src/store_dir.cc

   1
   2 /*
   3  * $Id$
   4  *
   5  * DEBUG: section 47    Store Directory Routines
   6  * AUTHOR: Duane Wessels
   7  *
   8  * SQUID Web Proxy Cache          http://www.squid-cache.org/
   9  * ----------------------------------------------------------
  10  *
  11  *  Squid is the result of efforts by numerous individuals from
  12  *  the Internet community; see the CONTRIBUTORS file for full
  13  *  details.   Many organizations have provided support for Squid's
  14  *  development; see the SPONSORS file for full details.  Squid is
  15  *  Copyrighted (C) 2001 by the Regents of the University of
  16  *  California; see the COPYRIGHT file for full details.  Squid
  17  *  incorporates software developed and/or copyrighted by other
  18  *  sources; see the CREDITS file for full details.
  19  *
  20  *  This program is free software; you can redistribute it and/or modify
  21  *  it under the terms of the GNU General Public License as published by
  22  *  the Free Software Foundation; either version 2 of the License, or
  23  *  (at your option) any later version.
  24  *
  25  *  This program is distributed in the hope that it will be useful,
  26  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  27  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  28  *  GNU General Public License for more details.
  29  *
  30  *  You should have received a copy of the GNU General Public License
  31  *  along with this program; if not, write to the Free Software
  32  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA.
  33  *
  34  */
  35
  36 #include "squid-old.h"
  37 #include "Store.h"
  38 #include "MemObject.h"
  39 #include "MemStore.h"
  40 #include "mem_node.h"
  41 #include "SquidMath.h"
  42 #include "SquidTime.h"
  43 #include "SwapDir.h"
  44 #include "swap_log_op.h"
  45
  46 #if HAVE_STATVFS
  47 #if HAVE_SYS_STATVFS_H
  48 #include <sys/statvfs.h>
  49 #endif
  50 #endif /* HAVE_STATVFS */
  51 /* statfs() needs <sys/param.h> and <sys/mount.h> on BSD systems */
  52 #if HAVE_SYS_PARAM_H
  53 #include <sys/param.h>
  54 #endif
  55 #if HAVE_SYS_MOUNT_H
  56 #include <sys/mount.h>
  57 #endif
  58 /* Windows and Linux use sys/vfs.h */
  59 #if HAVE_SYS_VFS_H
  60 #include <sys/vfs.h>
  61 #endif
  62
  63 #include "StoreHashIndex.h"
  64
  65 static STDIRSELECT storeDirSelectSwapDirRoundRobin;
  66 static STDIRSELECT storeDirSelectSwapDirLeastLoad;
  67
  68 /*
  69  * store_dirs_rebuilding is initialized to _1_ as a hack so that
  70  * storeDirWriteCleanLogs() doesn't try to do anything unless _all_
  71  * cache_dirs have been read.  For example, without this hack, Squid
  72  * will try to write clean log files if -kparse fails (becasue it
  73  * calls fatal()).
  74  */
  75 int StoreController::store_dirs_rebuilding = 1;
  76
  77 StoreController::StoreController() : swapDir (new StoreHashIndex())
  78         , memStore(NULL)
  79 {}
  80
  81 StoreController::~StoreController()
  82 {
  83     delete memStore;
  84 }
  85
  86 /*
  87  * This function pointer is set according to 'store_dir_select_algorithm'
  88  * in squid.conf.
  89  */
  90 STDIRSELECT *storeDirSelectSwapDir = storeDirSelectSwapDirLeastLoad;
  91
  92 void
  93 StoreController::init()
  94 {
  95     if (Config.memShared && IamWorkerProcess()) {
  96         memStore = new MemStore;
  97         memStore->init();
  98     }
  99
 100     swapDir->init();
 101
 102     if (0 == strcasecmp(Config.store_dir_select_algorithm, "round-robin")) {
 103         storeDirSelectSwapDir = storeDirSelectSwapDirRoundRobin;
 104         debugs(47, 1, "Using Round Robin store dir selection");
 105     } else {
 106         storeDirSelectSwapDir = storeDirSelectSwapDirLeastLoad;
 107         debugs(47, 1, "Using Least Load store dir selection");
 108     }
 109 }
 110
 111 void
 112 StoreController::createOneStore(Store &aStore)
 113 {
 114     /*
 115      * On Windows, fork() is not available.
 116      * The following is a workaround for create store directories sequentially
 117      * when running on native Windows port.
