src/store_swapout.cc

   1
   2 /*
   3  * $Id$
   4  *
   5  * DEBUG: section 20    Storage Manager Swapout Functions
   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.h"
  37 #include "cbdata.h"
  38 #include "StoreClient.h"
  39 #include "Store.h"
  40 /* FIXME: Abstract the use of this more */
  41 #include "mem_node.h"
  42 #include "MemObject.h"
  43 #include "protos.h"
  44 #include "SwapDir.h"
  45 #include "StatCounters.h"
  46 #include "swap_log_op.h"
  47
  48 static void storeSwapOutStart(StoreEntry * e);
  49 static StoreIOState::STIOCB storeSwapOutFileClosed;
  50 static StoreIOState::STFNCB storeSwapOutFileNotify;
  51
  52 // wrapper to cross C/C++ ABI boundary. xfree is extern "C" for libraries.
  53 static void xfree_cppwrapper(void *x)
  54 {
  55     xfree(x);
  56 }
  57
  58 /* start swapping object to disk */
  59 static void
  60 storeSwapOutStart(StoreEntry * e)
  61 {
  62     MemObject *mem = e->mem_obj;
  63     StoreIOState::Pointer sio;
  64     assert(mem);
  65     /* Build the swap metadata, so the filesystem will know how much
  66      * metadata there is to store
  67      */
  68     debugs(20, 5, "storeSwapOutStart: Begin SwapOut '" << e->url() << "' to dirno " <<
  69            e->swap_dirn << ", fileno " << std::hex << std::setw(8) << std::setfill('0') <<
  70            std::uppercase << e->swap_filen);
  71     e->swap_status = SWAPOUT_WRITING;
  72     /* If we start swapping out objects with OutOfBand Metadata,
  73      * then this code needs changing
  74      */
  75
  76     /* TODO: make some sort of data,size refcounted immutable buffer
  77      * and stop fooling ourselves with "const char*" buffers.
  78      */
  79
  80     // Create metadata now, possibly in vain: storeCreate needs swap_hdr_sz.
  81     const char *buf = e->getSerialisedMetaData ();
  82     assert(buf);
  83
  84     /* Create the swap file */
  85     generic_cbdata *c = new generic_cbdata(e);
  86     sio = storeCreate(e, storeSwapOutFileNotify, storeSwapOutFileClosed, c);
  87
  88     if (sio == NULL) {
  89         e->swap_status = SWAPOUT_NONE;
  90         mem->swapout.decision = MemObject::SwapOut::swImpossible;
  91         delete c;
  92         xfree((char*)buf);
  93         storeLog(STORE_LOG_SWAPOUTFAIL, e);
  94         return;
  95     }
  96
  97     mem->swapout.sio = sio;
  98     /* Don't lock until after create, or the replacement
  99      * code might get confused */
 100
 101     e->lock();
 102     /* Pick up the file number if it was assigned immediately */
 103     e->swap_filen = mem->swapout.sio->swap_filen;
 104
 105     e->swap_dirn = mem->swapout.sio->swap_dirn;
 106
 107     /* write out the swap metadata */
 108     storeIOWrite(mem->swapout.sio, buf, mem->swap_hdr_sz, 0, xfree_cppwrapper);
 109 }
 110
 111 static void
 112 storeSwapOutFileNotify(void *data, int errflag, StoreIOState::Pointer self)
 113 {
 114     generic_cbdata *c = (generic_cbdata *)data;
 115     StoreEntry *e = (StoreEntry *)c->data;
 116     MemObject *mem = e->mem_obj;
 117     assert(e->swap_status == SWAPOUT_WRITING);
 118     assert(mem);
 119     assert(mem->swapout.sio == self);
 120     assert(errflag == 0);
 121     assert(e->swap_filen < 0); // if this fails, call SwapDir::disconnect(e)
 122     e->swap_filen = mem->swapout.sio->swap_filen;
 123     e->swap_dirn = mem->swapout.sio->swap_dirn;
 124 }
 125
 126 static void
 127 doPages(StoreEntry *anEntry)
 128 {
 129     MemObject *mem = anEntry->mem_obj;
 130
 131     do {
 132         // find the page containing the first byte we have not swapped out yet
 133         mem_node *page =
 134             mem->data_hdr.getBlockContainingLocation(mem->swapout.queue_offset);
 135
 136         if (!page)
 137             return; // wait for more data to become available
 138
 139         // memNodeWriteComplete() and absence of buffer offset math below
 140         // imply that we always write from the very beginning of the page
 141         assert(page->start() == mem->swapout.queue_offset);
 142
 143         /*
 144          * Get the length of this buffer. We are assuming(!) that the buffer
 145          * length won't change on this buffer, or things are going to be very
 146          * strange. I think that after the copy to a buffer is done, the buffer
 147          * size should stay fixed regardless so that this code isn't confused,
 148          * but we can look at this at a later date or whenever the code results
 149          * in bad swapouts, whichever happens first. :-)
 150          */
 151         ssize_t swap_buf_len = page->nodeBuffer.length;
 152
 153         debugs(20, 3, "storeSwapOut: swap_buf_len = " << swap_buf_len);
 154
 155         assert(swap_buf_len > 0);
 156
 157         debugs(20, 3, "storeSwapOut: swapping out " << swap_buf_len << " bytes from " << mem->swapout.queue_offset);
 158
 159         mem->swapout.queue_offset += swap_buf_len;
 160
 161         storeIOWrite(mem->swapout.sio,
 162                      mem->data_hdr.NodeGet(page),
 163                      swap_buf_len,
 164                      -1,
 165                      memNodeWriteComplete);
 166
 167         /* the storeWrite() call might generate an error */
 168         if (anEntry->swap_status != SWAPOUT_WRITING)
 169             break;
 170
 171         int64_t swapout_size = mem->endOffset() - mem->swapout.queue_offset;
 172
 173         if (anEntry->store_status == STORE_PENDING)
 174             if (swapout_size < SM_PAGE_SIZE)
 175                 break;
 176
 177         if (swapout_size <= 0)
 178             return;
 179     } while (true);
 180 }
 181
 182 /* This routine is called every time data is sent to the client side.
