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 "SwapDir.h"
  44 #include "swap_log_op.h"
  45
  46 static void storeSwapOutStart(StoreEntry * e);
  47 static StoreIOState::STIOCB storeSwapOutFileClosed;
  48 static StoreIOState::STFNCB storeSwapOutFileNotify;
  49
  50 // wrapper to cross C/C++ ABI boundary. xfree is extern "C" for libraries.
  51 static void xfree_cppwrapper(void *x)
  52 {
  53     xfree(x);
  54 }
  55
  56 /* start swapping object to disk */
  57 static void
  58 storeSwapOutStart(StoreEntry * e)
  59 {
  60     MemObject *mem = e->mem_obj;
  61     StoreIOState::Pointer sio;
  62     assert(mem);
  63     /* Build the swap metadata, so the filesystem will know how much
  64      * metadata there is to store
  65      */
  66     debugs(20, 5, "storeSwapOutStart: Begin SwapOut '" << e->url() << "' to dirno " <<
  67            e->swap_dirn << ", fileno " << std::hex << std::setw(8) << std::setfill('0') <<
  68            std::uppercase << e->swap_filen);
  69     e->swap_status = SWAPOUT_WRITING;
  70     /* If we start swapping out objects with OutOfBand Metadata,
  71      * then this code needs changing
  72      */
  73
  74     /* TODO: make some sort of data,size refcounted immutable buffer
  75      * and stop fooling ourselves with "const char*" buffers.
  76      */
  77
  78     // Create metadata now, possibly in vain: storeCreate needs swap_hdr_sz.
  79     const char *buf = e->getSerialisedMetaData ();
  80     assert(buf);
  81
  82     /* Create the swap file */
  83     generic_cbdata *c = new generic_cbdata(e);
  84     sio = storeCreate(e, storeSwapOutFileNotify, storeSwapOutFileClosed, c);
  85
  86     if (sio == NULL) {
  87         e->swap_status = SWAPOUT_NONE;
  88         // our caller thinks SWAPOUT_NONE means swapping out has not started
  89         // yet so we better release here to avoid being called again and again
  90         e->releaseRequest();
  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
 183 /* This routine is called every time data is sent to the client side.
 184  * It's overhead is therefor, significant.
 185  */
 186 void
 187 StoreEntry::swapOut()
 188 {
 189     if (!mem_obj)
 190         return;
 191
 192     if (!swapoutPossible())
 193         return;
 194
 195     // Aborted entries have STORE_OK, but swapoutPossible rejects them. Thus,
 196     // store_status == STORE_OK below means we got everything we wanted.
 197
 198     debugs(20, 7, HERE << "storeSwapOut: mem->inmem_lo = " << mem_obj->inmem_lo);
 199     debugs(20, 7, HERE << "storeSwapOut: mem->endOffset() = " << mem_obj->endOffset());
 200     debugs(20, 7, HERE << "storeSwapOut: swapout.queue_offset = " << mem_obj->swapout.queue_offset);
 201
 202     if (mem_obj->swapout.sio != NULL)
 203         debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj->swapout.sio->offset()  );
 204
 205     // buffered bytes we have not swapped out yet
 206     int64_t swapout_maxsize = mem_obj->endOffset() - mem_obj->swapout.queue_offset;
 207
 208     assert(swapout_maxsize >= 0);
 209
 210     int64_t const lowest_offset = mem_obj->lowestMemReaderOffset();
 211
 212     debugs(20, 7, HERE << "storeSwapOut: lowest_offset = " << lowest_offset);
 213
 214     // Check to see whether we're going to defer the swapout based upon size
 215     if (store_status != STORE_OK) {
 216         const int64_t expectedSize = mem_obj->expectedReplySize();
 217         const int64_t maxKnownSize = expectedSize < 0 ?
 218             swapout_maxsize : expectedSize;
 219         debugs(20, 7, HERE << "storeSwapOut: maxKnownSize= " << maxKnownSize);
 220
 221         if (maxKnownSize < store_maxobjsize) {
 222             /*
 223              * NOTE: the store_maxobjsize here is the max of optional
 224              * max-size values from 'cache_dir' lines.  It is not the
 225              * same as 'maximum_object_size'.  By default, store_maxobjsize
 226              * will be set to -1.  However, I am worried that this
 227              * deferance may consume a lot of memory in some cases.
 228              * Should we add an option to limit this memory consumption?
 229              */
 230             debugs(20, 5, "storeSwapOut: Deferring swapout start for " <<
 231                 (store_maxobjsize - maxKnownSize) << " bytes");
 232             return;
 233          }
 234     }
 235
 236 // TODO: it is better to trim as soon as we swap something out, not before
 237     trimMemory();
 238 #if SIZEOF_OFF_T <= 4
 239
 240     if (mem_obj->endOffset() > 0x7FFF0000) {
 241         debugs(20, 0, "WARNING: preventing off_t overflow for " << url());
 242         abort();
 243         return;
 244     }
 245
 246 #endif
 247     if (swap_status == SWAPOUT_WRITING)
 248         assert(mem_obj->inmem_lo <=  mem_obj->objectBytesOnDisk() );
 249
 250     if (!swapOutAble())
 251         return;
 252
 253     debugs(20, 7, "storeSwapOut: swapout_size = " << swapout_maxsize);
 254
 255     if (swapout_maxsize == 0) { // swapped everything we got
 256         if (store_status == STORE_OK) { // got everything we wanted
 257             assert(mem_obj->object_sz >= 0);
 258             swapOutFileClose(StoreIOState::wroteAll);
 259         }
 260         // else need more data to swap out
 261         return;
 262     }
 263
 264     if (store_status == STORE_PENDING) {
 265         /* wait for a full block to write */
 266
 267         if (swapout_maxsize < SM_PAGE_SIZE)
 268             return;
 269
 270         /*
 271          * Wait until we are below the disk FD limit, only if the
 272          * next server-side read won't be deferred.
