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 = cl_offset(obj, start);
 110         policy->l_extent.end   = cl_offset(obj, end + 1) - 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< %#x %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_magic, 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         opg->ops_from = from;
 208         /* argument @to is exclusive, but @ops_to is inclusive */
 209         opg->ops_to   = to - 1;
 210         /* This isn't really necessary for transient pages, but we also don't
 211          * call clip on transient pages often, so it's OK.
 212          */
 213         spin_lock(&oap->oap_lock);
 214         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 215         spin_unlock(&oap->oap_lock);
 216 }
 217
 218 static int osc_page_flush(const struct lu_env *env,
 219                           const struct cl_page_slice *slice,
 220                           struct cl_io *io)
 221 {
 222         struct osc_page *opg = cl2osc_page(slice);
 223         int rc = 0;
 224         ENTRY;
 225         rc = osc_flush_async_page(env, io, opg);
 226         RETURN(rc);
 227 }
 228
 229 static void osc_page_touch(const struct lu_env *env,
 230                           const struct cl_page_slice *slice, size_t to)
 231 {
 232         struct osc_page *opg = cl2osc_page(slice);
 233         struct cl_object *obj = osc2cl(osc_page_object(opg));
 234
 235         osc_page_touch_at(env, obj, osc_index(opg), to);
 236 }
 237
 238 static const struct cl_page_operations osc_page_ops = {
 239         .cpo_print         = osc_page_print,
 240         .cpo_delete        = osc_page_delete,
 241         .cpo_clip           = osc_page_clip,
 242         .cpo_flush          = osc_page_flush,
 243         .cpo_page_touch    = osc_page_touch,
 244 };
 245
 246 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 247                   struct cl_page *cl_page, pgoff_t index)
 248 {
 249         struct osc_object *osc = cl2osc(obj);
 250         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 251         struct osc_io *oio = osc_env_io(env);
 252         int result;
 253
 254         opg->ops_from = 0;
 255         opg->ops_to = PAGE_SIZE - 1;
 256
 257         INIT_LIST_HEAD(&opg->ops_lru);
 258
 259         result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
 260         if (result != 0)
 261                 return result;
 262
 263         opg->ops_srvlock = osc_io_srvlock(oio);
 264         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 265
 266         /* reserve an LRU space for this page */
 267         if (cl_page->cp_type == CPT_CACHEABLE) {
 268                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 269                 if (result == 0) {
 270                         result = radix_tree_preload(GFP_NOFS);
 271                         if (result == 0) {
 272                                 spin_lock(&osc->oo_tree_lock);
 273                                 result = radix_tree_insert(&osc->oo_tree,
 274                                                            index, opg);
 275                                 if (result == 0) {
 276                                         ++osc->oo_npages;
 277                                         opg->ops_intree = 1;
 278                                 }
 279                                 spin_unlock(&osc->oo_tree_lock);
 280
 281                                 radix_tree_preload_end();
 282                         }
 283                 }
 284         }
 285
 286         return result;
 287 }
 288 EXPORT_SYMBOL(osc_page_init);
 289
 290 /**
 291  * Helper function called by osc_io_submit() for every page in an immediate
 292  * transfer (i.e., transferred synchronously).
 293  */
 294 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 295                      enum cl_req_type crt, int brw_flags)
 296 {
 297         struct osc_io *oio = osc_env_io(env);
 298         struct osc_async_page *oap = &opg->ops_oap;
 299
 300         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 301                  "magic 0x%x\n", oap, oap->oap_magic);
 302         LASSERT(oap->oap_async_flags & ASYNC_READY);
 303         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 304
 305         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 306         oap->oap_page_off = opg->ops_from;
 307         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 308         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 309
 310         if (oio->oi_cap_sys_resource) {
 311                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 312                 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
 313         }
 314
 315         osc_page_transfer_get(opg, "transfer\0imm");
 316         osc_page_transfer_add(env, opg, crt);
 317 }
 318
 319 /* --------------- LRU page management ------------------ */
 320
 321 /* OSC is a natural place to manage LRU pages as applications are specialized
 322  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 323  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 324  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 325  * for free LRU slots - this will be very bad so the algorithm requires each
 326  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 327  * at any time.
