5 * DEBUG: section 20 Storage Manager Swapout Functions
6 * AUTHOR: Duane Wessels
8 * SQUID Web Proxy Cache http://www.squid-cache.org/
9 * ----------------------------------------------------------
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.
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.
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.
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.
38 #include "StoreClient.h"
40 /* FIXME: Abstract the use of this more */
42 #include "MemObject.h"
45 #include "StatCounters.h"
46 #include "swap_log_op.h"
48 static void storeSwapOutStart(StoreEntry
* e
);
49 static StoreIOState::STIOCB storeSwapOutFileClosed
;
50 static StoreIOState::STFNCB storeSwapOutFileNotify
;
52 // wrapper to cross C/C++ ABI boundary. xfree is extern "C" for libraries.
53 static void xfree_cppwrapper(void *x
)
58 /* start swapping object to disk */
60 storeSwapOutStart(StoreEntry
* e
)
62 MemObject
*mem
= e
->mem_obj
;
63 StoreIOState::Pointer sio
;
65 /* Build the swap metadata, so the filesystem will know how much
66 * metadata there is to store
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
76 /* TODO: make some sort of data,size refcounted immutable buffer
77 * and stop fooling ourselves with "const char*" buffers.
80 // Create metadata now, possibly in vain: storeCreate needs swap_hdr_sz.
81 const char *buf
= e
->getSerialisedMetaData ();
84 /* Create the swap file */
85 generic_cbdata
*c
= new generic_cbdata(e
);
86 sio
= storeCreate(e
, storeSwapOutFileNotify
, storeSwapOutFileClosed
, c
);
89 e
->swap_status
= SWAPOUT_NONE
;
90 mem
->swapout
.decision
= MemObject::SwapOut::swImpossible
;
93 storeLog(STORE_LOG_SWAPOUTFAIL
, e
);
97 mem
->swapout
.sio
= sio
;
98 /* Don't lock until after create, or the replacement
99 * code might get confused */
102 /* Pick up the file number if it was assigned immediately */
103 e
->swap_filen
= mem
->swapout
.sio
->swap_filen
;
105 e
->swap_dirn
= mem
->swapout
.sio
->swap_dirn
;
107 /* write out the swap metadata */
108 storeIOWrite(mem
->swapout
.sio
, buf
, mem
->swap_hdr_sz
, 0, xfree_cppwrapper
);
112 storeSwapOutFileNotify(void *data
, int errflag
, StoreIOState::Pointer self
)
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
);
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
;
127 doPages(StoreEntry
*anEntry
)
129 MemObject
*mem
= anEntry
->mem_obj
;
132 // find the page containing the first byte we have not swapped out yet
134 mem
->data_hdr
.getBlockContainingLocation(mem
->swapout
.queue_offset
);
137 return; // wait for more data to become available
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
);
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. :-)
151 ssize_t swap_buf_len
= page
->nodeBuffer
.length
;
153 debugs(20, 3, "storeSwapOut: swap_buf_len = " << swap_buf_len
);
155 assert(swap_buf_len
> 0);
157 debugs(20, 3, "storeSwapOut: swapping out " << swap_buf_len
<< " bytes from " << mem
->swapout
.queue_offset
);
159 mem
->swapout
.queue_offset
+= swap_buf_len
;
161 storeIOWrite(mem
->swapout
.sio
,
162 mem
->data_hdr
.NodeGet(page
),
165 memNodeWriteComplete
);
167 /* the storeWrite() call might generate an error */
168 if (anEntry
->swap_status
!= SWAPOUT_WRITING
)
171 int64_t swapout_size
= mem
->endOffset() - mem
->swapout
.queue_offset
;
173 if (anEntry
->store_status
== STORE_PENDING
)
174 if (swapout_size
< SM_PAGE_SIZE
)
177 if (swapout_size
<= 0)
183 /* This routine is called every time data is sent to the client side.
184 * It's overhead is therefor, significant.
187 StoreEntry::swapOut()
192 // this flag may change so we must check even if we are swappingOut
193 if (EBIT_TEST(flags
, ENTRY_ABORTED
)) {
194 assert(EBIT_TEST(flags
, RELEASE_REQUEST
));
195 // StoreEntry::abort() already closed the swap out file, if any
196 // no trimming: data producer must stop production if ENTRY_ABORTED
200 const bool weAreOrMayBeSwappingOut
= swappingOut() || mayStartSwapOut();
202 Store::Root().maybeTrimMemory(*this, weAreOrMayBeSwappingOut
);
204 if (!weAreOrMayBeSwappingOut
)
205 return; // nothing else to do
207 // Aborted entries have STORE_OK, but swapoutPossible rejects them. Thus,
208 // store_status == STORE_OK below means we got everything we wanted.
