src/fs/rock/RockSwapDir.cc

   1 /*
   2  * Copyright (C) 1996-2016 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/RockHeaderUpdater.h"
  20 #include "fs/rock/RockIoRequests.h"
  21 #include "fs/rock/RockIoState.h"
  22 #include "fs/rock/RockRebuild.h"
  23 #include "fs/rock/RockSwapDir.h"
  24 #include "globals.h"
  25 #include "ipc/mem/Pages.h"
  26 #include "MemObject.h"
  27 #include "Parsing.h"
  28 #include "SquidConfig.h"
  29 #include "SquidMath.h"
  30 #include "tools.h"
  31
  32 #include <cstdlib>
  33 #include <iomanip>
  34 #include <limits>
  35
  36 #if HAVE_SYS_STAT_H
  37 #include <sys/stat.h>
  38 #endif
  39
  40 const int64_t Rock::SwapDir::HeaderSize = 16*1024;
  41
  42 Rock::SwapDir::SwapDir(): ::SwapDir("rock"),
  43     slotSize(HeaderSize), filePath(NULL), map(NULL), io(NULL),
  44     waitingForPage(NULL)
  45 {
  46 }
  47
  48 Rock::SwapDir::~SwapDir()
  49 {
  50     delete io;
  51     delete map;
  52     safe_free(filePath);
  53 }
  54
  55 // called when Squid core needs a StoreEntry with a given key
  56 StoreEntry *
  57 Rock::SwapDir::get(const cache_key *key)
  58 {
  59     if (!map || !theFile || !theFile->canRead())
  60         return NULL;
  61
  62     sfileno filen;
  63     const Ipc::StoreMapAnchor *const slot = map->openForReading(key, filen);
  64     if (!slot)
  65         return NULL;
  66
  67     // create a brand new store entry and initialize it with stored basics
  68     StoreEntry *e = new StoreEntry();
  69     anchorEntry(*e, filen, *slot);
  70
  71     e->hashInsert(key);
  72     trackReferences(*e);
  73
  74     return e;
  75     // the disk entry remains open for reading, protected from modifications
  76 }
  77
  78 bool
  79 Rock::SwapDir::anchorCollapsed(StoreEntry &collapsed, bool &inSync)
  80 {
  81     if (!map || !theFile || !theFile->canRead())
  82         return false;
  83
  84     sfileno filen;
  85     const Ipc::StoreMapAnchor *const slot = map->openForReading(
  86             reinterpret_cast<cache_key*>(collapsed.key), filen);
  87     if (!slot)
  88         return false;
  89
  90     anchorEntry(collapsed, filen, *slot);
  91     inSync = updateCollapsedWith(collapsed, *slot);
  92     return true; // even if inSync is false
  93 }
  94
  95 bool
  96 Rock::SwapDir::updateCollapsed(StoreEntry &collapsed)
  97 {
  98     if (!map || !theFile || !theFile->canRead())
  99         return false;
 100
 101     if (collapsed.swap_filen < 0) // no longer using a disk cache
 102         return true;
 103     assert(collapsed.swap_dirn == index);
 104
 105     const Ipc::StoreMapAnchor &s = map->readableEntry(collapsed.swap_filen);
 106     return updateCollapsedWith(collapsed, s);
 107 }
 108
 109 bool
 110 Rock::SwapDir::updateCollapsedWith(StoreEntry &collapsed, const Ipc::StoreMapAnchor &anchor)
 111 {
 112     collapsed.swap_file_sz = anchor.basics.swap_file_sz;
 113     return true;
 114 }
 115
 116 void
 117 Rock::SwapDir::anchorEntry(StoreEntry &e, const sfileno filen, const Ipc::StoreMapAnchor &anchor)
 118 {
 119     const Ipc::StoreMapAnchor::Basics &basics = anchor.basics;
 120
 121     e.swap_file_sz = basics.swap_file_sz;
 122     e.lastref = basics.lastref;
 123     e.timestamp = basics.timestamp;
 124     e.expires = basics.expires;
 125     e.lastmod = basics.lastmod;
 126     e.refcount = basics.refcount;
 127     e.flags = basics.flags;
 128
 129     if (anchor.complete()) {
 130         e.store_status = STORE_OK;
 131         e.swap_status = SWAPOUT_DONE;
 132     } else {
 133         e.store_status = STORE_PENDING;
 134         e.swap_status = SWAPOUT_WRITING; // even though another worker writes?
 135     }
 136
 137     e.ping_status = PING_NONE;
 138
 139     EBIT_CLR(e.flags, RELEASE_REQUEST);
 140     EBIT_CLR(e.flags, KEY_PRIVATE);
 141     EBIT_SET(e.flags, ENTRY_VALIDATED);
 142
 143     e.swap_dirn = index;
 144     e.swap_filen = filen;
 145 }
 146
 147 void Rock::SwapDir::disconnect(StoreEntry &e)
 148 {
 149     assert(e.swap_dirn == index);
 150     assert(e.swap_filen >= 0);
 151     // cannot have SWAPOUT_NONE entry with swap_filen >= 0
 152     assert(e.swap_status != SWAPOUT_NONE);
 153
 154     // do not rely on e.swap_status here because there is an async delay
 155     // before it switches from SWAPOUT_WRITING to SWAPOUT_DONE.
 156
 157     // since e has swap_filen, its slot is locked for reading and/or writing
 158     // but it is difficult to know whether THIS worker is reading or writing e,
 159     // especially since we may switch from writing to reading. This code relies
 160     // on Rock::IoState::writeableAnchor_ being set when we locked for writing.