 328  */
 329
 330 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 331
 332 /**
 333  * LRU pages are freed in batch mode. OSC should at least free this
 334  * number of pages to avoid running out of LRU slots.
 335  */
 336 static inline int lru_shrink_min(struct client_obd *cli)
 337 {
 338         return cli->cl_max_pages_per_rpc * 2;
 339 }
 340
 341 /**
 342  * free this number at most otherwise it will take too long time to finsih.
 343  */
 344 static inline int lru_shrink_max(struct client_obd *cli)
 345 {
 346         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 347 }
 348
 349 /**
 350  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 351  * we should free slots aggressively. In this way, slots are freed in a steady
 352  * step to maintain fairness among OSCs.
 353  *
 354  * Return how many LRU pages should be freed.
 355  */
 356 static int osc_cache_too_much(struct client_obd *cli)
 357 {
 358         struct cl_client_cache *cache = cli->cl_cache;
 359         long pages = atomic_long_read(&cli->cl_lru_in_list);
 360         unsigned long budget;
 361
 362         LASSERT(cache != NULL);
 363         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 364
 365         /* if it's going to run out LRU slots, we should free some, but not
 366          * too much to maintain faireness among OSCs. */
 367         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 368                 if (pages >= budget)
 369                         return lru_shrink_max(cli);
 370                 else if (pages >= budget / 2)
 371                         return lru_shrink_min(cli);
 372         } else {
 373                 time64_t duration = ktime_get_real_seconds();
 374                 long timediff;
 375
 376                 /* knock out pages by duration of no IO activity */
 377                 duration -= cli->cl_lru_last_used;
 378                 /*
 379                  * The difference shouldn't be more than 70 years
 380                  * so we can safely case to a long. Round to
 381                  * approximately 1 minute.
 382                  */
 383                 timediff = (long)(duration >> 6);
 384                 if (timediff > 0 && pages >= budget / timediff)
 385                         return lru_shrink_min(cli);
 386         }
 387         return 0;
 388 }
 389
 390 int lru_queue_work(const struct lu_env *env, void *data)
 391 {
 392         struct client_obd *cli = data;
 393         int count;
 394
 395         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 396         count = osc_cache_too_much(cli);
 397         if (count > 0) {
 398                 int rc = osc_lru_shrink(env, cli, count, false);
 399
 400                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 401                        cli_name(cli), rc, count);
 402                 if (rc >= count) {
 403                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 404                         ptlrpcd_queue_work(cli->cl_lru_work);
 405                 }
 406         }
 407
 408         RETURN(0);
 409 }
 410
 411 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 412 {
 413         LIST_HEAD(lru);
 414         struct osc_async_page *oap;
 415         long npages = 0;
 416
 417         list_for_each_entry(oap, plist, oap_pending_item) {
 418                 struct osc_page *opg = oap2osc_page(oap);
 419
 420                 if (!opg->ops_in_lru)
 421                         continue;
 422
 423                 ++npages;
 424                 LASSERT(list_empty(&opg->ops_lru));
 425                 list_add(&opg->ops_lru, &lru);
 426         }
 427
 428         if (npages > 0) {
 429                 spin_lock(&cli->cl_lru_list_lock);
 430                 list_splice_tail(&lru, &cli->cl_lru_list);
 431                 atomic_long_sub(npages, &cli->cl_lru_busy);
 432                 atomic_long_add(npages, &cli->cl_lru_in_list);
 433                 cli->cl_lru_last_used = ktime_get_real_seconds();
 434                 spin_unlock(&cli->cl_lru_list_lock);
 435
 436                 if (waitqueue_active(&osc_lru_waitq))
 437                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 438         }
 439 }
 440
 441 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 442 {
 443         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 444         list_del_init(&opg->ops_lru);
 445         atomic_long_dec(&cli->cl_lru_in_list);
 446 }
 447
 448 /**
 449  * Page is being destroyed. The page may be not in LRU list, if the transfer
 450  * has never finished(error occurred).
