src/fs/rock/RockSwapDir.cc

   1 /*
   2  * Copyright (C) 1996-2014 The Squid Software Foundation and contributors
   3  *
   4  * Squid software is distributed under GPLv2+ license and includes
   5  * contributions from numerous individuals and organizations.
   6  * Please see the COPYING and CONTRIBUTORS files for details.
   7  */
   8
   9 /* DEBUG: section 47    Store Directory Routines */
  10
  11 #include "squid.h"
  12 #include "cache_cf.h"
  13 #include "CollapsedForwarding.h"
  14 #include "ConfigOption.h"
  15 #include "DiskIO/DiskIOModule.h"
  16 #include "DiskIO/DiskIOStrategy.h"
  17 #include "DiskIO/ReadRequest.h"
  18 #include "DiskIO/WriteRequest.h"
  19 #include "fs/rock/RockIoRequests.h"
  20 #include "fs/rock/RockIoState.h"
  21 #include "fs/rock/RockRebuild.h"
  22 #include "fs/rock/RockSwapDir.h"
  23 #include "globals.h"
  24 #include "ipc/mem/Pages.h"
  25 #include "MemObject.h"
  26 #include "Parsing.h"
  27 #include "SquidConfig.h"
  28 #include "SquidMath.h"
  29 #include "tools.h"
  30
  31 #include <cstdlib>
  32 #include <iomanip>
  33 #include <limits>
  34
  35 #if HAVE_SYS_STAT_H
  36 #include <sys/stat.h>
  37 #endif
  38
  39 const int64_t Rock::SwapDir::HeaderSize = 16*1024;
  40
  41 Rock::SwapDir::SwapDir(): ::SwapDir("rock"),
  42     slotSize(HeaderSize), filePath(NULL), map(NULL), io(NULL),
  43     waitingForPage(NULL)
  44 {
  45 }
  46
  47 Rock::SwapDir::~SwapDir()
  48 {
  49     delete io;
  50     delete map;
  51     safe_free(filePath);
  52 }
  53
  54 StoreSearch *
  55 Rock::SwapDir::search(String const, HttpRequest *)
  56 {
  57     assert(false);
  58     return NULL; // XXX: implement
  59 }
  60
  61 void
  62 Rock::SwapDir::get(String const key, STOREGETCLIENT cb, void *data)
  63 {
  64     ::SwapDir::get(key, cb, data);
  65 }
  66
  67 // called when Squid core needs a StoreEntry with a given key
  68 StoreEntry *
  69 Rock::SwapDir::get(const cache_key *key)
  70 {
  71     if (!map || !theFile || !theFile->canRead())
  72         return NULL;
  73
  74     sfileno filen;
  75     const Ipc::StoreMapAnchor *const slot = map->openForReading(key, filen);
  76     if (!slot)
  77         return NULL;
  78
  79     // create a brand new store entry and initialize it with stored basics
  80     StoreEntry *e = new StoreEntry();
  81     anchorEntry(*e, filen, *slot);
  82
  83     e->hashInsert(key);
  84     trackReferences(*e);
  85
  86     return e;
  87     // the disk entry remains open for reading, protected from modifications
  88 }
  89
  90 bool
  91 Rock::SwapDir::anchorCollapsed(StoreEntry &collapsed, bool &inSync)
  92 {
  93     if (!map || !theFile || !theFile->canRead())
  94         return false;
  95
  96     sfileno filen;
  97     const Ipc::StoreMapAnchor *const slot = map->openForReading(
  98             reinterpret_cast<cache_key*>(collapsed.key), filen);
  99     if (!slot)
 100         return false;
 101
 102     anchorEntry(collapsed, filen, *slot);
 103     inSync = updateCollapsedWith(collapsed, *slot);
 104     return true; // even if inSync is false
 105 }
 106
 107 bool
 108 Rock::SwapDir::updateCollapsed(StoreEntry &collapsed)
 109 {
 110     if (!map || !theFile || !theFile->canRead())
 111         return false;
 112
 113     if (collapsed.swap_filen < 0) // no longer using a disk cache
 114         return true;
 115     assert(collapsed.swap_dirn == index);
 116
 117     const Ipc::StoreMapAnchor &s = map->readableEntry(collapsed.swap_filen);
 118     return updateCollapsedWith(collapsed, s);
 119 }
 120
 121 bool
 122 Rock::SwapDir::updateCollapsedWith(StoreEntry &collapsed, const Ipc::StoreMapAnchor &anchor)
 123 {
 124     collapsed.swap_file_sz = anchor.basics.swap_file_sz;
 125     return true;
 126 }
 127
 128 void
 129 Rock::SwapDir::anchorEntry(StoreEntry &e, const sfileno filen, const Ipc::StoreMapAnchor &anchor)
 130 {
 131     const Ipc::StoreMapAnchor::Basics &basics = anchor.basics;
 132
 133     e.swap_file_sz = basics.swap_file_sz;
 134     e.lastref = basics.lastref;
 135     e.timestamp = basics.timestamp;
 136     e.expires = basics.expires;
 137     e.lastmod = basics.lastmod;
 138     e.refcount = basics.refcount;
 139     e.flags = basics.flags;
 140
 141     if (anchor.complete()) {
 142         e.store_status = STORE_OK;
 143         e.swap_status = SWAPOUT_DONE;
 144     } else {
 145         e.store_status = STORE_PENDING;
 146         e.swap_status = SWAPOUT_WRITING; // even though another worker writes?
