Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Thu, May 21, 2009 at 02:39:57PM -0400, Jeff Trawick wrote: On Wed, May 20, 2009 at 8:53 AM, Joe Orton jor...@redhat.com wrote: Given that the semantics of the options has changed, I don't think it's worth changing httpd to maintain any pretence of compile-time or run-time compatibility here. Any code using the OPT_* constants as exposed by mod_perl cannot work as expected any more. Is the change in semantics required to fix the bug, or is it simply the current implementation? Attaching my original analysis for security@ which hopefully answers that question ;) Having thought about this longer, I do agree that it would be reasonable to provide OPT_INCNOEXEC as a noop integer for back-compat, but, it turns out we're out of bits - allow_options_t is an unsigned char and we're using 2^0 through 2^7 already. :( The only available option is to #define OPT_INCNOEXEC to some bogus string or something; not sure I like that much better than just a clean break. Worth noting: for any mod_perl-based code which tests only OPT_INCLUDE for is SSI enabled, that will continue to be compatible with the new implementation, modulo mod_perl build failures. The only issue is with code which needs to differentiate between SSI-with-exec and -without. Regards, Joe
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 05:26:07PM +0100, Joe Orton wrote: Attaching my original analysis for security@ which hopefully answers that question ;) attempt 2 I've now had a deeper look into this. I can't see a way to fix the problem without changing the semantics of the OPT_ bits used, as I mentioned briefly in my comment to Vincent. Status quo: a) OPT_INCLUDES is interpreted as SSI is allowed with exec= b) OPT_INCNOEXEC is interperted iff OPT_INCLUDES is also set as meaning SSI is allowed but exec= is not c) setting AllowOverride Options=IncludesNoExec results in both OPT_INCLUDES and OPT_INCNOEXEC being set in the -override_opts bitmask, i.e. either or both options can be overridden in .htaccess files From this leads the fact that an .htaccess file can set simply Options Includes in a context which inherits AllowOverride Options=IncludesNoExec. I'm presuming nobody will argue that's a feature not a bug? If so, I think this is the set of constraints which need to be satisfied: 1) the result of a config merge with only Options IncludesNoEXEC specified must not allow use of exec= in SSI 2) if AllowOverride Options= is used without Includes, notably, use of AllowOverride Options=IncludesNoExec, use of Options Includes in an .htaccess file must be an error 3) if AllowOverride Options=Includes is set, use of both Options Includes and Options IncludesNoExec must succeed and enable SSI with or without exec= respectively 4) if permitted by AllowOverride, setting Options Includes in a context from which Options IncludesNoExec is inherited, then the result must be one where exec= is allowed. Attached is a patch which passes the tests I have so far - Vincent, can you easily re-run your tests used to produce that lovely matrix, with this applied? Regards, Joe
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Thu, May 21, 2009 at 3:25 PM, Jeff Trawick traw...@gmail.com wrote: On Thu, May 21, 2009 at 3:08 PM, William A. Rowe, Jr. wr...@rowe-clan.net wrote: Jeff Trawick wrote: Does somebody else care to share their opinion on this? Which of these are okay? - existing mod_perl releases (and potentially other third-party modules) won't compile with 2.2.12 CORE_PRIVATE may be broken from release to release, it's a necessary concession to prevent utter stagnation :( The bits are not CORE_PRIVATE. You can find sample Perl code on the web that even tests these bits, though it isn't clear to me if that is a normal practice when using the Perl/mod_include interface. Hmmm, after trying to use what seems like a cool feature, I find that mod_perl was never taught to use the Apache 2's mod_include plug-in interface.
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 10:17 AM, Jeff Trawick traw...@gmail.com wrote: On Thu, May 21, 2009 at 3:25 PM, Jeff Trawick traw...@gmail.com wrote: On Thu, May 21, 2009 at 3:08 PM, William A. Rowe, Jr. wr...@rowe-clan.net wrote: Jeff Trawick wrote: Does somebody else care to share their opinion on this? Which of these are okay? - existing mod_perl releases (and potentially other third-party modules) won't compile with 2.2.12 CORE_PRIVATE may be broken from release to release, it's a necessary concession to prevent utter stagnation :( The bits are not CORE_PRIVATE. You can find sample Perl code on the web that even tests these bits, though it isn't clear to me if that is a normal practice when using the Perl/mod_include interface. Hmmm, after trying to use what seems like a cool feature, I find that mod_perl was never taught to use the Apache 2's mod_include plug-in interface. Jeff can you post the sample code and some details to d...@perl.apache.org?