 451  */
 452 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 453 {
 454         if (opg->ops_in_lru) {
 455                 spin_lock(&cli->cl_lru_list_lock);
 456                 if (!list_empty(&opg->ops_lru)) {
 457                         __osc_lru_del(cli, opg);
 458                 } else {
 459                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 460                         atomic_long_dec(&cli->cl_lru_busy);
 461                 }
 462                 spin_unlock(&cli->cl_lru_list_lock);
 463
 464                 atomic_long_inc(cli->cl_lru_left);
 465                 /* this is a great place to release more LRU pages if
 466                  * this osc occupies too many LRU pages and kernel is
 467                  * stealing one of them. */
 468                 if (osc_cache_too_much(cli)) {
 469                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 470                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 471                 }
 472                 wake_up(&osc_lru_waitq);
 473         } else {
 474                 LASSERT(list_empty(&opg->ops_lru));
 475         }
 476 }
 477
 478 /**
 479  * Delete page from LRU list for redirty.
 480  */
 481 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 482 {
 483         /* If page is being transferred for the first time,
 484          * ops_lru should be empty */
 485         if (opg->ops_in_lru) {
 486                 if (list_empty(&opg->ops_lru))
 487                         return;
 488                 spin_lock(&cli->cl_lru_list_lock);
 489                 if (!list_empty(&opg->ops_lru)) {
 490                         __osc_lru_del(cli, opg);
 491                         atomic_long_inc(&cli->cl_lru_busy);
 492                 }
 493                 spin_unlock(&cli->cl_lru_list_lock);
 494         }
 495 }
 496
 497 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 498                                 struct cl_page **pvec, int max_index)
 499 {
 500         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 501         int i;
 502
 503         ll_pagevec_init(pagevec, 0);
 504         for (i = 0; i < max_index; i++) {
 505                 struct cl_page *page = pvec[i];
 506
 507                 LASSERT(cl_page_is_owned(page, io));
 508                 cl_page_delete(env, page);
 509                 cl_page_discard(env, io, page);
 510                 cl_page_disown(env, io, page);
 511                 cl_pagevec_put(env, page, pagevec);
 512
 513                 pvec[i] = NULL;
 514         }
 515         pagevec_release(pagevec);
 516 }
 517
 518 /**
 519  * Check if a cl_page can be released, i.e, it's not being used.
 520  *
 521  * If unstable account is turned on, bulk transfer may hold one refcount
 522  * for recovery so we need to check vmpage refcount as well; otherwise,
 523  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 524  */
 525 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 526 {
 527         if (cl_page_in_use_noref(page))
 528                 return true;
 529
 530         if (cli->cl_cache->ccc_unstable_check) {
 531                 struct page *vmpage = cl_page_vmpage(page);
 532
 533                 /* this check is racy because the vmpage is not locked, but
 534                  * that's OK - the code which does the actual page release
 535                  * checks this again before releasing
 536                  *
 537                  * vmpage have two known users: cl_page and VM page cache
 538                  */
 539                 if (vmpage_in_use(vmpage, 0))
 540                         return true;
 541         }
 542         return false;
 543 }
 544
 545 /**
 546  * Drop @target of pages from LRU at most.
 547  */
 548 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 549                    long target, bool force)
 550 {
 551         struct cl_io *io;
 552         struct cl_object *clobj = NULL;
 553         struct cl_page **pvec;
 554         struct osc_page *opg;
 555         long count = 0;
 556         int maxscan = 0;
 557         int index = 0;
 558         int rc = 0;
 559         ENTRY;
 560
 561         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 562         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 563                 RETURN(0);
 564
 565         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 566                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 567         if (!force) {
 568                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 569                         RETURN(-EBUSY);
 570
 571                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 572                         atomic_dec(&cli->cl_lru_shrinkers);
 573                         RETURN(-EBUSY);
 574                 }
 575         } else {
 576                 atomic_inc(&cli->cl_lru_shrinkers);
 577         }
 578
 579         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 580         io = osc_env_thread_io(env);
 581
 582         spin_lock(&cli->cl_lru_list_lock);
 583         if (force)
 584                 cli->cl_lru_reclaim++;
 585         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 586         while (!list_empty(&cli->cl_lru_list)) {
 587                 struct cl_page *page;
 588                 bool will_free = false;
 589
 590                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 591                         break;
 592
 593                 if (--maxscan < 0)
 594                         break;
 595
