lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Page operations.
  54  */
  55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  56 {
  57         struct cl_page *page = opg->ops_cl.cpl_page;
  58
  59         LASSERT(!opg->ops_transfer_pinned);
  60         cl_page_get(page);
  61         lu_ref_add_atomic(&page->cp_reference, label, page);
  62         opg->ops_transfer_pinned = 1;
  63 }
  64
  65 static void osc_page_transfer_put(const struct lu_env *env,
  66                                   struct osc_page *opg)
  67 {
  68         struct cl_page *page = opg->ops_cl.cpl_page;
  69
  70         if (opg->ops_transfer_pinned) {
  71                 opg->ops_transfer_pinned = 0;
  72                 lu_ref_del(&page->cp_reference, "transfer", page);
  73                 cl_page_put(env, page);
  74         }
  75 }
  76
  77 /**
  78  * This is called once for every page when it is submitted for a transfer
  79  * either opportunistic (osc_page_cache_add()), or immediate
  80  * (osc_page_submit()).
  81  */
  82 static void osc_page_transfer_add(const struct lu_env *env,
  83                                   struct osc_page *opg, enum cl_req_type crt)
  84 {
  85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  86
  87         osc_lru_use(osc_cli(obj), opg);
  88 }
  89
  90 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  91                        struct cl_io *io, cl_commit_cbt cb)
  92 {
  93         int result;
  94         ENTRY;
  95
  96         osc_page_transfer_get(opg, "transfer\0cache");
  97         result = osc_queue_async_io(env, io, opg, cb);
  98         if (result != 0)
  99                 osc_page_transfer_put(env, opg);
 100         else
 101                 osc_page_transfer_add(env, opg, CRT_WRITE);
 102
 103         RETURN(result);
 104 }
 105
 106 void osc_index2policy(union ldlm_policy_data *policy,
 107                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 108 {
 109         memset(policy, 0, sizeof *policy);
 110         policy->l_extent.start = cl_offset(obj, start);
 111         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 112 }
 113
 114 static inline s64 osc_submit_duration(struct osc_page *opg)
 115 {
 116         if (ktime_to_ns(opg->ops_submit_time) == 0)
 117                 return 0;
 118
 119         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 120 }
 121
 122 static int osc_page_print(const struct lu_env *env,
 123                           const struct cl_page_slice *slice,
 124                           void *cookie, lu_printer_t printer)
 125 {
 126         struct osc_page *opg = cl2osc_page(slice);
 127         struct osc_async_page *oap = &opg->ops_oap;
 128         struct osc_object *obj = cl2osc(slice->cpl_obj);
 129         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 130
 131         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 132                           "1< %#x %d %u %c %c > "
 133                           "2< %lld %u %u %#x %#x | %p %p %p > "
 134                           "3< %d %lld %d > "
 135                           "4< %d %d %d %lu %c | %c %c %c %c > "
 136                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 137                           opg, osc_index(opg),
 138                           /* 1 */
 139                           oap->oap_magic, oap->oap_cmd,
 140                           oap->oap_interrupted,
 141                           list_empty_marker(&oap->oap_pending_item),
 142                           list_empty_marker(&oap->oap_rpc_item),
 143                           /* 2 */
 144                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 145                           oap->oap_async_flags, oap->oap_brw_flags,
 146                           oap->oap_request, oap->oap_cli, obj,
 147                           /* 3 */
 148                           opg->ops_transfer_pinned,
 149                           osc_submit_duration(opg), opg->ops_srvlock,
 150                           /* 4 */
 151                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 152                           cli->cl_max_rpcs_in_flight,
 153                           cli->cl_avail_grant,
 154                           list_empty_marker(&cli->cl_cache_waiters),
 155                           list_empty_marker(&cli->cl_loi_ready_list),
 156                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 157                           list_empty_marker(&cli->cl_loi_write_list),
 158                           list_empty_marker(&cli->cl_loi_read_list),
 159                           /* 5 */
 160                           list_empty_marker(&obj->oo_ready_item),
 161                           list_empty_marker(&obj->oo_hp_ready_item),
 162                           list_empty_marker(&obj->oo_write_item),
 163                           list_empty_marker(&obj->oo_read_item),
 164                           atomic_read(&obj->oo_nr_reads),
 165                           list_empty_marker(&obj->oo_reading_exts),
 166                           atomic_read(&obj->oo_nr_writes),
 167                           list_empty_marker(&obj->oo_hp_exts),
 168                           list_empty_marker(&obj->oo_urgent_exts));
 169 }
 170
 171 static void osc_page_delete(const struct lu_env *env,
 172                             const struct cl_page_slice *slice)
 173 {
 174         struct osc_page   *opg = cl2osc_page(slice);
