lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * Implementation of cl_page for OSC layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38 #include <lustre_osc.h>
  39
  40 #include "osc_internal.h"
  41
  42 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  43 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  44 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  45                          struct osc_page *opg);
  46
  47 /** \addtogroup osc
  48  *  @{
  49  */
  50
  51 /*
  52  * Page operations.
  53  */
  54 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  55 {
  56         struct cl_page *page = opg->ops_cl.cpl_page;
  57
  58         LASSERT(!opg->ops_transfer_pinned);
  59         cl_page_get(page);
  60         lu_ref_add_atomic(&page->cp_reference, label, page);
  61         opg->ops_transfer_pinned = 1;
  62 }
  63
  64 static void osc_page_transfer_put(const struct lu_env *env,
  65                                   struct osc_page *opg)
  66 {
  67         struct cl_page *page = opg->ops_cl.cpl_page;
  68
  69         if (opg->ops_transfer_pinned) {
  70                 opg->ops_transfer_pinned = 0;
  71                 lu_ref_del(&page->cp_reference, "transfer", page);
  72                 cl_page_put(env, page);
  73         }
  74 }
  75
  76 /**
  77  * This is called once for every page when it is submitted for a transfer
  78  * either opportunistic (osc_page_cache_add()), or immediate
  79  * (osc_page_submit()).
  80  */
  81 static void osc_page_transfer_add(const struct lu_env *env,
  82                                   struct osc_page *opg, enum cl_req_type crt)
  83 {
  84         struct osc_object *obj = osc_page_object(opg);
  85
  86         osc_lru_use(osc_cli(obj), opg);
  87 }
  88
  89 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  90                        struct cl_io *io, cl_commit_cbt cb)
  91 {
  92         int result;
  93         ENTRY;
  94
  95         osc_page_transfer_get(opg, "transfer\0cache");
  96         result = osc_queue_async_io(env, io, opg, cb);
  97         if (result != 0)
  98                 osc_page_transfer_put(env, opg);
  99         else
 100                 osc_page_transfer_add(env, opg, CRT_WRITE);
 101
 102         RETURN(result);
 103 }
 104
 105 void osc_index2policy(union ldlm_policy_data *policy,
 106                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 107 {
 108         memset(policy, 0, sizeof *policy);
 109         policy->l_extent.start = start << PAGE_SHIFT;
 110         policy->l_extent.end = ((end + 1) << PAGE_SHIFT) - 1;
 111 }
 112
 113 static int osc_page_print(const struct lu_env *env,
 114                           const struct cl_page_slice *slice,
 115                           void *cookie, lu_printer_t printer)
 116 {
 117         struct osc_page *opg = cl2osc_page(slice);
 118         struct osc_async_page *oap = &opg->ops_oap;
 119         struct osc_object *obj = osc_page_object(opg);
 120         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 121
 122         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 123                           "1< %d %c %c > "
 124                           "2< %lld %u %u %#x %#x | %p %p %p > "
 125                           "3< %d %d > "
 126                           "4< %d %d %d %lu %c | %c %c %c %c > "
 127                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 128                           opg, osc_index(opg),
 129                           /* 1 */
 130                           oap->oap_cmd,
 131                           list_empty_marker(&oap->oap_pending_item),
 132                           list_empty_marker(&oap->oap_rpc_item),
 133                           /* 2 */
 134                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 135                           oap->oap_async_flags, oap->oap_brw_flags,
 136                           oap->oap_request, cli, obj,
 137                           /* 3 */
 138                           opg->ops_transfer_pinned,
 139                           opg->ops_srvlock,
 140                           /* 4 */
 141                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 142                           cli->cl_max_rpcs_in_flight,
 143                           cli->cl_avail_grant,
 144                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 145                           list_empty_marker(&cli->cl_loi_ready_list),
 146                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 147                           list_empty_marker(&cli->cl_loi_write_list),
 148                           list_empty_marker(&cli->cl_loi_read_list),
 149                           /* 5 */
 150                           list_empty_marker(&obj->oo_ready_item),
 151                           list_empty_marker(&obj->oo_hp_ready_item),
 152                           list_empty_marker(&obj->oo_write_item),
 153                           list_empty_marker(&obj->oo_read_item),
 154                           atomic_read(&obj->oo_nr_reads),
 155                           list_empty_marker(&obj->oo_reading_exts),
 156                           atomic_read(&obj->oo_nr_writes),