 118      */
 119 #if !_SQUID_MSWIN_
 120
 121     if (fork())
 122         return;
 123
 124 #endif
 125
 126     aStore.create();
 127
 128 #if !_SQUID_MSWIN_
 129
 130     exit(0);
 131
 132 #endif
 133 }
 134
 135 void
 136 StoreController::create()
 137 {
 138     swapDir->create();
 139
 140 #if !_SQUID_MSWIN_
 141
 142     pid_t pid;
 143
 144     do {
 145         int status;
 146 #if _SQUID_NEXT_
 147
 148         pid = wait3(&status, WNOHANG, NULL);
 149 #else
 150
 151         pid = waitpid(-1, &status, 0);
 152 #endif
 153
 154     } while (pid > 0 || (pid < 0 && errno == EINTR));
 155
 156 #endif
 157 }
 158
 159 /**
 160  * Determine whether the given directory can handle this object
 161  * size
 162  *
 163  * Note: if the object size is -1, then the only swapdirs that
 164  * will return true here are ones that have min and max unset,
 165  * ie any-sized-object swapdirs. This is a good thing.
 166  */
 167 bool
 168 SwapDir::objectSizeIsAcceptable(int64_t objsize) const
 169 {
 170     // If the swapdir has no range limits, then it definitely can
 171     if (min_objsize <= 0 && max_objsize == -1)
 172         return true;
 173
 174     /*
 175      * If the object size is -1 and the storedir has limits we
 176      * can't store it there.
 177      */
 178     if (objsize == -1)
 179         return false;
 180
 181     // Else, make sure that the object size will fit.
 182     if (max_objsize == -1 && min_objsize <= objsize)
 183         return true;
 184     else
 185         return min_objsize <= objsize && max_objsize > objsize;
 186 }
 187
 188
 189 /*
 190  * This new selection scheme simply does round-robin on all SwapDirs.
 191  * A SwapDir is skipped if it is over the max_size (100%) limit, or
 192  * overloaded.
 193  */
 194 static int
 195 storeDirSelectSwapDirRoundRobin(const StoreEntry * e)
 196 {
 197     static int dirn = 0;
 198     int i;
 199     int load;
 200     RefCount<SwapDir> sd;
 201
 202     // e->objectLen() is negative at this point when we are still STORE_PENDING
 203     ssize_t objsize = e->mem_obj->expectedReplySize();
 204     if (objsize != -1)
 205         objsize += e->mem_obj->swap_hdr_sz;
 206
 207     for (i = 0; i < Config.cacheSwap.n_configured; i++) {
 208         if (++dirn >= Config.cacheSwap.n_configured)
 209             dirn = 0;
 210
 211         sd = dynamic_cast<SwapDir *>(INDEXSD(dirn));
 212
 213         if (!sd->canStore(*e, objsize, load))
 214             continue;
 215
 216         if (load < 0 || load > 1000) {
 217             continue;
 218         }
 219
 220         return dirn;
 221     }
 222
 223     return -1;
 224 }
 225
 226 /*
 227  * Spread load across all of the store directories
 228  *
 229  * Note: We should modify this later on to prefer sticking objects
 230  * in the *tightest fit* swapdir to conserve space, along with the
 231  * actual swapdir usage. But for now, this hack will do while
 232  * testing, so you should order your swapdirs in the config file
 233  * from smallest maxobjsize to unlimited (-1) maxobjsize.
 234  *
 235  * We also have to choose nleast == nconf since we need to consider
 236  * ALL swapdirs, regardless of state. Again, this is a hack while
 237  * we sort out the real usefulness of this algorithm.
 238  */
 239 static int
 240 storeDirSelectSwapDirLeastLoad(const StoreEntry * e)
 241 {
 242     int64_t most_free = 0;
 243     ssize_t least_objsize = -1;
 244     int least_load = INT_MAX;
 245     int load;
 246     int dirn = -1;
 247     int i;
 248     RefCount<SwapDir> SD;
 249
 250     // e->objectLen() is negative at this point when we are still STORE_PENDING
 251     ssize_t objsize = e->mem_obj->expectedReplySize();
 252
 253     if (objsize != -1)
 254         objsize += e->mem_obj->swap_hdr_sz;
 255
 256     for (i = 0; i < Config.cacheSwap.n_configured; i++) {
 257         SD = dynamic_cast<SwapDir *>(INDEXSD(i));
 258         SD->flags.selected = 0;
 259
 260         if (!SD->canStore(*e, objsize, load))
 261             continue;
 262
 263         if (load < 0 || load > 1000)
 264             continue;
 265
 266         if (load > least_load)
 267             continue;
 268
 269         const int64_t cur_free = SD->maxSize() - SD->currentSize();
 270
 271         /* If the load is equal, then look in more details */
 272         if (load == least_load) {
 273             /* closest max_objsize fit */
 274
 275             if (least_objsize != -1)
 276                 if (SD->max_objsize > least_objsize || SD->max_objsize == -1)
 277                     continue;
 278
 279             /* most free */
 280             if (cur_free < most_free)
 281                 continue;
 282         }
 283
 284         least_load = load;
 285         least_objsize = SD->max_objsize;
 286         most_free = cur_free;
 287         dirn = i;
 288     }
 289
 290     if (dirn >= 0)
 291         dynamic_cast<SwapDir *>(INDEXSD(dirn))->flags.selected = 1;
 292
 293     return dirn;
 294 }
 295
 296 /*
 297  * An entry written to the swap log MUST have the following
 298  * properties.
