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 /* start swapping object to disk */
  51 static void
  52 storeSwapOutStart(StoreEntry * e)
  53 {
  54     MemObject *mem = e->mem_obj;
  55     StoreIOState::Pointer sio;
  56     assert(mem);
  57     /* Build the swap metadata, so the filesystem will know how much
  58      * metadata there is to store
  59      */
  60     debugs(20, 5, "storeSwapOutStart: Begin SwapOut '" << e->url() << "' to dirno " <<
  61            e->swap_dirn << ", fileno " << std::hex << std::setw(8) << std::setfill('0') <<
  62            std::uppercase << e->swap_filen);
  63     e->swap_status = SWAPOUT_WRITING;
  64     /* If we start swapping out objects with OutOfBand Metadata,
  65      * then this code needs changing
  66      */
  67     /* Create the swap file */
  68     generic_cbdata *c = new generic_cbdata(e);
  69     sio = storeCreate(e, storeSwapOutFileNotify, storeSwapOutFileClosed, c);
  70
  71     if (sio == NULL) {
  72         e->swap_status = SWAPOUT_NONE;
  73         delete c;
  74         storeLog(STORE_LOG_SWAPOUTFAIL, e);
  75         return;
  76     }
  77
  78     mem->swapout.sio = sio;
  79     /* Don't lock until after create, or the replacement
  80      * code might get confused */
  81
  82     e->lock();
  83     /* Pick up the file number if it was assigned immediately */
  84     e->swap_filen = mem->swapout.sio->swap_filen;
  85
  86     e->swap_dirn = mem->swapout.sio->swap_dirn;
  87
  88     /* write out the swap metadata */
  89     /* TODO: make some sort of data,size refcounted immutable buffer
  90      * for use by this sort of function.
  91      */
  92     char const *buf = e->getSerialisedMetaData ();
  93
  94     /* If we start swapping out with out of band metadata, this assert
  95      * will catch it - this code needs to be adjusted if that happens
  96      */
  97     assert (buf);
  98
  99     storeIOWrite(mem->swapout.sio, buf, mem->swap_hdr_sz, 0, xfree);
 100 }
 101
 102 static void
 103 storeSwapOutFileNotify(void *data, int errflag, StoreIOState::Pointer self)
 104 {
 105     generic_cbdata *c = (generic_cbdata *)data;
 106     StoreEntry *e = (StoreEntry *)c->data;
 107     MemObject *mem = e->mem_obj;
 108     assert(e->swap_status == SWAPOUT_WRITING);
 109     assert(mem);
 110     assert(mem->swapout.sio == self);
 111     assert(errflag == 0);
 112     e->swap_filen = mem->swapout.sio->swap_filen;
 113     e->swap_dirn = mem->swapout.sio->swap_dirn;
 114 }
 115
 116 static void
 117 doPages(StoreEntry *anEntry)
 118 {
 119     MemObject *mem = anEntry->mem_obj;
 120
 121     do {
 122         /*
 123          * Evil hack time.
 124          * We are paging out to disk in page size chunks. however, later on when
 125          * we update the queue position, we might not have a page (I *think*),
 126          * so we do the actual page update here.
 127          */
 128
 129         if (mem->swapout.memnode == NULL) {
 130             /* We need to swap out the first page */
 131             mem->swapout.memnode = const_cast<mem_node *>(mem->data_hdr.start());
 132         } else {
 133             /* We need to swap out the next page */
 134             /* 20030636 RBC - we don't have ->next anymore.
 135              * But we do have the next location */
 136             mem->swapout.memnode = mem->data_hdr.getBlockContainingLocation (mem->swapout.memnode->end());
 137         }
 138
 139         /*
 140          * Get the length of this buffer. We are assuming(!) that the buffer
 141          * length won't change on this buffer, or things are going to be very
 142          * strange. I think that after the copy to a buffer is done, the buffer
 143          * size should stay fixed regardless so that this code isn't confused,
 144          * but we can look at this at a later date or whenever the code results
 145          * in bad swapouts, whichever happens first. :-)
 146          */
 147         ssize_t swap_buf_len = mem->swapout.memnode->nodeBuffer.length;
 148
 149         debugs(20, 3, "storeSwapOut: swap_buf_len = " << swap_buf_len);
 150
 151         assert(swap_buf_len > 0);
 152
 153         debugs(20, 3, "storeSwapOut: swapping out " << swap_buf_len << " bytes from " << mem->swapout.queue_offset);
 154
 155         mem->swapout.queue_offset += swap_buf_len;
 156
 157         storeIOWrite(mem->swapout.sio,
 158                      mem->data_hdr.NodeGet(mem->swapout.memnode),
 159                      swap_buf_len,
 160                      -1,
 161                      memNodeWriteComplete);
 162
 163         /* the storeWrite() call might generate an error */
 164         if (anEntry->swap_status != SWAPOUT_WRITING)
 165             break;
 166
 167         int64_t swapout_size = mem->endOffset() - mem->swapout.queue_offset;
 168
 169         if (anEntry->store_status == STORE_PENDING)
 170             if (swapout_size < SM_PAGE_SIZE)
 171                 break;
 172
 173         if (swapout_size <= 0)
 174             return;
 175     } while (true);
 176 }
 177
 178
 179 /* This routine is called every time data is sent to the client side.