 157                           list_empty_marker(&obj->oo_hp_exts),
 158                           list_empty_marker(&obj->oo_urgent_exts));
 159 }
 160
 161 static void osc_page_delete(const struct lu_env *env,
 162                             const struct cl_page_slice *slice)
 163 {
 164         struct osc_page   *opg = cl2osc_page(slice);
 165         struct osc_object *obj = osc_page_object(opg);
 166         int rc;
 167
 168         ENTRY;
 169         CDEBUG(D_TRACE, "%p\n", opg);
 170         osc_page_transfer_put(env, opg);
 171         rc = osc_teardown_async_page(env, obj, opg);
 172         if (rc) {
 173                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 174                               "Trying to teardown failed: %d\n", rc);
 175                 LASSERT(0);
 176         }
 177
 178         osc_lru_del(osc_cli(obj), opg);
 179
 180         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 181                 void *value = NULL;
 182
 183                 spin_lock(&obj->oo_tree_lock);
 184                 if (opg->ops_intree) {
 185                         value = radix_tree_delete(&obj->oo_tree,
 186                                                   osc_index(opg));
 187                         if (value != NULL) {
 188                                 --obj->oo_npages;
 189                                 opg->ops_intree = 0;
 190                         }
 191                 }
 192                 spin_unlock(&obj->oo_tree_lock);
 193
 194                 LASSERT(ergo(value != NULL, value == opg));
 195         }
 196
 197         EXIT;
 198 }
 199
 200 static void osc_page_clip(const struct lu_env *env,
 201                           const struct cl_page_slice *slice,
 202                           int from, int to)
 203 {
 204         struct osc_page       *opg = cl2osc_page(slice);
 205         struct osc_async_page *oap = &opg->ops_oap;
 206
 207         CDEBUG(D_CACHE, "from %d, to %d\n", from, to);
 208
 209         opg->ops_from = from;
 210         /* argument @to is exclusive, but @ops_to is inclusive */
 211         opg->ops_to   = to - 1;
 212         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 213 }
 214
 215 static int osc_page_flush(const struct lu_env *env,
 216                           const struct cl_page_slice *slice,
 217                           struct cl_io *io)
 218 {
 219         struct osc_page *opg = cl2osc_page(slice);
 220         int rc = 0;
 221         ENTRY;
 222         rc = osc_flush_async_page(env, io, opg);
 223         RETURN(rc);
 224 }
 225
 226 static void osc_page_touch(const struct lu_env *env,
 227                           const struct cl_page_slice *slice, size_t to)
 228 {
 229         struct osc_page *opg = cl2osc_page(slice);
 230         struct cl_object *obj = osc2cl(osc_page_object(opg));
 231
 232         osc_page_touch_at(env, obj, osc_index(opg), to);
 233 }
 234
 235 static const struct cl_page_operations osc_page_ops = {
 236         .cpo_print         = osc_page_print,
 237         .cpo_delete        = osc_page_delete,
 238         .cpo_clip           = osc_page_clip,
 239         .cpo_flush          = osc_page_flush,
 240         .cpo_page_touch    = osc_page_touch,
 241 };
 242
 243 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 244                   struct cl_page *cl_page, pgoff_t index)
 245 {
 246         struct osc_object *osc = cl2osc(obj);
 247         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 248         struct osc_io *oio = osc_env_io(env);
 249         int result;
 250
 251         opg->ops_from = 0;
 252         opg->ops_to = PAGE_SIZE - 1;
 253
 254         INIT_LIST_HEAD(&opg->ops_lru);
 255
 256         result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
 257         if (result != 0)
 258                 return result;
 259
 260         opg->ops_srvlock = osc_io_srvlock(oio);
 261         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 262
 263         /* reserve an LRU space for this page */
 264         if (cl_page->cp_type == CPT_CACHEABLE) {
 265                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 266                 if (result == 0) {
 267                         result = radix_tree_preload(GFP_NOFS);
 268                         if (result == 0) {
 269                                 spin_lock(&osc->oo_tree_lock);
 270                                 result = radix_tree_insert(&osc->oo_tree,
 271                                                            index, opg);
 272                                 if (result == 0) {
 273                                         ++osc->oo_npages;
 274                                         opg->ops_intree = 1;
 275                                 }
 276                                 spin_unlock(&osc->oo_tree_lock);
 277
 278                                 radix_tree_preload_end();
 279                         }
 280                 }
 281         }
 282
 283         return result;
 284 }
 285 EXPORT_SYMBOL(osc_page_init);
 286
 287 /**
 288  * Helper function called by osc_io_submit() for every page in an immediate
 289  * transfer (i.e., transferred synchronously).