 161     if (e.mem_obj && e.mem_obj->swapout.sio != NULL &&
 162             dynamic_cast<IoState&>(*e.mem_obj->swapout.sio).writeableAnchor_) {
 163         map->abortWriting(e.swap_filen);
 164         e.swap_dirn = -1;
 165         e.swap_filen = -1;
 166         e.swap_status = SWAPOUT_NONE;
 167         dynamic_cast<IoState&>(*e.mem_obj->swapout.sio).writeableAnchor_ = NULL;
 168         Store::Root().transientsAbandon(e); // broadcasts after the change
 169     } else {
 170         map->closeForReading(e.swap_filen);
 171         e.swap_dirn = -1;
 172         e.swap_filen = -1;
 173         e.swap_status = SWAPOUT_NONE;
 174     }
 175 }
 176
 177 uint64_t
 178 Rock::SwapDir::currentSize() const
 179 {
 180     const uint64_t spaceSize = !freeSlots ?
 181                                maxSize() : (slotSize * freeSlots->size());
 182     // everything that is not free is in use
 183     return maxSize() - spaceSize;
 184 }
 185
 186 uint64_t
 187 Rock::SwapDir::currentCount() const
 188 {
 189     return map ? map->entryCount() : 0;
 190 }
 191
 192 /// In SMP mode only the disker process reports stats to avoid
 193 /// counting the same stats by multiple processes.
 194 bool
 195 Rock::SwapDir::doReportStat() const
 196 {
 197     return ::SwapDir::doReportStat() && (!UsingSmp() || IamDiskProcess());
 198 }
 199
 200 void
 201 Rock::SwapDir::swappedOut(const StoreEntry &)
 202 {
 203     // stats are not stored but computed when needed
 204 }
 205
 206 int64_t
 207 Rock::SwapDir::slotLimitAbsolute() const
 208 {
 209     // the max value is an invalid one; all values must be below the limit
 210     assert(std::numeric_limits<Ipc::StoreMapSliceId>::max() ==
 211            std::numeric_limits<SlotId>::max());
 212     return std::numeric_limits<SlotId>::max();
 213 }
 214
 215 int64_t
 216 Rock::SwapDir::slotLimitActual() const
 217 {
 218     const int64_t sWanted = (maxSize() - HeaderSize)/slotSize;
 219     const int64_t sLimitLo = map ? map->sliceLimit() : 0; // dynamic shrinking unsupported
 220     const int64_t sLimitHi = slotLimitAbsolute();
 221     return min(max(sLimitLo, sWanted), sLimitHi);
 222 }
 223
 224 int64_t
 225 Rock::SwapDir::entryLimitActual() const
 226 {
 227     return min(slotLimitActual(), entryLimitAbsolute());
 228 }
 229
 230 // TODO: encapsulate as a tool
 231 void
 232 Rock::SwapDir::create()
 233 {
 234     assert(path);
 235     assert(filePath);
 236
 237     if (UsingSmp() && !IamDiskProcess()) {
 238         debugs (47,3, HERE << "disker will create in " << path);
 239         return;
 240     }
 241
 242     debugs (47,3, HERE << "creating in " << path);
 243
 244     struct stat dir_sb;
 245     if (::stat(path, &dir_sb) == 0) {
 246         struct stat file_sb;
 247         if (::stat(filePath, &file_sb) == 0) {
 248             debugs (47, DBG_IMPORTANT, "Skipping existing Rock db: " << filePath);
 249             return;
 250         }
 251         // else the db file is not there or is not accessible, and we will try
 252         // to create it later below, generating a detailed error on failures.
 253     } else { // path does not exist or is inaccessible
 254         // If path exists but is not accessible, mkdir() below will fail, and
 255         // the admin should see the error and act accordingly, so there is
 256         // no need to distinguish ENOENT from other possible stat() errors.
 257         debugs (47, DBG_IMPORTANT, "Creating Rock db directory: " << path);
 258         const int res = mkdir(path, 0700);
 259         if (res != 0)
 260             createError("mkdir");
 261     }
 262
 263     debugs (47, DBG_IMPORTANT, "Creating Rock db: " << filePath);
 264     const int swap = open(filePath, O_WRONLY|O_CREAT|O_TRUNC|O_BINARY, 0600);
 265     if (swap < 0)
 266         createError("create");
 267
 268 #if SLOWLY_FILL_WITH_ZEROS
 269     char block[1024];
 270     Must(maxSize() % sizeof(block) == 0);
 271     memset(block, '\0', sizeof(block));
 272
 273     for (off_t offset = 0; offset < maxSize(); offset += sizeof(block)) {
 274         if (write(swap, block, sizeof(block)) != sizeof(block))
 275             createError("write");
 276     }
 277 #else
 278     if (ftruncate(swap, maxSize()) != 0)
 279         createError("truncate");
 280
 281     char header[HeaderSize];
 282     memset(header, '\0', sizeof(header));
 283     if (write(swap, header, sizeof(header)) != sizeof(header))
 284         createError("write");
 285 #endif
 286
 287     close(swap);
 288 }
 289
 290 // report Rock DB creation error and exit
 291 void
 292 Rock::SwapDir::createError(const char *const msg)
 293 {
 294     debugs(47, DBG_CRITICAL, "ERROR: Failed to initialize Rock Store db in " <<
 295            filePath << "; " << msg << " error: " << xstrerror());
 296     fatal("Rock Store db creation error");
 297 }
 298
 299 void
 300 Rock::SwapDir::init()
 301 {
 302     debugs(47,2, HERE);
 303
 304     // XXX: SwapDirs aren't refcounted. We make IORequestor calls, which
 305     // are refcounted. We up our count once to avoid implicit delete's.