 299  *   1.  It MUST be a public key.  It does no good to log
 300  *       a public ADD, change the key, then log a private
 301  *       DEL.  So we need to log a DEL before we change a
 302  *       key from public to private.
 303  *   2.  It MUST have a valid (> -1) swap_filen.
 304  */
 305 void
 306 storeDirSwapLog(const StoreEntry * e, int op)
 307 {
 308     assert (e);
 309     assert(!EBIT_TEST(e->flags, KEY_PRIVATE));
 310     assert(e->swap_filen >= 0);
 311     /*
 312      * icons and such; don't write them to the swap log
 313      */
 314
 315     if (EBIT_TEST(e->flags, ENTRY_SPECIAL))
 316         return;
 317
 318     assert(op > SWAP_LOG_NOP && op < SWAP_LOG_MAX);
 319
 320     debugs(20, 3, "storeDirSwapLog: " <<
 321            swap_log_op_str[op] << " " <<
 322            e->getMD5Text() << " " <<
 323            e->swap_dirn << " " <<
 324            std::hex << std::uppercase << std::setfill('0') << std::setw(8) << e->swap_filen);
 325
 326     dynamic_cast<SwapDir *>(INDEXSD(e->swap_dirn))->logEntry(*e, op);
 327 }
 328
 329 void
 330 StoreController::getStats(StoreInfoStats &stats) const
 331 {
 332     if (memStore)
 333         memStore->getStats(stats);
 334     else {
 335         // move this code to a non-shared memory cache class when we have it
 336         stats.mem.shared = false;
 337         stats.mem.capacity = Config.memMaxSize;
 338         stats.mem.size = mem_node::StoreMemSize();
 339         stats.mem.count = hot_obj_count;
 340     }
 341
 342     swapDir->getStats(stats);
 343
 344     // low-level info not specific to memory or disk cache
 345     stats.store_entry_count = StoreEntry::inUseCount();
 346     stats.mem_object_count = MemObject::inUseCount();
 347 }
 348
 349 void
 350 StoreController::stat(StoreEntry &output) const
 351 {
 352     storeAppendPrintf(&output, "Store Directory Statistics:\n");
 353     storeAppendPrintf(&output, "Store Entries          : %lu\n",
 354                       (unsigned long int)StoreEntry::inUseCount());
 355     storeAppendPrintf(&output, "Maximum Swap Size      : %"PRIu64" KB\n",
 356                       maxSize() >> 10);
 357     storeAppendPrintf(&output, "Current Store Swap Size: %.2f KB\n",
 358                       currentSize() / 1024.0);
 359     storeAppendPrintf(&output, "Current Capacity       : %.2f%% used, %.2f%% free\n",
 360                       Math::doublePercent(currentSize(), maxSize()),
 361                       Math::doublePercent((maxSize() - currentSize()), maxSize()));
 362
 363     if (memStore)
 364         memStore->stat(output);
 365
 366     /* now the swapDir */
 367     swapDir->stat(output);
 368 }
 369
 370 /* if needed, this could be taught to cache the result */
 371 uint64_t
 372 StoreController::maxSize() const
 373 {
 374     /* TODO: include memory cache ? */
 375     return swapDir->maxSize();
 376 }
 377
 378 uint64_t
 379 StoreController::minSize() const
 380 {
 381     /* TODO: include memory cache ? */
 382     return swapDir->minSize();
 383 }
 384
 385 uint64_t
 386 StoreController::currentSize() const
 387 {
 388     return swapDir->currentSize();
 389 }
 390
 391 uint64_t
 392 StoreController::currentCount() const
 393 {
 394     return swapDir->currentCount();
 395 }
 396
 397 int64_t
 398 StoreController::maxObjectSize() const
 399 {
 400     return swapDir->maxObjectSize();
 401 }
 402
 403 void
 404 SwapDir::diskFull()
 405 {
 406     if (currentSize() >= maxSize())
 407         return;
 408
 409     max_size = currentSize();
 410
 411     debugs(20, 1, "WARNING: Shrinking cache_dir #" << index << " to " << currentSize() / 1024.0 << " KB");
 412 }
 413
 414 void
 415 storeDirOpenSwapLogs(void)
 416 {
 417     for (int dirn = 0; dirn < Config.cacheSwap.n_configured; ++dirn)
 418         dynamic_cast<SwapDir *>(INDEXSD(dirn))->openLog();
 419 }
 420
 421 void
 422 storeDirCloseSwapLogs(void)
 423 {
 424     for (int dirn = 0; dirn < Config.cacheSwap.n_configured; ++dirn)
 425         dynamic_cast<SwapDir *>(INDEXSD(dirn))->closeLog();
 426 }
 427
 428 /*
 429  *  storeDirWriteCleanLogs
 430  *
 431  *  Writes a "clean" swap log file from in-memory metadata.