 147     }
 148
 149     e.ping_status = PING_NONE;
 150
 151     EBIT_CLR(e.flags, RELEASE_REQUEST);
 152     EBIT_CLR(e.flags, KEY_PRIVATE);
 153     EBIT_SET(e.flags, ENTRY_VALIDATED);
 154
 155     e.swap_dirn = index;
 156     e.swap_filen = filen;
 157 }
 158
 159 void Rock::SwapDir::disconnect(StoreEntry &e)
 160 {
 161     assert(e.swap_dirn == index);
 162     assert(e.swap_filen >= 0);
 163     // cannot have SWAPOUT_NONE entry with swap_filen >= 0
 164     assert(e.swap_status != SWAPOUT_NONE);
 165
 166     // do not rely on e.swap_status here because there is an async delay
 167     // before it switches from SWAPOUT_WRITING to SWAPOUT_DONE.
 168
 169     // since e has swap_filen, its slot is locked for reading and/or writing
 170     // but it is difficult to know whether THIS worker is reading or writing e,
 171     // especially since we may switch from writing to reading. This code relies
 172     // on Rock::IoState::writeableAnchor_ being set when we locked for writing.
 173     if (e.mem_obj && e.mem_obj->swapout.sio != NULL &&
 174             dynamic_cast<IoState&>(*e.mem_obj->swapout.sio).writeableAnchor_) {
 175         map->abortWriting(e.swap_filen);
 176         e.swap_dirn = -1;
 177         e.swap_filen = -1;
 178         e.swap_status = SWAPOUT_NONE;
 179         dynamic_cast<IoState&>(*e.mem_obj->swapout.sio).writeableAnchor_ = NULL;
 180         Store::Root().transientsAbandon(e); // broadcasts after the change
 181     } else {
 182         map->closeForReading(e.swap_filen);
 183         e.swap_dirn = -1;
 184         e.swap_filen = -1;
 185         e.swap_status = SWAPOUT_NONE;
 186     }
 187 }
 188
 189 uint64_t
 190 Rock::SwapDir::currentSize() const
 191 {
 192     const uint64_t spaceSize = !freeSlots ?
 193                                maxSize() : (slotSize * freeSlots->size());
 194     // everything that is not free is in use
 195     return maxSize() - spaceSize;
 196 }
 197
 198 uint64_t
 199 Rock::SwapDir::currentCount() const
 200 {
 201     return map ? map->entryCount() : 0;
 202 }
 203
 204 /// In SMP mode only the disker process reports stats to avoid
 205 /// counting the same stats by multiple processes.
 206 bool
 207 Rock::SwapDir::doReportStat() const
 208 {
 209     return ::SwapDir::doReportStat() && (!UsingSmp() || IamDiskProcess());
 210 }
 211
 212 void
 213 Rock::SwapDir::swappedOut(const StoreEntry &)
 214 {
 215     // stats are not stored but computed when needed
 216 }
 217
 218 int64_t
 219 Rock::SwapDir::slotLimitAbsolute() const
 220 {
 221     // the max value is an invalid one; all values must be below the limit
 222     assert(std::numeric_limits<Ipc::StoreMapSliceId>::max() ==
 223            std::numeric_limits<SlotId>::max());
 224     return std::numeric_limits<SlotId>::max();
 225 }
 226
 227 int64_t
 228 Rock::SwapDir::slotLimitActual() const
 229 {
 230     const int64_t sWanted = (maxSize() - HeaderSize)/slotSize;
 231     const int64_t sLimitLo = map ? map->sliceLimit() : 0; // dynamic shrinking unsupported
 232     const int64_t sLimitHi = slotLimitAbsolute();
 233     return min(max(sLimitLo, sWanted), sLimitHi);
 234 }
 235
 236 int64_t
 237 Rock::SwapDir::entryLimitActual() const
 238 {
 239     return min(slotLimitActual(), entryLimitAbsolute());
 240 }
 241
 242 // TODO: encapsulate as a tool
 243 void
 244 Rock::SwapDir::create()
 245 {
 246     assert(path);
 247     assert(filePath);
 248
 249     if (UsingSmp() && !IamDiskProcess()) {
 250         debugs (47,3, HERE << "disker will create in " << path);
 251         return;
 252     }
 253
 254     debugs (47,3, HERE << "creating in " << path);
 255
 256     struct stat dir_sb;
 257     if (::stat(path, &dir_sb) == 0) {
 258         struct stat file_sb;
 259         if (::stat(filePath, &file_sb) == 0) {
 260             debugs (47, DBG_IMPORTANT, "Skipping existing Rock db: " << filePath);
 261             return;
 262         }
 263         // else the db file is not there or is not accessible, and we will try
 264         // to create it later below, generating a detailed error on failures.
 265     } else { // path does not exist or is inaccessible
 266         // If path exists but is not accessible, mkdir() below will fail, and
 267         // the admin should see the error and act accordingly, so there is
 268         // no need to distinguish ENOENT from other possible stat() errors.