 183  * It's overhead is therefor, significant.
 184  */
 185 void
 186 StoreEntry::swapOut()
 187 {
 188     if (!mem_obj)
 189         return;
 190
 191     // this flag may change so we must check even if we are swappingOut
 192     if (EBIT_TEST(flags, ENTRY_ABORTED)) {
 193         assert(EBIT_TEST(flags, RELEASE_REQUEST));
 194         // StoreEntry::abort() already closed the swap out file, if any
 195         // no trimming: data producer must stop production if ENTRY_ABORTED
 196         return;
 197     }
 198
 199     const bool weAreOrMayBeSwappingOut = swappingOut() || mayStartSwapOut();
 200
 201     Store::Root().maybeTrimMemory(*this, weAreOrMayBeSwappingOut);
 202
 203     if (!weAreOrMayBeSwappingOut)
 204         return; // nothing else to do
 205
 206     // Aborted entries have STORE_OK, but swapoutPossible rejects them. Thus,
 207     // store_status == STORE_OK below means we got everything we wanted.
 208
 209     debugs(20, 7, HERE << "storeSwapOut: mem->inmem_lo = " << mem_obj->inmem_lo);
 210     debugs(20, 7, HERE << "storeSwapOut: mem->endOffset() = " << mem_obj->endOffset());
 211     debugs(20, 7, HERE << "storeSwapOut: swapout.queue_offset = " << mem_obj->swapout.queue_offset);
 212
 213     if (mem_obj->swapout.sio != NULL)
 214         debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj->swapout.sio->offset()  );
 215
 216     int64_t const lowest_offset = mem_obj->lowestMemReaderOffset();
 217
 218     debugs(20, 7, HERE << "storeSwapOut: lowest_offset = " << lowest_offset);
 219
 220 #if SIZEOF_OFF_T <= 4
 221
 222     if (mem_obj->endOffset() > 0x7FFF0000) {
 223         debugs(20, DBG_CRITICAL, "WARNING: preventing off_t overflow for " << url());
 224         abort();
 225         return;
 226     }
 227
 228 #endif
 229     if (swap_status == SWAPOUT_WRITING)
 230         assert(mem_obj->inmem_lo <=  mem_obj->objectBytesOnDisk() );
 231
 232     // buffered bytes we have not swapped out yet
 233     const int64_t swapout_maxsize = mem_obj->availableForSwapOut();
 234     assert(swapout_maxsize >= 0);
 235     debugs(20, 7, "storeSwapOut: swapout_size = " << swapout_maxsize);
 236
 237     if (swapout_maxsize == 0) { // swapped everything we got
 238         if (store_status == STORE_OK) { // got everything we wanted
 239             assert(mem_obj->object_sz >= 0);
 240             swapOutFileClose(StoreIOState::wroteAll);
 241         }
 242         // else need more data to swap out
 243         return;
 244     }
 245
 246     if (store_status == STORE_PENDING) {
 247         /* wait for a full block to write */
 248
 249         if (swapout_maxsize < SM_PAGE_SIZE)
 250             return;
 251
 252         /*
 253          * Wait until we are below the disk FD limit, only if the
 254          * next server-side read won't be deferred.