 290  */
 291 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 292                      enum cl_req_type crt, int brw_flags)
 293 {
 294         struct osc_io *oio = osc_env_io(env);
 295         struct osc_async_page *oap = &opg->ops_oap;
 296
 297         LASSERT(oap->oap_async_flags & ASYNC_READY);
 298         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 299
 300         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 301         oap->oap_page_off = opg->ops_from;
 302         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 303         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 304
 305         if (oio->oi_cap_sys_resource)
 306                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 307
 308         osc_page_transfer_get(opg, "transfer\0imm");
 309         osc_page_transfer_add(env, opg, crt);
 310 }
 311
 312 /* --------------- LRU page management ------------------ */
 313
 314 /* OSC is a natural place to manage LRU pages as applications are specialized
 315  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 316  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 317  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 318  * for free LRU slots - this will be very bad so the algorithm requires each
 319  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 320  * at any time.
 321  */
 322
 323 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 324
 325 /**
 326  * LRU pages are freed in batch mode. OSC should at least free this
 327  * number of pages to avoid running out of LRU slots.
 328  */
 329 static inline int lru_shrink_min(struct client_obd *cli)
 330 {
 331         return cli->cl_max_pages_per_rpc * 2;
 332 }
 333
 334 /**
 335  * free this number at most otherwise it will take too long time to finsih.
 336  */
 337 static inline int lru_shrink_max(struct client_obd *cli)
 338 {
 339         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 340 }
 341
 342 /**
 343  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 344  * we should free slots aggressively. In this way, slots are freed in a steady
 345  * step to maintain fairness among OSCs.
 346  *
 347  * Return how many LRU pages should be freed.
 348  */
 349 static int osc_cache_too_much(struct client_obd *cli)
 350 {
 351         struct cl_client_cache *cache = cli->cl_cache;
 352         long pages = atomic_long_read(&cli->cl_lru_in_list);
 353         unsigned long budget;
 354
 355         LASSERT(cache != NULL);
 356         budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
 357
 358         /* if it's going to run out LRU slots, we should free some, but not
 359          * too much to maintain faireness among OSCs. */
 360         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 361                 if (pages >= budget)
 362                         return lru_shrink_max(cli);
 363                 else if (pages >= budget / 2)
 364                         return lru_shrink_min(cli);
 365         } else {
 366                 time64_t duration = ktime_get_real_seconds();
 367                 long timediff;
 368
 369                 /* knock out pages by duration of no IO activity */
 370                 duration -= cli->cl_lru_last_used;
 371                 /*
 372                  * The difference shouldn't be more than 70 years
 373                  * so we can safely case to a long. Round to
 374                  * approximately 1 minute.
 375                  */
 376                 timediff = (long)(duration >> 6);
 377                 if (timediff > 0 && pages >= budget / timediff)
 378                         return lru_shrink_min(cli);
 379         }
 380         return 0;
 381 }
 382
 383 int lru_queue_work(const struct lu_env *env, void *data)
 384 {
 385         struct client_obd *cli = data;
 386         int count;
 387
 388         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 389         count = osc_cache_too_much(cli);
 390         if (count > 0) {
 391                 int rc = osc_lru_shrink(env, cli, count, false);
 392
 393                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 394                        cli_name(cli), rc, count);
 395                 if (rc >= count) {
 396                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 397                         ptlrpcd_queue_work(cli->cl_lru_work);
 398                 }
 399         }
 400
 401         RETURN(0);
 402 }
 403
 404 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 405 {
 406         LIST_HEAD(lru);
 407         struct osc_async_page *oap;
 408         long npages = 0;
 409
 410         list_for_each_entry(oap, plist, oap_pending_item) {
 411                 struct osc_page *opg = oap2osc_page(oap);
 412
 413                 if (!opg->ops_in_lru)
 414                         continue;
 415
 416                 ++npages;
 417                 LASSERT(list_empty(&opg->ops_lru));
 418                 list_add(&opg->ops_lru, &lru);
 419         }
 420
 421         if (npages > 0) {
 422                 spin_lock(&cli->cl_lru_list_lock);
 423                 list_splice_tail(&lru, &cli->cl_lru_list);
 424                 atomic_long_sub(npages, &cli->cl_lru_busy);
 425                 atomic_long_add(npages, &cli->cl_lru_in_list);
 426                 cli->cl_lru_last_used = ktime_get_real_seconds();
 427                 spin_unlock(&cli->cl_lru_list_lock);
 428
 429                 if (waitqueue_active(&osc_lru_waitq))
 430                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 431         }
 432 }
 433
 434 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 435 {
 436         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 437         list_del_init(&opg->ops_lru);
 438         atomic_long_dec(&cli->cl_lru_in_list);
 439 }
 440
 441 /**
 442  * Page is being destroyed. The page may be not in LRU list, if the transfer
 443  * has never finished(error occurred).
