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 > "
 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                           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_discard(env, io, page);
 502                 cl_page_disown(env, io, page);
 503                 cl_pagevec_put(env, page, pagevec);
 504
 505                 pvec[i] = NULL;
 506         }
 507         pagevec_release(pagevec);
 508 }
 509
 510 /**
 511  * Check if a cl_page can be released, i.e, it's not being used.
 512  *
 513  * If unstable account is turned on, bulk transfer may hold one refcount
 514  * for recovery so we need to check vmpage refcount as well; otherwise,
 515  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 516  */
 517 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 518 {
 519         if (cl_page_in_use_noref(page))
 520                 return true;
 521
 522         if (cli->cl_cache->ccc_unstable_check) {
 523                 struct page *vmpage = cl_page_vmpage(page);
 524
 525                 /* vmpage have two known users: cl_page and VM page cache */
 526                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 527                         return true;
 528         }
 529         return false;
 530 }
 531
 532 /**
 533  * Drop @target of pages from LRU at most.
 534  */
 535 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 536                    long target, bool force)
 537 {
 538         struct cl_io *io;
 539         struct cl_object *clobj = NULL;
 540         struct cl_page **pvec;
 541         struct osc_page *opg;
 542         long count = 0;
 543         int maxscan = 0;
 544         int index = 0;
 545         int rc = 0;
 546         ENTRY;
 547
 548         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 549         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 550                 RETURN(0);
 551
 552         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 553                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 554         if (!force) {
 555                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 556                         RETURN(-EBUSY);
 557
 558                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 559                         atomic_dec(&cli->cl_lru_shrinkers);
 560                         RETURN(-EBUSY);
 561                 }
 562         } else {
 563                 atomic_inc(&cli->cl_lru_shrinkers);
 564         }
 565
 566         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 567         io = osc_env_thread_io(env);
 568
 569         spin_lock(&cli->cl_lru_list_lock);
 570         if (force)
 571                 cli->cl_lru_reclaim++;
 572         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 573         while (!list_empty(&cli->cl_lru_list)) {
 574                 struct cl_page *page;
 575                 bool will_free = false;
 576
 577                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 578                         break;
 579
 580                 if (--maxscan < 0)
 581                         break;
 582
 583                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 584                                        ops_lru);
 585                 page = opg->ops_cl.cpl_page;
 586                 if (lru_page_busy(cli, page)) {
 587                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 588                         continue;
 589                 }
 590
 591                 LASSERT(page->cp_obj != NULL);
 592                 if (clobj != page->cp_obj) {
 593                         struct cl_object *tmp = page->cp_obj;
 594
 595                         cl_object_get(tmp);
 596                         spin_unlock(&cli->cl_lru_list_lock);
 597
 598                         if (clobj != NULL) {
 599                                 discard_pagevec(env, io, pvec, index);
 600                                 index = 0;
 601
 602                                 cl_io_fini(env, io);
 603                                 cl_object_put(env, clobj);
 604                                 clobj = NULL;
 605                         }
 606
 607                         clobj = tmp;
 608                         io->ci_obj = clobj;
 609                         io->ci_ignore_layout = 1;
 610                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 611
 612                         spin_lock(&cli->cl_lru_list_lock);
 613
 614                         if (rc != 0)
 615                                 break;
 616
 617                         ++maxscan;
 618                         continue;
 619                 }
 620
 621                 if (cl_page_own_try(env, io, page) == 0) {
 622                         if (!lru_page_busy(cli, page)) {
 623                                 /* remove it from lru list earlier to avoid
 624                                  * lock contention */
 625                                 __osc_lru_del(cli, opg);
 626                                 opg->ops_in_lru = 0; /* will be discarded */
 627
 628                                 cl_page_get(page);