 432  *  This is a rewrite of the original function to troll each
 433  *  StoreDir and write the logs, and flush at the end of
 434  *  the run. Thanks goes to Eric Stern, since this solution
 435  *  came out of his COSS code.
 436  */
 437 int
 438 storeDirWriteCleanLogs(int reopen)
 439 {
 440     const StoreEntry *e = NULL;
 441     int n = 0;
 442
 443     struct timeval start;
 444     double dt;
 445     RefCount<SwapDir> sd;
 446     int dirn;
 447     int notdone = 1;
 448
 449     if (StoreController::store_dirs_rebuilding) {
 450         debugs(20, 1, "Not currently OK to rewrite swap log.");
 451         debugs(20, 1, "storeDirWriteCleanLogs: Operation aborted.");
 452         return 0;
 453     }
 454
 455     debugs(20, 1, "storeDirWriteCleanLogs: Starting...");
 456     getCurrentTime();
 457     start = current_time;
 458
 459     for (dirn = 0; dirn < Config.cacheSwap.n_configured; dirn++) {
 460         sd = dynamic_cast<SwapDir *>(INDEXSD(dirn));
 461
 462         if (sd->writeCleanStart() < 0) {
 463             debugs(20, 1, "log.clean.start() failed for dir #" << sd->index);
 464             continue;
 465         }
 466     }
 467
 468     /*
 469      * This may look inefficient as CPU wise it is more efficient to do this
 470      * sequentially, but I/O wise the parallellism helps as it allows more
 471      * hdd spindles to be active.
 472      */
 473     while (notdone) {
 474         notdone = 0;
 475
 476         for (dirn = 0; dirn < Config.cacheSwap.n_configured; dirn++) {
 477             sd = dynamic_cast<SwapDir *>(INDEXSD(dirn));
 478
 479             if (NULL == sd->cleanLog)
 480                 continue;
 481
 482             e = sd->cleanLog->nextEntry();
 483
 484             if (!e)
 485                 continue;
 486
 487             notdone = 1;
 488
 489             if (!sd->canLog(*e))
 490                 continue;
 491
 492             sd->cleanLog->write(*e);
 493
 494             if ((++n & 0xFFFF) == 0) {
 495                 getCurrentTime();
 496                 debugs(20, 1, "  " << std::setw(7) << n  <<
 497                        " entries written so far.");
 498             }
 499         }
 500     }
 501
 502     /* Flush */
 503     for (dirn = 0; dirn < Config.cacheSwap.n_configured; dirn++)
 504         dynamic_cast<SwapDir *>(INDEXSD(dirn))->writeCleanDone();
 505
 506     if (reopen)
 507         storeDirOpenSwapLogs();
 508
 509     getCurrentTime();
 510
 511     dt = tvSubDsec(start, current_time);
 512
 513     debugs(20, 1, "  Finished.  Wrote " << n << " entries.");
 514     debugs(20, 1, "  Took "<< std::setw(3)<< std::setprecision(2) << dt <<
 515            " seconds ("<< std::setw(6) << ((double) n / (dt > 0.0 ? dt : 1.0)) << " entries/sec).");
 516
 517
 518     return n;
 519 }
 520
 521 StoreSearch *
 522 StoreController::search(String const url, HttpRequest *request)
 523 {
 524     /* cheat, for now you can't search the memory hot cache */
 525     return swapDir->search(url, request);
 526 }
 527
 528 StorePointer
 529 StoreHashIndex::store(int const x) const
 530 {
 531     return INDEXSD(x);
 532 }
 533
 534 SwapDir &
 535 StoreHashIndex::dir(const int i) const
 536 {
 537     SwapDir *sd = dynamic_cast<SwapDir*>(INDEXSD(i));
 538     assert(sd);
 539     return *sd;
 540 }
 541
 542 void
 543 StoreController::sync(void)
 544 {
 545     if (memStore)
 546         memStore->sync();
 547     swapDir->sync();
 548 }
 549
 550 /*
 551  * handle callbacks all avaliable fs'es
 552  */
 553 int
 554 StoreController::callback()
 555 {
 556     /* This will likely double count. Thats ok. */
 557     PROF_start(storeDirCallback);
 558
 559     /* mem cache callbacks ? */
 560     int result = swapDir->callback();
 561
 562     PROF_stop(storeDirCallback);
 563
 564     return result;
 565 }
 566
 567 int
 568 storeDirGetBlkSize(const char *path, int *blksize)
 569 {
 570 #if HAVE_STATVFS
 571
 572     struct statvfs sfs;
 573
 574     if (statvfs(path, &sfs)) {
 575         debugs(50, 1, "" << path << ": " << xstrerror());
 576         *blksize = 2048;
 577         return 1;
 578     }
 579
 580     *blksize = (int) sfs.f_frsize;
 581 #else
 582
 583     struct statfs sfs;
 584
 585     if (statfs(path, &sfs)) {
 586         debugs(50, 1, "" << path << ": " << xstrerror());
 587         *blksize = 2048;
 588         return 1;
 589     }
 590
 591     *blksize = (int) sfs.f_bsize;
 592 #endif
 593     /*
 594      * Sanity check; make sure we have a meaningful value.