 444  */
 445 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 446 {
 447         if (opg->ops_in_lru) {
 448                 spin_lock(&cli->cl_lru_list_lock);
 449                 if (!list_empty(&opg->ops_lru)) {
 450                         __osc_lru_del(cli, opg);
 451                 } else {
 452                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 453                         atomic_long_dec(&cli->cl_lru_busy);
 454                 }
 455                 spin_unlock(&cli->cl_lru_list_lock);
 456
 457                 atomic_long_inc(cli->cl_lru_left);
 458                 /* this is a great place to release more LRU pages if
 459                  * this osc occupies too many LRU pages and kernel is
 460                  * stealing one of them. */
 461                 if (osc_cache_too_much(cli)) {
 462                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 463                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 464                 }
 465                 wake_up(&osc_lru_waitq);
 466         } else {
 467                 LASSERT(list_empty(&opg->ops_lru));
 468         }
 469 }
 470
 471 /**
 472  * Delete page from LRU list for redirty.
 473  */
 474 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 475 {
 476         /* If page is being transferred for the first time,
 477          * ops_lru should be empty */
 478         if (opg->ops_in_lru) {
 479                 if (list_empty(&opg->ops_lru))
 480                         return;
 481                 spin_lock(&cli->cl_lru_list_lock);
 482                 if (!list_empty(&opg->ops_lru)) {
 483                         __osc_lru_del(cli, opg);
 484                         atomic_long_inc(&cli->cl_lru_busy);
 485                 }
 486                 spin_unlock(&cli->cl_lru_list_lock);
 487         }
 488 }
 489
 490 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 491                                 struct cl_page **pvec, int max_index)
 492 {
 493         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 494         int i;
 495
 496         ll_pagevec_init(pagevec, 0);
 497         for (i = 0; i < max_index; i++) {
 498                 struct cl_page *page = pvec[i];
 499
 500                 LASSERT(cl_page_is_owned(page, io));
 501                 cl_page_delete(env, page);
 502                 cl_page_discard(env, io, page);
 503                 cl_page_disown(env, io, page);
 504                 cl_pagevec_put(env, page, pagevec);
 505
 506                 pvec[i] = NULL;
 507         }
 508         pagevec_release(pagevec);
 509 }
 510
 511 /**
 512  * Check if a cl_page can be released, i.e, it's not being used.
 513  *
 514  * If unstable account is turned on, bulk transfer may hold one refcount
 515  * for recovery so we need to check vmpage refcount as well; otherwise,
 516  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 517  */
 518 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 519 {
 520         if (cl_page_in_use_noref(page))
 521                 return true;
 522
 523         if (cli->cl_cache->ccc_unstable_check) {
 524                 struct page *vmpage = cl_page_vmpage(page);
 525
 526                 /* vmpage have two known users: cl_page and VM page cache */
 527                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 528                         return true;
 529         }
 530         return false;
 531 }
 532
 533 /**
 534  * Drop @target of pages from LRU at most.