 306     lock();
 307
 308     freeSlots = shm_old(Ipc::Mem::PageStack)(freeSlotsPath());
 309
 310     Must(!map);
 311     map = new DirMap(inodeMapPath());
 312     map->cleaner = this;
 313
 314     const char *ioModule = needsDiskStrand() ? "IpcIo" : "Blocking";
 315     if (DiskIOModule *m = DiskIOModule::Find(ioModule)) {
 316         debugs(47,2, HERE << "Using DiskIO module: " << ioModule);
 317         io = m->createStrategy();
 318         io->init();
 319     } else {
 320         debugs(47, DBG_CRITICAL, "FATAL: Rock store is missing DiskIO module: " <<
 321                ioModule);
 322         fatal("Rock Store missing a required DiskIO module");
 323     }
 324
 325     theFile = io->newFile(filePath);
 326     theFile->configure(fileConfig);
 327     theFile->open(O_RDWR, 0644, this);
 328
 329     // Increment early. Otherwise, if one SwapDir finishes rebuild before
 330     // others start, storeRebuildComplete() will think the rebuild is over!
 331     // TODO: move store_dirs_rebuilding hack to store modules that need it.
 332     ++StoreController::store_dirs_rebuilding;
 333 }
 334
 335 bool
 336 Rock::SwapDir::needsDiskStrand() const
 337 {
 338     const bool wontEvenWorkWithoutDisker = Config.workers > 1;
 339     const bool wouldWorkBetterWithDisker = DiskIOModule::Find("IpcIo");
 340     return InDaemonMode() && (wontEvenWorkWithoutDisker ||
 341                               wouldWorkBetterWithDisker);
 342 }
 343
 344 void
 345 Rock::SwapDir::parse(int anIndex, char *aPath)
 346 {
 347     index = anIndex;
 348
 349     path = xstrdup(aPath);
 350
 351     // cache store is located at path/db
 352     String fname(path);
 353     fname.append("/rock");
 354     filePath = xstrdup(fname.termedBuf());
 355
 356     parseSize(false);
 357     parseOptions(0);
 358
 359     // Current openForWriting() code overwrites the old slot if needed
 360     // and possible, so proactively removing old slots is probably useless.
 361     assert(!repl); // repl = createRemovalPolicy(Config.replPolicy);
 362
 363     validateOptions();
 364 }
 365
 366 void
 367 Rock::SwapDir::reconfigure()
 368 {
 369     parseSize(true);
 370     parseOptions(1);
 371     // TODO: can we reconfigure the replacement policy (repl)?
 372     validateOptions();
 373 }
 374
 375 /// parse maximum db disk size
 376 void
 377 Rock::SwapDir::parseSize(const bool reconfig)
 378 {
 379     const int i = GetInteger();
 380     if (i < 0)
 381         fatal("negative Rock cache_dir size value");
 382     const uint64_t new_max_size =
 383         static_cast<uint64_t>(i) << 20; // MBytes to Bytes
 384     if (!reconfig)
 385         max_size = new_max_size;
 386     else if (new_max_size != max_size) {
 387         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir '" << path << "' size "
 388                "cannot be changed dynamically, value left unchanged (" <<
 389                (max_size >> 20) << " MB)");
 390     }
 391 }
 392
 393 ConfigOption *
 394 Rock::SwapDir::getOptionTree() const
 395 {
 396     ConfigOption *copt = ::SwapDir::getOptionTree();
 397     ConfigOptionVector *vector = dynamic_cast<ConfigOptionVector*>(copt);
 398     if (vector) {
 399         // if copt is actually a ConfigOptionVector
 400         vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseSizeOption, &SwapDir::dumpSizeOption));
 401         vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseTimeOption, &SwapDir::dumpTimeOption));
 402         vector->options.push_back(new ConfigOptionAdapter<SwapDir>(*const_cast<SwapDir *>(this), &SwapDir::parseRateOption, &SwapDir::dumpRateOption));
 403     } else {
 404         // we don't know how to handle copt, as it's not a ConfigOptionVector.
 405         // free it (and return nullptr)
 406         delete copt;
 407         copt = nullptr;
 408     }
 409     return copt;
 410 }
 411
 412 bool
 413 Rock::SwapDir::allowOptionReconfigure(const char *const option) const
 414 {
 415     return strcmp(option, "slot-size") != 0 &&
 416            ::SwapDir::allowOptionReconfigure(option);
 417 }
 418
 419 /// parses time-specific options; mimics ::SwapDir::optionObjectSizeParse()
 420 bool
 421 Rock::SwapDir::parseTimeOption(char const *option, const char *value, int reconfig)
 422 {
 423     // TODO: ::SwapDir or, better, Config should provide time-parsing routines,
 424     // including time unit handling. Same for size and rate.