 595      */
 596
 597     if (*blksize < 512)
 598         *blksize = 2048;
 599
 600     return 0;
 601 }
 602
 603 #define fsbtoblk(num, fsbs, bs) \
 604     (((fsbs) != 0 && (fsbs) < (bs)) ? \
 605             (num) / ((bs) / (fsbs)) : (num) * ((fsbs) / (bs)))
 606 int
 607 storeDirGetUFSStats(const char *path, int *totl_kb, int *free_kb, int *totl_in, int *free_in)
 608 {
 609 #if HAVE_STATVFS
 610
 611     struct statvfs sfs;
 612
 613     if (statvfs(path, &sfs)) {
 614         debugs(50, 1, "" << path << ": " << xstrerror());
 615         return 1;
 616     }
 617
 618     *totl_kb = (int) fsbtoblk(sfs.f_blocks, sfs.f_frsize, 1024);
 619     *free_kb = (int) fsbtoblk(sfs.f_bfree, sfs.f_frsize, 1024);
 620     *totl_in = (int) sfs.f_files;
 621     *free_in = (int) sfs.f_ffree;
 622 #else
 623
 624     struct statfs sfs;
 625
 626     if (statfs(path, &sfs)) {
 627         debugs(50, 1, "" << path << ": " << xstrerror());
 628         return 1;
 629     }
 630
 631     *totl_kb = (int) fsbtoblk(sfs.f_blocks, sfs.f_bsize, 1024);
 632     *free_kb = (int) fsbtoblk(sfs.f_bfree, sfs.f_bsize, 1024);
 633     *totl_in = (int) sfs.f_files;
 634     *free_in = (int) sfs.f_ffree;
 635 #endif
 636
 637     return 0;
 638 }
 639
 640 void
 641 allocate_new_swapdir(SquidConfig::_cacheSwap * swap)
 642 {
 643     if (swap->swapDirs == NULL) {
 644         swap->n_allocated = 4;
 645         swap->swapDirs = static_cast<SwapDir::Pointer *>(xcalloc(swap->n_allocated, sizeof(SwapDir::Pointer)));
 646     }
 647
 648     if (swap->n_allocated == swap->n_configured) {
 649         swap->n_allocated <<= 1;
 650         SwapDir::Pointer *const tmp = static_cast<SwapDir::Pointer *>(xcalloc(swap->n_allocated, sizeof(SwapDir::Pointer)));
 651         memcpy(tmp, swap->swapDirs, swap->n_configured * sizeof(SwapDir *));
 652         xfree(swap->swapDirs);
 653         swap->swapDirs = tmp;
 654     }
 655 }
 656
 657 void
 658 free_cachedir(SquidConfig::_cacheSwap * swap)
 659 {
 660     int i;
 661     /* DON'T FREE THESE FOR RECONFIGURE */
 662
 663     if (reconfiguring)
 664         return;
 665
 666     for (i = 0; i < swap->n_configured; i++) {
 667         /* TODO XXX this lets the swapdir free resources asynchronously
 668         * swap->swapDirs[i]->deactivate();
 669         * but there may be such a means already.