 535  */
 536 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 537                    long target, bool force)
 538 {
 539         struct cl_io *io;
 540         struct cl_object *clobj = NULL;
 541         struct cl_page **pvec;
 542         struct osc_page *opg;
 543         long count = 0;
 544         int maxscan = 0;
 545         int index = 0;
 546         int rc = 0;
 547         ENTRY;
 548
 549         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 550         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 551                 RETURN(0);
 552
 553         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 554                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 555         if (!force) {
 556                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 557                         RETURN(-EBUSY);
 558
 559                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 560                         atomic_dec(&cli->cl_lru_shrinkers);
 561                         RETURN(-EBUSY);
 562                 }
 563         } else {
 564                 atomic_inc(&cli->cl_lru_shrinkers);
 565         }
 566
 567         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 568         io = osc_env_thread_io(env);
 569
 570         spin_lock(&cli->cl_lru_list_lock);
 571         if (force)
 572                 cli->cl_lru_reclaim++;
 573         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 574         while (!list_empty(&cli->cl_lru_list)) {
 575                 struct cl_page *page;
 576                 bool will_free = false;
 577
 578                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 579                         break;
 580
 581                 if (--maxscan < 0)
 582                         break;
 583
 584                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 585                                        ops_lru);
 586                 page = opg->ops_cl.cpl_page;
 587                 if (lru_page_busy(cli, page)) {
 588                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 589                         continue;
 590                 }
 591
 592                 LASSERT(page->cp_obj != NULL);
 593                 if (clobj != page->cp_obj) {
 594                         struct cl_object *tmp = page->cp_obj;
 595
 596                         cl_object_get(tmp);
 597                         spin_unlock(&cli->cl_lru_list_lock);
 598
 599                         if (clobj != NULL) {
 600                                 discard_pagevec(env, io, pvec, index);
 601                                 index = 0;
 602
 603                                 cl_io_fini(env, io);
 604                                 cl_object_put(env, clobj);
 605                                 clobj = NULL;
 606                         }
 607
 608                         clobj = tmp;
 609                         io->ci_obj = clobj;
 610                         io->ci_ignore_layout = 1;
 611                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 612
 613                         spin_lock(&cli->cl_lru_list_lock);
 614
 615                         if (rc != 0)
 616                                 break;
 617
 618                         ++maxscan;
 619                         continue;
 620                 }
 621
 622                 if (cl_page_own_try(env, io, page) == 0) {
 623                         if (!lru_page_busy(cli, page)) {
 624                                 /* remove it from lru list earlier to avoid
 625                                  * lock contention */
 626                                 __osc_lru_del(cli, opg);
 627                                 opg->ops_in_lru = 0; /* will be discarded */
 628
 629                                 cl_page_get(page);
 630                                 will_free = true;
 631                         } else {
 632                                 cl_page_disown(env, io, page);
 633                         }
 634                 }
 635
 636                 if (!will_free) {
 637                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 638                         continue;
 639                 }
 640
 641                 /* Don't discard and free the page with cl_lru_list held */
 642                 pvec[index++] = page;
 643                 if (unlikely(index == OTI_PVEC_SIZE)) {
 644                         spin_unlock(&cli->cl_lru_list_lock);
 645                         discard_pagevec(env, io, pvec, index);
 646                         index = 0;
 647
 648                         spin_lock(&cli->cl_lru_list_lock);
 649                 }
 650
 651                 if (++count >= target)
 652                         break;
 653         }
 654         spin_unlock(&cli->cl_lru_list_lock);
 655
 656         if (clobj != NULL) {
 657                 discard_pagevec(env, io, pvec, index);
 658
 659                 cl_io_fini(env, io);
 660                 cl_object_put(env, clobj);
 661         }
 662
 663         atomic_dec(&cli->cl_lru_shrinkers);
 664         if (count > 0) {
 665                 atomic_long_add(count, cli->cl_lru_left);
 666                 wake_up(&osc_lru_waitq);
 667         }
 668         RETURN(count > 0 ? count : rc);
 669 }
 670 EXPORT_SYMBOL(osc_lru_shrink);
 671
 672 /**
 673  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 674  * \@npages of LRU slots. For performance consideration, it's better to drop
 675  * LRU pages in batch. Therefore, the actual number is adjusted at least
 676  * max_pages_per_rpc.