 269         debugs (47, DBG_IMPORTANT, "Creating Rock db directory: " << path);
 270         const int res = mkdir(path, 0700);
 271         if (res != 0)
 272             createError("mkdir");
 273     }
 274
 275     debugs (47, DBG_IMPORTANT, "Creating Rock db: " << filePath);
 276     const int swap = open(filePath, O_WRONLY|O_CREAT|O_TRUNC|O_BINARY, 0600);
 277     if (swap < 0)
 278         createError("create");
 279
 280 #if SLOWLY_FILL_WITH_ZEROS
 281     char block[1024];
 282     Must(maxSize() % sizeof(block) == 0);
 283     memset(block, '\0', sizeof(block));
 284
 285     for (off_t offset = 0; offset < maxSize(); offset += sizeof(block)) {
 286         if (write(swap, block, sizeof(block)) != sizeof(block))
 287             createError("write");
 288     }
 289 #else
 290     if (ftruncate(swap, maxSize()) != 0)
 291         createError("truncate");
 292
 293     char header[HeaderSize];
 294     memset(header, '\0', sizeof(header));
 295     if (write(swap, header, sizeof(header)) != sizeof(header))
 296         createError("write");
 297 #endif
 298
 299     close(swap);
 300 }
 301
 302 // report Rock DB creation error and exit
 303 void
 304 Rock::SwapDir::createError(const char *const msg)
 305 {
 306     debugs(47, DBG_CRITICAL, "ERROR: Failed to initialize Rock Store db in " <<
 307            filePath << "; " << msg << " error: " << xstrerror());
 308     fatal("Rock Store db creation error");
 309 }
 310
 311 void
 312 Rock::SwapDir::init()
 313 {
 314     debugs(47,2, HERE);
 315
 316     // XXX: SwapDirs aren't refcounted. We make IORequestor calls, which
 317     // are refcounted. We up our count once to avoid implicit delete's.
 318     lock();
 319
 320     freeSlots = shm_old(Ipc::Mem::PageStack)(freeSlotsPath());
 321
 322     Must(!map);
 323     map = new DirMap(inodeMapPath());
 324     map->cleaner = this;
 325
 326     const char *ioModule = needsDiskStrand() ? "IpcIo" : "Blocking";
 327     if (DiskIOModule *m = DiskIOModule::Find(ioModule)) {
 328         debugs(47,2, HERE << "Using DiskIO module: " << ioModule);
 329         io = m->createStrategy();
 330         io->init();
 331     } else {
 332         debugs(47, DBG_CRITICAL, "FATAL: Rock store is missing DiskIO module: " <<
 333                ioModule);
 334         fatal("Rock Store missing a required DiskIO module");
 335     }
 336
 337     theFile = io->newFile(filePath);
 338     theFile->configure(fileConfig);
 339     theFile->open(O_RDWR, 0644, this);
 340
 341     // Increment early. Otherwise, if one SwapDir finishes rebuild before
 342     // others start, storeRebuildComplete() will think the rebuild is over!
 343     // TODO: move store_dirs_rebuilding hack to store modules that need it.
 344     ++StoreController::store_dirs_rebuilding;
 345 }
 346
 347 bool
 348 Rock::SwapDir::needsDiskStrand() const
 349 {
 350     const bool wontEvenWorkWithoutDisker = Config.workers > 1;
 351     const bool wouldWorkBetterWithDisker = DiskIOModule::Find("IpcIo");
 352     return InDaemonMode() && (wontEvenWorkWithoutDisker ||
 353                               wouldWorkBetterWithDisker);
 354 }
 355
 356 void
 357 Rock::SwapDir::parse(int anIndex, char *aPath)
 358 {
 359     index = anIndex;
 360
 361     path = xstrdup(aPath);
 362
 363     // cache store is located at path/db
 364     String fname(path);
 365     fname.append("/rock");
 366     filePath = xstrdup(fname.termedBuf());
 367
 368     parseSize(false);
 369     parseOptions(0);
 370
 371     // Current openForWriting() code overwrites the old slot if needed
 372     // and possible, so proactively removing old slots is probably useless.
 373     assert(!repl); // repl = createRemovalPolicy(Config.replPolicy);
 374
 375     validateOptions();
 376 }
 377
 378 void
 379 Rock::SwapDir::reconfigure()
 380 {
 381     parseSize(true);
 382     parseOptions(1);
 383     // TODO: can we reconfigure the replacement policy (repl)?
 384     validateOptions();
 385 }
 386
 387 /// parse maximum db disk size
 388 void
 389 Rock::SwapDir::parseSize(const bool reconfig)
 390 {
 391     const int i = GetInteger();
 392     if (i < 0)
 393         fatal("negative Rock cache_dir size value");
 394     const uint64_t new_max_size =
 395         static_cast<uint64_t>(i) << 20; // MBytes to Bytes
 396     if (!reconfig)
 397         max_size = new_max_size;
 398     else if (new_max_size != max_size) {
 399         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir '" << path << "' size "
 400                "cannot be changed dynamically, value left unchanged (" <<
 401                (max_size >> 20) << " MB)");
 402     }
 403 }
 404
 405 ConfigOption *
 406 Rock::SwapDir::getOptionTree() const
 407 {
 408     ConfigOptionVector *vector = dynamic_cast<ConfigOptionVector*>(::SwapDir::getOptionTree());
 409     assert(vector);
 410     vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseSizeOption, &SwapDir::dumpSizeOption));
 411     vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseTimeOption, &SwapDir::dumpTimeOption));
 412     vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseRateOption, &SwapDir::dumpRateOption));
 413     return vector;
 414 }
 415
 416 bool
 417 Rock::SwapDir::allowOptionReconfigure(const char *const option) const
 418 {
 419     return strcmp(option, "slot-size") != 0 &&
 420            ::SwapDir::allowOptionReconfigure(option);
 421 }
 422
 423 /// parses time-specific options; mimics ::SwapDir::optionObjectSizeParse()
 424 bool
 425 Rock::SwapDir::parseTimeOption(char const *option, const char *value, int reconfig)
 426 {
 427     // TODO: ::SwapDir or, better, Config should provide time-parsing routines,
 428     // including time unit handling. Same for size and rate.