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri 22 May 2009, Jeff Trawick wrote: Hmmm, after trying to use what seems like a cool feature, I find that mod_perl was never taught to use the Apache 2's mod_include plug-in interface. AFAIK, that is provided by Geoff's CPAN module Apache::IncludeHook or so. Torsten -- Need professional mod_perl support? Just hire me: torsten.foert...@gmx.net
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
Joe Orton wrote: Having thought about this longer, I do agree that it would be reasonable to provide OPT_INCNOEXEC as a noop integer for back-compat, but, it turns out we're out of bits - allow_options_t is an unsigned char and we're using 2^0 through 2^7 already. :( The C langauge promotes char - int for comparison. 256 should work fine, no? It would devolve to 0, of course, but 256 255 should test fine. Thoughts? The only available option is to #define OPT_INCNOEXEC to some bogus string or something; not sure I like that much better than just a clean break.
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 2:46 PM, Torsten Foertsch torsten.foert...@gmx.netwrote: On Fri 22 May 2009, Jeff Trawick wrote: Hmmm, after trying to use what seems like a cool feature, I find that mod_perl was never taught to use the Apache 2's mod_include plug-in interface. AFAIK, that is provided by Geoff's CPAN module Apache::IncludeHook or so. Neat, and the use of [SSI_]FLAG_NO_EXEC in the filter context isn't affected by this change. FWIW, it looks like it won't work with httpd 2.2 (IncludeHook.xs uses FLAG_NO_EXEC instead of SSI_FLAG_NO_EXEC) unless there's some compile-time mapping going on somewhere that I don't see.
How to use modules?
My script has a use POSIX; in it, which is filling my error log with several hundred error lines: Constant subroutine ModPerl ::ROOT ::ModPerl ::PerlRunPrefork ::home_oraweb_perl_badgers_makebadges_2epl::_POSIX_CHILD_MAX redefined at /opt/lampp/lib/perl5/site_perl/5.10.0/i686-linux/ModPerl/Util.pm line 69. What's the proper way to do this? -- Bruce Johnson University of Arizona College of Pharmacy Information Technology Group Institutions do not have opinions, merely customs
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 2:59 PM, William A. Rowe, Jr. wr...@rowe-clan.netwrote: Joe Orton wrote: Having thought about this longer, I do agree that it would be reasonable to provide OPT_INCNOEXEC as a noop integer for back-compat, but, it turns out we're out of bits - allow_options_t is an unsigned char and we're using 2^0 through 2^7 already. :( The C langauge promotes char - int for comparison. 256 should work fine, no? It would devolve to 0, of course, but 256 255 should test fine. Thoughts? Backing up a bit... I originally thought we could map bit values in 2.2.x to avoid affecting modules, but that isn't possible since includes-with-exec is two bits instead of one. Mapping OPT_INCNOEXEC to a no-op integer is something that takes place at compile time, and helps applications which reference the symbol but don't use it in any important way. (IOW, let mod_perl and other similar tarballs compile.) It is good in that it lets mod_perl compile, but bad in that mod_perl continues to export the Perl mapping of OPT_INCNOEXEC even after httpd has been upgraded and at some point later mod_perl is upgraded. Failing the compile is our only opportunity to catch some affected modules (though it is a rather late opportunity since the modules will likely be rebuilt later since they're supposed to work as-is when upgrading httpd; somebody will grumble though). I don't think we should try to preserve compilability if we can't preserve compatibility. The only available option is to #define OPT_INCNOEXEC to some bogus string or something; not sure I like that much better than just a clean break. /* * #define OPT_INCNOEXEC 32 * Apache 2.2.12 and later no longer provide this. * Applications which distinguish between includes-without-exec and includes-with-exec * must use different logic for 2.2.12 and 2.2.=12. * Prior to 2.2.12: * includes-without-exec: OPT_INCNOEXEC flag on, OPT_INCLUDES flag off * includes-with-exec: OPT_INCNOEXEC flag off, OPT_INCLUDES flag on * As of 2.2.12: * includes-without-exec: OPT_INCLUDES flag on, OPT_INC_WITH_EXEC flag off * includes-with-exec: OPT_INCLUDES flag on, OPT_INC_WITH_EXEC flag on * */
Re: How to use modules?