 425
 426     time_msec_t *storedTime;
 427     if (strcmp(option, "swap-timeout") == 0)
 428         storedTime = &fileConfig.ioTimeout;
 429     else
 430         return false;
 431
 432     if (!value)
 433         self_destruct();
 434
 435     // TODO: handle time units and detect parsing errors better
 436     const int64_t parsedValue = strtoll(value, NULL, 10);
 437     if (parsedValue < 0) {
 438         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << parsedValue);
 439         self_destruct();
 440     }
 441
 442     const time_msec_t newTime = static_cast<time_msec_t>(parsedValue);
 443
 444     if (!reconfig)
 445         *storedTime = newTime;
 446     else if (*storedTime != newTime) {
 447         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 448                << " cannot be changed dynamically, value left unchanged: " <<
 449                *storedTime);
 450     }
 451
 452     return true;
 453 }
 454
 455 /// reports time-specific options; mimics ::SwapDir::optionObjectSizeDump()
 456 void
 457 Rock::SwapDir::dumpTimeOption(StoreEntry * e) const
 458 {
 459     if (fileConfig.ioTimeout)
 460         storeAppendPrintf(e, " swap-timeout=%" PRId64,
 461                           static_cast<int64_t>(fileConfig.ioTimeout));
 462 }
 463
 464 /// parses rate-specific options; mimics ::SwapDir::optionObjectSizeParse()
 465 bool
 466 Rock::SwapDir::parseRateOption(char const *option, const char *value, int isaReconfig)
 467 {
 468     int *storedRate;
 469     if (strcmp(option, "max-swap-rate") == 0)
 470         storedRate = &fileConfig.ioRate;
 471     else
 472         return false;
 473
 474     if (!value)
 475         self_destruct();
 476
 477     // TODO: handle time units and detect parsing errors better
 478     const int64_t parsedValue = strtoll(value, NULL, 10);
 479     if (parsedValue < 0) {
 480         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << parsedValue);
 481         self_destruct();
 482     }
 483
 484     const int newRate = static_cast<int>(parsedValue);
 485
 486     if (newRate < 0) {
 487         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must not be negative but is: " << newRate);
 488         self_destruct();
 489     }
 490
 491     if (!isaReconfig)
 492         *storedRate = newRate;
 493     else if (*storedRate != newRate) {
 494         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 495                << " cannot be changed dynamically, value left unchanged: " <<
 496                *storedRate);
 497     }
 498
 499     return true;
 500 }
 501
 502 /// reports rate-specific options; mimics ::SwapDir::optionObjectSizeDump()
 503 void
 504 Rock::SwapDir::dumpRateOption(StoreEntry * e) const
 505 {
 506     if (fileConfig.ioRate >= 0)
 507         storeAppendPrintf(e, " max-swap-rate=%d", fileConfig.ioRate);
 508 }
 509
 510 /// parses size-specific options; mimics ::SwapDir::optionObjectSizeParse()
 511 bool
 512 Rock::SwapDir::parseSizeOption(char const *option, const char *value, int reconfig)
 513 {
 514     uint64_t *storedSize;
 515     if (strcmp(option, "slot-size") == 0)
 516         storedSize = &slotSize;
 517     else
 518         return false;
 519
 520     if (!value)
 521         self_destruct();
 522
 523     // TODO: handle size units and detect parsing errors better
 524     const uint64_t newSize = strtoll(value, NULL, 10);
 525     if (newSize <= 0) {
 526         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must be positive; got: " << newSize);
 527         self_destruct();
 528     }
 529
 530     if (newSize <= sizeof(DbCellHeader)) {
 531         debugs(3, DBG_CRITICAL, "FATAL: cache_dir " << path << ' ' << option << " must exceed " << sizeof(DbCellHeader) << "; got: " << newSize);
 532         self_destruct();
 533     }
 534
 535     if (!reconfig)
 536         *storedSize = newSize;
 537     else if (*storedSize != newSize) {
 538         debugs(3, DBG_IMPORTANT, "WARNING: cache_dir " << path << ' ' << option
 539                << " cannot be changed dynamically, value left unchanged: " <<
 540                *storedSize);
 541     }
 542
 543     return true;
 544 }
 545
 546 /// reports size-specific options; mimics ::SwapDir::optionObjectSizeDump()
 547 void
 548 Rock::SwapDir::dumpSizeOption(StoreEntry * e) const
 549 {
 550     storeAppendPrintf(e, " slot-size=%" PRId64, slotSize);
 551 }
 552
 553 /// check the results of the configuration; only level-0 debugging works here
 554 void
 555 Rock::SwapDir::validateOptions()
 556 {
 557     if (slotSize <= 0)
 558         fatal("Rock store requires a positive slot-size");
 559
 560     const int64_t maxSizeRoundingWaste = 1024 * 1024; // size is configured in MB
 561     const int64_t slotSizeRoundingWaste = slotSize;
 562     const int64_t maxRoundingWaste =
 563         max(maxSizeRoundingWaste, slotSizeRoundingWaste);
 564
 565     // an entry consumes at least one slot; round up to reduce false warnings
 566     const int64_t blockSize = static_cast<int64_t>(slotSize);
 567     const int64_t maxObjSize = max(blockSize,
 568                                    ((maxObjectSize()+blockSize-1)/blockSize)*blockSize);
 569
 570     // Does the "sfileno*max-size" limit match configured db capacity?