 175         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 176         int rc;
 177
 178         ENTRY;
 179         CDEBUG(D_TRACE, "%p\n", opg);
 180         osc_page_transfer_put(env, opg);
 181         rc = osc_teardown_async_page(env, obj, opg);
 182         if (rc) {
 183                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 184                               "Trying to teardown failed: %d\n", rc);
 185                 LASSERT(0);
 186         }
 187
 188         osc_lru_del(osc_cli(obj), opg);
 189
 190         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 191                 void *value = NULL;
 192
 193                 spin_lock(&obj->oo_tree_lock);
 194                 if (opg->ops_intree) {
 195                         value = radix_tree_delete(&obj->oo_tree,
 196                                                   osc_index(opg));
 197                         if (value != NULL) {
 198                                 --obj->oo_npages;
 199                                 opg->ops_intree = 0;
 200                         }
 201                 }
 202                 spin_unlock(&obj->oo_tree_lock);
 203
 204                 LASSERT(ergo(value != NULL, value == opg));
 205         }
 206
 207         EXIT;
 208 }
 209
 210 static void osc_page_clip(const struct lu_env *env,
 211                           const struct cl_page_slice *slice,
 212                           int from, int to)
 213 {
 214         struct osc_page       *opg = cl2osc_page(slice);
 215         struct osc_async_page *oap = &opg->ops_oap;
 216
 217         opg->ops_from = from;
 218         opg->ops_to   = to;
 219         spin_lock(&oap->oap_lock);
 220         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 221         spin_unlock(&oap->oap_lock);
 222 }
 223
 224 static int osc_page_cancel(const struct lu_env *env,
 225                            const struct cl_page_slice *slice)
 226 {
 227         struct osc_page *opg = cl2osc_page(slice);
 228         int rc = 0;
 229
 230         /* Check if the transferring against this page
 231          * is completed, or not even queued. */
 232         if (opg->ops_transfer_pinned)
 233                 /* FIXME: may not be interrupted.. */
 234                 rc = osc_cancel_async_page(env, opg);
 235         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 236         return rc;
 237 }
 238
 239 static int osc_page_flush(const struct lu_env *env,
 240                           const struct cl_page_slice *slice,
 241                           struct cl_io *io)
 242 {
 243         struct osc_page *opg = cl2osc_page(slice);
 244         int rc = 0;
 245         ENTRY;
 246         rc = osc_flush_async_page(env, io, opg);
 247         RETURN(rc);
 248 }
 249
 250 static void osc_page_touch(const struct lu_env *env,
 251                           const struct cl_page_slice *slice, size_t to)
 252 {
 253         struct osc_page *opg = cl2osc_page(slice);
 254         struct cl_object *obj = opg->ops_cl.cpl_obj;
 255
 256         osc_page_touch_at(env, obj, osc_index(opg), to);
 257 }
 258
 259 static const struct cl_page_operations osc_page_ops = {
 260         .cpo_print         = osc_page_print,
 261         .cpo_delete        = osc_page_delete,
 262         .cpo_clip           = osc_page_clip,
 263         .cpo_cancel         = osc_page_cancel,
 264         .cpo_flush          = osc_page_flush,
 265         .cpo_page_touch    = osc_page_touch,
 266 };
 267
 268 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 269                   struct cl_page *page, pgoff_t index)
 270 {
 271         struct osc_object *osc = cl2osc(obj);
 272         struct osc_page   *opg = cl_object_page_slice(obj, page);
 273         struct osc_io *oio = osc_env_io(env);
 274         int result;
 275
 276         opg->ops_from = 0;
 277         opg->ops_to   = PAGE_SIZE;
 278
 279         INIT_LIST_HEAD(&opg->ops_lru);
 280
 281         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 282                                      cl_offset(obj, index));
 283         if (result != 0)
 284                 return result;
 285
 286         opg->ops_srvlock = osc_io_srvlock(oio);
 287         cl_page_slice_add(page, &opg->ops_cl, obj, index,
 288                           &osc_page_ops);
 289
 290
 291         /* reserve an LRU space for this page */
 292         if (page->cp_type == CPT_CACHEABLE) {
 293                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 294                 if (result == 0) {
 295                         result = radix_tree_preload(GFP_NOFS);
 296                         if (result == 0) {
 297                                 spin_lock(&osc->oo_tree_lock);
 298                                 result = radix_tree_insert(&osc->oo_tree,
 299                                                            index, opg);
 300                                 if (result == 0) {
 301                                         ++osc->oo_npages;
 302                                         opg->ops_intree = 1;
 303                                 }
 304                                 spin_unlock(&osc->oo_tree_lock);
 305
 306                                 radix_tree_preload_end();
 307                         }
 308                 }
 309         }
 310
 311         return result;
 312 }
 313 EXPORT_SYMBOL(osc_page_init);
 314
 315 /**
 316  * Helper function called by osc_io_submit() for every page in an immediate
 317  * transfer (i.e., transferred synchronously).
