#2558 apc_fetch() memory leak
Submitted: 2004-10-18 14:18 UTC
From: xuefer at 21cn dot com Assigned: rasmus
Status: Closed Package:
PHP Version: Irrelevant OS: linux
Roadmaps: (Not assigned)    
Subscription  


 [2004-10-18 14:18 UTC] xuefer at 21cn dot com
Description: ------------ echo "<pre>": var_dump(apc_sma_info()); $k = "a...................................................."; apc_store($k, 1); for ($i = 0; $i < 10000; $i ++) { apc_fetch($k); } var_dump(apc_sma_info()); reload the page 2+ times to see the result i guess it's the key memory leak

Comments

 [2004-10-19 02:29 UTC] xuefer at 21cn dot com
i've move filename from entry to key.data, renamed as identifier now both file/user have "char *" which have exactly same behavior and function(usage) then, i merge union struct into key struct(just make it simple) copy pointer instead of apc_xstrdup in apc_make_*key identifier but do apc_xstrdup in make_slot, free it in free_slot everything goes fine, including my APC_USE_INODE seems have no core dump again the merge of filename&identifyer make it easy to inode free
 [2004-10-26 20:49 UTC] gschlossnagle at php dot net
Can you post a patch,please?
 [2004-10-27 01:22 UTC] xuefer at 21cn dot com
sure, i was hunting for space to upload ftp://211.92.88.40/pub/moo/apc_win32.patch (with other bug fixed) apc_cache.c apc_cache.h: fixed for APC_USE_INODE the struct of key is changed, maybe u don't agree. never do strdup for key before insert!(make_slot) "find" on the cache is more frequent than insert apc_compile.c: don't care about builtin_functions apc_fcntl.c: we have 2 locks but why only 1 file used? apc_mmap.c: mmap depends on shm? "#ifdef shm_open" works? maybe u can make marco in config.m4 apc_sem.c: remove redefine warnning, maybe wrong fix apc_shm.c: IPC_PRIVATE of TSRM seems buggy, the 2nd process will not able to create the shm other changes: for apc_cache apc_user_cache, don't put it in GLOBALS ------------ ftp://211.92.88.40/pub/moo/apc_optimizer.c.patch some of this patch is posted in another bug report this time have a new fix: function rewrite_const_cast - zval_dtor(&cur->op1.u.constant); + /*zval_dtor(&cur->op1.u.constant);*/ i never used (string)'1' but phpmyadmin do, so i found this corrupt when i test with win32 (which i have installed phpmyadmin) although i modified constant_fold, it's still not fixed :( optimizer is written by cute gschlossnagle right? i saw a stantalone-but-very-old version in cvs/pecl :P maybe u can have a review on it any question or complain about the patch is welcome
 [2004-10-27 01:29 UTC] xuefer at 21cn dot com
"fixed" for APC_USE_INODE should be: "improved" for APC_USE_INODE btw: this "const_cast" make phpmyadmin fine but before "const_cast" fix(and after my other optimizer fix), it's still quite stable it's used in production server(100MB traffic), only quite a few core dumps(for builtin_function in apc_compile) maybe 1 dump for 4months i'd suggest make apc.optimizer to NOT EXPERIMENTAL
 [2004-10-27 18:30 UTC] rasmus
I can't get to that ftp site. Could you email me the patches directly? (rasmus@php.net)
 [2004-10-28 01:48 UTC] xuefer at 21cn dot com
host was down and is up now. i sent many emails to u but get no reply except the request for confirm. i confirmed it and no response from u(or server)
 [2004-10-28 17:42 UTC] rasmus
I have committed a variation of your apc_fetch memory leak fix. You changed all sorts of other stuff in your supplied patches and we will look at these separately. One bug per report please.