 670         * RBC 20041225
 671         */
 672         swap->swapDirs[i] = NULL;
 673     }
 674
 675     safe_free(swap->swapDirs);
 676     swap->swapDirs = NULL;
 677     swap->n_allocated = 0;
 678     swap->n_configured = 0;
 679 }
 680
 681 /* this should be a virtual method on StoreEntry,
 682  * i.e. e->referenced()
 683  * so that the entry can notify the creating Store
 684  */
 685 void
 686 StoreController::reference(StoreEntry &e)
 687 {
 688     /* Notify the fs that we're referencing this object again */
 689
 690     if (e.swap_dirn > -1)
 691         swapDir->reference(e);
 692
 693     // Notify the memory cache that we're referencing this object again
 694     if (memStore && e.mem_status == IN_MEMORY)
 695         memStore->reference(e);
 696
 697     // TODO: move this code to a non-shared memory cache class when we have it
 698     if (e.mem_obj) {
 699         if (mem_policy->Referenced)
 700             mem_policy->Referenced(mem_policy, &e, &e.mem_obj->repl);
 701     }
 702 }
 703
 704 bool
 705 StoreController::dereference(StoreEntry & e)
 706 {
 707     bool keepInStoreTable = true; // keep if there are no objections
 708
 709     /* Notify the fs that we're not referencing this object any more */
 710
 711     if (e.swap_filen > -1)
 712         keepInStoreTable = swapDir->dereference(e) && keepInStoreTable;
 713
 714     // Notify the memory cache that we're not referencing this object any more
 715     if (memStore && e.mem_status == IN_MEMORY)
 716         keepInStoreTable = memStore->dereference(e) && keepInStoreTable;
 717
 718     // TODO: move this code to a non-shared memory cache class when we have it
 719     if (e.mem_obj) {
 720         if (mem_policy->Dereferenced)
 721             mem_policy->Dereferenced(mem_policy, &e, &e.mem_obj->repl);
 722     }
 723
 724     return keepInStoreTable;
 725 }
 726
 727 StoreEntry *
 728 StoreController::get(const cache_key *key)
 729 {
 730     if (StoreEntry *e = swapDir->get(key)) {
 731         // TODO: ignore and maybe handleIdleEntry() unlocked intransit entries
 732         // because their backing store slot may be gone already.
 733         debugs(20, 3, HERE << "got in-transit entry: " << *e);
 734         return e;
 735     }
 736
 737     if (memStore) {
 738         if (StoreEntry *e = memStore->get(key)) {
 739             debugs(20, 3, HERE << "got mem-cached entry: " << *e);
 740             return e;
 741         }
 742     }
 743
 744     // TODO: this disk iteration is misplaced; move to StoreHashIndex when
 745     // the global store_table is no longer used for in-transit objects.
 746     if (const int cacheDirs = Config.cacheSwap.n_configured) {
 747         // ask each cache_dir until the entry is found; use static starting
 748         // point to avoid asking the same subset of disks more often
 749         // TODO: coordinate with put() to be able to guess the right disk often
 750         static int idx = 0;
 751         for (int n = 0; n < cacheDirs; ++n) {
 752             idx = (idx + 1) % cacheDirs;
 753             SwapDir *sd = dynamic_cast<SwapDir*>(INDEXSD(idx));
 754             if (!sd->active())
 755                 continue;
 756
 757             if (StoreEntry *e = sd->get(key)) {
 758                 debugs(20, 3, HERE << "cache_dir " << idx <<
 759                        " got cached entry: " << *e);
 760                 return e;
 761             }
 762         }
 763     }
 764
 765     debugs(20, 4, HERE << "none of " << Config.cacheSwap.n_configured <<
 766            " cache_dirs have " << storeKeyText(key));
 767     return NULL;
 768 }
 769
 770 void
 771 StoreController::get(String const key, STOREGETCLIENT aCallback, void *aCallbackData)
 772 {
 773     fatal("not implemented");
 774 }
 775
 776 void
 777 StoreController::handleIdleEntry(StoreEntry &e)
 778 {
 779     bool keepInLocalMemory = false;
 780     if (memStore) {
 781         memStore->considerKeeping(e);
 782         // leave keepInLocalMemory false; memStore maintains its own cache
 783     } else {
 784         keepInLocalMemory = e.memoryCachable() && // entry is in good shape and
 785                             // the local memory cache is not overflowing
 786                             (mem_node::InUseCount() <= store_pages_max);
 787     }
 788
 789     // An idle, unlocked entry that belongs to a SwapDir which controls
 790     // its own index, should not stay in the global store_table.