 318  */
 319 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 320                      enum cl_req_type crt, int brw_flags)
 321 {
 322         struct osc_io *oio = osc_env_io(env);
 323         struct osc_async_page *oap = &opg->ops_oap;
 324
 325         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 326                  "magic 0x%x\n", oap, oap->oap_magic);
 327         LASSERT(oap->oap_async_flags & ASYNC_READY);
 328         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 329
 330         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 331         oap->oap_page_off  = opg->ops_from;
 332         oap->oap_count     = opg->ops_to - opg->ops_from;
 333         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 334
 335         if (oio->oi_cap_sys_resource) {
 336                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 337                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 338         }
 339
 340         opg->ops_submit_time = ktime_get();
 341         osc_page_transfer_get(opg, "transfer\0imm");
 342         osc_page_transfer_add(env, opg, crt);
 343 }
 344
 345 /* --------------- LRU page management ------------------ */
 346
 347 /* OSC is a natural place to manage LRU pages as applications are specialized
 348  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 349  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 350  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 351  * for free LRU slots - this will be very bad so the algorithm requires each
 352  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 353  * at any time.
 354  */
 355
 356 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 357
 358 /**
 359  * LRU pages are freed in batch mode. OSC should at least free this
 360  * number of pages to avoid running out of LRU slots.
 361  */
 362 static inline int lru_shrink_min(struct client_obd *cli)
 363 {
 364         return cli->cl_max_pages_per_rpc * 2;
 365 }
 366
 367 /**
 368  * free this number at most otherwise it will take too long time to finsih.
 369  */
 370 static inline int lru_shrink_max(struct client_obd *cli)
 371 {
 372         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 373 }
 374
 375 /**
 376  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 377  * we should free slots aggressively. In this way, slots are freed in a steady
 378  * step to maintain fairness among OSCs.
 379  *
 380  * Return how many LRU pages should be freed.
 381  */
 382 static int osc_cache_too_much(struct client_obd *cli)
 383 {
 384         struct cl_client_cache *cache = cli->cl_cache;
 385         long pages = atomic_long_read(&cli->cl_lru_in_list);
 386         unsigned long budget;
 387
 388         LASSERT(cache != NULL);
 389         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 390
 391         /* if it's going to run out LRU slots, we should free some, but not
 392          * too much to maintain faireness among OSCs. */
 393         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 394                 if (pages >= budget)
 395                         return lru_shrink_max(cli);
 396                 else if (pages >= budget / 2)
 397                         return lru_shrink_min(cli);
 398         } else {
 399                 time64_t duration = ktime_get_real_seconds();
 400                 long timediff;
 401
 402                 /* knock out pages by duration of no IO activity */
 403                 duration -= cli->cl_lru_last_used;
 404                 /*
 405                  * The difference shouldn't be more than 70 years
 406                  * so we can safely case to a long. Round to
 407                  * approximately 1 minute.