 180  * It's overhead is therefor, significant.
 181  */
 182 void
 183 StoreEntry::swapOut()
 184 {
 185     if (!mem_obj)
 186         return;
 187
 188     if (!swapoutPossible())
 189         return;
 190
 191     debugs(20, 7, HERE << "storeSwapOut: mem->inmem_lo = " << mem_obj->inmem_lo);
 192     debugs(20, 7, HERE << "storeSwapOut: mem->endOffset() = " << mem_obj->endOffset());
 193     debugs(20, 7, HERE << "storeSwapOut: swapout.queue_offset = " << mem_obj->swapout.queue_offset);
 194
 195     if (mem_obj->swapout.sio != NULL)
 196         debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj->swapout.sio->offset()  );
 197
 198     int64_t swapout_maxsize = mem_obj->endOffset() - mem_obj->swapout.queue_offset;
 199
 200     assert(swapout_maxsize >= 0);
 201
 202     int64_t const lowest_offset = mem_obj->lowestMemReaderOffset();
 203
 204     debugs(20, 7, HERE << "storeSwapOut: lowest_offset = " << lowest_offset);
 205
 206     /*
 207      * Grab the swapout_size and check to see whether we're going to defer
 208      * the swapout based upon size
 209      */
 210     if ((store_status != STORE_OK) && (swapout_maxsize < store_maxobjsize)) {
 211         /*
 212          * NOTE: the store_maxobjsize here is the max of optional
 213          * max-size values from 'cache_dir' lines.  It is not the
 214          * same as 'maximum_object_size'.  By default, store_maxobjsize
 215          * will be set to -1.  However, I am worried that this
 216          * deferance may consume a lot of memory in some cases.
 217          * It would be good to make this decision based on reply
 218          * content-length, rather than wait to accumulate huge
 219          * amounts of object data in memory.
 220          */
 221         debugs(20, 5, "storeSwapOut: Deferring starting swapping out");
 222         return;
 223     }
 224
 225     trimMemory();
 226 #if SIZEOF_OFF_T <= 4
 227
 228     if (mem_obj->endOffset() > 0x7FFF0000) {
 229         debugs(20, 0, "WARNING: preventing off_t overflow for " << url());
 230         abort();
 231         return;
 232     }
 233
 234 #endif
 235     if (swap_status == SWAPOUT_WRITING)
 236         assert(mem_obj->inmem_lo <=  mem_obj->objectBytesOnDisk() );
 237
 238     if (!swapOutAble())
 239         return;
 240
 241     debugs(20, 7, "storeSwapOut: swapout_size = " << swapout_maxsize);
 242
 243     if (swapout_maxsize == 0) {
 244         if (store_status == STORE_OK)
 245             swapOutFileClose();
 246
 247         return;                 /* Nevermore! */
 248     }
 249
 250     if (store_status == STORE_PENDING) {
 251         /* wait for a full block to write */
 252
 253         if (swapout_maxsize < SM_PAGE_SIZE)
 254             return;
 255
 256         /*
 257          * Wait until we are below the disk FD limit, only if the
 258          * next server-side read won't be deferred.
 259          */
 260         if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
 261             return;
 262     }
 263
 264     /* Ok, we have stuff to swap out.  Is there a swapout.sio open? */
 265     if (swap_status == SWAPOUT_NONE) {
 266         assert(mem_obj->swapout.sio == NULL);
 267         assert(mem_obj->inmem_lo == 0);
 268
 269         if (checkCachable())
 270             storeSwapOutStart(this);
 271         else
 272             return;
 273
 274         /* ENTRY_CACHABLE will be cleared and we'll never get here again */
 275     }
 276
 277     if (mem_obj->swapout.sio == NULL)
 278         return;
 279
 280     doPages(this);
 281
 282     if (mem_obj->swapout.sio == NULL)
 283         /* oops, we're not swapping out any more */
 284         return;
 285
 286     if (store_status == STORE_OK) {
 287         /*
 288          * If the state is STORE_OK, then all data must have been given
 289          * to the filesystem at this point because storeSwapOut() is
 290          * not going to be called again for this entry.