 429
 430     time_msec_t *storedTime;
 431     if (strcmp(option, "swap-timeout") == 0)
 432         storedTime = &fileConfig.ioTimeout;
 433     else
 434         return false;
 435
 436     if (!value)
 437         self_destruct();
 438
 439     // TODO: handle time units and detect parsing errors better
 440     const int64_t parsedValue = strtoll(value, NULL, 10);
 441     if (parsedValue < 0) {
 442         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << parsedValue);
 443         self_destruct();
 444     }
 445
 446     const time_msec_t newTime = static_cast<time_msec_t>(parsedValue);
 447
 448     if (!reconfig)
 449         *storedTime = newTime;
 450     else if (*storedTime != newTime) {
 451         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 452                << " cannot be changed dynamically, value left unchanged: " <<
 453                *storedTime);
 454     }
 455
 456     return true;
 457 }
 458
 459 /// reports time-specific options; mimics ::SwapDir::optionObjectSizeDump()
 460 void
 461 Rock::SwapDir::dumpTimeOption(StoreEntry * e) const
 462 {
 463     if (fileConfig.ioTimeout)
 464         storeAppendPrintf(e, " swap-timeout=%" PRId64,
 465                           static_cast<int64_t>(fileConfig.ioTimeout));
 466 }
 467
 468 /// parses rate-specific options; mimics ::SwapDir::optionObjectSizeParse()
 469 bool
 470 Rock::SwapDir::parseRateOption(char const *option, const char *value, int isaReconfig)
 471 {
 472     int *storedRate;
 473     if (strcmp(option, "max-swap-rate") == 0)
 474         storedRate = &fileConfig.ioRate;
 475     else
 476         return false;
 477
 478     if (!value)
 479         self_destruct();
 480
 481     // TODO: handle time units and detect parsing errors better
 482     const int64_t parsedValue = strtoll(value, NULL, 10);
 483     if (parsedValue < 0) {
 484         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << parsedValue);
 485         self_destruct();
 486     }
 487
 488     const int newRate = static_cast<int>(parsedValue);
 489
 490     if (newRate < 0) {
 491         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << newRate);
 492         self_destruct();
 493     }
 494
 495     if (!isaReconfig)
 496         *storedRate = newRate;
 497     else if (*storedRate != newRate) {
 498         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 499                << " cannot be changed dynamically, value left unchanged: " <<
 500                *storedRate);
 501     }
 502
 503     return true;
 504 }
 505
 506 /// reports rate-specific options; mimics ::SwapDir::optionObjectSizeDump()
 507 void
 508 Rock::SwapDir::dumpRateOption(StoreEntry * e) const
 509 {
 510     if (fileConfig.ioRate >= 0)
 511         storeAppendPrintf(e, " max-swap-rate=%d", fileConfig.ioRate);
 512 }
 513
 514 /// parses size-specific options; mimics ::SwapDir::optionObjectSizeParse()
 515 bool
 516 Rock::SwapDir::parseSizeOption(char const *option, const char *value, int reconfig)
 517 {
 518     uint64_t *storedSize;
 519     if (strcmp(option, "slot-size") == 0)
 520         storedSize = &slotSize;
 521     else
 522         return false;
 523
 524     if (!value)
 525         self_destruct();
 526
 527     // TODO: handle size units and detect parsing errors better
 528     const uint64_t newSize = strtoll(value, NULL, 10);
 529     if (newSize <= 0) {
 530         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must be positive; got: " << newSize);
 531         self_destruct();
 532     }
 533
 534     if (newSize <= sizeof(DbCellHeader)) {
 535         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must exceed " << sizeof(DbCellHeader) << "; got: " << newSize);
 536         self_destruct();
 537     }
 538
 539     if (!reconfig)
 540         *storedSize = newSize;
 541     else if (*storedSize != newSize) {
 542         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 543                << " cannot be changed dynamically, value left unchanged: " <<
 544                *storedSize);
 545     }
 546
 547     return true;
 548 }
 549
 550 /// reports size-specific options; mimics ::SwapDir::optionObjectSizeDump()
 551 void
 552 Rock::SwapDir::dumpSizeOption(StoreEntry * e) const
 553 {
 554     storeAppendPrintf(e, " slot-size=%" PRId64, slotSize);
 555 }
 556
 557 /// check the results of the configuration; only level-0 debugging works here
 558 void
 559 Rock::SwapDir::validateOptions()
 560 {
 561     if (slotSize <= 0)
 562         fatal("Rock store requires a positive slot-size");
 563
 564     const int64_t maxSizeRoundingWaste = 1024 * 1024; // size is configured in MB
 565     const int64_t slotSizeRoundingWaste = slotSize;
 566     const int64_t maxRoundingWaste =
 567         max(maxSizeRoundingWaste, slotSizeRoundingWaste);
 568
 569     // an entry consumes at least one slot; round up to reduce false warnings
 570     const int64_t blockSize = static_cast<int64_t>(slotSize);
 571     const int64_t maxObjSize = max(blockSize,
 572                                    ((maxObjectSize()+blockSize-1)/blockSize)*blockSize);
 573
 574     // Does the "sfileno*max-size" limit match configured db capacity?
