_count;
 977         long                     unstable_count;
 978
 979         if (desc)
 980                 page_count = desc->bd_iov_count;
 981         else
 982                 page_count = aa->aa_page_count;
 983
 984         LASSERT(page_count >= 0);
 985
 986         dec_unstable_page_accounting(desc, aa);
 987
 988         unstable_count = atomic_long_sub_return(page_count,
 989                                                 &cli->cl_unstable_count);
 990         LASSERT(unstable_count >= 0);
 991
 992         unstable_count = atomic_long_sub_return(page_count,
 993                                            &cli->cl_cache->ccc_unstable_nr);
 994         LASSERT(unstable_count >= 0);
 995         if (unstable_count == 0)
 996                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 997
 998         if (waitqueue_active(&osc_lru_waitq))
 999                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1000 }
1001
1002 /**
1003  * "unstable" page accounting. See: osc_dec_unstable_pages.
1004  */
1005 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1006 {
1007         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1008         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1009         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1010         long                     page_count;
1011
1012         /* No unstable page tracking */
1013         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1014                 return;
1015
1016         if (desc)
1017                 page_count = desc->bd_iov_count;
1018         else
1019                 page_count = aa->aa_page_count;
1020
1021         add_unstable_page_accounting(desc, aa);
1022         atomic_long_add(page_count, &cli->cl_unstable_count);
1023         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1024
1025         /* If the request has already been committed (i.e. brw_commit
1026          * called via rq_commit_cb), we need to undo the unstable page
1027          * increments we just performed because rq_commit_cb wont be
1028          * called again. */
1029         spin_lock(&req->rq_lock);
1030         if (unlikely(req->rq_committed)) {
1031                 spin_unlock(&req->rq_lock);
1032
1033                 osc_dec_unstable_pages(req);
1034         } else {
1035                 req->rq_unstable = 1;
1036                 spin_unlock(&req->rq_lock);
1037         }
1038 }
1039
1040 /**
1041  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1042  * This function will be called by every BRW RPC so it's critical
1043  * to make this function fast.
1044  */
1045 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1046 {
1047         long unstable_nr, osc_unstable_count;
1048
1049         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1050         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1051                 return false;
1052
1053         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1054         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1055
1056         CDEBUG(D_CACHE,
1057                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1058                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1059
1060         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1061          * has one full RPC window of unstable pages, it's a good chance
1062          * to piggyback a SOFT_SYNC flag.
1063          * Please notice that the OST won't take immediate response for the
1064          * SOFT_SYNC request so active OSCs will have more chance to carry
1065          * the flag, this is reasonable. */
1066         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1067                osc_unstable_count > cli->cl_max_pages_per_rpc *
1068                                     cli->cl_max_rpcs_in_flight;
1069 }
1070
1071 /**
1072  * Return how many LRU pages in the cache of all OSC devices
1073  *
1074  * \retval      return # of cached LRU pages times reclaimation tendency
1075  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1076  */
1077 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1078                                      struct shrink_control *sc)
1079 {
1080         struct client_obd *cli;
1081         unsigned long cached = 0;
1082
1083         spin_lock(&osc_shrink_lock);
1084         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1085                 cached += atomic_long_read(&cli->cl_lru_in_list);
1086         spin_unlock(&osc_shrink_lock);
1087
1088         return (cached  * sysctl_vfs_cache_pressure) / 100;
1089 }
1090
1091 /**
1092  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1093  *
1094  * \retval      number of cached LRU pages reclaimed
1095  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1096  *
1097  * Linux kernel will loop calling this shrinker scan routine with
1098  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1099  *
1100  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1101  * to scan and try to reclaim LRU pages, just return 0 and
1102  * osc_cache_shrink_count() will report the LRU page number.
1103  */
1104 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1105                                     struct shrink_control *sc)
1106 {
1107         struct client_obd *cli;
1108         struct client_obd *stop_anchor = NULL;
1109         struct lu_env *env;
1110         long shrank = 0;
1111         int rc;
1112         __u16 refcheck;
1113
1114         if (sc->nr_to_scan == 0)
1115                 return 0;
1116
1117         if (!(sc->gfp_mask & __GFP_FS))
1118                 return SHRINK_STOP;
1119
1120         env = cl_env_get(&refcheck);
1121         if (IS_ERR(env))
1122                 return SHRINK_STOP;
1123
1124         spin_lock(&osc_shrink_lock);
1125         while (!list_empty(&osc_shrink_list)) {
1126                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1127                                  cl_shrink_list);
1128
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 */