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         delete c;
  89         xfree((char*)buf);
  90         storeLog(STORE_LOG_SWAPOUTFAIL, e);
  91         return;
  92     }
  93
  94     mem->swapout.sio = sio;
  95     /* Don't lock until after create, or the replacement
  96      * code might get confused */
  97
  98     e->lock();
  99     /* Pick up the file number if it was assigned immediately */
 100     e->swap_filen = mem->swapout.sio->swap_filen;
 101
 102     e->swap_dirn = mem->swapout.sio->swap_dirn;
 103
 104     /* write out the swap metadata */
 105     storeIOWrite(mem->swapout.sio, buf, mem->swap_hdr_sz, 0, xfree_cppwrapper);
 106 }
 107
 108 static void
 109 storeSwapOutFileNotify(void *data, int errflag, StoreIOState::Pointer self)
 110 {
 111     generic_cbdata *c = (generic_cbdata *)data;
 112     StoreEntry *e = (StoreEntry *)c->data;
 113     MemObject *mem = e->mem_obj;
 114     assert(e->swap_status == SWAPOUT_WRITING);
 115     assert(mem);
 116     assert(mem->swapout.sio == self);
 117     assert(errflag == 0);
 118     e->swap_filen = mem->swapout.sio->swap_filen;
 119     e->swap_dirn = mem->swapout.sio->swap_dirn;
 120 }
 121
 122 static void
 123 doPages(StoreEntry *anEntry)
 124 {
 125     MemObject *mem = anEntry->mem_obj;
 126
 127     do {
 128         /*
 129          * Evil hack time.
 130          * We are paging out to disk in page size chunks. however, later on when
 131          * we update the queue position, we might not have a page (I *think*),
 132          * so we do the actual page update here.
 133          */
 134
 135         if (mem->swapout.memnode == NULL) {
 136             /* We need to swap out the first page */
 137             mem->swapout.memnode = const_cast<mem_node *>(mem->data_hdr.start());
 138         } else {
 139             /* We need to swap out the next page */
 140             /* 20030636 RBC - we don't have ->next anymore.
 141              * But we do have the next location */
 142             mem->swapout.memnode = mem->data_hdr.getBlockContainingLocation (mem->swapout.memnode->end());
 143         }
 144
 145         /*
 146          * Get the length of this buffer. We are assuming(!) that the buffer
 147          * length won't change on this buffer, or things are going to be very
 148          * strange. I think that after the copy to a buffer is done, the buffer
 149          * size should stay fixed regardless so that this code isn't confused,
 150          * but we can look at this at a later date or whenever the code results
 151          * in bad swapouts, whichever happens first. :-)
 152          */
 153         ssize_t swap_buf_len = mem->swapout.memnode->nodeBuffer.length;
 154
 155         debugs(20, 3, "storeSwapOut: swap_buf_len = " << swap_buf_len);
 156
 157         assert(swap_buf_len > 0);
 158
 159         debugs(20, 3, "storeSwapOut: swapping out " << swap_buf_len << " bytes from " << mem->swapout.queue_offset);
 160
 161         mem->swapout.queue_offset += swap_buf_len;
 162
 163         storeIOWrite(mem->swapout.sio,
 164                      mem->data_hdr.NodeGet(mem->swapout.memnode),
 165                      swap_buf_len,
 166                      -1,
 167                      memNodeWriteComplete);
 168
 169         /* the storeWrite() call might generate an error */
 170         if (anEntry->swap_status != SWAPOUT_WRITING)
 171             break;
 172
 173         int64_t swapout_size = mem->endOffset() - mem->swapout.queue_offset;
 174
 175         if (anEntry->store_status == STORE_PENDING)
 176             if (swapout_size < SM_PAGE_SIZE)
 177                 break;
 178
 179         if (swapout_size <= 0)
 180             return;
 181     } while (true);
 182 }
 183
 184
 185 /* This routine is called every time data is sent to the client side.
 186  * It's overhead is therefor, significant.