 629                                 will_free = true;
 630                         } else {
 631                                 cl_page_disown(env, io, page);
 632                         }
 633                 }
 634
 635                 if (!will_free) {
 636                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 637                         continue;
 638                 }
 639
 640                 /* Don't discard and free the page with cl_lru_list held */
 641                 pvec[index++] = page;
 642                 if (unlikely(index == OTI_PVEC_SIZE)) {
 643                         spin_unlock(&cli->cl_lru_list_lock);
 644                         discard_pagevec(env, io, pvec, index);
 645                         index = 0;
 646
 647                         spin_lock(&cli->cl_lru_list_lock);
 648                 }
 649
 650                 if (++count >= target)
 651                         break;
 652         }
 653         spin_unlock(&cli->cl_lru_list_lock);
 654
 655         if (clobj != NULL) {
 656                 discard_pagevec(env, io, pvec, index);
 657
 658                 cl_io_fini(env, io);
 659                 cl_object_put(env, clobj);
 660         }
 661
 662         atomic_dec(&cli->cl_lru_shrinkers);
 663         if (count > 0) {
 664                 atomic_long_add(count, cli->cl_lru_left);
 665                 wake_up(&osc_lru_waitq);
 666         }
 667         RETURN(count > 0 ? count : rc);
 668 }
 669 EXPORT_SYMBOL(osc_lru_shrink);
 670
 671 /**
 672  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 673  * \@npages of LRU slots. For performance consideration, it's better to drop
 674  * LRU pages in batch. Therefore, the actual number is adjusted at least
 675  * max_pages_per_rpc.
 676  */
 677 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 678 {
 679         struct lu_env *env;
 680         struct cl_client_cache *cache = cli->cl_cache;
 681         struct client_obd *scan;
 682         int max_scans;
 683         __u16 refcheck;
 684         long rc = 0;
 685         ENTRY;
 686
 687         LASSERT(cache != NULL);
 688
 689         env = cl_env_get(&refcheck);
 690         if (IS_ERR(env))
 691                 RETURN(rc);
 692
 693         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 694         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 695                cli_name(cli), npages);
 696         rc = osc_lru_shrink(env, cli, npages, true);
 697         if (rc >= npages) {
 698                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 699                        cli_name(cli), rc, npages);
 700                 if (osc_cache_too_much(cli) > 0)
 701                         ptlrpcd_queue_work(cli->cl_lru_work);
 702                 GOTO(out, rc);
 703         } else if (rc > 0) {
 704                 npages -= rc;
 705         }
 706
 707         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 708                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 709                 atomic_long_read(&cli->cl_lru_busy), npages);
 710
 711         /* Reclaim LRU slots from other client_obd as it can't free enough
 712          * from its own. This should rarely happen. */
 713         spin_lock(&cache->ccc_lru_lock);
 714         LASSERT(!list_empty(&cache->ccc_lru));
 715
 716         cache->ccc_lru_shrinkers++;
 717         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 718
 719         max_scans = refcount_read(&cache->ccc_users) - 2;
 720         while (--max_scans > 0 &&
 721                (scan = list_first_entry_or_null(&cache->ccc_lru,
 722                                                   struct client_obd,
 723                                                   cl_lru_osc)) != NULL) {
 724                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 725                        cli_name(scan), scan,
 726                        atomic_long_read(&scan->cl_lru_in_list),
 727                        atomic_long_read(&scan->cl_lru_busy));
 728
 729                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 730                 if (osc_cache_too_much(scan) > 0) {
 731                         spin_unlock(&cache->ccc_lru_lock);
 732
 733                         rc = osc_lru_shrink(env, scan, npages, true);
 734                         spin_lock(&cache->ccc_lru_lock);
 735                         if (rc >= npages)
 736                                 break;
 737                         if (rc > 0)
 738                                 npages -= rc;
 739                 }
 740         }
 741         spin_unlock(&cache->ccc_lru_lock);
 742
 743 out:
 744         cl_env_put(env, &refcheck);
 745         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 746                cli_name(cli), cli, rc);
 747         return rc;
 748 }
 749
 750 /**
 751  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 752  *
 753  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 754  * Only in the case that the LRU slots are in extreme shortage, it should
 755  * have reserved enough slots for an IO.