 273          */
 274         if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
 275             return;
 276     }
 277
 278     /* Ok, we have stuff to swap out.  Is there a swapout.sio open? */
 279     if (swap_status == SWAPOUT_NONE) {
 280         assert(mem_obj->swapout.sio == NULL);
 281         assert(mem_obj->inmem_lo == 0);
 282
 283         if (checkCachable())
 284             storeSwapOutStart(this);
 285         else
 286             return;
 287
 288         /* ENTRY_CACHABLE will be cleared and we'll never get here again */
 289     }
 290
 291     if (mem_obj->swapout.sio == NULL)
 292         return;
 293
 294     doPages(this);
 295
 296     if (mem_obj->swapout.sio == NULL)
 297         /* oops, we're not swapping out any more */
 298         return;
 299
 300     if (store_status == STORE_OK) {
 301         /*
 302          * If the state is STORE_OK, then all data must have been given
 303          * to the filesystem at this point because storeSwapOut() is
 304          * not going to be called again for this entry.
 305          */
 306         assert(mem_obj->object_sz >= 0);
 307         assert(mem_obj->endOffset() == mem_obj->swapout.queue_offset);
 308         swapOutFileClose(StoreIOState::wroteAll);
 309     }
 310 }
 311
 312 void
 313 StoreEntry::swapOutFileClose(int how)
 314 {
 315     assert(mem_obj != NULL);
 316     debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text() << " how=" << how);
 317     debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj->swapout.sio.getRaw());
 318
 319     if (mem_obj->swapout.sio == NULL)
 320         return;
 321
 322     storeClose(mem_obj->swapout.sio, how);
 323 }
 324
 325 static void
 326 storeSwapOutFileClosed(void *data, int errflag, StoreIOState::Pointer self)
 327 {
 328     generic_cbdata *c = (generic_cbdata *)data;
 329     StoreEntry *e = (StoreEntry *)c->data;
 330     MemObject *mem = e->mem_obj;
 331     assert(mem->swapout.sio == self);
 332     assert(e->swap_status == SWAPOUT_WRITING);
 333     cbdataFree(c);
 334
 335     // if object_size is still unknown, the entry was probably aborted
 336     if (errflag || e->objectLen() < 0) {
 337         debugs(20, 2, "storeSwapOutFileClosed: dirno " << e->swap_dirn << ", swapfile " <<
 338                std::hex << std::setw(8) << std::setfill('0') << std::uppercase <<
 339                e->swap_filen << ", errflag=" << errflag);
 340
 341         if (errflag == DISK_NO_SPACE_LEFT) {
 342             /* FIXME: this should be handle by the link from store IO to
 343              * Store, rather than being a top level API call.
 344              */
 345             e->store()->diskFull();
 346             storeConfigure();
 347         }
 348
 349         if (e->swap_filen >= 0)
 350             e->unlink();
 351
 352         assert(e->swap_status == SWAPOUT_NONE);
 353
 354         e->releaseRequest();
 355     } else {
 356         /* swapping complete */
 357         debugs(20, 3, "storeSwapOutFileClosed: SwapOut complete: '" << e->url() << "' to " <<
 358                e->swap_dirn  << ", " << std::hex << std::setw(8) << std::setfill('0') <<
 359                std::uppercase << e->swap_filen);
 360         debugs(20, 5, HERE << "swap_file_sz = " <<
 361                e->objectLen() << " + " << mem->swap_hdr_sz);
 362         assert(e->objectLen() >= 0); // we checked that above
 363         e->swap_file_sz = e->objectLen() + mem->swap_hdr_sz;
 364         e->swap_status = SWAPOUT_DONE;
 365         e->store()->updateSize(e->swap_file_sz, 1);
 366
 367         if (e->checkCachable()) {
 368             storeLog(STORE_LOG_SWAPOUT, e);
 369             storeDirSwapLog(e, SWAP_LOG_ADD);
 370         }
 371
 372         statCounter.swap.outs++;
 373     }
 374
 375     debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__ << ":" << __LINE__);
 376     mem->swapout.sio = NULL;
 377     e->unlock();
 378 }
 379
 380 /*
 381  * Is this entry a candidate for writing to disk?
 382  */
 383 bool
 384 StoreEntry::swapOutAble() const
 385 {
 386     dlink_node *node;
 387
 388     if (mem_obj->swapout.sio != NULL)
 389         return true;
 390
 391     if (mem_obj->inmem_lo > 0)
 392         return false;
 393
 394     /*
 395      * If there are DISK clients, we must write to disk
 396      * even if its not cachable
 397      * RBC: Surely we should not create disk client on non cacheable objects?
 398      * therefore this should be an assert?
 399      * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
 400      * an assert.
 401      */
 402     for (node = mem_obj->clients.head; node; node = node->next) {
 403         if (((store_client *) node->data)->getType() == STORE_DISK_CLIENT)
 404             return true;
 405     }
 406
 407     /* Don't pollute the disk with icons and other special entries */
 408     if (EBIT_TEST(flags, ENTRY_SPECIAL))
 409         return false;
 410
 411     if (!EBIT_TEST(flags, ENTRY_CACHABLE))
 412         return false;
 413
 414     if (!mem_obj->isContiguous())
 415         return false;
 416
 417     return true;
 418 }