 255          */
 256         if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
 257             return;
 258     }
 259
 260     /* Ok, we have stuff to swap out.  Is there a swapout.sio open? */
 261     if (swap_status == SWAPOUT_NONE) {
 262         assert(mem_obj->swapout.sio == NULL);
 263         assert(mem_obj->inmem_lo == 0);
 264         storeSwapOutStart(this); // sets SwapOut::swImpossible on failures
 265     }
 266
 267     if (mem_obj->swapout.sio == NULL)
 268         return;
 269
 270     doPages(this);
 271
 272     if (mem_obj->swapout.sio == NULL)
 273         /* oops, we're not swapping out any more */
 274         return;
 275
 276     if (store_status == STORE_OK) {
 277         /*
 278          * If the state is STORE_OK, then all data must have been given
 279          * to the filesystem at this point because storeSwapOut() is
 280          * not going to be called again for this entry.
 281          */
 282         assert(mem_obj->object_sz >= 0);
 283         assert(mem_obj->endOffset() == mem_obj->swapout.queue_offset);
 284         swapOutFileClose(StoreIOState::wroteAll);
 285     }
 286 }
 287
 288 void
 289 StoreEntry::swapOutFileClose(int how)
 290 {
 291     assert(mem_obj != NULL);
 292     debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text() << " how=" << how);
 293     debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj->swapout.sio.getRaw());
 294
 295     if (mem_obj->swapout.sio == NULL)
 296         return;
 297
 298     storeClose(mem_obj->swapout.sio, how);
 299 }
 300
 301 static void
 302 storeSwapOutFileClosed(void *data, int errflag, StoreIOState::Pointer self)
 303 {
 304     generic_cbdata *c = (generic_cbdata *)data;
 305     StoreEntry *e = (StoreEntry *)c->data;
 306     MemObject *mem = e->mem_obj;
 307     assert(mem->swapout.sio == self);
 308     assert(e->swap_status == SWAPOUT_WRITING);
 309     cbdataFree(c);
 310
 311     // if object_size is still unknown, the entry was probably aborted
 312     if (errflag || e->objectLen() < 0) {
 313         debugs(20, 2, "storeSwapOutFileClosed: dirno " << e->swap_dirn << ", swapfile " <<
 314                std::hex << std::setw(8) << std::setfill('0') << std::uppercase <<
 315                e->swap_filen << ", errflag=" << errflag);
 316
 317         if (errflag == DISK_NO_SPACE_LEFT) {
 318             /* FIXME: this should be handle by the link from store IO to
 319              * Store, rather than being a top level API call.
 320              */
 321             e->store()->diskFull();
 322             storeConfigure();
 323         }
 324
 325         if (e->swap_filen >= 0)
 326             e->unlink();
 327
 328         assert(e->swap_status == SWAPOUT_NONE);
 329
 330         e->releaseRequest();
 331     } else {
 332         /* swapping complete */
 333         debugs(20, 3, "storeSwapOutFileClosed: SwapOut complete: '" << e->url() << "' to " <<
 334                e->swap_dirn  << ", " << std::hex << std::setw(8) << std::setfill('0') <<
 335                std::uppercase << e->swap_filen);
 336         debugs(20, 5, HERE << "swap_file_sz = " <<
 337                e->objectLen() << " + " << mem->swap_hdr_sz);
 338
 339         e->swap_file_sz = e->objectLen() + mem->swap_hdr_sz;
 340         e->swap_status = SWAPOUT_DONE;
 341         e->store()->swappedOut(*e);
 342
 343         // XXX: For some Stores, it is pointless to re-check cachability here
 344         // and it leads to double counts in store_check_cachable_hist. We need
 345         // another way to signal a completed but failed swapout. Or, better,
 346         // each Store should handle its own logging and LOG state setting.