 791     if (!dereference(e)) {
 792         debugs(20, 5, HERE << "destroying unlocked entry: " << &e << ' ' << e);
 793         destroyStoreEntry(static_cast<hash_link*>(&e));
 794         return;
 795     }
 796
 797     // TODO: move this into [non-shared] memory cache class when we have one
 798     if (keepInLocalMemory) {
 799         e.setMemStatus(IN_MEMORY);
 800         e.mem_obj->unlinkRequest();
 801     } else {
 802         e.purgeMem(); // may free e
 803     }
 804 }
 805
 806 StoreHashIndex::StoreHashIndex()
 807 {
 808     if (store_table)
 809         abort();
 810     assert (store_table == NULL);
 811 }
 812
 813 StoreHashIndex::~StoreHashIndex()
 814 {
 815     if (store_table) {
 816         hashFreeItems(store_table, destroyStoreEntry);
 817         hashFreeMemory(store_table);
 818         store_table = NULL;
 819     }
 820 }
 821
 822 int
 823 StoreHashIndex::callback()
 824 {
 825     int result = 0;
 826     int j;
 827     static int ndir = 0;
 828
 829     do {
 830         j = 0;
 831
 832         for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 833             if (ndir >= Config.cacheSwap.n_configured)
 834                 ndir = ndir % Config.cacheSwap.n_configured;
 835
 836             int temp_result = store(ndir)->callback();
 837
 838             ++ndir;
 839
 840             j += temp_result;
 841
 842             result += temp_result;
 843
 844             if (j > 100)
 845                 fatal ("too much io\n");
 846         }
 847     } while (j > 0);
 848
 849     ndir++;
 850
 851     return result;
 852 }
 853
 854 void
 855 StoreHashIndex::create()
 856 {
 857     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 858         if (dir(i).active())
 859             store(i)->create();
 860     }
 861 }
 862
 863 /* Lookup an object in the cache.
 864  * return just a reference to object, don't start swapping in yet. */
 865 StoreEntry *
 866 StoreHashIndex::get(const cache_key *key)
 867 {
 868     PROF_start(storeGet);
 869     debugs(20, 3, "storeGet: looking up " << storeKeyText(key));
 870     StoreEntry *p = static_cast<StoreEntry *>(hash_lookup(store_table, key));
 871     PROF_stop(storeGet);
 872     return p;
 873 }
 874
 875 void
 876 StoreHashIndex::get(String const key, STOREGETCLIENT aCallback, void *aCallbackData)
 877 {
 878     fatal("not implemented");
 879 }
 880
 881 void
 882 StoreHashIndex::init()
 883 {
 884     /* Calculate size of hash table (maximum currently 64k buckets).  */
 885     /* this is very bogus, its specific to the any Store maintaining an
 886      * in-core index, not global */
 887     size_t buckets = (Store::Root().maxSize() + Config.memMaxSize) / Config.Store.avgObjectSize;
 888     debugs(20, 1, "Swap maxSize " << (Store::Root().maxSize() >> 10) <<
 889            " + " << ( Config.memMaxSize >> 10) << " KB, estimated " << buckets << " objects");
 890     buckets /= Config.Store.objectsPerBucket;
 891     debugs(20, 1, "Target number of buckets: " << buckets);
 892     /* ideally the full scan period should be configurable, for the
 893      * moment it remains at approximately 24 hours.  */
 894     store_hash_buckets = storeKeyHashBuckets(buckets);
 895     debugs(20, 1, "Using " << store_hash_buckets << " Store buckets");
 896     debugs(20, 1, "Max Mem  size: " << ( Config.memMaxSize >> 10) << " KB" <<
 897            (Config.memShared ? " [shared]" : ""));
 898     debugs(20, 1, "Max Swap size: " << (Store::Root().maxSize() >> 10) << " KB");
 899
 900     store_table = hash_create(storeKeyHashCmp,
 901                               store_hash_buckets, storeKeyHashHash);
 902
 903     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 904         /* this starts a search of the store dirs, loading their
 905          * index. under the new Store api this should be
 906          * driven by the StoreHashIndex, not by each store.
 907         *
 908         * That is, the HashIndex should perform a search of each dir it is
 909         * indexing to do the hash insertions. The search is then able to
 910         * decide 'from-memory', or 'from-clean-log' or 'from-dirty-log' or
 911         * 'from-no-log'.
 912         *
 913          * Step 1: make the store rebuilds use a search internally
 914         * Step 2: change the search logic to use the four modes described
 915         *         above
 916         * Step 3: have the hash index walk the searches itself.