 187  */
 188 void
 189 StoreEntry::swapOut()
 190 {
 191     if (!mem_obj)
 192         return;
 193
 194     if (!swapoutPossible())
 195         return;
 196
 197     debugs(20, 7, HERE << "storeSwapOut: mem->inmem_lo = " << mem_obj->inmem_lo);
 198     debugs(20, 7, HERE << "storeSwapOut: mem->endOffset() = " << mem_obj->endOffset());
 199     debugs(20, 7, HERE << "storeSwapOut: swapout.queue_offset = " << mem_obj->swapout.queue_offset);
 200
 201     if (mem_obj->swapout.sio != NULL)
 202         debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj->swapout.sio->offset()  );
 203
 204     int64_t swapout_maxsize = mem_obj->endOffset() - mem_obj->swapout.queue_offset;
 205
 206     assert(swapout_maxsize >= 0);
 207
 208     int64_t const lowest_offset = mem_obj->lowestMemReaderOffset();
 209
 210     debugs(20, 7, HERE << "storeSwapOut: lowest_offset = " << lowest_offset);
 211
 212     /*
 213      * Grab the swapout_size and check to see whether we're going to defer
 214      * the swapout based upon size
 215      */
 216     if ((store_status != STORE_OK) && (swapout_maxsize < store_maxobjsize)) {
 217         /*
 218          * NOTE: the store_maxobjsize here is the max of optional
 219          * max-size values from 'cache_dir' lines.  It is not the
 220          * same as 'maximum_object_size'.  By default, store_maxobjsize
 221          * will be set to -1.  However, I am worried that this
 222          * deferance may consume a lot of memory in some cases.
 223          * It would be good to make this decision based on reply
 224          * content-length, rather than wait to accumulate huge
 225          * amounts of object data in memory.
 226          */
 227         debugs(20, 5, "storeSwapOut: Deferring starting swapping out");
 228         return;
 229     }
 230
 231     trimMemory();
 232 #if SIZEOF_OFF_T <= 4
 233
 234     if (mem_obj->endOffset() > 0x7FFF0000) {
 235         debugs(20, 0, "WARNING: preventing off_t overflow for " << url());
 236         abort();
 237         return;
 238     }
 239
 240 #endif
 241     if (swap_status == SWAPOUT_WRITING)
 242         assert(mem_obj->inmem_lo <=  mem_obj->objectBytesOnDisk() );
 243
 244     if (!swapOutAble())
 245         return;
 246
 247     debugs(20, 7, "storeSwapOut: swapout_size = " << swapout_maxsize);
 248
 249     if (swapout_maxsize == 0) {
 250         if (store_status == STORE_OK)
 251             swapOutFileClose();
 252
 253         return;                 /* Nevermore! */
 254     }
 255
 256     if (store_status == STORE_PENDING) {
 257         /* wait for a full block to write */
 258
 259         if (swapout_maxsize < SM_PAGE_SIZE)
 260             return;
 261
 262         /*
 263          * Wait until we are below the disk FD limit, only if the
 264          * next server-side read won't be deferred.
 265          */
 266         if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
 267             return;
 268     }
 269
 270     /* Ok, we have stuff to swap out.  Is there a swapout.sio open? */
 271     if (swap_status == SWAPOUT_NONE) {
 272         assert(mem_obj->swapout.sio == NULL);
 273         assert(mem_obj->inmem_lo == 0);
 274
 275         if (checkCachable())
 276             storeSwapOutStart(this);
 277         else
 278             return;
 279
 280         /* ENTRY_CACHABLE will be cleared and we'll never get here again */
 281     }
 282
 283     if (mem_obj->swapout.sio == NULL)
 284         return;
 285
 286     doPages(this);
 287
 288     if (mem_obj->swapout.sio == NULL)
 289         /* oops, we're not swapping out any more */
 290         return;
 291
 292     if (store_status == STORE_OK) {
 293         /*
 294          * If the state is STORE_OK, then all data must have been given
 295          * to the filesystem at this point because storeSwapOut() is
 296          * not going to be called again for this entry.