On Fri, May 22, 2009 at 4:20 PM, Bruce Johnson john...@pharmacy.arizona.edu wrote: My script has a use POSIX; in it, which is filling my error log with several hundred error lines: Constant subroutine ModPerl::ROOT::ModPerl::PerlRunPrefork::home_oraweb_perl_badgers_makebadges_2epl::_POSIX_CHILD_MAX redefined at /opt/lampp/lib/perl5/site_perl/5.10.0/i686-linux/ModPerl/Util.pm line 69. You don't need to import all those constants. use POSIX (); - Perrin
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 4:21 PM, Jeff Trawick traw...@gmail.com wrote: On Fri, May 22, 2009 at 2:59 PM, William A. Rowe, Jr. wr...@rowe-clan.net wrote: Joe Orton wrote: Having thought about this longer, I do agree that it would be reasonable to provide OPT_INCNOEXEC as a noop integer for back-compat, but, it turns out we're out of bits - allow_options_t is an unsigned char and we're using 2^0 through 2^7 already. :( The C langauge promotes char - int for comparison. 256 should work fine, no? It would devolve to 0, of course, but 256 255 should test fine. Thoughts? Backing up a bit... I originally thought we could map bit values in 2.2.x to avoid affecting modules, but that isn't possible since includes-with-exec is two bits instead of one. Mapping OPT_INCNOEXEC to a no-op integer is something that takes place at compile time, and helps applications which reference the symbol but don't use it in any important way. (IOW, let mod_perl and other similar tarballs compile.) It is good in that it lets mod_perl compile, but bad in that mod_perl continues to export the Perl mapping of OPT_INCNOEXEC even after httpd has been upgraded and at some point later mod_perl is upgraded. Failing the compile is our only opportunity to catch some affected modules (though it is a rather late opportunity since the modules will likely be rebuilt later since they're supposed to work as-is when upgrading httpd; somebody will grumble though). I don't think we should try to preserve compilability if we can't preserve compatibility. The only available option is to #define OPT_INCNOEXEC to some bogus string or something; not sure I like that much better than just a clean break. /* * #define OPT_INCNOEXEC 32 * Apache 2.2.12 and later no longer provide this. * Applications which distinguish between includes-without-exec and includes-with-exec * must use different logic for 2.2.12 and 2.2.=12. * Prior to 2.2.12: * includes-without-exec: OPT_INCNOEXEC flag on, OPT_INCLUDES flag off oops, both flags were on here * includes-with-exec: OPT_INCNOEXEC flag off, OPT_INCLUDES flag on * As of 2.2.12: * includes-without-exec: OPT_INCLUDES flag on, OPT_INC_WITH_EXEC flag off * includes-with-exec: OPT_INCLUDES flag on, OPT_INC_WITH_EXEC flag on * */
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
Jeff Trawick wrote: Backing up a bit... I originally thought we could map bit values in 2.2.x to avoid affecting modules, but that isn't possible since includes-with-exec is two bits instead of one. Hold on... I think this can still work; * Retain new true 'Includes' bit as old IncludesNoExec macro value Keep ancient Includes flag bit as 256, never true. - all httpd modules testing for including but not executing permission see the permission as allowed - old httpd modules testing for includes with exec permission see the permission as denied, until they update the module - httpd modules which force/override the includes without exec permission would still work - httpd modules which force/override the includes exec behavior would just fail to update anything (256 0xff == 00), so it becomes a noop until they update the module So it has no negative security consequences, still would require an update to the rare module, but lets us ship something without really nasty side effects.
Re: svn commit: r773881 - in /httpd/httpd/branches/2.2.x: CHANGES STATUS include/http_core.h modules/filters/mod_include.c server/config.c server/core.c
On Fri, May 22, 2009 at 5:10 PM, William A. Rowe, Jr. wr...@rowe-clan.netwrote: Jeff Trawick wrote: Backing up a bit... I originally thought we could map bit values in 2.2.x to avoid affecting modules, but that isn't possible since includes-with-exec is two bits instead of one. Hold on... I think this can still work; * Retain new true 'Includes' bit as old IncludesNoExec macro value Keep ancient Includes flag bit as 256, never true. - all httpd modules testing for including but not executing permission see the permission as allowed - old httpd modules testing for includes with exec permission see the permission as denied, until they update the module - httpd modules which force/override the includes without exec permission would still work - httpd modules which force/override the includes exec behavior would just fail to update anything (256 0xff == 00), so it becomes a noop until they update the module So it has no negative security consequences, still would require an update to the rare module, but lets us ship something without really nasty side effects. I'll think harder about this once my latest proposal gets shot down ;)