 571     const double entriesMayOccupy = entryLimitAbsolute()*static_cast<double>(maxObjSize);
 572     if (entriesMayOccupy + maxRoundingWaste < maxSize()) {
 573         const int64_t diskWasteSize = maxSize() - static_cast<int64_t>(entriesMayOccupy);
 574         debugs(47, DBG_CRITICAL, "WARNING: Rock cache_dir " << path << " wastes disk space due to entry limits:" <<
 575                "\n\tconfigured db capacity: " << maxSize() << " bytes" <<
 576                "\n\tconfigured db slot size: " << slotSize << " bytes" <<
 577                "\n\tconfigured maximum entry size: " << maxObjectSize() << " bytes" <<
 578                "\n\tmaximum number of cache_dir entries supported by Squid: " << entryLimitAbsolute() <<
 579                "\n\tdisk space all entries may use: " << entriesMayOccupy << " bytes" <<
 580                "\n\tdisk space wasted: " << diskWasteSize << " bytes");
 581     }
 582
 583     // Does the "absolute slot count" limit match configured db capacity?
 584     const double slotsMayOccupy = slotLimitAbsolute()*static_cast<double>(slotSize);
 585     if (slotsMayOccupy + maxRoundingWaste < maxSize()) {
 586         const int64_t diskWasteSize = maxSize() - static_cast<int64_t>(entriesMayOccupy);
 587         debugs(47, DBG_CRITICAL, "WARNING: Rock cache_dir " << path << " wastes disk space due to slot limits:" <<
 588                "\n\tconfigured db capacity: " << maxSize() << " bytes" <<
 589                "\n\tconfigured db slot size: " << slotSize << " bytes" <<
 590                "\n\tmaximum number of rock cache_dir slots supported by Squid: " << slotLimitAbsolute() <<
 591                "\n\tdisk space all slots may use: " << slotsMayOccupy << " bytes" <<
 592                "\n\tdisk space wasted: " << diskWasteSize << " bytes");
 593     }
 594 }
 595
 596 void
 597 Rock::SwapDir::rebuild()
 598 {
 599     //++StoreController::store_dirs_rebuilding; // see Rock::SwapDir::init()
 600     AsyncJob::Start(new Rebuild(this));
 601 }
 602
 603 bool
 604 Rock::SwapDir::canStore(const StoreEntry &e, int64_t diskSpaceNeeded, int &load) const
 605 {
 606     if (!::SwapDir::canStore(e, sizeof(DbCellHeader)+diskSpaceNeeded, load))
 607         return false;
 608
 609     if (!theFile || !theFile->canWrite())
 610         return false;
 611
 612     if (!map)
 613         return false;
 614
 615     // Do not start I/O transaction if there are less than 10% free pages left.
 616     // TODO: reserve page instead
 617     if (needsDiskStrand() &&
 618             Ipc::Mem::PageLevel(Ipc::Mem::PageId::ioPage) >= 0.9 * Ipc::Mem::PageLimit(Ipc::Mem::PageId::ioPage)) {
 619         debugs(47, 5, HERE << "too few shared pages for IPC I/O left");
 620         return false;
 621     }
 622
 623     if (io->shedLoad())
 624         return false;
 625
 626     load = io->load();
 627     return true;
 628 }
 629
 630 StoreIOState::Pointer
 631 Rock::SwapDir::createStoreIO(StoreEntry &e, StoreIOState::STFNCB *cbFile, StoreIOState::STIOCB *cbIo, void *data)
 632 {
 633     if (!theFile || theFile->error()) {
 634         debugs(47,4, HERE << theFile);
 635         return NULL;
 636     }
 637
 638     sfileno filen;
 639     Ipc::StoreMapAnchor *const slot =
 640         map->openForWriting(reinterpret_cast<const cache_key *>(e.key), filen);
 641     if (!slot) {
 642         debugs(47, 5, HERE << "map->add failed");
 643         return NULL;
 644     }
 645
 646     assert(filen >= 0);
 647     slot->set(e);
 648
 649     // XXX: We rely on our caller, storeSwapOutStart(), to set e.fileno.
 650     // If that does not happen, the entry will not decrement the read level!
 651
 652     Rock::SwapDir::Pointer self(this);
 653     IoState *sio = new IoState(self, &e, cbFile, cbIo, data);
 654
 655     sio->swap_dirn = index;
 656     sio->swap_filen = filen;
 657     sio->writeableAnchor_ = slot;
 658
 659     debugs(47,5, HERE << "dir " << index << " created new filen " <<
 660            std::setfill('0') << std::hex << std::uppercase << std::setw(8) <<
 661            sio->swap_filen << std::dec << " starting at " <<
 662            diskOffset(sio->swap_filen));
 663
 664     sio->file(theFile);
 665
 666     trackReferences(e);
 667     return sio;
 668 }
 669
 670 StoreIOState::Pointer
 671 Rock::SwapDir::createUpdateIO(const Ipc::StoreMapUpdate &update, StoreIOState::STFNCB *cbFile, StoreIOState::STIOCB *cbIo, void *data)
 672 {
 673     if (!theFile || theFile->error()) {