 408                  */
 409                 timediff = (long)(duration >> 6);
 410                 if (timediff > 0 && pages >= budget / timediff)
 411                         return lru_shrink_min(cli);
 412         }
 413         return 0;
 414 }
 415
 416 int lru_queue_work(const struct lu_env *env, void *data)
 417 {
 418         struct client_obd *cli = data;
 419         int count;
 420
 421         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 422         count = osc_cache_too_much(cli);
 423         if (count > 0) {
 424                 int rc = osc_lru_shrink(env, cli, count, false);
 425
 426                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 427                        cli_name(cli), rc, count);
 428                 if (rc >= count) {
 429                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 430                         ptlrpcd_queue_work(cli->cl_lru_work);
 431                 }
 432         }
 433
 434         RETURN(0);
 435 }
 436
 437 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 438 {
 439         LIST_HEAD(lru);
 440         struct osc_async_page *oap;
 441         long npages = 0;
 442
 443         list_for_each_entry(oap, plist, oap_pending_item) {
 444                 struct osc_page *opg = oap2osc_page(oap);
 445
 446                 if (!opg->ops_in_lru)
 447                         continue;
 448
 449                 ++npages;
 450                 LASSERT(list_empty(&opg->ops_lru));
 451                 list_add(&opg->ops_lru, &lru);
 452         }
 453
 454         if (npages > 0) {
 455                 spin_lock(&cli->cl_lru_list_lock);
 456                 list_splice_tail(&lru, &cli->cl_lru_list);
 457                 atomic_long_sub(npages, &cli->cl_lru_busy);
 458                 atomic_long_add(npages, &cli->cl_lru_in_list);
 459                 cli->cl_lru_last_used = ktime_get_real_seconds();
 460                 spin_unlock(&cli->cl_lru_list_lock);
 461
 462                 if (waitqueue_active(&osc_lru_waitq))
 463                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 464         }
 465 }
 466
 467 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 468 {
 469         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 470         list_del_init(&opg->ops_lru);
 471         atomic_long_dec(&cli->cl_lru_in_list);
 472 }
 473
 474 /**
 475  * Page is being destroyed. The page may be not in LRU list, if the transfer
 476  * has never finished(error occurred).
 477  */
 478 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 479 {
 480         if (opg->ops_in_lru) {
 481                 spin_lock(&cli->cl_lru_list_lock);
 482                 if (!list_empty(&opg->ops_lru)) {
 483                         __osc_lru_del(cli, opg);
 484                 } else {
 485                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 486                         atomic_long_dec(&cli->cl_lru_busy);
 487                 }
 488                 spin_unlock(&cli->cl_lru_list_lock);
 489
 490                 atomic_long_inc(cli->cl_lru_left);
 491                 /* this is a great place to release more LRU pages if
 492                  * this osc occupies too many LRU pages and kernel is
 493                  * stealing one of them. */
 494                 if (osc_cache_too_much(cli)) {
 495                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 496                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 497                 }
 498                 wake_up(&osc_lru_waitq);
 499         } else {
 500                 LASSERT(list_empty(&opg->ops_lru));
 501         }
 502 }
 503
 504 /**
 505  * Delete page from LRU list for redirty.
 506  */
 507 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 508 {
 509         /* If page is being transferred for the first time,
 510          * ops_lru should be empty */
 511         if (opg->ops_in_lru) {
 512                 if (list_empty(&opg->ops_lru))
 513                         return;
 514                 spin_lock(&cli->cl_lru_list_lock);
 515                 if (!list_empty(&opg->ops_lru)) {
 516                         __osc_lru_del(cli, opg);
 517                         atomic_long_inc(&cli->cl_lru_busy);
 518                 }
 519                 spin_unlock(&cli->cl_lru_list_lock);
 520         }
 521 }
 522
 523 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 524                                 struct cl_page **pvec, int max_index)
 525 {
 526         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 527         int i;
 528
 529         ll_pagevec_init(pagevec, 0);
 530         for (i = 0; i < max_index; i++) {
 531                 struct cl_page *page = pvec[i];
 532
 533                 LASSERT(cl_page_is_owned(page, io));
 534                 cl_page_delete(env, page);
 535                 cl_page_discard(env, io, page);
 536                 cl_page_disown(env, io, page);
 537                 cl_pagevec_put(env, page, pagevec);
 538
 539                 pvec[i] = NULL;
 540         }
 541         pagevec_release(pagevec);
 542 }
 543
 544 /**
 545  * Check if a cl_page can be released, i.e, it's not being used.
