On Friday 16 June 2006 17:23, Perrin Harkins wrote: > Dave Rolsky wrote: > > Apache::SizeLimit has a long-standing bug on Linux where it never > > actually kills a child, because of Perl's caching of ppid info > > internally, and the way this interacts with Apache's forking. > > > > See this thread for details: > > http://mail-archives.apache.org/mod_mbox/perl-modperl/200505.mbox/%3C51EE > >[EMAIL PROTECTED] > > You're saying that it always thinks it is in the main process even > though it's actually in a child? I haven't seen this behavior. Is it > specific to certain versions of Perl or Linux?
Yes, since 5.8.1. The problem is that the now old linux-threads implementation returns different values for getpid() and getppid() for different threads. Hence Perl decides better to cache these values and update the cache on fork() to make it more compatible with other thread implementations. Since Apache does its own fork the cache is not updated. Mp2 contains code that handles the situation but mp1 does not. For more information see also http://www.gossamer-threads.com/lists/perl/porters/193162 http://www.gossamer-threads.com/lists/perl/porters/175806 Dave, have you tried the proposed approach from the thread you mentioned? > PerlModule Perl::AfterFork > PerlModule Apache::Constants > PerlInitHandler "sub {Perl::AfterFork::reinit(); return Apache::OK;}" > > or > > PerlModule Perl::AfterFork > PerlInitHandler "sub {Perl::AfterFork::reinit();}" If you are working on Apache::SizeLimit maybe you could incorporate also Linux::Smaps for newer kernels. Also the problem of counting out of core pages could be addressed at least in the docs, see http://www.gossamer-threads.com/lists/modperl/dev/84375 Torsten
pgp09H0VNlyeb.pgp
Description: PGP signature