 297          */
 298         assert(mem_obj->endOffset() == mem_obj->swapout.queue_offset);
 299         swapOutFileClose();
 300     }
 301 }
 302
 303 void
 304 StoreEntry::swapOutFileClose()
 305 {
 306     assert(mem_obj != NULL);
 307     debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text());
 308     debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj->swapout.sio.getRaw());
 309
 310     if (mem_obj->swapout.sio == NULL)
 311         return;
 312
 313     storeClose(mem_obj->swapout.sio);
 314 }
 315
 316 static void
 317 storeSwapOutFileClosed(void *data, int errflag, StoreIOState::Pointer self)
 318 {
 319     generic_cbdata *c = (generic_cbdata *)data;
 320     StoreEntry *e = (StoreEntry *)c->data;
 321     MemObject *mem = e->mem_obj;
 322     assert(mem->swapout.sio == self);
 323     assert(e->swap_status == SWAPOUT_WRITING);
 324     cbdataFree(c);
 325
 326     if (errflag) {
 327         debugs(20, 1, "storeSwapOutFileClosed: dirno " << e->swap_dirn << ", swapfile " <<
 328                std::hex << std::setw(8) << std::setfill('0') << std::uppercase <<
 329                e->swap_filen << ", errflag=" << errflag);
 330         debugs(20, 1, "\t" << xstrerror());
 331
 332         if (errflag == DISK_NO_SPACE_LEFT) {
 333             /* FIXME: this should be handle by the link from store IO to
 334              * Store, rather than being a top level API call.
 335              */
 336             e->store()->diskFull();
 337             storeConfigure();
 338         }
 339
 340         if (e->swap_filen > 0)
 341             e->unlink();
 342
 343         e->swap_filen = -1;
 344
 345         e->swap_dirn = -1;
 346
 347         e->swap_status = SWAPOUT_NONE;
 348
 349         e->releaseRequest();
 350     } else {
 351         /* swapping complete */
 352         debugs(20, 3, "storeSwapOutFileClosed: SwapOut complete: '" << e->url() << "' to " <<
 353                e->swap_dirn  << ", " << std::hex << std::setw(8) << std::setfill('0') <<
 354                std::uppercase << e->swap_filen);
 355         e->swap_file_sz = e->objectLen() + mem->swap_hdr_sz;
 356         e->swap_status = SWAPOUT_DONE;
 357         e->store()->updateSize(e->swap_file_sz, 1);
 358
 359         if (e->checkCachable()) {
 360             storeLog(STORE_LOG_SWAPOUT, e);
 361             storeDirSwapLog(e, SWAP_LOG_ADD);
 362         }
 363
 364         statCounter.swap.outs++;
 365     }
 366
 367     debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__ << ":" << __LINE__);
 368     mem->swapout.sio = NULL;
 369     e->unlock();
 370 }
 371
 372 /*
 373  * Is this entry a candidate for writing to disk?
 374  */
 375 bool
 376 StoreEntry::swapOutAble() const
 377 {
 378     dlink_node *node;
 379
 380     if (mem_obj->swapout.sio != NULL)
 381         return true;
 382
 383     if (mem_obj->inmem_lo > 0)
 384         return false;
 385
 386     /*
 387      * If there are DISK clients, we must write to disk
 388      * even if its not cachable
 389      * RBC: Surely we should not create disk client on non cacheable objects?
 390      * therefore this should be an assert?
 391      * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
 392      * an assert.
 393      */
 394     for (node = mem_obj->clients.head; node; node = node->next) {
 395         if (((store_client *) node->data)->getType() == STORE_DISK_CLIENT)
 396             return true;
 397     }
 398
 399     /* Don't pollute the disk with icons and other special entries */
 400     if (EBIT_TEST(flags, ENTRY_SPECIAL))
 401         return false;
 402
 403     if (!EBIT_TEST(flags, ENTRY_CACHABLE))
 404         return false;
 405
 406     if (!mem_obj->isContiguous())
 407         return false;
 408
 409     return true;
 410 }