 756  */
 757 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 758                          struct osc_page *opg)
 759 {
 760         struct osc_io *oio = osc_env_io(env);
 761         int rc = 0;
 762
 763         ENTRY;
 764
 765         if (cli->cl_cache == NULL) /* shall not be in LRU */
 766                 RETURN(0);
 767
 768         if (oio->oi_lru_reserved > 0) {
 769                 --oio->oi_lru_reserved;
 770                 goto out;
 771         }
 772
 773         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 774         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 775                 /* run out of LRU spaces, try to drop some by itself */
 776                 rc = osc_lru_reclaim(cli, 1);
 777                 if (rc < 0)
 778                         break;
 779                 if (rc > 0)
 780                         continue;
 781                 /* IO issued by readahead, don't try hard */
 782                 if (oio->oi_is_readahead) {
 783                         if (atomic_long_read(cli->cl_lru_left) > 0)
 784                                 continue;
 785                         rc = -EBUSY;
 786                         break;
 787                 }
 788
 789                 cond_resched();
 790                 rc = l_wait_event_abortable(
 791                         osc_lru_waitq,
 792                         atomic_long_read(cli->cl_lru_left) > 0);
 793                 if (rc < 0) {
 794                         rc = -EINTR;
 795                         break;
 796                 }
 797         }
 798
 799 out:
 800         if (rc >= 0) {
 801                 atomic_long_inc(&cli->cl_lru_busy);
 802                 opg->ops_in_lru = 1;
 803                 rc = 0;
 804         }
 805
 806         RETURN(rc);
 807 }
 808
 809 /**
 810  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 811  *
 812  * The benefit of doing this is to reduce contention against atomic counter
 813  * cl_lru_left by changing it from per-page access to per-IO access.
 814  */
 815 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 816 {
 817         unsigned long reserved = 0;
 818         unsigned long max_pages;
 819         unsigned long c;
 820         int rc;
 821
 822 again:
 823         c = atomic_long_read(cli->cl_lru_left);
 824         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 825                 c = atomic_long_read(cli->cl_lru_left);
 826
 827         if (c < npages) {
 828                 /*
 829                  * Trigger writeback in the hope some LRU slot could
 830                  * be freed.
 831                  */
 832                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 833                 if (rc)
 834                         return 0;
 835         }
 836
 837         while (c >= npages) {
 838                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 839                         reserved = npages;
 840                         break;
 841                 }
 842                 c = atomic_long_read(cli->cl_lru_left);
 843         }
 844
 845         if (reserved != npages) {
 846                 cond_resched();
 847                 rc = l_wait_event_abortable(
 848                         osc_lru_waitq,
 849                         atomic_long_read(cli->cl_lru_left) > 0);
 850                 goto again;
 851         }
 852
 853         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 854         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 855                 /* If there aren't enough pages in the per-OSC LRU then
 856                  * wake up the LRU thread to try and clear out space, so
 857                  * we don't block if pages are being dirtied quickly. */
 858                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 859                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 860                        max_pages);
 861                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 862         }
 863
 864         return reserved;
 865 }
 866
 867 /**
 868  * osc_lru_unreserve() is called to unreserve LRU slots.
 869  *
 870  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 871  * reasons such as page already existing or I/O error. Those reserved slots
 872  * should be freed by calling this function.
 873  */
 874 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 875 {
 876         atomic_long_add(npages, cli->cl_lru_left);
 877         wake_up(&osc_lru_waitq);
 878 }
 879
 880 /**
 881  * Atomic operations are expensive. We accumulate the accounting for the
 882  * same page zone to get better performance.
 883  * In practice this can work pretty good because the pages in the same RPC
 884  * are likely from the same page zone.