 347         if (e->checkCachable()) {
 348             storeLog(STORE_LOG_SWAPOUT, e);
 349             storeDirSwapLog(e, SWAP_LOG_ADD);
 350         }
 351
 352         ++statCounter.swap.outs;
 353     }
 354
 355     debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__ << ":" << __LINE__);
 356     mem->swapout.sio = NULL;
 357     e->unlock();
 358 }
 359
 360 bool
 361 StoreEntry::mayStartSwapOut()
 362 {
 363     dlink_node *node;
 364
 365     // must be checked in the caller
 366     assert(!EBIT_TEST(flags, ENTRY_ABORTED));
 367     assert(!swappingOut());
 368
 369     if (!Config.cacheSwap.n_configured)
 370         return false;
 371
 372     assert(mem_obj);
 373     MemObject::SwapOut::Decision &decision = mem_obj->swapout.decision;
 374
 375     // if we decided that swapout is not possible, do not repeat same checks
 376     if (decision == MemObject::SwapOut::swImpossible) {
 377         debugs(20, 3, HERE << " already rejected");
 378         return false;
 379     }
 380
 381     // if we decided that swapout is possible, do not repeat same checks
 382     if (decision == MemObject::SwapOut::swPossible) {
 383         debugs(20, 3,  HERE << "already allowed");
 384         return true;
 385     }
 386
 387     // if we swapped out already, do not start over
 388     if (swap_status == SWAPOUT_DONE) {
 389         debugs(20, 3,  HERE << "already did");
 390         decision = MemObject::SwapOut::swImpossible;
 391         return false;
 392     }
 393
 394     if (!checkCachable()) {
 395         debugs(20, 3,  HERE << "not cachable");
 396         decision = MemObject::SwapOut::swImpossible;
 397         return false;
 398     }
 399
 400     if (EBIT_TEST(flags, ENTRY_SPECIAL)) {
 401         debugs(20, 3,  HERE  << url() << " SPECIAL");
 402         decision = MemObject::SwapOut::swImpossible;
 403         return false;
 404     }
 405
 406     // check cache_dir max-size limit if all cache_dirs have it
 407     if (store_maxobjsize >= 0) {
 408         // TODO: add estimated store metadata size to be conservative
 409
 410         // use guaranteed maximum if it is known
 411         const int64_t expectedEnd = mem_obj->expectedReplySize();
 412         debugs(20, 7,  HERE << "expectedEnd = " << expectedEnd);
 413         if (expectedEnd > store_maxobjsize) {
 414             debugs(20, 3,  HERE << "will not fit: " << expectedEnd <<
 415                    " > " << store_maxobjsize);
 416             decision = MemObject::SwapOut::swImpossible;
 417             return false; // known to outgrow the limit eventually
 418         }
 419
 420         // use current minimum (always known)
 421         const int64_t currentEnd = mem_obj->endOffset();
 422         if (currentEnd > store_maxobjsize) {
 423             debugs(20, 3,  HERE << "does not fit: " << currentEnd <<
 424                    " > " << store_maxobjsize);
 425             decision = MemObject::SwapOut::swImpossible;
 426             return false; // already does not fit and may only get bigger
 427         }
 428
 429         // prevent default swPossible answer for yet unknown length
 430         if (expectedEnd < 0) {
 431             debugs(20, 3,  HERE << "wait for more info: " <<
 432                    store_maxobjsize);
 433             return false; // may fit later, but will be rejected now
 434         }
 435
 436         if (store_status != STORE_OK) {
 437             const int64_t maxKnownSize = expectedEnd < 0 ?
 438                                          mem_obj->availableForSwapOut() : expectedEnd;
 439             debugs(20, 7, HERE << "maxKnownSize= " << maxKnownSize);
 440             if (maxKnownSize < store_maxobjsize) {
 441                 /*
 442                  * NOTE: the store_maxobjsize here is the max of optional
 443                  * max-size values from 'cache_dir' lines.  It is not the
 444                  * same as 'maximum_object_size'.  By default, store_maxobjsize
 445                  * will be set to -1.  However, I am worried that this
 446                  * deferance may consume a lot of memory in some cases.
 447                  * Should we add an option to limit this memory consumption?
 448                  */
 449                 debugs(20, 5,  HERE << "Deferring swapout start for " <<
 450                        (store_maxobjsize - maxKnownSize) << " bytes");
 451                 return false;
 452             }
 453         }
 454     }
 455
 456     if (mem_obj->inmem_lo > 0) {
 457         debugs(20, 3, "storeSwapOut: (inmem_lo > 0)  imem_lo:" <<  mem_obj->inmem_lo);
 458         decision = MemObject::SwapOut::swImpossible;
 459         return false;
 460     }
 461
 462     /*
 463      * If there are DISK clients, we must write to disk
 464      * even if its not cachable
 465      * RBC: Surely we should not create disk client on non cacheable objects?
 466      * therefore this should be an assert?
 467      * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
 468      * an assert.
 469      *
 470      * XXX: Not clear what "mem races" the above refers to, especially when
 471      * dealing with non-cachable objects that cannot have multiple clients.
 472      *
 473      * XXX: If STORE_DISK_CLIENT needs SwapOut::swPossible, we have to check
 474      * for that flag earlier, but forcing swapping may contradict max-size or
 475      * other swapability restrictions. Change storeClientType() and/or its
 476      * callers to take swap-in availability into account.
 477      */
 478     for (node = mem_obj->clients.head; node; node = node->next) {
 479         if (((store_client *) node->data)->getType() == STORE_DISK_CLIENT) {
 480             debugs(20, 3, HERE << "DISK client found");
 481             decision = MemObject::SwapOut::swPossible;
 482             return true;
 483         }
 484     }
 485
 486     if (!mem_obj->isContiguous()) {
 487         debugs(20, 3, "storeSwapOut: not Contiguous");
 488         decision = MemObject::SwapOut::swImpossible;
 489         return false;
 490     }
 491
 492     decision = MemObject::SwapOut::swPossible;
 493     return true;
 494 }