 546  *
 547  * If unstable account is turned on, bulk transfer may hold one refcount
 548  * for recovery so we need to check vmpage refcount as well; otherwise,
 549  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 550  */
 551 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 552 {
 553         if (cl_page_in_use_noref(page))
 554                 return true;
 555
 556         if (cli->cl_cache->ccc_unstable_check) {
 557                 struct page *vmpage = cl_page_vmpage(page);
 558
 559                 /* vmpage have two known users: cl_page and VM page cache */
 560                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 561                         return true;
 562         }
 563         return false;
 564 }
 565
 566 /**
 567  * Drop @target of pages from LRU at most.
 568  */
 569 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 570                    long target, bool force)
 571 {
 572         struct cl_io *io;
 573         struct cl_object *clobj = NULL;
 574         struct cl_page **pvec;
 575         struct osc_page *opg;
 576         long count = 0;
 577         int maxscan = 0;
 578         int index = 0;
 579         int rc = 0;
 580         ENTRY;
 581
 582         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 583         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 584                 RETURN(0);
 585
 586         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 587                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 588         if (!force) {
 589                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 590                         RETURN(-EBUSY);
 591
 592                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 593                         atomic_dec(&cli->cl_lru_shrinkers);
 594                         RETURN(-EBUSY);
 595                 }
 596         } else {
 597                 atomic_inc(&cli->cl_lru_shrinkers);
 598         }
 599
 600         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 601         io = osc_env_thread_io(env);
 602
 603         spin_lock(&cli->cl_lru_list_lock);
 604         if (force)
 605                 cli->cl_lru_reclaim++;
 606         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 607         while (!list_empty(&cli->cl_lru_list)) {
 608                 struct cl_page *page;
 609                 bool will_free = false;
 610
 611                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 612                         break;
 613
 614                 if (--maxscan < 0)
 615                         break;
 616
 617                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 618                                  ops_lru);
 619                 page = opg->ops_cl.cpl_page;
 620                 if (lru_page_busy(cli, page)) {
 621                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 622                         continue;
 623                 }
 624
 625                 LASSERT(page->cp_obj != NULL);
 626                 if (clobj != page->cp_obj) {
 627                         struct cl_object *tmp = page->cp_obj;
 628
 629                         cl_object_get(tmp);
 630                         spin_unlock(&cli->cl_lru_list_lock);
 631
 632                         if (clobj != NULL) {
 633                                 discard_pagevec(env, io, pvec, index);
 634                                 index = 0;
 635
 636                                 cl_io_fini(env, io);
 637                                 cl_object_put(env, clobj);
 638                                 clobj = NULL;
 639                         }
 640
 641                         clobj = tmp;
 642                         io->ci_obj = clobj;
 643                         io->ci_ignore_layout = 1;
 644                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 645
 646                         spin_lock(&cli->cl_lru_list_lock);
 647
 648                         if (rc != 0)
 649                                 break;
 650
 651                         ++maxscan;
 652                         continue;
 653                 }
 654
 655                 if (cl_page_own_try(env, io, page) == 0) {
 656                         if (!lru_page_busy(cli, page)) {
 657                                 /* remove it from lru list earlier to avoid
 658                                  * lock contention */
 659                                 __osc_lru_del(cli, opg);
 660                                 opg->ops_in_lru = 0; /* will be discarded */
 661
 662                                 cl_page_get(page);
 663                                 will_free = true;
 664                         } else {
 665                                 cl_page_disown(env, io, page);
 666                         }
 667                 }
 668
 669                 if (!will_free) {
 670                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 671                         continue;
 672                 }
 673
 674                 /* Don't discard and free the page with cl_lru_list held */
 675                 pvec[index++] = page;
 676                 if (unlikely(index == OTI_PVEC_SIZE)) {
 677                         spin_unlock(&cli->cl_lru_list_lock);
 678                         discard_pagevec(env, io, pvec, index);
 679                         index = 0;
 680
 681                         spin_lock(&cli->cl_lru_list_lock);
 682                 }
 683
 684                 if (++count >= target)
 685                         break;
 686         }
 687         spin_unlock(&cli->cl_lru_list_lock);
 688
 689         if (clobj != NULL) {
 690                 discard_pagevec(env, io, pvec, index);
 691
 692                 cl_io_fini(env, io);
 693                 cl_object_put(env, clobj);
 694         }
 695
 696         atomic_dec(&cli->cl_lru_shrinkers);
 697         if (count > 0) {
 698                 atomic_long_add(count, cli->cl_lru_left);
 699                 wake_up_all(&osc_lru_waitq);
 700         }
 701         RETURN(count > 0 ? count : rc);
 702 }
 703 EXPORT_SYMBOL(osc_lru_shrink);
 704
 705 /**
 706  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 707  * \@npages of LRU slots. For performance consideration, it's better to drop
 708  * LRU pages in batch. Therefore, the actual number is adjusted at least
 709  * max_pages_per_rpc.