 596                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 597                                        ops_lru);
 598                 page = opg->ops_cl.cpl_page;
 599                 if (lru_page_busy(cli, page)) {
 600                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 601                         continue;
 602                 }
 603
 604                 LASSERT(page->cp_obj != NULL);
 605                 if (clobj != page->cp_obj) {
 606                         struct cl_object *tmp = page->cp_obj;
 607
 608                         cl_object_get(tmp);
 609                         spin_unlock(&cli->cl_lru_list_lock);
 610
 611                         if (clobj != NULL) {
 612                                 discard_pagevec(env, io, pvec, index);
 613                                 index = 0;
 614
 615                                 cl_io_fini(env, io);
 616                                 cl_object_put(env, clobj);
 617                                 clobj = NULL;
 618                         }
 619
 620                         clobj = tmp;
 621                         io->ci_obj = clobj;
 622                         io->ci_ignore_layout = 1;
 623                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 624
 625                         spin_lock(&cli->cl_lru_list_lock);
 626
 627                         if (rc != 0)
 628                                 break;
 629
 630                         ++maxscan;
 631                         continue;
 632                 }
 633
 634                 if (cl_page_own_try(env, io, page) == 0) {
 635                         if (!lru_page_busy(cli, page)) {
 636                                 /* remove it from lru list earlier to avoid
 637                                  * lock contention */
 638                                 __osc_lru_del(cli, opg);
 639                                 opg->ops_in_lru = 0; /* will be discarded */
 640
 641                                 cl_page_get(page);
 642                                 will_free = true;
 643                         } else {
 644                                 cl_page_disown(env, io, page);
 645                         }
 646                 }
 647
 648                 if (!will_free) {
 649                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 650                         continue;
 651                 }
 652
 653                 /* Don't discard and free the page with cl_lru_list held */
 654                 pvec[index++] = page;
 655                 if (unlikely(index == OTI_PVEC_SIZE)) {
 656                         spin_unlock(&cli->cl_lru_list_lock);
 657                         discard_pagevec(env, io, pvec, index);
 658                         index = 0;
 659
 660                         spin_lock(&cli->cl_lru_list_lock);
 661                 }
 662
 663                 if (++count >= target)
 664                         break;
 665         }
 666         spin_unlock(&cli->cl_lru_list_lock);
 667
 668         if (clobj != NULL) {
 669                 discard_pagevec(env, io, pvec, index);
 670
 671                 cl_io_fini(env, io);
 672                 cl_object_put(env, clobj);
 673         }
 674
 675         atomic_dec(&cli->cl_lru_shrinkers);
 676         if (count > 0) {
 677                 atomic_long_add(count, cli->cl_lru_left);
 678                 wake_up(&osc_lru_waitq);
 679         }
 680         RETURN(count > 0 ? count : rc);
 681 }
 682 EXPORT_SYMBOL(osc_lru_shrink);
 683
 684 /**
 685  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 686  * \@npages of LRU slots. For performance consideration, it's better to drop
 687  * LRU pages in batch. Therefore, the actual number is adjusted at least
 688  * max_pages_per_rpc.
 689  */
 690 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 691 {
 692         struct lu_env *env;
 693         struct cl_client_cache *cache = cli->cl_cache;
 694         struct client_obd *scan;
 695         int max_scans;
 696         __u16 refcheck;
 697         long rc = 0;
 698         ENTRY;
 699
 700         LASSERT(cache != NULL);
 701
 702         env = cl_env_get(&refcheck);
 703         if (IS_ERR(env))
 704                 RETURN(rc);
 705
 706         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 707         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 708                cli_name(cli), npages);
 709         rc = osc_lru_shrink(env, cli, npages, true);
 710         if (rc >= npages) {
 711                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 712                        cli_name(cli), rc, npages);
 713                 if (osc_cache_too_much(cli) > 0)
 714                         ptlrpcd_queue_work(cli->cl_lru_work);
 715                 GOTO(out, rc);
 716         } else if (rc > 0) {
 717                 npages -= rc;
 718         }
 719
 720         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 721                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 722                 atomic_long_read(&cli->cl_lru_busy), npages);
 723
 724         /* Reclaim LRU slots from other client_obd as it can't free enough
 725          * from its own. This should rarely happen. */
 726         spin_lock(&cache->ccc_lru_lock);
 727         LASSERT(!list_empty(&cache->ccc_lru));
 728
 729         cache->ccc_lru_shrinkers++;
 730         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 731
 732         max_scans = atomic_read(&cache->ccc_users) - 2;
 733         while (--max_scans > 0 &&
 734                (scan = list_first_entry_or_null(&cache->ccc_lru,
