Re: BalancerPersist "relaxable" checks (was: svn commit: r1733283 - /httpd/httpd/branches/2.4.x/STATUS)

2016-03-02 Thread Yann Ylavic 
On Wed, Mar 2, 2016 at 11:47 PM, Yann Ylavic  wrote:
> On Wed, Mar 2, 2016 at 8:08 PM, Jim Jagielski  wrote:
>>
>> Hmmm... iirc, if there is a mismatch, the old files are
>> automagically removed.
>
> You are right, I was still influenced by PR58024 triggering AH02599
> (painful fix), but that was due to the missing unlink() semantic on
> Windows...
>
> I still think AH02599 can be triggered on unclean shutdown, with name
> based SHMs when the cleanup is not called (this has nothing to do with
> mod_proxy_hcheck, though).
>
> Couldn't we use anonymous SHMs for slotmem_shm(s) since we
> create/destroy them for each restart anyway?
> I don't think the apr_shm_attach() path in slotmem_create() should
> ever be hit (thanks to the caching of in 'globallistmem' of the
> process' created SHMs, and other processes should use slotmem_attach()
> anyway).
> Hmm, I may miss some Windows MPM thingy again...

Nevermind, we still need to attach() in children processes!

Re: BalancerPersist "relaxable" checks (was: svn commit: r1733283 - /httpd/httpd/branches/2.4.x/STATUS)

2016-03-02 Thread Yann Ylavic 
On Wed, Mar 2, 2016 at 8:08 PM, Jim Jagielski  wrote:
>
> Hmmm... iirc, if there is a mismatch, the old files are
> automagically removed.

You are right, I was still influenced by PR58024 triggering AH02599
(painful fix), but that was due to the missing unlink() semantic on
Windows...

I still think AH02599 can be triggered on unclean shutdown, with name
based SHMs when the cleanup is not called (this has nothing to do with
mod_proxy_hcheck, though).

Couldn't we use anonymous SHMs for slotmem_shm(s) since we
create/destroy them for each restart anyway?
I don't think the apr_shm_attach() path in slotmem_create() should
ever be hit (thanks to the caching of in 'globallistmem' of the
process' created SHMs, and other processes should use slotmem_attach()
anyway).
Hmm, I may miss some Windows MPM thingy again...

Re: BalancerPersist "relaxable" checks (was: svn commit: r1733283 - /httpd/httpd/branches/2.4.x/STATUS)

2016-03-02 Thread Jim Jagielski 

> On Mar 2, 2016, at 11:13 AM, Yann Ylavic  wrote:
> 
> re: mod_proxy_hcheck backport
> 
>>  ylavic: Looks like the changes on struct proxy_worker_shared would break
>>  startup with "BalancerPersist on" due to the strict checks on
>>  the sizes of existing slotmems (slotmem_create/attach)?
>> + jim: Yes, that is right (re: breakage)... this would be noted at
>> +  release.
> 
> Maybe we could have a "BalancerPersist try" option which would
> remove/reset the existing persistent files/SHMs when the sizes change
> (either corrupted files or this kind of upgrade change).
> 
> That could save the admin from having to locate the relevent
> files/SHMs (on disk/ipc/dev) and remove them, which is not obvious
> when you have multiple httpd instances running and only some of them
> need this administrative cleanage (the others keep running), since the
> names of the files/SHMs is not "configurable" (the ID based on
> servername/line/.. is finally ap_proxy_hashfunc()ed for practical
> reasons) and can hardly be related to an httpd instance...

Hmmm... iirc, if there is a mismatch, the old files are
automagically removed.