 674         debugs(47,4, theFile);
 675         return nullptr;
 676     }
 677
 678     Must(update.fresh);
 679     Must(update.fresh.fileNo >= 0);
 680
 681     Rock::SwapDir::Pointer self(this);
 682     IoState *sio = new IoState(self, update.entry, cbFile, cbIo, data);
 683
 684     sio->swap_dirn = index;
 685     sio->swap_filen = update.fresh.fileNo;
 686     sio->writeableAnchor_ = update.fresh.anchor;
 687
 688     debugs(47,5, "dir " << index << " updating filen " <<
 689            std::setfill('0') << std::hex << std::uppercase << std::setw(8) <<
 690            sio->swap_filen << std::dec << " starting at " <<
 691            diskOffset(sio->swap_filen));
 692
 693     sio->file(theFile);
 694     return sio;
 695 }
 696
 697 int64_t
 698 Rock::SwapDir::diskOffset(const SlotId sid) const
 699 {
 700     assert(sid >= 0);
 701     return HeaderSize + slotSize*sid;
 702 }
 703
 704 int64_t
 705 Rock::SwapDir::diskOffset(Ipc::Mem::PageId &pageId) const
 706 {
 707     assert(pageId);
 708     return diskOffset(pageId.number - 1);
 709 }
 710
 711 int64_t
 712 Rock::SwapDir::diskOffsetLimit() const
 713 {
 714     assert(map);
 715     return diskOffset(map->sliceLimit());
 716 }
 717
 718 bool
 719 Rock::SwapDir::useFreeSlot(Ipc::Mem::PageId &pageId)
 720 {
 721     if (freeSlots->pop(pageId)) {
 722         debugs(47, 5, "got a previously free slot: " << pageId);
 723         return true;
 724     }
 725
 726     // catch free slots delivered to noteFreeMapSlice()
 727     assert(!waitingForPage);
 728     waitingForPage = &pageId;
 729     if (map->purgeOne()) {
 730         assert(!waitingForPage); // noteFreeMapSlice() should have cleared it
 731         assert(pageId.set());
 732         debugs(47, 5, "got a previously busy slot: " << pageId);
 733         return true;
 734     }
 735     assert(waitingForPage == &pageId);
 736     waitingForPage = NULL;
 737
 738     debugs(47, 3, "cannot get a slot; entries: " << map->entryCount());
 739     return false;
 740 }
 741
 742 bool
 743 Rock::SwapDir::validSlotId(const SlotId slotId) const
 744 {
 745     return 0 <= slotId && slotId < slotLimitActual();
 746 }
 747
 748 void
 749 Rock::SwapDir::noteFreeMapSlice(const Ipc::StoreMapSliceId sliceId)
 750 {
 751     Ipc::Mem::PageId pageId;
 752     pageId.pool = index+1;
 753     pageId.number = sliceId+1;
 754     if (waitingForPage) {
 755         *waitingForPage = pageId;
 756         waitingForPage = NULL;
 757     } else {
 758         freeSlots->push(pageId);
 759     }
 760 }
 761
 762 // tries to open an old entry with swap_filen for reading
 763 StoreIOState::Pointer
 764 Rock::SwapDir::openStoreIO(StoreEntry &e, StoreIOState::STFNCB *cbFile, StoreIOState::STIOCB *cbIo, void *data)
 765 {
 766     if (!theFile || theFile->error()) {
 767         debugs(47,4, HERE << theFile);
 768         return NULL;
 769     }
 770
 771     if (e.swap_filen < 0) {
 772         debugs(47,4, HERE << e);
 773         return NULL;
 774     }
 775
 776     // Do not start I/O transaction if there are less than 10% free pages left.
 777     // TODO: reserve page instead
 778     if (needsDiskStrand() &&
 779             Ipc::Mem::PageLevel(Ipc::Mem::PageId::ioPage) >= 0.9 * Ipc::Mem::PageLimit(Ipc::Mem::PageId::ioPage)) {
 780         debugs(47, 5, HERE << "too few shared pages for IPC I/O left");
 781         return NULL;
 782     }
 783
 784     // The are two ways an entry can get swap_filen: our get() locked it for
 785     // reading or our storeSwapOutStart() locked it for writing. Peeking at our
 786     // locked entry is safe, but no support for reading the entry we swap out.
 787     const Ipc::StoreMapAnchor *slot = map->peekAtReader(e.swap_filen);
 788     if (!slot)
 789         return NULL; // we were writing afterall
 790
 791     Rock::SwapDir::Pointer self(this);
 792     IoState *sio = new IoState(self, &e, cbFile, cbIo, data);
 793
 794     sio->swap_dirn = index;
 795     sio->swap_filen = e.swap_filen;
 796     sio->readableAnchor_ = slot;
 797     sio->file(theFile);
 798
 799     debugs(47,5, HERE << "dir " << index << " has old filen: " <<
 800            std::setfill('0') << std::hex << std::uppercase << std::setw(8) <<
 801            sio->swap_filen);
 802
 803     assert(slot->sameKey(static_cast<const cache_key*>(e.key)));
 804     // For collapsed disk hits: e.swap_file_sz and slot->basics.swap_file_sz
 805     // may still be zero and basics.swap_file_sz may grow.