210 debugs(20, 7, HERE
<< "storeSwapOut: mem->inmem_lo = " << mem_obj
->inmem_lo
);
211 debugs(20, 7, HERE
<< "storeSwapOut: mem->endOffset() = " << mem_obj
->endOffset());
212 debugs(20, 7, HERE
<< "storeSwapOut: swapout.queue_offset = " << mem_obj
->swapout
.queue_offset
);
214 if (mem_obj
->swapout
.sio
!= NULL
)
215 debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj
->swapout
.sio
->offset() );
217 int64_t const lowest_offset
= mem_obj
->lowestMemReaderOffset();
219 debugs(20, 7, HERE
<< "storeSwapOut: lowest_offset = " << lowest_offset
);
221 #if SIZEOF_OFF_T <= 4
223 if (mem_obj
->endOffset() > 0x7FFF0000) {
224 debugs(20, DBG_CRITICAL
, "WARNING: preventing off_t overflow for " << url());
230 if (swap_status
== SWAPOUT_WRITING
)
231 assert(mem_obj
->inmem_lo
<= mem_obj
->objectBytesOnDisk() );
233 // buffered bytes we have not swapped out yet
234 const int64_t swapout_maxsize
= mem_obj
->availableForSwapOut();
235 assert(swapout_maxsize
>= 0);
236 debugs(20, 7, "storeSwapOut: swapout_size = " << swapout_maxsize
);
238 if (swapout_maxsize
== 0) { // swapped everything we got
239 if (store_status
== STORE_OK
) { // got everything we wanted
240 assert(mem_obj
->object_sz
>= 0);
241 swapOutFileClose(StoreIOState::wroteAll
);
243 // else need more data to swap out
247 if (store_status
== STORE_PENDING
) {
248 /* wait for a full block to write */
250 if (swapout_maxsize
< SM_PAGE_SIZE
)
254 * Wait until we are below the disk FD limit, only if the
255 * next server-side read won't be deferred.
257 if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
261 /* Ok, we have stuff to swap out. Is there a swapout.sio open? */
262 if (swap_status
== SWAPOUT_NONE
) {
263 assert(mem_obj
->swapout
.sio
== NULL
);
264 assert(mem_obj
->inmem_lo
== 0);
265 storeSwapOutStart(this); // sets SwapOut::swImpossible on failures
268 if (mem_obj
->swapout
.sio
== NULL
)
273 if (mem_obj
->swapout
.sio
== NULL
)
274 /* oops, we're not swapping out any more */
277 if (store_status
== STORE_OK
) {
279 * If the state is STORE_OK, then all data must have been given
280 * to the filesystem at this point because storeSwapOut() is
281 * not going to be called again for this entry.
283 assert(mem_obj
->object_sz
>= 0);
284 assert(mem_obj
->endOffset() == mem_obj
->swapout
.queue_offset
);
285 swapOutFileClose(StoreIOState::wroteAll
);
290 StoreEntry::swapOutFileClose(int how
)
292 assert(mem_obj
!= NULL
);
293 debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text() << " how=" << how
);
294 debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj
->swapout
.sio
.getRaw());
296 if (mem_obj
->swapout
.sio
== NULL
)
299 storeClose(mem_obj
->swapout
.sio
, how
);
303 storeSwapOutFileClosed(void *data
, int errflag
, StoreIOState::Pointer self
)
305 generic_cbdata
*c
= (generic_cbdata
*)data
;
306 StoreEntry
*e
= (StoreEntry
*)c
->data
;
307 MemObject
*mem
= e
->mem_obj
;
308 assert(mem
->swapout
.sio
== self
);
309 assert(e
->swap_status
== SWAPOUT_WRITING
);
312 // if object_size is still unknown, the entry was probably aborted
313 if (errflag
|| e
->objectLen() < 0) {
314 debugs(20, 2, "storeSwapOutFileClosed: dirno " << e
->swap_dirn
<< ", swapfile " <<
315 std::hex
<< std::setw(8) << std::setfill('0') << std::uppercase
<<
316 e
->swap_filen
<< ", errflag=" << errflag
);
318 if (errflag
== DISK_NO_SPACE_LEFT
) {
319 /* FIXME: this should be handle by the link from store IO to
320 * Store, rather than being a top level API call.
322 e
->store()->diskFull();
326 if (e
->swap_filen
>= 0)
329 assert(e
->swap_status
== SWAPOUT_NONE
);
333 /* swapping complete */
334 debugs(20, 3, "storeSwapOutFileClosed: SwapOut complete: '" << e
->url() << "' to " <<
335 e
->swap_dirn
<< ", " << std::hex
<< std::setw(8) << std::setfill('0') <<
336 std::uppercase
<< e
->swap_filen
);
337 debugs(20, 5, HERE
<< "swap_file_sz = " <<
338 e
->objectLen() << " + " << mem
->swap_hdr_sz
);
340 e
->swap_file_sz
= e
->objectLen() + mem
->swap_hdr_sz
;
341 e
->swap_status
= SWAPOUT_DONE
;
342 e
->store()->swappedOut(*e
);
344 // XXX: For some Stores, it is pointless to re-check cachability here
345 // and it leads to double counts in store_check_cachable_hist. We need
346 // another way to signal a completed but failed swapout. Or, better,
347 // each Store should handle its own logging and LOG state setting.
