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"
44 #include "StatCounters.h"
45 #include "swap_log_op.h"
47 static void storeSwapOutStart(StoreEntry
* e
);
48 static StoreIOState::STIOCB storeSwapOutFileClosed
;
49 static StoreIOState::STFNCB storeSwapOutFileNotify
;
51 // wrapper to cross C/C++ ABI boundary. xfree is extern "C" for libraries.
52 static void xfree_cppwrapper(void *x
)
57 /* start swapping object to disk */
59 storeSwapOutStart(StoreEntry
* e
)
61 MemObject
*mem
= e
->mem_obj
;
62 StoreIOState::Pointer sio
;
64 /* Build the swap metadata, so the filesystem will know how much
65 * metadata there is to store
67 debugs(20, 5, "storeSwapOutStart: Begin SwapOut '" << e
->url() << "' to dirno " <<
68 e
->swap_dirn
<< ", fileno " << std::hex
<< std::setw(8) << std::setfill('0') <<
69 std::uppercase
<< e
->swap_filen
);
70 e
->swap_status
= SWAPOUT_WRITING
;
71 /* If we start swapping out objects with OutOfBand Metadata,
72 * then this code needs changing
75 /* TODO: make some sort of data,size refcounted immutable buffer
76 * and stop fooling ourselves with "const char*" buffers.
79 // Create metadata now, possibly in vain: storeCreate needs swap_hdr_sz.
80 const char *buf
= e
->getSerialisedMetaData ();
83 /* Create the swap file */
84 generic_cbdata
*c
= new generic_cbdata(e
);
85 sio
= storeCreate(e
, storeSwapOutFileNotify
, storeSwapOutFileClosed
, c
);
88 e
->swap_status
= SWAPOUT_NONE
;
89 mem
->swapout
.decision
= MemObject::SwapOut::swImpossible
;
92 storeLog(STORE_LOG_SWAPOUTFAIL
, e
);
96 mem
->swapout
.sio
= sio
;
97 /* Don't lock until after create, or the replacement
98 * code might get confused */
101 /* Pick up the file number if it was assigned immediately */
102 e
->swap_filen
= mem
->swapout
.sio
->swap_filen
;
104 e
->swap_dirn
= mem
->swapout
.sio
->swap_dirn
;
106 /* write out the swap metadata */
107 storeIOWrite(mem
->swapout
.sio
, buf
, mem
->swap_hdr_sz
, 0, xfree_cppwrapper
);
111 storeSwapOutFileNotify(void *data
, int errflag
, StoreIOState::Pointer self
)
113 generic_cbdata
*c
= (generic_cbdata
*)data
;
114 StoreEntry
*e
= (StoreEntry
*)c
->data
;
115 MemObject
*mem
= e
->mem_obj
;
116 assert(e
->swap_status
== SWAPOUT_WRITING
);
118 assert(mem
->swapout
.sio
== self
);
119 assert(errflag
== 0);
120 assert(e
->swap_filen
< 0); // if this fails, call SwapDir::disconnect(e)
121 e
->swap_filen
= mem
->swapout
.sio
->swap_filen
;
122 e
->swap_dirn
= mem
->swapout
.sio
->swap_dirn
;
126 doPages(StoreEntry
*anEntry
)
128 MemObject
*mem
= anEntry
->mem_obj
;
131 // find the page containing the first byte we have not swapped out yet
133 mem
->data_hdr
.getBlockContainingLocation(mem
->swapout
.queue_offset
);
136 return; // wait for more data to become available
138 // memNodeWriteComplete() and absence of buffer offset math below
139 // imply that we always write from the very beginning of the page
140 assert(page
->start() == mem
->swapout
.queue_offset
);
143 * Get the length of this buffer. We are assuming(!) that the buffer
144 * length won't change on this buffer, or things are going to be very
145 * strange. I think that after the copy to a buffer is done, the buffer
146 * size should stay fixed regardless so that this code isn't confused,
147 * but we can look at this at a later date or whenever the code results
148 * in bad swapouts, whichever happens first. :-)
150 ssize_t swap_buf_len
= page
->nodeBuffer
.length
;
152 debugs(20, 3, "storeSwapOut: swap_buf_len = " << swap_buf_len
);
154 assert(swap_buf_len
> 0);
156 debugs(20, 3, "storeSwapOut: swapping out " << swap_buf_len
<< " bytes from " << mem
->swapout
.queue_offset
);
158 mem
->swapout
.queue_offset
+= swap_buf_len
;
160 storeIOWrite(mem
->swapout
.sio
,
161 mem
->data_hdr
.NodeGet(page
),
164 memNodeWriteComplete
);
166 /* the storeWrite() call might generate an error */
167 if (anEntry
->swap_status
!= SWAPOUT_WRITING
)
170 int64_t swapout_size
= mem
->endOffset() - mem
->swapout
.queue_offset
;
172 if (anEntry
->store_status
== STORE_PENDING
)
173 if (swapout_size
< SM_PAGE_SIZE
)
176 if (swapout_size
<= 0)
182 /* This routine is called every time data is sent to the client side.