 735                                                   struct client_obd,
 736                                                   cl_lru_osc)) != NULL) {
 737                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 738                        cli_name(scan), scan,
 739                        atomic_long_read(&scan->cl_lru_in_list),
 740                        atomic_long_read(&scan->cl_lru_busy));
 741
 742                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 743                 if (osc_cache_too_much(scan) > 0) {
 744                         spin_unlock(&cache->ccc_lru_lock);
 745
 746                         rc = osc_lru_shrink(env, scan, npages, true);
 747                         spin_lock(&cache->ccc_lru_lock);
 748                         if (rc >= npages)
 749                                 break;
 750                         if (rc > 0)
 751                                 npages -= rc;
 752                 }
 753         }
 754         spin_unlock(&cache->ccc_lru_lock);
 755
 756 out:
 757         cl_env_put(env, &refcheck);
 758         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 759                cli_name(cli), cli, rc);
 760         return rc;
 761 }
 762
 763 /**
 764  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 765  *
 766  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 767  * Only in the case that the LRU slots are in extreme shortage, it should
 768  * have reserved enough slots for an IO.
 769  */
 770 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 771                          struct osc_page *opg)
 772 {
 773         struct osc_io *oio = osc_env_io(env);
 774         int rc = 0;
 775
 776         ENTRY;
 777
 778         if (cli->cl_cache == NULL) /* shall not be in LRU */
 779                 RETURN(0);
 780
 781         if (oio->oi_lru_reserved > 0) {
 782                 --oio->oi_lru_reserved;
 783                 goto out;
 784         }
 785
 786         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 787         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 788                 /* run out of LRU spaces, try to drop some by itself */
 789                 rc = osc_lru_reclaim(cli, 1);
 790                 if (rc < 0)
 791                         break;
 792                 if (rc > 0)
 793                         continue;
 794                 /* IO issued by readahead, don't try hard */
 795                 if (oio->oi_is_readahead) {
 796                         if (atomic_long_read(cli->cl_lru_left) > 0)
 797                                 continue;
 798                         rc = -EBUSY;
 799                         break;
 800                 }
 801
 802                 cond_resched();
 803                 rc = l_wait_event_abortable(
 804                         osc_lru_waitq,
 805                         atomic_long_read(cli->cl_lru_left) > 0);
 806                 if (rc < 0) {
 807                         rc = -EINTR;
 808                         break;
 809                 }
 810         }
 811
 812 out:
 813         if (rc >= 0) {
 814                 atomic_long_inc(&cli->cl_lru_busy);
 815                 opg->ops_in_lru = 1;
 816                 rc = 0;
 817         }
 818
 819         RETURN(rc);
 820 }
 821
 822 /**
 823  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 824  *
 825  * The benefit of doing this is to reduce contention against atomic counter
 826  * cl_lru_left by changing it from per-page access to per-IO access.
 827  */
 828 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 829 {
 830         unsigned long reserved = 0;
 831         unsigned long max_pages;
 832         unsigned long c;
 833         int rc;
 834
 835 again:
 836         c = atomic_long_read(cli->cl_lru_left);
 837         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 838                 c = atomic_long_read(cli->cl_lru_left);
 839
 840         if (c < npages) {
 841                 /*
 842                  * Trigger writeback in the hope some LRU slot could
 843                  * be freed.
 844                  */
 845                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 846                 if (rc)
 847                         return 0;
 848         }
 849
 850         while (c >= npages) {
 851                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 852                         reserved = npages;
 853                         break;
 854                 }
 855                 c = atomic_long_read(cli->cl_lru_left);
 856         }
 857
 858         if (reserved != npages) {
 859                 cond_resched();
 860                 rc = l_wait_event_abortable(
 861                         osc_lru_waitq,
 862                         atomic_long_read(cli->cl_lru_left) > 0);
 863                 goto again;
 864         }
 865
 866         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 867         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 868                 /* If there aren't enough pages in the per-OSC LRU then
 869                  * wake up the LRU thread to try and clear out space, so
 870                  * we don't block if pages are being dirtied quickly. */
 871                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 872                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 873                        max_pages);
 874                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 875         }
 876
 877         return reserved;
 878 }
 879
 880 /**
 881  * osc_lru_unreserve() is called to unreserve LRU slots.