 917          */
 918         if (dir(i).active())
 919             store(i)->init();
 920     }
 921 }
 922
 923 uint64_t
 924 StoreHashIndex::maxSize() const
 925 {
 926     uint64_t result = 0;
 927
 928     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 929         if (dir(i).doReportStat())
 930             result += store(i)->maxSize();
 931     }
 932
 933     return result;
 934 }
 935
 936 uint64_t
 937 StoreHashIndex::minSize() const
 938 {
 939     uint64_t result = 0;
 940
 941     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 942         if (dir(i).doReportStat())
 943             result += store(i)->minSize();
 944     }
 945
 946     return result;
 947 }
 948
 949 uint64_t
 950 StoreHashIndex::currentSize() const
 951 {
 952     uint64_t result = 0;
 953
 954     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 955         if (dir(i).doReportStat())
 956             result += store(i)->currentSize();
 957     }
 958
 959     return result;
 960 }
 961
 962 uint64_t
 963 StoreHashIndex::currentCount() const
 964 {
 965     uint64_t result = 0;
 966
 967     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 968         if (dir(i).doReportStat())
 969             result += store(i)->currentCount();
 970     }
 971
 972     return result;
 973 }
 974
 975 int64_t
 976 StoreHashIndex::maxObjectSize() const
 977 {
 978     int64_t result = -1;
 979
 980     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 981         if (dir(i).active() && store(i)->maxObjectSize() > result)
 982             result = store(i)->maxObjectSize();
 983     }
 984
 985     return result;
 986 }
 987
 988 void
 989 StoreHashIndex::getStats(StoreInfoStats &stats) const
 990 {
 991     // accumulate per-disk cache stats
 992     for (int i = 0; i < Config.cacheSwap.n_configured; i++) {
 993         StoreInfoStats dirStats;
 994         store(i)->getStats(dirStats);
 995         stats += dirStats;
 996     }
 997
 998     // common to all disks
 999     stats.swap.open_disk_fd = store_open_disk_fd;
1000
1001     // memory cache stats are collected in StoreController::getStats(), for now
1002 }
1003
1004 void
1005 StoreHashIndex::stat(StoreEntry & output) const
1006 {
1007     int i;
1008
1009     /* Now go through each store, calling its stat routine */
1010
1011     for (i = 0; i < Config.cacheSwap.n_configured; i++) {
1012         storeAppendPrintf(&output, "\n");
1013         store(i)->stat(output);
1014     }
1015 }
1016
1017 void
1018 StoreHashIndex::reference(StoreEntry &e)
1019 {
1020     e.store()->reference(e);
1021 }
1022
1023 bool
1024 StoreHashIndex::dereference(StoreEntry &e)
1025 {
1026     return e.store()->dereference(e);
1027 }
1028
1029 void
1030 StoreHashIndex::maintain()
1031 {
1032     int i;
1033     /* walk each fs */
1034
1035     for (i = 0; i < Config.cacheSwap.n_configured; i++) {
1036         /* XXX FixMe: This should be done "in parallell" on the different
1037          * cache_dirs, not one at a time.
1038          */
1039         /* call the maintain function .. */
1040         store(i)->maintain();
1041     }
1042 }
1043
1044 void
1045 StoreHashIndex::sync()
1046 {
1047     for (int i = 0; i < Config.cacheSwap.n_configured; ++i)
1048         store(i)->sync();
1049 }
1050
1051 StoreSearch *
1052 StoreHashIndex::search(String const url, HttpRequest *)
1053 {
1054     if (url.size())
1055         fatal ("Cannot search by url yet\n");
1056
1057     return new StoreSearchHashIndex (this);
1058 }
1059
1060 CBDATA_CLASS_INIT(StoreSearchHashIndex);
1061
1062 StoreSearchHashIndex::StoreSearchHashIndex(RefCount<StoreHashIndex> aSwapDir) : sd(aSwapDir), _done (false), bucket (0)
1063 {}
1064
1065 /* do not link
1066 StoreSearchHashIndex::StoreSearchHashIndex(StoreSearchHashIndex const &);
1067 */
1068
1069 StoreSearchHashIndex::~StoreSearchHashIndex()
1070 {}
1071
1072 void
1073 StoreSearchHashIndex::next(void (aCallback)(void *), void *aCallbackData)
1074 {
1075     next();
1076     aCallback (aCallbackData);
1077 }
1078
1079 bool
1080 StoreSearchHashIndex::next()
1081 {
1082     if (entries.size())
1083         entries.pop_back();
1084
1085     while (!isDone() && !entries.size())
1086         copyBucket();
1087
1088     return currentItem() != NULL;
1089 }
1090
1091 bool
1092 StoreSearchHashIndex::error() const
1093 {
1094     return false;
1095 }
1096
1097 bool
1098 StoreSearchHashIndex::isDone() const
1099 {
1100     return bucket >= store_hash_buckets || _done;
1101 }
1102
1103 StoreEntry *
1104 StoreSearchHashIndex::currentItem()
1105 {
1106     if (!entries.size())
1107         return NULL;
1108
1109     return entries.back();
1110 }
1111
1112 void
1113 StoreSearchHashIndex::copyBucket()
1114 {
1115     /* probably need to lock the store entries...
1116      * we copy them all to prevent races on the links. */
1117     debugs(47, 3, "StoreSearchHashIndex::copyBucket #" << bucket);
1118     assert (!entries.size());
1119     hash_link *link_ptr = NULL;
1120     hash_link *link_next = NULL;
1121     link_next = hash_get_bucket(store_table, bucket);
1122
1123     while (NULL != (link_ptr = link_next)) {
1124         link_next = link_ptr->next;
1125         StoreEntry *e = (StoreEntry *) link_ptr;
1126
1127         entries.push_back(e);
1128     }
1129
1130     bucket++;
1131     debugs(47,3, "got entries: " << entries.size());
1132 }