 806     assert(slot->basics.swap_file_sz >= e.swap_file_sz);
 807
 808     return sio;
 809 }
 810
 811 void
 812 Rock::SwapDir::ioCompletedNotification()
 813 {
 814     if (!theFile)
 815         fatalf("Rock cache_dir failed to initialize db file: %s", filePath);
 816
 817     if (theFile->error())
 818         fatalf("Rock cache_dir at %s failed to open db file: %s", filePath,
 819                xstrerror());
 820
 821     debugs(47, 2, "Rock cache_dir[" << index << "] limits: " <<
 822            std::setw(12) << maxSize() << " disk bytes, " <<
 823            std::setw(7) << map->entryLimit() << " entries, and " <<
 824            std::setw(7) << map->sliceLimit() << " slots");
 825
 826     rebuild();
 827 }
 828
 829 void
 830 Rock::SwapDir::closeCompleted()
 831 {
 832     theFile = NULL;
 833 }
 834
 835 void
 836 Rock::SwapDir::readCompleted(const char *, int rlen, int errflag, RefCount< ::ReadRequest> r)
 837 {
 838     ReadRequest *request = dynamic_cast<Rock::ReadRequest*>(r.getRaw());
 839     assert(request);
 840     IoState::Pointer sio = request->sio;
 841
 842     if (errflag == DISK_OK && rlen > 0)
 843         sio->offset_ += rlen;
 844
 845     sio->callReaderBack(r->buf, rlen);
 846 }
 847
 848 void
 849 Rock::SwapDir::writeCompleted(int errflag, size_t, RefCount< ::WriteRequest> r)
 850 {
 851     Rock::WriteRequest *request = dynamic_cast<Rock::WriteRequest*>(r.getRaw());
 852     assert(request);
 853     assert(request->sio !=  NULL);
 854     IoState &sio = *request->sio;
 855
 856     // quit if somebody called IoState::close() while we were waiting
 857     if (!sio.stillWaiting()) {
 858         debugs(79, 3, "ignoring closed entry " << sio.swap_filen);
 859         noteFreeMapSlice(request->sidNext);
 860         return;
 861     }
 862
 863     debugs(79, 7, "errflag=" << errflag << " rlen=" << request->len << " eof=" << request->eof);
 864
 865     // TODO: Fail if disk dropped one of the previous write requests.
 866
 867     if (errflag == DISK_OK) {
 868         // do not increment sio.offset_ because we do it in sio->write()
 869
 870         // finalize the shared slice info after writing slice contents to disk
 871         Ipc::StoreMap::Slice &slice =
 872             map->writeableSlice(sio.swap_filen, request->sidCurrent);
 873         slice.size = request->len - sizeof(DbCellHeader);
 874         slice.next = request->sidNext;
 875
 876         if (request->eof) {
 877             assert(sio.e);
 878             assert(sio.writeableAnchor_);
 879             if (sio.touchingStoreEntry()) {
 880                 sio.e->swap_file_sz = sio.writeableAnchor_->basics.swap_file_sz =
 881                                           sio.offset_;
 882
 883                 // close, the entry gets the read lock
 884                 map->closeForWriting(sio.swap_filen, true);
 885             }
 886             sio.writeableAnchor_ = NULL;
 887             sio.splicingPoint = request->sidCurrent;
 888             sio.finishedWriting(errflag);
 889         }
 890     } else {
 891         noteFreeMapSlice(request->sidNext);
 892
 893         writeError(sio);
 894         sio.finishedWriting(errflag);
 895         // and hope that Core will call disconnect() to close the map entry
 896     }
 897
 898     if (sio.touchingStoreEntry())
 899         CollapsedForwarding::Broadcast(*sio.e);
 900 }
 901
 902 void
 903 Rock::SwapDir::writeError(StoreIOState &sio)
 904 {
 905     // Do not abortWriting here. The entry should keep the write lock
 906     // instead of losing association with the store and confusing core.
 907     map->freeEntry(sio.swap_filen); // will mark as unusable, just in case
 908
 909     if (sio.touchingStoreEntry())
 910         Store::Root().transientsAbandon(*sio.e);
 911     // else noop: a fresh entry update error does not affect stale entry readers
 912
 913     // All callers must also call IoState callback, to propagate the error.
 914 }
 915
 916 void
 917 Rock::SwapDir::updateHeaders(StoreEntry *updatedE)
 918 {
 919     if (!map)
 920         return;
 921
 922     Ipc::StoreMapUpdate update(updatedE);
 923     if (!map->openForUpdating(update, updatedE->swap_filen))
 924         return;
 925
 926     try {
 927         AsyncJob::Start(new HeaderUpdater(this, update));
 928     } catch (const std::exception &ex) {
 929         debugs(20, 2, "error starting to update entry " << *updatedE << ": " << ex.what());
 930         map->abortUpdating(update);
 931     }
 932 }
 933
 934 bool
 935 Rock::SwapDir::full() const
 936 {
 937     return freeSlots != NULL && !freeSlots->size();
 938 }
 939
 940 // storeSwapOutFileClosed calls this nethod on DISK_NO_SPACE_LEFT,
 941 // but it should not happen for us
 942 void
 943 Rock::SwapDir::diskFull()
 944 {
 945     debugs(20, DBG_IMPORTANT, "BUG: No space left with rock cache_dir: " <<
 946            filePath);
 947 }
 948
 949 /// purge while full(); it should be sufficient to purge just one
 950 void
 951 Rock::SwapDir::maintain()
 952 {
 953     // The Store calls this to free some db space, but there is nothing wrong
 954     // with a full() db, except when db has to shrink after reconfigure, and
 955     // we do not support shrinking yet (it would have to purge specific slots).
 956     // TODO: Disable maintain() requests when they are pointless.