 710  */
 711 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 712 {
 713         struct lu_env *env;
 714         struct cl_client_cache *cache = cli->cl_cache;
 715         int max_scans;
 716         __u16 refcheck;
 717         long rc = 0;
 718         ENTRY;
 719
 720         LASSERT(cache != NULL);
 721
 722         env = cl_env_get(&refcheck);
 723         if (IS_ERR(env))
 724                 RETURN(rc);
 725
 726         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 727         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 728                cli_name(cli), npages);
 729         rc = osc_lru_shrink(env, cli, npages, true);
 730         if (rc >= npages) {
 731                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 732                        cli_name(cli), rc, npages);
 733                 if (osc_cache_too_much(cli) > 0)
 734                         ptlrpcd_queue_work(cli->cl_lru_work);
 735                 GOTO(out, rc);
 736         } else if (rc > 0) {
 737                 npages -= rc;
 738         }
 739
 740         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 741                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 742                 atomic_long_read(&cli->cl_lru_busy), npages);
 743
 744         /* Reclaim LRU slots from other client_obd as it can't free enough
 745          * from its own. This should rarely happen. */
 746         spin_lock(&cache->ccc_lru_lock);
 747         LASSERT(!list_empty(&cache->ccc_lru));
 748
 749         cache->ccc_lru_shrinkers++;
 750         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 751
 752         max_scans = atomic_read(&cache->ccc_users) - 2;
 753         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 754                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 755                                  cl_lru_osc);
 756
 757                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 758                         cli_name(cli), cli,
 759                         atomic_long_read(&cli->cl_lru_in_list),
 760                         atomic_long_read(&cli->cl_lru_busy));
 761
 762                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 763                 if (osc_cache_too_much(cli) > 0) {
 764                         spin_unlock(&cache->ccc_lru_lock);
 765
 766                         rc = osc_lru_shrink(env, cli, npages, true);
 767                         spin_lock(&cache->ccc_lru_lock);
 768                         if (rc >= npages)
 769                                 break;
 770                         if (rc > 0)
 771                                 npages -= rc;
 772                 }
 773         }
 774         spin_unlock(&cache->ccc_lru_lock);
 775
 776 out:
 777         cl_env_put(env, &refcheck);
 778         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 779                 cli_name(cli), cli, rc);
 780         return rc;
 781 }
 782
 783 /**
 784  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 785  *
 786  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 787  * Only in the case that the LRU slots are in extreme shortage, it should
 788  * have reserved enough slots for an IO.
 789  */
 790 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 791                          struct osc_page *opg)
 792 {
 793         struct osc_io *oio = osc_env_io(env);
 794         int rc = 0;
 795
 796         ENTRY;
 797
 798         if (cli->cl_cache == NULL) /* shall not be in LRU */
 799                 RETURN(0);
 800
 801         if (oio->oi_lru_reserved > 0) {
 802                 --oio->oi_lru_reserved;
 803                 goto out;
 804         }
 805
 806         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 807         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 808                 /* run out of LRU spaces, try to drop some by itself */
 809                 rc = osc_lru_reclaim(cli, 1);
 810                 if (rc < 0)
 811                         break;
 812                 if (rc > 0)
 813                         continue;
 814
 815                 cond_resched();
 816                 rc = l_wait_event_abortable(
 817                         osc_lru_waitq,
 818                         atomic_long_read(cli->cl_lru_left) > 0);
 819                 if (rc < 0) {
 820                         rc = -EINTR;
 821                         break;
 822                 }
 823         }
 824
 825 out:
 826         if (rc >= 0) {
 827                 atomic_long_inc(&cli->cl_lru_busy);
 828                 opg->ops_in_lru = 1;
 829                 rc = 0;
 830         }
 831
 832         RETURN(rc);
 833 }
 834
 835 /**
 836  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 837  *
 838  * The benefit of doing this is to reduce contention against atomic counter
 839  * cl_lru_left by changing it from per-page access to per-IO access.