 291          */
 292         assert(mem_obj->endOffset() == mem_obj->swapout.queue_offset);
 293         swapOutFileClose();
 294     }
 295 }
 296
 297 void
 298 StoreEntry::swapOutFileClose()
 299 {
 300     assert(mem_obj != NULL);
 301     debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text());
 302     debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj->swapout.sio.getRaw());
 303
 304     if (mem_obj->swapout.sio == NULL)
 305         return;
 306
 307     storeClose(mem_obj->swapout.sio);
 308 }
 309
 310 static void
 311 storeSwapOutFileClosed(void *data, int errflag, StoreIOState::Pointer self)
 312 {
 313     generic_cbdata *c = (generic_cbdata *)data;
 314     StoreEntry *e = (StoreEntry *)c->data;
 315     MemObject *mem = e->mem_obj;
 316     assert(mem->swapout.sio == self);
 317     assert(e->swap_status == SWAPOUT_WRITING);
 318     cbdataFree(c);
 319
 320     if (errflag) {
 321         debugs(20, 1, "storeSwapOutFileClosed: dirno " << e->swap_dirn << ", swapfile " <<
 322                std::hex << std::setw(8) << std::setfill('0') << std::uppercase <<
 323                e->swap_filen << ", errflag=" << errflag);
 324         debugs(20, 1, "\t" << xstrerror());
 325
 326         if (errflag == DISK_NO_SPACE_LEFT) {
 327             /* FIXME: this should be handle by the link from store IO to
 328              * Store, rather than being a top level API call.
 329              */
 330             e->store()->diskFull();
 331             storeConfigure();
 332         }
 333
 334         if (e->swap_filen > 0)
 335             e->unlink();
 336
 337         e->swap_filen = -1;
 338
 339         e->swap_dirn = -1;
 340
 341         e->swap_status = SWAPOUT_NONE;
 342
 343         e->releaseRequest();
 344     } else {
 345         /* swapping complete */
 346         debugs(20, 3, "storeSwapOutFileClosed: SwapOut complete: '" << e->url() << "' to " <<
 347                e->swap_dirn  << ", " << std::hex << std::setw(8) << std::setfill('0') <<
 348                std::uppercase << e->swap_filen);
 349         e->swap_file_sz = e->objectLen() + mem->swap_hdr_sz;
 350         e->swap_status = SWAPOUT_DONE;
 351         e->store()->updateSize(e->swap_file_sz, 1);
 352
 353         if (e->checkCachable()) {
 354             storeLog(STORE_LOG_SWAPOUT, e);
 355             storeDirSwapLog(e, SWAP_LOG_ADD);
 356         }
 357
 358         statCounter.swap.outs++;
 359     }
 360
 361     debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__ << ":" << __LINE__);
 362     mem->swapout.sio = NULL;
 363     e->unlock();
 364 }
 365
 366 /*
 367  * Is this entry a candidate for writing to disk?
 368  */
 369 bool
 370 StoreEntry::swapOutAble() const
 371 {
 372     dlink_node *node;
 373
 374     if (mem_obj->swapout.sio != NULL)
 375         return true;
 376
 377     if (mem_obj->inmem_lo > 0)
 378         return false;
 379
 380     /*
 381      * If there are DISK clients, we must write to disk
 382      * even if its not cachable
 383      * RBC: Surely we should not create disk client on non cacheable objects?
 384      * therefore this should be an assert?
 385      * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
 386      * an assert.
 387      */
 388     for (node = mem_obj->clients.head; node; node = node->next) {
 389         if (((store_client *) node->data)->getType() == STORE_DISK_CLIENT)
 390             return true;
 391     }
 392
 393     /* Don't pollute the disk with icons and other special entries */
 394     if (EBIT_TEST(flags, ENTRY_SPECIAL))
 395         return false;
 396
 397     if (!EBIT_TEST(flags, ENTRY_CACHABLE))
 398         return false;
 399
 400     if (!mem_obj->isContiguous())
 401         return false;
 402
 403     return true;
 404 }