 885  */
 886 #ifdef HAVE_NR_UNSTABLE_NFS
 887 /* Old kernels use a separate counter for unstable pages,
 888  * newer kernels treat them like any other writeback.
 889  * (see Linux commit: v5.7-467-g8d92890bd6b8)
 890  */
 891 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_UNSTABLE_NFS)
 892 #elif !defined(HAVE_NR_ZONE_WRITE_PENDING)
 893 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_WRITEBACK)
 894 #endif
 895
 896 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 897                                             int factor)
 898 {
 899         int page_count;
 900         void *zone = NULL;
 901         int count = 0;
 902         int i;
 903
 904         ENTRY;
 905
 906         page_count = desc->bd_iov_count;
 907
 908         CDEBUG(D_PAGE, "%s %d unstable pages\n",
 909                factor == 1 ? "adding" : "removing", page_count);
 910
 911         for (i = 0; i < page_count; i++) {
 912                 void *pz = page_zone(desc->bd_vec[i].bv_page);
 913
 914                 if (likely(pz == zone)) {
 915                         ++count;
 916                         continue;
 917                 }
 918
 919                 if (count > 0) {
 920                         mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 921                                             factor * count);
 922                         count = 0;
 923                 }
 924                 zone = pz;
 925                 ++count;
 926         }
 927         if (count > 0)
 928                 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 929                                     factor * count);
 930
 931         EXIT;
 932 }
 933
 934 static inline void add_unstable_pages(struct ptlrpc_bulk_desc *desc)
 935 {
 936         unstable_page_accounting(desc, 1);
 937 }
 938
 939 static inline void dec_unstable_pages(struct ptlrpc_bulk_desc *desc)
 940 {
 941         unstable_page_accounting(desc, -1);
 942 }
 943
 944 /**
 945  * Performs "unstable" page accounting. This function balances the
 946  * increment operations performed in osc_inc_unstable_pages. It is
 947  * registered as the RPC request callback, and is executed when the
 948  * bulk RPC is committed on the server. Thus at this point, the pages
 949  * involved in the bulk transfer are no longer considered unstable.
 950  *
 951  * If this function is called, the request should have been committed
 952  * or req:rq_unstable must have been set; it implies that the unstable
 953  * statistic have been added.
 954  */
 955 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 956 {
 957         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 958         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 959         int                      page_count;
 960         long                     unstable_count;
 961
 962         /* no desc means short io, which doesn't have separate unstable pages,
 963          * it's just using space inside the RPC itself
 964          */
 965         if (!desc)
 966                 return;
 967
 968         page_count = desc->bd_iov_count;
 969
 970         LASSERT(page_count >= 0);
 971
 972         dec_unstable_pages(desc);
 973
 974         unstable_count = atomic_long_sub_return(page_count,
 975                                                 &cli->cl_unstable_count);
 976         LASSERT(unstable_count >= 0);
 977
 978         unstable_count = atomic_long_sub_return(page_count,
 979                                            &cli->cl_cache->ccc_unstable_nr);
 980         LASSERT(unstable_count >= 0);
 981
 982         if (waitqueue_active(&osc_lru_waitq))
 983                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 984 }
 985
 986 /**
 987  * "unstable" page accounting. See: osc_dec_unstable_pages.
 988  */
 989 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 990 {
 991         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 992         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 993         long                     page_count;
 994
 995         /* No unstable page tracking */
 996         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 997                 return;
 998
 999         /* no desc means short io, which doesn't have separate unstable pages,
1000          * it's just using space inside the RPC itself
1001          */
1002         if (!desc)
1003                 return;
1004
1005         page_count = desc->bd_iov_count;
1006
1007         add_unstable_pages(desc);
1008         atomic_long_add(page_count, &cli->cl_unstable_count);
1009         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1010
1011         /* If the request has already been committed (i.e. brw_commit
1012          * called via rq_commit_cb), we need to undo the unstable page
1013          * increments we just performed because rq_commit_cb wont be
1014          * called again. */
1015         spin_lock(&req->rq_lock);
1016         if (unlikely(req->rq_committed)) {
1017                 spin_unlock(&req->rq_lock);
1018
1019                 osc_dec_unstable_pages(req);
1020         } else {
1021                 req->rq_unstable = 1;
1022                 spin_unlock(&req->rq_lock);
1023         }
1024 }
1025
1026 /**
1027  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1028  * This function will be called by every BRW RPC so it's critical
1029  * to make this function fast.
