2016-06-08 16:14 GMT+02:00 Vacelet, Manuel <manuel.vace...@enalean.com>:
> On Tue, Jun 7, 2016 at 11:02 PM, Luca Toscano <toscano.l...@gmail.com> > wrote: > >> >> >> 2016-06-07 10:55 GMT+02:00 Vacelet, Manuel <manuel.vace...@enalean.com>: >> >>> >>> >>> On Mon, Jun 6, 2016 at 5:32 PM, Vacelet, Manuel < >>> manuel.vace...@enalean.com> wrote: >>> >>>> dOn Mon, Jun 6, 2016 at 5:00 PM, Vacelet, Manuel < >>>> manuel.vace...@enalean.com> wrote: >>>> >>>>> On Mon, Jun 6, 2016 at 4:09 PM, Luca Toscano <toscano.l...@gmail.com> >>>>> wrote: >>>>> >>>>>> >>>>>> I was able to repro building httpd from 2.4.x branch and following >>>>>> your configuration files on github. I am almost sure that somewhere httpd >>>>>> sets the Last-Modified header translating "foo" to the first Jan 1970 >>>>>> date. >>>>>> I realized though that I didn't recall the real issue, since passing >>>>>> value >>>>>> not following the RFC can lead to inconsistencies, so I went back and >>>>>> checked the correspondence. Quoting: >>>>>> >>>>>> "Actually I wrote this snippet to highlight the behaviour (the >>>>>> original code sent the date in iso8601 instead of rfc1123) because it was >>>>>> more obvious. >>>>>> During my tests (this is extracted from an automated test suite), >>>>>> even after having converted dates to rfc1123, I continued to get some >>>>>> sparse errors. What I got is that the value I sent was sometimes slightly >>>>>> modified (a second or 2) depending on the machine load." >>>>>> >>>>>> So my understanding is that you would like to know why a >>>>>> Last-Modified header with a legitimate date/time set by a PHP app gets >>>>>> "delayed" by a couple of seconds from httpd, right? >>>>>> >>>>> >>>>> Yes for sure, this is the primary issue. >>>>> However, the (undocumented) difference of behavior from one version to >>>>> another (2.2 -> 2.4 and more surprisingly from between two 2.4 versions) >>>>> is >>>>> also in question here. >>>>> Even more strange, 2.4 built for other distrib doesn't highlight the >>>>> behaviour ! >>>>> >>>>> >>>> >>>> I made another series of test and it seems to be linked to fastcgi. >>>> >>>> I took the stock apache (2.4.6 plus tons of patches) & php-fpm (5.4.16 >>>> + tons of patches) from RHEL7 and I get the exact same behaviour (headers >>>> rewritten to EPOCH) >>>> However, if I server the very same php script from mod_php (instead of >>>> fcgi) it "works" (the headers are not modified). >>>> >>>> >>> For the record, I also have the same behaviour (headers rewritten when >>> using php-fpm + fastcgi) on alpine linux 3.4 that ships apache2-2.4.20. >>> So AFAICT, it doesn't seem distro specific. >>> >>> On the root of the problem, from my point of view: >>> - the difference between mod_php vs. php-fpm + fcgi is understandable >>> (even if not desired and not documented). >>> - the fact that fcgi handler parse & rewrite headers seems to lead to >>> inconsistencies (I'll try to build a test case for that). >>> - however, even if the headers are wrong, I think apache default (use >>> EPOCH) is wrong as it leads to very inconsistent behaviour (the resource >>> will never expire). I would prefer either: >>> -- do not touch the header >>> -- raise a warning and discard the header >>> >>> What do you think ? >>> >> >> >> From my tests the following snippet of code should be responsible for the >> switch from 'foo' to unix epoch: >> >> *https://github.com/apache/httpd/blob/2.4.x/server/util_script.c#L663 >> <https://github.com/apache/httpd/blob/2.4.x/server/util_script.c#L663>* >> >> The function that contains the code, ap_scan_script_header_err_core_ex, >> is wrapped by a lot of other functions eventually called by modules like >> mod-proxy-fcgi. A more verbose description of the function in: >> >> https://github.com/apache/httpd/blob/2.4.x/include/util_script.h#L200 >> >> Not sure what would be the best thing to do, but probably we could follow >> up in a official apache bugzilla task? >> https://bz.apache.org/bugzilla/enter_bug.cgi?product=Apache%20httpd-2 >> >> > Wow, thanks for the investigation ! > Sorry for the delay! I submitted a patch for trunk with a possible fix, namely dropping (and logging at trace1 level) any non compliant date/time set in a Last-Modified header returned by a FCGI/CGI script: http://svn.apache.org/r1748379 The fix is also in the list of proposal for backport to the 2.4.x branch, we'll see what other people think about this solution. We should also do a follow up for the other main issue, namely the fact that you see a different/delayed Last-Modified header sometimes among your FCGI/httpd responses. Can you give me an example of Last-Modified header value before/after the "delay" and a way to repro it? Thanks! Luca