 882  *
 883  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 884  * reasons such as page already existing or I/O error. Those reserved slots
 885  * should be freed by calling this function.
 886  */
 887 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 888 {
 889         atomic_long_add(npages, cli->cl_lru_left);
 890         wake_up(&osc_lru_waitq);
 891 }
 892
 893 /**
 894  * Atomic operations are expensive. We accumulate the accounting for the
 895  * same page zone to get better performance.
 896  * In practice this can work pretty good because the pages in the same RPC
 897  * are likely from the same page zone.
 898  */
 899 #ifdef HAVE_NR_UNSTABLE_NFS
 900 /* Old kernels use a separate counter for unstable pages,
 901  * newer kernels treat them like any other writeback.
 902  */
 903 #define NR_WRITEBACK NR_UNSTABLE_NFS
 904 #endif
 905
 906 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 907                                             struct osc_brw_async_args *aa,
 908                                             int factor)
 909 {
 910         int page_count;
 911         void *zone = NULL;
 912         int count = 0;
 913         int i;
 914
 915         if (desc != NULL) {
 916                 page_count = desc->bd_iov_count;
 917         } else {
 918                 page_count = aa->aa_page_count;
 919         }
 920
 921         for (i = 0; i < page_count; i++) {
 922                 void *pz;
 923                 if (desc)
 924                         pz = page_zone(desc->bd_vec[i].bv_page);
 925                 else
 926                         pz = page_zone(aa->aa_ppga[i]->pg);
 927
 928                 if (likely(pz == zone)) {
 929                         ++count;
 930                         continue;
 931                 }
 932
 933                 if (count > 0) {
 934                         mod_zone_page_state(zone, NR_WRITEBACK,
 935                                             factor * count);
 936                         count = 0;
 937                 }
 938                 zone = pz;
 939                 ++count;
 940         }
 941         if (count > 0)
 942                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 943 }
 944
 945 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 946                                                 struct osc_brw_async_args *aa)
 947 {
 948         unstable_page_accounting(desc, aa, 1);
 949 }
 950
 951 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 952                                                 struct osc_brw_async_args *aa)
 953 {
 954         unstable_page_accounting(desc, aa, -1);
 955 }
 956
 957 /**
 958  * Performs "unstable" page accounting. This function balances the
 959  * increment operations performed in osc_inc_unstable_pages. It is
 960  * registered as the RPC request callback, and is executed when the
 961  * bulk RPC is committed on the server. Thus at this point, the pages
 962  * involved in the bulk transfer are no longer considered unstable.
 963  *
 964  * If this function is called, the request should have been committed
 965  * or req:rq_unstable must have been set; it implies that the unstable
 966  * statistic have been added.
 967  */
 968 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 969 {
 970         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 971         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 972         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 973         int                      page_count;
 974         long                     unstable_count;
 975
 976         if (desc)
 977                 page_count = desc->bd_iov_count;
 978         else
 979                 page_count = aa->aa_page_count;
 980
 981         LASSERT(page_count >= 0);
 982
 983         dec_unstable_page_accounting(desc, aa);
 984
 985         unstable_count = atomic_long_sub_return(page_count,
 986                                                 &cli->cl_unstable_count);
 987         LASSERT(unstable_count >= 0);
 988
 989         unstable_count = atomic_long_sub_return(page_count,
 990                                            &cli->cl_cache->ccc_unstable_nr);
 991         LASSERT(unstable_count >= 0);
 992         if (unstable_count == 0)
 993                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 994
 995         if (waitqueue_active(&osc_lru_waitq))
 996                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 997 }
 998
 999 /**
1000  * "unstable" page accounting. See: osc_dec_unstable_pages.