 575     const double entriesMayOccupy = entryLimitAbsolute()*static_cast<double>(maxObjSize);
 576     if (entriesMayOccupy + maxRoundingWaste < maxSize()) {
 577         const int64_t diskWasteSize = maxSize() - static_cast<int64_t>(entriesMayOccupy);
 578         debugs(47, DBG_CRITICAL, "WARNING: Rock cache_dir " << path << " wastes disk space due to entry limits:" <<
 579                "\n\tconfigured db capacity: " << maxSize() << " bytes" <<
 580                "\n\tconfigured db slot size: " << slotSize << " bytes" <<
 581                "\n\tconfigured maximum entry size: " << maxObjectSize() << " bytes" <<
 582                "\n\tmaximum number of cache_dir entries supported by Squid: " << entryLimitAbsolute() <<
 583                "\n\tdisk space all entries may use: " << entriesMayOccupy << " bytes" <<
 584                "\n\tdisk space wasted: " << diskWasteSize << " bytes");
 585     }
 586
 587     // Does the "absolute slot count" limit match configured db capacity?
 588     const double slotsMayOccupy = slotLimitAbsolute()*static_cast<double>(slotSize);
 589     if (slotsMayOccupy + maxRoundingWaste < maxSize()) {
 590         const int64_t diskWasteSize = maxSize() - static_cast<int64_t>(entriesMayOccupy);
 591         debugs(47, DBG_CRITICAL, "WARNING: Rock cache_dir " << path << " wastes disk space due to slot limits:" <<
 592                "\n\tconfigured db capacity: " << maxSize() << " bytes" <<
 593                "\n\tconfigured db slot size: " << slotSize << " bytes" <<
 594                "\n\tmaximum number of rock cache_dir slots supported by Squid: " << slotLimitAbsolute() <<
 595                "\n\tdisk space all slots may use: " << slotsMayOccupy << " bytes" <<
 596                "\n\tdisk space wasted: " << diskWasteSize << " bytes");
 597     }
 598 }
 599
 600 void
 601 Rock::SwapDir::rebuild()
 602 {
 603     //++StoreController::store_dirs_rebuilding; // see Rock::SwapDir::init()
 604     AsyncJob::Start(new Rebuild(this));
 605 }
 606
 607 bool
 608 Rock::SwapDir::canStore(const StoreEntry &e, int64_t diskSpaceNeeded, int &load) const
 609 {
 610     if (!::SwapDir::canStore(e, sizeof(DbCellHeader)+diskSpaceNeeded, load))
 611         return false;
 612
 613     if (!theFile || !theFile->canWrite())
 614         return false;
 615
 616     if (!map)
 617         return false;
 618
 619     // Do not start I/O transaction if there are less than 10% free pages left.
 620     // TODO: reserve page instead
 621     if (needsDiskStrand() &&
 622             Ipc::Mem::PageLevel(Ipc::Mem::PageId::ioPage) >= 0.9 * Ipc::Mem::PageLimit(Ipc::Mem::PageId::ioPage)) {
 623         debugs(47, 5, HERE << "too few shared pages for IPC I/O left");
 624         return false;
 625     }
 626
 627     if (io->shedLoad())
 628         return false;
 629
 630     load = io->load();
 631     return true;
 632 }
 633
 634 StoreIOState::Pointer
 635 Rock::SwapDir::createStoreIO(StoreEntry &e, StoreIOState::STFNCB *cbFile, StoreIOState::STIOCB *cbIo, void *data)
 636 {
 637     if (!theFile || theFile->error()) {
 638         debugs(47,4, HERE << theFile);
 639         return NULL;
 640     }
 641
 642     sfileno filen;
 643     Ipc::StoreMapAnchor *const slot =
 644         map->openForWriting(reinterpret_cast<const cache_key *>(e.key), filen);
 645     if (!slot) {
 646         debugs(47, 5, HERE << "map->add failed");
 647         return NULL;
 648     }
 649
 650     assert(filen >= 0);
 651     slot->set(e);
 652
 653     // XXX: We rely on our caller, storeSwapOutStart(), to set e.fileno.
 654     // If that does not happen, the entry will not decrement the read level!