348 if (e
->checkCachable()) {
349 storeLog(STORE_LOG_SWAPOUT
, e
);
350 storeDirSwapLog(e
, SWAP_LOG_ADD
);
353 ++statCounter
.swap
.outs
;
356 debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__
<< ":" << __LINE__
);
357 mem
->swapout
.sio
= NULL
;
362 StoreEntry::mayStartSwapOut()
366 // must be checked in the caller
367 assert(!EBIT_TEST(flags
, ENTRY_ABORTED
));
368 assert(!swappingOut());
370 if (!Config
.cacheSwap
.n_configured
)
374 MemObject::SwapOut::Decision
&decision
= mem_obj
->swapout
.decision
;
376 // if we decided that swapout is not possible, do not repeat same checks
377 if (decision
== MemObject::SwapOut::swImpossible
) {
378 debugs(20, 3, HERE
<< " already rejected");
382 // if we decided that swapout is possible, do not repeat same checks
383 if (decision
== MemObject::SwapOut::swPossible
) {
384 debugs(20, 3, HERE
<< "already allowed");
388 // if we swapped out already, do not start over
389 if (swap_status
== SWAPOUT_DONE
) {
390 debugs(20, 3, HERE
<< "already did");
391 decision
= MemObject::SwapOut::swImpossible
;
395 if (!checkCachable()) {
396 debugs(20, 3, HERE
<< "not cachable");
397 decision
= MemObject::SwapOut::swImpossible
;
401 if (EBIT_TEST(flags
, ENTRY_SPECIAL
)) {
402 debugs(20, 3, HERE
<< url() << " SPECIAL");
403 decision
= MemObject::SwapOut::swImpossible
;
407 // check cache_dir max-size limit if all cache_dirs have it
408 if (store_maxobjsize
>= 0) {
409 // TODO: add estimated store metadata size to be conservative
411 // use guaranteed maximum if it is known
412 const int64_t expectedEnd
= mem_obj
->expectedReplySize();
413 debugs(20, 7, HERE
<< "expectedEnd = " << expectedEnd
);
414 if (expectedEnd
> store_maxobjsize
) {
415 debugs(20, 3, HERE
<< "will not fit: " << expectedEnd
<<
416 " > " << store_maxobjsize
);
417 decision
= MemObject::SwapOut::swImpossible
;
418 return false; // known to outgrow the limit eventually
421 // use current minimum (always known)
422 const int64_t currentEnd
= mem_obj
->endOffset();
423 if (currentEnd
> store_maxobjsize
) {
424 debugs(20, 3, HERE
<< "does not fit: " << currentEnd
<<
425 " > " << store_maxobjsize
);
426 decision
= MemObject::SwapOut::swImpossible
;
427 return false; // already does not fit and may only get bigger
430 // prevent default swPossible answer for yet unknown length
431 if (expectedEnd
< 0) {
432 debugs(20, 3, HERE
<< "wait for more info: " <<
434 return false; // may fit later, but will be rejected now
437 if (store_status
!= STORE_OK
) {
438 const int64_t maxKnownSize
= expectedEnd
< 0 ?
439 mem_obj
->availableForSwapOut() : expectedEnd
;
440 debugs(20, 7, HERE
<< "maxKnownSize= " << maxKnownSize
);
441 if (maxKnownSize
< store_maxobjsize
) {
443 * NOTE: the store_maxobjsize here is the max of optional
444 * max-size values from 'cache_dir' lines. It is not the
445 * same as 'maximum_object_size'. By default, store_maxobjsize
446 * will be set to -1. However, I am worried that this
447 * deferance may consume a lot of memory in some cases.
448 * Should we add an option to limit this memory consumption?
450 debugs(20, 5, HERE
<< "Deferring swapout start for " <<
451 (store_maxobjsize
- maxKnownSize
) << " bytes");
457 if (mem_obj
->inmem_lo
> 0) {
458 debugs(20, 3, "storeSwapOut: (inmem_lo > 0) imem_lo:" << mem_obj
->inmem_lo
);
459 decision
= MemObject::SwapOut::swImpossible
;
464 * If there are DISK clients, we must write to disk
465 * even if its not cachable
466 * RBC: Surely we should not create disk client on non cacheable objects?
467 * therefore this should be an assert?
468 * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
471 * XXX: Not clear what "mem races" the above refers to, especially when
472 * dealing with non-cachable objects that cannot have multiple clients.
474 * XXX: If STORE_DISK_CLIENT needs SwapOut::swPossible, we have to check
475 * for that flag earlier, but forcing swapping may contradict max-size or
476 * other swapability restrictions. Change storeClientType() and/or its
477 * callers to take swap-in availability into account.
479 for (node
= mem_obj
->clients
.head
; node
; node
= node
->next
) {
480 if (((store_client
*) node
->data
)->getType() == STORE_DISK_CLIENT
) {
481 debugs(20, 3, HERE
<< "DISK client found");
482 decision
= MemObject::SwapOut::swPossible
;
487 if (!mem_obj
->isContiguous()) {
488 debugs(20, 3, "storeSwapOut: not Contiguous");
489 decision
= MemObject::SwapOut::swImpossible
;
493 decision
= MemObject::SwapOut::swPossible
;