 677  */
 678 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 679 {
 680         struct lu_env *env;
 681         struct cl_client_cache *cache = cli->cl_cache;
 682         struct client_obd *scan;
 683         int max_scans;
 684         __u16 refcheck;
 685         long rc = 0;
 686         ENTRY;
 687
 688         LASSERT(cache != NULL);
 689
 690         env = cl_env_get(&refcheck);
 691         if (IS_ERR(env))
 692                 RETURN(rc);
 693
 694         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 695         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 696                cli_name(cli), npages);
 697         rc = osc_lru_shrink(env, cli, npages, true);
 698         if (rc >= npages) {
 699                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 700                        cli_name(cli), rc, npages);
 701                 if (osc_cache_too_much(cli) > 0)
 702                         ptlrpcd_queue_work(cli->cl_lru_work);
 703                 GOTO(out, rc);
 704         } else if (rc > 0) {
 705                 npages -= rc;
 706         }
 707
 708         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 709                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 710                 atomic_long_read(&cli->cl_lru_busy), npages);
 711
 712         /* Reclaim LRU slots from other client_obd as it can't free enough
 713          * from its own. This should rarely happen. */
 714         spin_lock(&cache->ccc_lru_lock);
 715         LASSERT(!list_empty(&cache->ccc_lru));
 716
 717         cache->ccc_lru_shrinkers++;
 718         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 719
 720         max_scans = refcount_read(&cache->ccc_users) - 2;
 721         while (--max_scans > 0 &&
 722                (scan = list_first_entry_or_null(&cache->ccc_lru,
 723                                                   struct client_obd,
 724                                                   cl_lru_osc)) != NULL) {
 725                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 726                        cli_name(scan), scan,
 727                        atomic_long_read(&scan->cl_lru_in_list),
 728                        atomic_long_read(&scan->cl_lru_busy));
 729
 730                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 731                 if (osc_cache_too_much(scan) > 0) {
 732                         spin_unlock(&cache->ccc_lru_lock);
 733
 734                         rc = osc_lru_shrink(env, scan, npages, true);
 735                         spin_lock(&cache->ccc_lru_lock);
 736                         if (rc >= npages)
 737                                 break;
 738                         if (rc > 0)
 739                                 npages -= rc;
 740                 }
 741         }
 742         spin_unlock(&cache->ccc_lru_lock);
 743
 744 out:
 745         cl_env_put(env, &refcheck);
 746         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 747                cli_name(cli), cli, rc);
 748         return rc;
 749 }
 750
 751 /**
 752  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 753  *
 754  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 755  * Only in the case that the LRU slots are in extreme shortage, it should
 756  * have reserved enough slots for an IO.
 757  */
 758 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 759                          struct osc_page *opg)
 760 {
 761         struct osc_io *oio = osc_env_io(env);
 762         int rc = 0;
 763
 764         ENTRY;
 765
 766         if (cli->cl_cache == NULL) /* shall not be in LRU */
 767                 RETURN(0);
 768
 769         if (oio->oi_lru_reserved > 0) {
 770                 --oio->oi_lru_reserved;
 771                 goto out;
 772         }
 773
 774         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 775         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 776                 /* run out of LRU spaces, try to drop some by itself */
 777                 rc = osc_lru_reclaim(cli, 1);
 778                 if (rc < 0)
 779                         break;
 780                 if (rc > 0)
 781                         continue;
 782                 /* IO issued by readahead, don't try hard */
 783                 if (oio->oi_is_readahead) {
 784                         if (atomic_long_read(cli->cl_lru_left) > 0)
 785                                 continue;
 786                         rc = -EBUSY;
 787                         break;
 788                 }
 789
 790                 cond_resched();
 791                 rc = l_wait_event_abortable(
 792                         osc_lru_waitq,
 793                         atomic_long_read(cli->cl_lru_left) > 0);
 794                 if (rc < 0) {
 795                         rc = -EINTR;
 796                         break;
 797                 }
 798         }
 799
 800 out:
 801         if (rc >= 0) {
 802                 atomic_long_inc(&cli->cl_lru_busy);
 803                 opg->ops_in_lru = 1;
 804                 rc = 0;
 805         }
 806
 807         RETURN(rc);
 808 }
 809
 810 /**
 811  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 812  *
 813  * The benefit of doing this is to reduce contention against atomic counter
 814  * cl_lru_left by changing it from per-page access to per-IO access.
 815  */
 816 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 817 {
 818         unsigned long reserved = 0;
 819         unsigned long max_pages;
 820         unsigned long c;
 821         int rc;
 822
 823 again:
 824         c = atomic_long_read(cli->cl_lru_left);
 825         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 826                 c = atomic_long_read(cli->cl_lru_left);
 827
 828         if (c < npages) {
 829                 /*
 830                  * Trigger writeback in the hope some LRU slot could
 831                  * be freed.
 832                  */
 833                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 834                 if (rc)
 835                         return 0;
 836         }
 837
 838         while (c >= npages) {
 839                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 840                         reserved = npages;
 841                         break;
 842                 }
 843                 c = atomic_long_read(cli->cl_lru_left);
 844         }
 845
 846         if (reserved != npages) {
 847                 cond_resched();
 848                 rc = l_wait_event_abortable(
 849                         osc_lru_waitq,
 850                         atomic_long_read(cli->cl_lru_left) > 0);
 851                 goto again;
 852         }
 853
 854         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 855         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 856                 /* If there aren't enough pages in the per-OSC LRU then
 857                  * wake up the LRU thread to try and clear out space, so
 858                  * we don't block if pages are being dirtied quickly. */
 859                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 860                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 861                        max_pages);
 862                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 863         }
 864
 865         return reserved;
 866 }
 867
 868 /**
 869  * osc_lru_unreserve() is called to unreserve LRU slots.