1001  */
1002 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1003 {
1004         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1005         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1006         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1007         long                     page_count;
1008
1009         /* No unstable page tracking */
1010         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1011                 return;
1012
1013         if (desc)
1014                 page_count = desc->bd_iov_count;
1015         else
1016                 page_count = aa->aa_page_count;
1017
1018         add_unstable_page_accounting(desc, aa);
1019         atomic_long_add(page_count, &cli->cl_unstable_count);
1020         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1021
1022         /* If the request has already been committed (i.e. brw_commit
1023          * called via rq_commit_cb), we need to undo the unstable page
1024          * increments we just performed because rq_commit_cb wont be
1025          * called again. */
1026         spin_lock(&req->rq_lock);
1027         if (unlikely(req->rq_committed)) {
1028                 spin_unlock(&req->rq_lock);
1029
1030                 osc_dec_unstable_pages(req);
1031         } else {
1032                 req->rq_unstable = 1;
1033                 spin_unlock(&req->rq_lock);
1034         }
1035 }
1036
1037 /**
1038  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1039  * This function will be called by every BRW RPC so it's critical
1040  * to make this function fast.
1041  */
1042 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1043 {
1044         long unstable_nr, osc_unstable_count;
1045
1046         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1047         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1048                 return false;
1049
1050         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1051         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1052
1053         CDEBUG(D_CACHE,
1054                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1055                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1056
1057         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1058          * has one full RPC window of unstable pages, it's a good chance
1059          * to piggyback a SOFT_SYNC flag.
1060          * Please notice that the OST won't take immediate response for the
1061          * SOFT_SYNC request so active OSCs will have more chance to carry
1062          * the flag, this is reasonable. */
1063         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1064                osc_unstable_count > cli->cl_max_pages_per_rpc *
1065                                     cli->cl_max_rpcs_in_flight;
1066 }
1067
1068 /**
1069  * Return how many LRU pages in the cache of all OSC devices
1070  *
1071  * \retval      return # of cached LRU pages times reclaimation tendency
1072  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1073  */
1074 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1075                                      struct shrink_control *sc)
1076 {
1077         struct client_obd *cli;
1078         unsigned long cached = 0;
1079
1080         spin_lock(&osc_shrink_lock);
1081         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1082                 cached += atomic_long_read(&cli->cl_lru_in_list);
1083         spin_unlock(&osc_shrink_lock);
1084
1085         return (cached  * sysctl_vfs_cache_pressure) / 100;
1086 }
1087
1088 /**
1089  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1090  *
1091  * \retval      number of cached LRU pages reclaimed
1092  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1093  *
1094  * Linux kernel will loop calling this shrinker scan routine with
1095  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1096  *
1097  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1098  * to scan and try to reclaim LRU pages, just return 0 and
1099  * osc_cache_shrink_count() will report the LRU page number.
1100  */
1101 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1102                                     struct shrink_control *sc)
1103 {
1104         struct client_obd *cli;
1105         struct client_obd *stop_anchor = NULL;
1106         struct lu_env *env;
1107         long shrank = 0;
1108         int rc;
1109         __u16 refcheck;
1110
1111         if (sc->nr_to_scan == 0)
1112                 return 0;
1113
1114         if (!(sc->gfp_mask & __GFP_FS))
1115                 return SHRINK_STOP;
1116
1117         env = cl_env_get(&refcheck);
1118         if (IS_ERR(env))
1119                 return SHRINK_STOP;
1120
1121         spin_lock(&osc_shrink_lock);
1122         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1123                                                struct client_obd,
1124                                                cl_shrink_list)) != NULL) {
1125                 if (stop_anchor == NULL)
1126                         stop_anchor = cli;
1127                 else if (cli == stop_anchor)
1128                         break;
1129
1130                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1131                 spin_unlock(&osc_shrink_lock);
1132
1133                 /* shrink no more than max_pages_per_rpc for an OSC */
1134                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1135                                     cli->cl_max_pages_per_rpc ?
1136                                     cli->cl_max_pages_per_rpc :
1137                                     sc->nr_to_scan - shrank, true);
1138                 if (rc > 0)
1139                         shrank += rc;
1140
1141                 if (shrank >= sc->nr_to_scan)
1142                         goto out;
1143
1144                 spin_lock(&osc_shrink_lock);
1145         }
1146         spin_unlock(&osc_shrink_lock);
1147
1148 out:
1149         cl_env_put(env, &refcheck);
1150
1151         return shrank;
1152 }
1153
1154 /** @} osc */