 840  */
 841 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 842 {
 843         unsigned long reserved = 0;
 844         unsigned long max_pages;
 845         unsigned long c;
 846
 847         /* reserve a full RPC window at most to avoid that a thread accidentally
 848          * consumes too many LRU slots */
 849         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 850         if (npages > max_pages)
 851                 npages = max_pages;
 852
 853         c = atomic_long_read(cli->cl_lru_left);
 854         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 855                 c = atomic_long_read(cli->cl_lru_left);
 856         while (c >= npages) {
 857                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 858                         reserved = npages;
 859                         break;
 860                 }
 861                 c = atomic_long_read(cli->cl_lru_left);
 862         }
 863         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 864                 /* If there aren't enough pages in the per-OSC LRU then
 865                  * wake up the LRU thread to try and clear out space, so
 866                  * we don't block if pages are being dirtied quickly. */
 867                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 868                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 869                        max_pages);
 870                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 871         }
 872
 873         return reserved;
 874 }
 875
 876 /**
 877  * osc_lru_unreserve() is called to unreserve LRU slots.
 878  *
 879  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 880  * reasons such as page already existing or I/O error. Those reserved slots
 881  * should be freed by calling this function.
 882  */
 883 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 884 {
 885         atomic_long_add(npages, cli->cl_lru_left);
 886         wake_up_all(&osc_lru_waitq);
 887 }
 888
 889 /**
 890  * Atomic operations are expensive. We accumulate the accounting for the
 891  * same page zone to get better performance.
 892  * In practice this can work pretty good because the pages in the same RPC
 893  * are likely from the same page zone.
 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                 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 906                 page_count = desc->bd_iov_count;
 907         } else {
 908                 page_count = aa->aa_page_count;
 909         }
 910
 911         for (i = 0; i < page_count; i++) {
 912                 void *pz;
 913                 if (desc)
 914                         pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 915                 else
 916                         pz = page_zone(aa->aa_ppga[i]->pg);
 917
 918                 if (likely(pz == zone)) {
 919                         ++count;
 920                         continue;
 921                 }
 922
 923                 if (count > 0) {
 924                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 925                                             factor * count);
 926                         count = 0;
 927                 }
 928                 zone = pz;
 929                 ++count;
 930         }
 931         if (count > 0)
 932                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 933 }
 934
 935 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 936                                                 struct osc_brw_async_args *aa)
 937 {
 938         unstable_page_accounting(desc, aa, 1);
 939 }
 940
 941 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 942                                                 struct osc_brw_async_args *aa)
 943 {
 944         unstable_page_accounting(desc, aa, -1);
 945 }
 946
 947 /**
 948  * Performs "unstable" page accounting. This function balances the
 949  * increment operations performed in osc_inc_unstable_pages. It is
 950  * registered as the RPC request callback, and is executed when the
 951  * bulk RPC is committed on the server. Thus at this point, the pages
 952  * involved in the bulk transfer are no longer considered unstable.
 953  *
 954  * If this function is called, the request should have been committed
 955  * or req:rq_unstable must have been set; it implies that the unstable
 956  * statistic have been added.