183 * It's overhead is therefor, significant.
186 StoreEntry::swapOut()
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
199 const bool weAreOrMayBeSwappingOut
= swappingOut() || mayStartSwapOut();
201 trimMemory(weAreOrMayBeSwappingOut
);
203 if (!weAreOrMayBeSwappingOut
)
204 return; // nothing else to do
206 // Aborted entries have STORE_OK, but swapoutPossible rejects them. Thus,
207 // store_status == STORE_OK below means we got everything we wanted.
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
);
213 if (mem_obj
->swapout
.sio
!= NULL
)
214 debugs(20, 7, "storeSwapOut: storeOffset() = " << mem_obj
->swapout
.sio
->offset() );
216 int64_t const lowest_offset
= mem_obj
->lowestMemReaderOffset();
218 debugs(20, 7, HERE
<< "storeSwapOut: lowest_offset = " << lowest_offset
);
220 #if SIZEOF_OFF_T <= 4
222 if (mem_obj
->endOffset() > 0x7FFF0000) {
223 debugs(20, 0, "WARNING: preventing off_t overflow for " << url());
229 if (swap_status
== SWAPOUT_WRITING
)
230 assert(mem_obj
->inmem_lo
<= mem_obj
->objectBytesOnDisk() );
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
);
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
);
242 // else need more data to swap out
246 if (store_status
== STORE_PENDING
) {
247 /* wait for a full block to write */
249 if (swapout_maxsize
< SM_PAGE_SIZE
)
253 * Wait until we are below the disk FD limit, only if the
254 * next server-side read won't be deferred.
256 if (storeTooManyDiskFilesOpen() && !checkDeferRead(-1))
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
267 if (mem_obj
->swapout
.sio
== NULL
)
272 if (mem_obj
->swapout
.sio
== NULL
)
273 /* oops, we're not swapping out any more */
276 if (store_status
== STORE_OK
) {
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.
282 assert(mem_obj
->object_sz
>= 0);
283 assert(mem_obj
->endOffset() == mem_obj
->swapout
.queue_offset
);
284 swapOutFileClose(StoreIOState::wroteAll
);
289 StoreEntry::swapOutFileClose(int how
)
291 assert(mem_obj
!= NULL
);
292 debugs(20, 3, "storeSwapOutFileClose: " << getMD5Text() << " how=" << how
);
293 debugs(20, 3, "storeSwapOutFileClose: sio = " << mem_obj
->swapout
.sio
.getRaw());
295 if (mem_obj
->swapout
.sio
== NULL
)
298 storeClose(mem_obj
->swapout
.sio
, how
);
302 storeSwapOutFileClosed(void *data
, int errflag
, StoreIOState::Pointer self
)
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
);
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
);
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.