 655
 656     Rock::SwapDir::Pointer self(this);
 657     IoState *sio = new IoState(self, &e, cbFile, cbIo, data);
 658
 659     sio->swap_dirn = index;
 660     sio->swap_filen = filen;
 661     sio->writeableAnchor_ = slot;
 662
 663     debugs(47,5, HERE << "dir " << index << " created new filen " <<
 664            std::setfill('0') << std::hex << std::uppercase << std::setw(8) <<
 665            sio->swap_filen << std::dec << " starting at " <<
 666            diskOffset(sio->swap_filen));
 667
 668     sio->file(theFile);
 669
 670     trackReferences(e);
 671     return sio;
 672 }
 673
 674 int64_t
 675 Rock::SwapDir::diskOffset(const SlotId sid) const
 676 {
 677     assert(sid >= 0);
 678     return HeaderSize + slotSize*sid;
 679 }
 680
 681 int64_t
 682 Rock::SwapDir::diskOffset(Ipc::Mem::PageId &pageId) const
 683 {
 684     assert(pageId);
 685     return diskOffset(pageId.number - 1);
 686 }
 687
 688 int64_t
 689 Rock::SwapDir::diskOffsetLimit() const
 690 {
 691     assert(map);
 692     return diskOffset(map->sliceLimit());
 693 }
 694
 695 bool
 696 Rock::SwapDir::useFreeSlot(Ipc::Mem::PageId &pageId)
 697 {
 698     if (freeSlots->pop(pageId)) {
 699         debugs(47, 5, "got a previously free slot: " << pageId);
 700         return true;
 701     }
 702
 703     // catch free slots delivered to noteFreeMapSlice()
 704     assert(!waitingForPage);
 705     waitingForPage = &pageId;
 706     if (map->purgeOne()) {
 707         assert(!waitingForPage); // noteFreeMapSlice() should have cleared it
 708         assert(pageId.set());
 709         debugs(47, 5, "got a previously busy slot: " << pageId);
 710         return true;
 711     }
 712     assert(waitingForPage == &pageId);
 713     waitingForPage = NULL;
 714
 715     debugs(47, 3, "cannot get a slot; entries: " << map->entryCount());
 716     return false;
 717 }
 718
 719 bool
 720 Rock::SwapDir::validSlotId(const SlotId slotId) const
 721 {
 722     return 0 <= slotId && slotId < slotLimitActual();
 723 }
 724
 725 void
 726 Rock::SwapDir::noteFreeMapSlice(const Ipc::StoreMapSliceId sliceId)
 727 {
 728     Ipc::Mem::PageId pageId;
 729     pageId.pool = index+1;
 730     pageId.number = sliceId+1;
 731     if (waitingForPage) {
 732         *waitingForPage = pageId;
 733         waitingForPage = NULL;
 734     } else {
 735         freeSlots->push(pageId);
 736     }
 737 }
 738
 739 // tries to open an old entry with swap_filen for reading
 740 StoreIOState::Pointer
 741 Rock::SwapDir::openStoreIO(StoreEntry &e, StoreIOState::STFNCB *cbFile, StoreIOState::STIOCB *cbIo, void *data)
 742 {
 743     if (!theFile || theFile->error()) {
 744         debugs(47,4, HERE << theFile);
 745         return NULL;
 746     }
 747
 748     if (e.swap_filen < 0) {
 749         debugs(47,4, HERE << e);
 750         return NULL;
 751     }
 752
 753     // Do not start I/O transaction if there are less than 10% free pages left.
 754     // TODO: reserve page instead
 755     if (needsDiskStrand() &&
 756             Ipc::Mem::PageLevel(Ipc::Mem::PageId::ioPage) >= 0.9 * Ipc::Mem::PageLimit(Ipc::Mem::PageId::ioPage)) {
 757         debugs(47, 5, HERE << "too few shared pages for IPC I/O left");
 758         return NULL;
 759     }
 760
 761     // The are two ways an entry can get swap_filen: our get() locked it for
 762     // reading or our storeSwapOutStart() locked it for writing. Peeking at our
 763     // locked entry is safe, but no support for reading the entry we swap out.
 764     const Ipc::StoreMapAnchor *slot = map->peekAtReader(e.swap_filen);
 765     if (!slot)
 766         return NULL; // we were writing afterall
 767
 768     Rock::SwapDir::Pointer self(this);
 769     IoState *sio = new IoState(self, &e, cbFile, cbIo, data);
 770
 771     sio->swap_dirn = index;
 772     sio->swap_filen = e.swap_filen;
 773     sio->readableAnchor_ = slot;
 774     sio->file(theFile);
 775
 776     debugs(47,5, HERE << "dir " << index << " has old filen: " <<
 777            std::setfill('0') << std::hex << std::uppercase << std::setw(8) <<
 778            sio->swap_filen);
 779
 780     assert(slot->sameKey(static_cast<const cache_key*>(e.key)));
 781     // For collapsed disk hits: e.swap_file_sz and slot->basics.swap_file_sz
 782     // may still be zero and basics.swap_file_sz may grow.
