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