 957 }
 958
 959 void
 960 Rock::SwapDir::reference(StoreEntry &e)
 961 {
 962     debugs(47, 5, HERE << &e << ' ' << e.swap_dirn << ' ' << e.swap_filen);
 963     if (repl && repl->Referenced)
 964         repl->Referenced(repl, &e, &e.repl);
 965 }
 966
 967 bool
 968 Rock::SwapDir::dereference(StoreEntry &e)
 969 {
 970     debugs(47, 5, HERE << &e << ' ' << e.swap_dirn << ' ' << e.swap_filen);
 971     if (repl && repl->Dereferenced)
 972         repl->Dereferenced(repl, &e, &e.repl);
 973
 974     // no need to keep e in the global store_table for us; we have our own map
 975     return false;
 976 }
 977
 978 bool
 979 Rock::SwapDir::unlinkdUseful() const
 980 {
 981     // no entry-specific files to unlink
 982     return false;
 983 }
 984
 985 void
 986 Rock::SwapDir::unlink(StoreEntry &e)
 987 {
 988     debugs(47, 5, HERE << e);
 989     ignoreReferences(e);
 990     map->freeEntry(e.swap_filen);
 991     disconnect(e);
 992 }
 993
 994 void
 995 Rock::SwapDir::markForUnlink(StoreEntry &e)
 996 {
 997     debugs(47, 5, e);
 998     map->freeEntry(e.swap_filen);
 999 }
1000
1001 void
1002 Rock::SwapDir::trackReferences(StoreEntry &e)
1003 {
1004     debugs(47, 5, HERE << e);
1005     if (repl)
1006         repl->Add(repl, &e, &e.repl);
1007 }
1008
1009 void
1010 Rock::SwapDir::ignoreReferences(StoreEntry &e)
1011 {
1012     debugs(47, 5, HERE << e);
1013     if (repl)
1014         repl->Remove(repl, &e, &e.repl);
1015 }
1016
1017 void
1018 Rock::SwapDir::statfs(StoreEntry &e) const
1019 {
1020     storeAppendPrintf(&e, "\n");
1021     storeAppendPrintf(&e, "Maximum Size: %" PRIu64 " KB\n", maxSize() >> 10);
1022     storeAppendPrintf(&e, "Current Size: %.2f KB %.2f%%\n",
1023                       currentSize() / 1024.0,
1024                       Math::doublePercent(currentSize(), maxSize()));
1025
1026     const int entryLimit = entryLimitActual();
1027     const int slotLimit = slotLimitActual();
1028     storeAppendPrintf(&e, "Maximum entries: %9d\n", entryLimit);
1029     if (map && entryLimit > 0) {
1030         const int entryCount = map->entryCount();
1031         storeAppendPrintf(&e, "Current entries: %9d %.2f%%\n",
1032                           entryCount, (100.0 * entryCount / entryLimit));
1033     }
1034
1035     storeAppendPrintf(&e, "Maximum slots:   %9d\n", slotLimit);
1036     if (map && slotLimit > 0) {
1037         const unsigned int slotsFree = !freeSlots ? 0 : freeSlots->size();
1038         if (slotsFree <= static_cast<const unsigned int>(slotLimit)) {
1039             const int usedSlots = slotLimit - static_cast<const int>(slotsFree);
1040             storeAppendPrintf(&e, "Used slots:      %9d %.2f%%\n",
1041                               usedSlots, (100.0 * usedSlots / slotLimit));
1042         }
1043         if (slotLimit < 100) { // XXX: otherwise too expensive to count
1044             Ipc::ReadWriteLockStats stats;
1045             map->updateStats(stats);
1046             stats.dump(e);
1047         }
1048     }
1049
1050     storeAppendPrintf(&e, "Pending operations: %d out of %d\n",
1051                       store_open_disk_fd, Config.max_open_disk_fds);
1052
1053     storeAppendPrintf(&e, "Flags:");
1054
1055     if (flags.selected)
1056         storeAppendPrintf(&e, " SELECTED");
1057
1058     if (flags.read_only)
1059         storeAppendPrintf(&e, " READ-ONLY");
1060
1061     storeAppendPrintf(&e, "\n");
1062
1063 }
1064
1065 SBuf
1066 Rock::SwapDir::inodeMapPath() const
1067 {
1068     return Ipc::Mem::Segment::Name(SBuf(path), "map");
1069 }
1070
1071 const char *
1072 Rock::SwapDir::freeSlotsPath() const
1073 {
1074     static String spacesPath;
1075     spacesPath = path;
1076     spacesPath.append("_spaces");
1077     return spacesPath.termedBuf();
1078 }
1079
1080 namespace Rock
1081 {
1082 RunnerRegistrationEntry(SwapDirRr);
1083 }
1084
1085 void Rock::SwapDirRr::create()
1086 {
1087     Must(mapOwners.empty() && freeSlotsOwners.empty());
1088     for (int i = 0; i < Config.cacheSwap.n_configured; ++i) {
1089         if (const Rock::SwapDir *const sd = dynamic_cast<Rock::SwapDir *>(INDEXSD(i))) {
1090             const int64_t capacity = sd->slotLimitActual();
1091
1092             SwapDir::DirMap::Owner *const mapOwner =
1093                 SwapDir::DirMap::Init(sd->inodeMapPath(), capacity);
1094             mapOwners.push_back(mapOwner);
1095
1096             // TODO: somehow remove pool id and counters from PageStack?
1097             Ipc::Mem::Owner<Ipc::Mem::PageStack> *const freeSlotsOwner =
1098                 shm_new(Ipc::Mem::PageStack)(sd->freeSlotsPath(),
1099                                              i+1, capacity, 0);
1100             freeSlotsOwners.push_back(freeSlotsOwner);
1101
1102             // TODO: add method to initialize PageStack with no free pages
1103             while (true) {
1104                 Ipc::Mem::PageId pageId;
1105                 if (!freeSlotsOwner->object()->pop(pageId))
1106                     break;
1107             }
1108         }
1109     }
1110 }
1111
1112 Rock::SwapDirRr::~SwapDirRr()
1113 {
1114     for (size_t i = 0; i < mapOwners.size(); ++i) {
1115         delete mapOwners[i];
1116         delete freeSlotsOwners[i];
1117     }
1118 }
1119