1030  */
1031 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1032 {
1033         long unstable_nr, osc_unstable_count;
1034
1035         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1036         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1037                 return false;
1038
1039         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1040         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1041
1042         CDEBUG(D_CACHE,
1043                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1044                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1045
1046         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1047          * has one full RPC window of unstable pages, it's a good chance
1048          * to piggyback a SOFT_SYNC flag.
1049          * Please notice that the OST won't take immediate response for the
1050          * SOFT_SYNC request so active OSCs will have more chance to carry
1051          * the flag, this is reasonable. */
1052         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1053                osc_unstable_count > cli->cl_max_pages_per_rpc *
1054                                     cli->cl_max_rpcs_in_flight;
1055 }
1056
1057 /**
1058  * Return how many LRU pages in the cache of all OSC devices
1059  *
1060  * \retval      return # of cached LRU pages times reclaimation tendency
1061  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1062  */
1063 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1064                                      struct shrink_control *sc)
1065 {
1066         struct client_obd *cli;
1067         unsigned long cached = 0;
1068
1069         spin_lock(&osc_shrink_lock);
1070         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1071                 cached += atomic_long_read(&cli->cl_lru_in_list);
1072         spin_unlock(&osc_shrink_lock);
1073
1074         return (cached  * sysctl_vfs_cache_pressure) / 100;
1075 }
1076
1077 /**
1078  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1079  *
1080  * \retval      number of cached LRU pages reclaimed
1081  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1082  *
1083  * Linux kernel will loop calling this shrinker scan routine with
1084  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1085  *
1086  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1087  * to scan and try to reclaim LRU pages, just return 0 and
1088  * osc_cache_shrink_count() will report the LRU page number.
1089  */
1090 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1091                                     struct shrink_control *sc)
1092 {
1093         struct client_obd *cli;
1094         struct client_obd *stop_anchor = NULL;
1095         struct lu_env *env;
1096         long shrank = 0;
1097         int rc;
1098         __u16 refcheck;
1099
1100         if (sc->nr_to_scan == 0)
1101                 return 0;
1102
1103         if (!(sc->gfp_mask & __GFP_FS))
1104                 return SHRINK_STOP;
1105
1106         env = cl_env_get(&refcheck);
1107         if (IS_ERR(env))
1108                 return SHRINK_STOP;
1109
1110         spin_lock(&osc_shrink_lock);
1111         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1112                                                struct client_obd,
1113                                                cl_shrink_list)) != NULL) {
1114                 if (stop_anchor == NULL)
1115                         stop_anchor = cli;
1116                 else if (cli == stop_anchor)
1117                         break;
1118
1119                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1120                 spin_unlock(&osc_shrink_lock);
1121
1122                 /* shrink no more than max_pages_per_rpc for an OSC */
1123                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1124                                     cli->cl_max_pages_per_rpc ?
1125                                     cli->cl_max_pages_per_rpc :
1126                                     sc->nr_to_scan - shrank, true);
1127                 if (rc > 0)
1128                         shrank += rc;
1129
1130                 if (shrank >= sc->nr_to_scan)
1131                         goto out;
1132
1133                 spin_lock(&osc_shrink_lock);
1134         }
1135         spin_unlock(&osc_shrink_lock);
1136
1137 out:
1138         cl_env_put(env, &refcheck);
1139
1140         return shrank;
1141 }
1142
1143 /** @} osc */