 783     assert(slot->basics.swap_file_sz >= e.swap_file_sz);
 784
 785     return sio;
 786 }
 787
 788 void
 789 Rock::SwapDir::ioCompletedNotification()
 790 {
 791     if (!theFile)
 792         fatalf("Rock cache_dir failed to initialize db file: %s", filePath);
 793
 794     if (theFile->error())
 795         fatalf("Rock cache_dir at %s failed to open db file: %s", filePath,
 796                xstrerror());
 797
 798     debugs(47, 2, "Rock cache_dir[" << index << "] limits: " <<
 799            std::setw(12) << maxSize() << " disk bytes, " <<
 800            std::setw(7) << map->entryLimit() << " entries, and " <<
 801            std::setw(7) << map->sliceLimit() << " slots");
 802
 803     rebuild();
 804 }
 805
 806 void
 807 Rock::SwapDir::closeCompleted()
 808 {
 809     theFile = NULL;
 810 }
 811
 812 void
 813 Rock::SwapDir::readCompleted(const char *, int rlen, int errflag, RefCount< ::ReadRequest> r)
 814 {
 815     ReadRequest *request = dynamic_cast<Rock::ReadRequest*>(r.getRaw());
 816     assert(request);
 817     IoState::Pointer sio = request->sio;
 818
 819     if (errflag == DISK_OK && rlen > 0)
 820         sio->offset_ += rlen;
 821
 822     sio->callReaderBack(r->buf, rlen);
 823 }
 824
 825 void
 826 Rock::SwapDir::writeCompleted(int errflag, size_t, RefCount< ::WriteRequest> r)
 827 {
 828     Rock::WriteRequest *request = dynamic_cast<Rock::WriteRequest*>(r.getRaw());
 829     assert(request);
 830     assert(request->sio !=  NULL);
 831     IoState &sio = *request->sio;
 832
 833     // quit if somebody called IoState::close() while we were waiting
 834     if (!sio.stillWaiting()) {
 835         debugs(79, 3, "ignoring closed entry " << sio.swap_filen);
 836         noteFreeMapSlice(request->sidNext);
 837         return;
 838     }
 839
 840     // TODO: Fail if disk dropped one of the previous write requests.
 841
 842     if (errflag == DISK_OK) {
 843         // do not increment sio.offset_ because we do it in sio->write()
 844
 845         // finalize the shared slice info after writing slice contents to disk
 846         Ipc::StoreMap::Slice &slice =
 847             map->writeableSlice(sio.swap_filen, request->sidCurrent);
 848         slice.size = request->len - sizeof(DbCellHeader);
 849         slice.next = request->sidNext;
 850
 851         if (request->eof) {
 852             assert(sio.e);
 853             assert(sio.writeableAnchor_);
 854             sio.e->swap_file_sz = sio.writeableAnchor_->basics.swap_file_sz =
 855                                       sio.offset_;
 856
 857             // close, the entry gets the read lock
 858             map->closeForWriting(sio.swap_filen, true);
 859             sio.writeableAnchor_ = NULL;
 860             sio.finishedWriting(errflag);
 861         }
 862     } else {
 863         noteFreeMapSlice(request->sidNext);
 864
 865         writeError(*sio.e);
 866         sio.finishedWriting(errflag);
 867         // and hope that Core will call disconnect() to close the map entry
 868     }
 869
 870     CollapsedForwarding::Broadcast(*sio.e);
 871 }
 872
 873 void
 874 Rock::SwapDir::writeError(StoreEntry &e)
 875 {
 876     // Do not abortWriting here. The entry should keep the write lock
 877     // instead of losing association with the store and confusing core.
 878     map->freeEntry(e.swap_filen); // will mark as unusable, just in case
 879
 880     Store::Root().transientsAbandon(e);
 881
 882     // All callers must also call IoState callback, to propagate the error.
 883 }
 884
 885 bool
 886 Rock::SwapDir::full() const
 887 {
 888     return freeSlots != NULL && !freeSlots->size();
 889 }
 890
 891 // storeSwapOutFileClosed calls this nethod on DISK_NO_SPACE_LEFT,
 892 // but it should not happen for us
 893 void
 894 Rock::SwapDir::diskFull()
 895 {
 896     debugs(20, DBG_IMPORTANT, "BUG: No space left with rock cache_dir: " <<
 897            filePath);
 898 }
 899
 900 /// purge while full(); it should be sufficient to purge just one
 901 void
 902 Rock::SwapDir::maintain()
 903 {
 904     // The Store calls this to free some db space, but there is nothing wrong
 905     // with a full() db, except when db has to shrink after reconfigure, and
 906     // we do not support shrinking yet (it would have to purge specific slots).
 907     // TODO: Disable maintain() requests when they are pointless.