 957  */
 958 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 959 {
 960         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 961         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 962         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 963         int                      page_count;
 964         long                     unstable_count;
 965
 966         if (desc)
 967                 page_count = desc->bd_iov_count;
 968         else
 969                 page_count = aa->aa_page_count;
 970
 971         LASSERT(page_count >= 0);
 972
 973         dec_unstable_page_accounting(desc, aa);
 974
 975         unstable_count = atomic_long_sub_return(page_count,
 976                                                 &cli->cl_unstable_count);
 977         LASSERT(unstable_count >= 0);
 978
 979         unstable_count = atomic_long_sub_return(page_count,
 980                                            &cli->cl_cache->ccc_unstable_nr);
 981         LASSERT(unstable_count >= 0);
 982         if (unstable_count == 0)
 983                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 984
 985         if (waitqueue_active(&osc_lru_waitq))
 986                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 987 }
 988
 989 /**
 990  * "unstable" page accounting. See: osc_dec_unstable_pages.
 991  */
 992 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 993 {
 994         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 995         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 996         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 997         long                     page_count;
 998
 999         /* No unstable page tracking */
1000         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1001                 return;
1002
1003         if (desc)
1004                 page_count = desc->bd_iov_count;
1005         else
1006                 page_count = aa->aa_page_count;
1007
1008         add_unstable_page_accounting(desc, aa);
1009         atomic_long_add(page_count, &cli->cl_unstable_count);
1010         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1011
1012         /* If the request has already been committed (i.e. brw_commit
1013          * called via rq_commit_cb), we need to undo the unstable page
1014          * increments we just performed because rq_commit_cb wont be
1015          * called again. */
1016         spin_lock(&req->rq_lock);
1017         if (unlikely(req->rq_committed)) {
1018                 spin_unlock(&req->rq_lock);
1019
1020                 osc_dec_unstable_pages(req);
1021         } else {
1022                 req->rq_unstable = 1;
1023                 spin_unlock(&req->rq_lock);
1024         }
1025 }
1026
1027 /**
1028  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1029  * This function will be called by every BRW RPC so it's critical
1030  * to make this function fast.
1031  */
1032 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1033 {
1034         long unstable_nr, osc_unstable_count;
1035
1036         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1037         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1038                 return false;
1039
1040         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1041         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1042
1043         CDEBUG(D_CACHE,
1044                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1045                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1046
1047         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1048          * has one full RPC window of unstable pages, it's a good chance
1049          * to piggyback a SOFT_SYNC flag.
1050          * Please notice that the OST won't take immediate response for the
1051          * SOFT_SYNC request so active OSCs will have more chance to carry
1052          * the flag, this is reasonable. */
1053         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1054                osc_unstable_count > cli->cl_max_pages_per_rpc *
1055                                     cli->cl_max_rpcs_in_flight;
1056 }
1057
1058 /**
1059  * Return how many LRU pages in the cache of all OSC devices
1060  *
1061  * \retval      return # of cached LRU pages times reclaimation tendency
1062  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1063  */
1064 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1065                                      struct shrink_control *sc)
1066 {
1067         struct client_obd *cli;
1068         unsigned long cached = 0;
1069
1070         spin_lock(&osc_shrink_lock);
1071         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1072                 cached += atomic_long_read(&cli->cl_lru_in_list);
1073         spin_unlock(&osc_shrink_lock);
1074
1075         return (cached  * sysctl_vfs_cache_pressure) / 100;
1076 }
1077
1078 /**
1079  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1080  *
1081  * \retval      number of cached LRU pages reclaimed
1082  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1083  *
1084  * Linux kernel will loop calling this shrinker scan routine with
1085  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1086  *
1087  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1088  * to scan and try to reclaim LRU pages, just return 0 and
1089  * osc_cache_shrink_count() will report the LRU page number.
1090  */
1091 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1092                                     struct shrink_control *sc)
1093 {
1094         struct client_obd *cli;
1095         struct client_obd *stop_anchor = NULL;
1096         struct lu_env *env;
1097         long shrank = 0;
1098         int rc;
1099         __u16 refcheck;
1100
1101         if (sc->nr_to_scan == 0)
1102                 return 0;
1103
1104         if (!(sc->gfp_mask & __GFP_FS))
1105                 return SHRINK_STOP;
1106
1107         env = cl_env_get(&refcheck);
1108         if (IS_ERR(env))
1109                 return SHRINK_STOP;
1110
1111         spin_lock(&osc_shrink_lock);
1112         while (!list_empty(&osc_shrink_list)) {
1113                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1114                                  cl_shrink_list);
1115
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 */