 870  *
 871  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 872  * reasons such as page already existing or I/O error. Those reserved slots
 873  * should be freed by calling this function.
 874  */
 875 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 876 {
 877         atomic_long_add(npages, cli->cl_lru_left);
 878         wake_up(&osc_lru_waitq);
 879 }
 880
 881 /**
 882  * Atomic operations are expensive. We accumulate the accounting for the
 883  * same page zone to get better performance.
 884  * In practice this can work pretty good because the pages in the same RPC
 885  * are likely from the same page zone.
 886  */
 887 #ifdef HAVE_NR_UNSTABLE_NFS
 888 /* Old kernels use a separate counter for unstable pages,
 889  * newer kernels treat them like any other writeback.
 890  * (see Linux commit: v5.7-467-g8d92890bd6b8)
 891  */
 892 #define NR_WRITEBACK NR_UNSTABLE_NFS
 893 #endif
 894
 895 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 896                                             struct osc_brw_async_args *aa,
 897                                             int factor)
 898 {
 899         int page_count;
 900         void *zone = NULL;
 901         int count = 0;
 902         int i;
 903
 904         if (desc != NULL) {
 905                 page_count = desc->bd_iov_count;
 906         } else {
 907                 page_count = aa->aa_page_count;
 908         }
 909
 910         for (i = 0; i < page_count; i++) {
 911                 void *pz;
 912                 if (desc)
 913                         pz = page_zone(desc->bd_vec[i].bv_page);
 914                 else
 915                         pz = page_zone(aa->aa_ppga[i]->pg);
 916
 917                 if (likely(pz == zone)) {
 918                         ++count;
 919                         continue;
 920                 }
 921
 922                 if (count > 0) {
 923                         mod_zone_page_state(zone,
 924                                             (enum zone_stat_item)NR_WRITEBACK,
 925                                             factor * count);
 926                         count = 0;
 927                 }
 928                 zone = pz;
 929                 ++count;
 930         }
 931         if (count > 0)
 932                 mod_zone_page_state(zone, (enum zone_stat_item)NR_WRITEBACK,
 933                                     factor * count);
 934 }
 935
 936 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 937                                                 struct osc_brw_async_args *aa)
 938 {
 939         unstable_page_accounting(desc, aa, 1);
 940 }
 941
 942 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 943                                                 struct osc_brw_async_args *aa)
 944 {
 945         unstable_page_accounting(desc, aa, -1);
 946 }
 947
 948 /**
 949  * Performs "unstable" page accounting. This function balances the
 950  * increment operations performed in osc_inc_unstable_pages. It is
 951  * registered as the RPC request callback, and is executed when the
 952  * bulk RPC is committed on the server. Thus at this point, the pages
 953  * involved in the bulk transfer are no longer considered unstable.
 954  *
 955  * If this function is called, the request should have been committed
 956  * or req:rq_unstable must have been set; it implies that the unstable
 957  * statistic have been added.
 958  */
 959 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 960 {
 961         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 962         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 963         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 964         int                      page_count;
 965         long                     unstable_count;
 966
 967         if (desc)
 968                 page_count = desc->bd_iov_count;
 969         else
 970                 page_count = aa->aa_page_count;
 971
 972         LASSERT(page_count >= 0);
 973
 974         dec_unstable_page_accounting(desc, aa);
 975
 976         unstable_count = atomic_long_sub_return(page_count,
 977                                                 &cli->cl_unstable_count);
 978         LASSERT(unstable_count >= 0);
 979
 980         unstable_count = atomic_long_sub_return(page_count,
 981                                            &cli->cl_cache->ccc_unstable_nr);
 982         LASSERT(unstable_count >= 0);
 983         if (unstable_count == 0)
 984                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 985
 986         if (waitqueue_active(&osc_lru_waitq))
 987                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 988 }
 989
 990 /**
 991  * "unstable" page accounting. See: osc_dec_unstable_pages.