 908 }
 909
 910 void
 911 Rock::SwapDir::reference(StoreEntry &e)
 912 {
 913     debugs(47, 5, HERE << &e << ' ' << e.swap_dirn << ' ' << e.swap_filen);
 914     if (repl && repl->Referenced)
 915         repl->Referenced(repl, &e, &e.repl);
 916 }
 917
 918 bool
 919 Rock::SwapDir::dereference(StoreEntry &e, bool)
 920 {
 921     debugs(47, 5, HERE << &e << ' ' << e.swap_dirn << ' ' << e.swap_filen);
 922     if (repl && repl->Dereferenced)
 923         repl->Dereferenced(repl, &e, &e.repl);
 924
 925     // no need to keep e in the global store_table for us; we have our own map
 926     return false;
 927 }
 928
 929 bool
 930 Rock::SwapDir::unlinkdUseful() const
 931 {
 932     // no entry-specific files to unlink
 933     return false;
 934 }
 935
 936 void
 937 Rock::SwapDir::unlink(StoreEntry &e)
 938 {
 939     debugs(47, 5, HERE << e);
 940     ignoreReferences(e);
 941     map->freeEntry(e.swap_filen);
 942     disconnect(e);
 943 }
 944
 945 void
 946 Rock::SwapDir::markForUnlink(StoreEntry &e)
 947 {
 948     debugs(47, 5, e);
 949     map->freeEntry(e.swap_filen);
 950 }
 951
 952 void
 953 Rock::SwapDir::trackReferences(StoreEntry &e)
 954 {
 955     debugs(47, 5, HERE << e);
 956     if (repl)
 957         repl->Add(repl, &e, &e.repl);
 958 }
 959
 960 void
 961 Rock::SwapDir::ignoreReferences(StoreEntry &e)
 962 {
 963     debugs(47, 5, HERE << e);
 964     if (repl)
 965         repl->Remove(repl, &e, &e.repl);
 966 }
 967
 968 void
 969 Rock::SwapDir::statfs(StoreEntry &e) const
 970 {
 971     storeAppendPrintf(&e, "\n");
 972     storeAppendPrintf(&e, "Maximum Size: %" PRIu64 " KB\n", maxSize() >> 10);
 973     storeAppendPrintf(&e, "Current Size: %.2f KB %.2f%%\n",
 974                       currentSize() / 1024.0,
 975                       Math::doublePercent(currentSize(), maxSize()));
 976
 977     const int entryLimit = entryLimitActual();
 978     const int slotLimit = slotLimitActual();
 979     storeAppendPrintf(&e, "Maximum entries: %9d\n", entryLimit);
 980     if (map && entryLimit > 0) {
 981         const int entryCount = map->entryCount();
 982         storeAppendPrintf(&e, "Current entries: %9d %.2f%%\n",
 983                           entryCount, (100.0 * entryCount / entryLimit));
 984     }
 985
 986     storeAppendPrintf(&e, "Maximum slots:   %9d\n", slotLimit);
 987     if (map && slotLimit > 0) {
 988         const unsigned int slotsFree = !freeSlots ? 0 : freeSlots->size();
 989         if (slotsFree <= static_cast<const unsigned int>(slotLimit)) {
 990             const int usedSlots = slotLimit - static_cast<const int>(slotsFree);
 991             storeAppendPrintf(&e, "Used slots:      %9d %.2f%%\n",
 992                               usedSlots, (100.0 * usedSlots / slotLimit));
 993         }
 994         if (slotLimit < 100) { // XXX: otherwise too expensive to count
 995             Ipc::ReadWriteLockStats stats;
 996             map->updateStats(stats);
 997             stats.dump(e);
 998         }
 999     }
1000
1001     storeAppendPrintf(&e, "Pending operations: %d out of %d\n",
1002                       store_open_disk_fd, Config.max_open_disk_fds);
1003
1004     storeAppendPrintf(&e, "Flags:");
1005
1006     if (flags.selected)
1007         storeAppendPrintf(&e, " SELECTED");
1008
1009     if (flags.read_only)
1010         storeAppendPrintf(&e, " READ-ONLY");
1011
1012     storeAppendPrintf(&e, "\n");
1013
1014 }
1015
1016 SBuf
1017 Rock::SwapDir::inodeMapPath() const
1018 {
1019     return Ipc::Mem::Segment::Name(SBuf(path), "map");
1020 }
1021
1022 const char *
1023 Rock::SwapDir::freeSlotsPath() const
1024 {
1025     static String spacesPath;
1026     spacesPath = path;
1027     spacesPath.append("_spaces");
1028     return spacesPath.termedBuf();
1029 }
1030
1031 namespace Rock
1032 {
1033 RunnerRegistrationEntry(SwapDirRr);
1034 }
1035
1036 void Rock::SwapDirRr::create()
1037 {
1038     Must(mapOwners.empty() && freeSlotsOwners.empty());
1039     for (int i = 0; i < Config.cacheSwap.n_configured; ++i) {
1040         if (const Rock::SwapDir *const sd = dynamic_cast<Rock::SwapDir *>(INDEXSD(i))) {
1041             const int64_t capacity = sd->slotLimitActual();
1042
1043             SwapDir::DirMap::Owner *const mapOwner =
1044                 SwapDir::DirMap::Init(sd->inodeMapPath(), capacity);
1045             mapOwners.push_back(mapOwner);
1046
1047             // TODO: somehow remove pool id and counters from PageStack?
1048             Ipc::Mem::Owner<Ipc::Mem::PageStack> *const freeSlotsOwner =
1049                 shm_new(Ipc::Mem::PageStack)(sd->freeSlotsPath(),
1050                                              i+1, capacity, 0);
1051             freeSlotsOwners.push_back(freeSlotsOwner);
1052
1053             // TODO: add method to initialize PageStack with no free pages
1054             while (true) {
1055                 Ipc::Mem::PageId pageId;
1056                 if (!freeSlotsOwner->object()->pop(pageId))
1057                     break;
1058             }
1059         }
1060     }
1061 }
1062
1063 Rock::SwapDirRr::~SwapDirRr()
1064 {
1065     for (size_t i = 0; i < mapOwners.size(); ++i) {
1066         delete mapOwners[i];
1067         delete freeSlotsOwners[i];
1068     }
1069 }
1070