321 e
->store()->diskFull();
325 if (e
->swap_filen
>= 0)
328 assert(e
->swap_status
== SWAPOUT_NONE
);
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
);
339 e
->swap_file_sz
= e
->objectLen() + mem
->swap_hdr_sz
;
340 e
->swap_status
= SWAPOUT_DONE
;
341 e
->store()->swappedOut(*e
);
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
);
352 ++statCounter
.swap
.outs
;
355 debugs(20, 3, "storeSwapOutFileClosed: " << __FILE__
<< ":" << __LINE__
);
356 mem
->swapout
.sio
= NULL
;
361 StoreEntry::mayStartSwapOut()
365 // must be checked in the caller
366 assert(!EBIT_TEST(flags
, ENTRY_ABORTED
));
368 if (!Config
.cacheSwap
.n_configured
)
372 MemObject::SwapOut::Decision
&decision
= mem_obj
->swapout
.decision
;
374 // if we decided that swapout is not possible, do not repeat same checks
375 if (decision
== MemObject::SwapOut::swImpossible
) {
376 debugs(20, 3, HERE
<< " already rejected");
380 // if we decided that swapout is possible, do not repeat same checks
381 if (decision
== MemObject::SwapOut::swPossible
) {
382 debugs(20, 3, HERE
<< "already allowed");
386 // if we are swapping out already, do not repeat same checks
387 if (swap_status
!= SWAPOUT_NONE
) {
388 debugs(20, 3, HERE
<< " already started");
389 decision
= MemObject::SwapOut::swPossible
;
393 if (!checkCachable()) {
394 debugs(20, 3, HERE
<< "not cachable");
395 decision
= MemObject::SwapOut::swImpossible
;
399 if (EBIT_TEST(flags
, ENTRY_SPECIAL
)) {
400 debugs(20, 3, HERE
<< url() << " SPECIAL");
401 decision
= MemObject::SwapOut::swImpossible
;
405 // check cache_dir max-size limit if all cache_dirs have it
406 if (store_maxobjsize
>= 0) {
407 // TODO: add estimated store metadata size to be conservative
409 // use guaranteed maximum if it is known
410 const int64_t expectedEnd
= mem_obj
->expectedReplySize();
411 debugs(20, 7, HERE
<< "expectedEnd = " << expectedEnd
);
412 if (expectedEnd
> store_maxobjsize
) {
413 debugs(20, 3, HERE
<< "will not fit: " << expectedEnd
<<
414 " > " << store_maxobjsize
);
415 decision
= MemObject::SwapOut::swImpossible
;
416 return false; // known to outgrow the limit eventually
419 // use current minimum (always known)
420 const int64_t currentEnd
= mem_obj
->endOffset();
421 if (currentEnd
> store_maxobjsize
) {
422 debugs(20, 3, HERE
<< "does not fit: " << currentEnd
<<
423 " > " << store_maxobjsize
);
424 decision
= MemObject::SwapOut::swImpossible
;
425 return false; // already does not fit and may only get bigger
428 // prevent default swPossible answer for yet unknown length
429 if (expectedEnd
< 0) {
430 debugs(20, 3, HERE
<< "wait for more info: " <<
432 return false; // may fit later, but will be rejected now
435 if (store_status
!= STORE_OK
) {
436 const int64_t maxKnownSize
= expectedEnd
< 0 ?
437 mem_obj
->availableForSwapOut() : expectedEnd
;
438 debugs(20, 7, HERE
<< "maxKnownSize= " << maxKnownSize
);
439 if (maxKnownSize
< store_maxobjsize
) {
441 * NOTE: the store_maxobjsize here is the max of optional
442 * max-size values from 'cache_dir' lines. It is not the
443 * same as 'maximum_object_size'. By default, store_maxobjsize
444 * will be set to -1. However, I am worried that this
445 * deferance may consume a lot of memory in some cases.
446 * Should we add an option to limit this memory consumption?
448 debugs(20, 5, HERE
<< "Deferring swapout start for " <<
449 (store_maxobjsize
- maxKnownSize
) << " bytes");
455 if (mem_obj
->inmem_lo
> 0) {
456 debugs(20, 3, "storeSwapOut: (inmem_lo > 0) imem_lo:" << mem_obj
->inmem_lo
);
457 decision
= MemObject::SwapOut::swImpossible
;
462 * If there are DISK clients, we must write to disk
463 * even if its not cachable
464 * RBC: Surely we should not create disk client on non cacheable objects?
465 * therefore this should be an assert?
466 * RBC 20030708: We can use disk to avoid mem races, so this shouldn't be
469 * XXX: Not clear what "mem races" the above refers to, especially when
470 * dealing with non-cachable objects that cannot have multiple clients.
472 * XXX: If STORE_DISK_CLIENT needs SwapOut::swPossible, we have to check
473 * for that flag earlier, but forcing swapping may contradict max-size or
474 * other swapability restrictions. Change storeClientType() and/or its
475 * callers to take swap-in availability into account.
477 for (node
= mem_obj
->clients
.head
; node
; node
= node
->next
) {
478 if (((store_client
*) node
->data
)->getType() == STORE_DISK_CLIENT
) {
479 debugs(20, 3, HERE
<< "DISK client found");
480 decision
= MemObject::SwapOut::swPossible
;
485 if (!mem_obj
->isContiguous()) {
486 debugs(20, 3, "storeSwapOut: not Contiguous");
487 decision
= MemObject::SwapOut::swImpossible
;
491 decision
= MemObject::SwapOut::swPossible
;