 992  */
 993 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 994 {
 995         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 996         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 997         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 998         long                     page_count;
 999
1000         /* No unstable page tracking */
1001         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1002                 return;
1003
1004         if (desc)
1005                 page_count = desc->bd_iov_count;
1006         else
1007                 page_count = aa->aa_page_count;
1008
1009         add_unstable_page_accounting(desc, aa);
1010         atomic_long_add(page_count, &cli->cl_unstable_count);
1011         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1012
1013         /* If the request has already been committed (i.e. brw_commit
1014          * called via rq_commit_cb), we need to undo the unstable page
1015          * increments we just performed because rq_commit_cb wont be
1016          * called again. */
1017         spin_lock(&req->rq_lock);
1018         if (unlikely(req->rq_committed)) {
1019                 spin_unlock(&req->rq_lock);
1020
1021                 osc_dec_unstable_pages(req);
1022         } else {
1023                 req->rq_unstable = 1;
1024                 spin_unlock(&req->rq_lock);
1025         }
1026 }
1027
1028 /**
1029  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1030  * This function will be called by every BRW RPC so it's critical
1031  * to make this function fast.
1032  */
1033 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1034 {
1035         long unstable_nr, osc_unstable_count;
1036
1037         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1038         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1039                 return false;
1040
1041         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1042         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1043
1044         CDEBUG(D_CACHE,
1045                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1046                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1047
1048         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1049          * has one full RPC window of unstable pages, it's a good chance
1050          * to piggyback a SOFT_SYNC flag.
1051          * Please notice that the OST won't take immediate response for the
1052          * SOFT_SYNC request so active OSCs will have more chance to carry
1053          * the flag, this is reasonable. */
1054         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1055                osc_unstable_count > cli->cl_max_pages_per_rpc *
1056                                     cli->cl_max_rpcs_in_flight;
1057 }
1058
1059 /**
1060  * Return how many LRU pages in the cache of all OSC devices
1061  *
1062  * \retval      return # of cached LRU pages times reclaimation tendency
1063  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1064  */
1065 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1066                                      struct shrink_control *sc)
1067 {
1068         struct client_obd *cli;
1069         unsigned long cached = 0;
1070
1071         spin_lock(&osc_shrink_lock);
1072         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1073                 cached += atomic_long_read(&cli->cl_lru_in_list);
1074         spin_unlock(&osc_shrink_lock);
1075
1076         return (cached  * sysctl_vfs_cache_pressure) / 100;
1077 }
1078
1079 /**
1080  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1081  *
1082  * \retval      number of cached LRU pages reclaimed
1083  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1084  *
1085  * Linux kernel will loop calling this shrinker scan routine with
1086  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1087  *
1088  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1089  * to scan and try to reclaim LRU pages, just return 0 and
1090  * osc_cache_shrink_count() will report the LRU page number.
1091  */
1092 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1093                                     struct shrink_control *sc)
1094 {
1095         struct client_obd *cli;
1096         struct client_obd *stop_anchor = NULL;
1097         struct lu_env *env;
1098         long shrank = 0;
1099         int rc;
1100         __u16 refcheck;
1101
1102         if (sc->nr_to_scan == 0)
1103                 return 0;
1104
1105         if (!(sc->gfp_mask & __GFP_FS))
1106                 return SHRINK_STOP;
1107
1108         env = cl_env_get(&refcheck);
1109         if (IS_ERR(env))
1110                 return SHRINK_STOP;
1111
1112         spin_lock(&osc_shrink_lock);
1113         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1114                                                struct client_obd,
1115                                                cl_shrink_list)) != NULL) {
1116                 if (stop_anchor == NULL)
1117                         stop_anchor = cli;
1118                 else if (cli == stop_anchor)
1119                         break;
1120
1121                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1122                 spin_unlock(&osc_shrink_lock);
1123
1124                 /* shrink no more than max_pages_per_rpc for an OSC */
1125                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1126                                     cli->cl_max_pages_per_rpc ?
1127                                     cli->cl_max_pages_per_rpc :
1128                                     sc->nr_to_scan - shrank, true);
1129                 if (rc > 0)
1130                         shrank += rc;
1131
1132                 if (shrank >= sc->nr_to_scan)
1133                         goto out;
1134
1135                 spin_lock(&osc_shrink_lock);
1136         }
1137         spin_unlock(&osc_shrink_lock);
1138
1139 out:
1140         cl_env_put(env, &refcheck);
1141
1142         return shrank;
1143 }
1144
1145 /** @} osc */