Hello Alex,
Thank you for sharing your thoughts.
The A-example looks easier to implement and more reliable.
Kind regards,
Ankor.
вт, 11 мар. 2025 г. в 17:12, Alex Rousskov <rousskov@xxxxxxxxxxxxxxxxxxxxxxx>:
On 2025-03-10 23:56, Andrey K wrote:
> > Alex: FWIW, related future Squid improvements may include:
> > * Detecting such shared memory segments clashes; refusing to start.
> > * Disabling shared memory use when caching is completely disabled.
> But ... segments may remain from the previous
> Squid crash. Thus, in my opinion, it is impractical to refuse to launch
> the program.
"Squid should overwrite segments left by a crashed Squid" does not imply
that "detecting segment clashes among independent Squid instances is
impractical".
The following rough examples are _not_ implementation blueprints, but
they illustrate the lack of the above implication:
A. If Squid were to incorporate PID filename[^1] into segment names,
then segments from different Squid instances will not clash.
B. If Squid were to write its PID filename[^1] into a shared memory
segment, then Squid would be able to detect that another running
instance is using the same shared memory segment (while still
overwriting segments left by an earlier crash).
[^1]: Full Squid PID filename are already guaranteed to be unique across
Squid instances that may overwrite each other segments because Squid
refuses to start if another Squid instance is using its PID file. The
only red flag I can think of here is esoteric chrooted instances, but I
would not be surprised if those instances cannot share segments.
> I think it's enough to put a warning in the code when detecting old
> segments.
I disagree that a level-0 warning during regular/uneventful startups is
a good idea, especially if there will be one warning for every segment.
We should try to address the actual problem (i.e. refuse to start a
segment-clashing concurrent instance) rather than allowing clashes but
adding a yet another warning about a usually benign event.
Alex.
> Something like that:
> void
> Ipc::Mem::Segment::create(const off_t aSize)
> {
> assert(aSize > 0);
> assert(theFD < 0);
>
> int xerrno = 0;
>
> // Why a brand new segment? A Squid crash may leave a reusable
> segment, but
> // our placement-new code requires an all-0s segment. We could
> truncate and
> // resize the old segment, but OS X does not allow using O_TRUNC with
> // shm_open() and does not support ftruncate() for old segments.
> if (!createFresh(xerrno) && xerrno == EEXIST) {
> *debugs(54, DBG_CRITICAL, "WARNING: there is an old shared memory
> segment: '" << theName << "'. Perhaps it was left after the previous
> crash of the Squid. We will remove it. Or it may be a sign that another
> instance of the Squid is running. In this case, you must launch the
> program with the -n option and specify the unique name of the service.");*
> unlink();
> createFresh(xerrno);
> }
> Kind regards,
> Ankor.
>
>
> сб, 8 мар. 2025 г. в 08:14, Andrey K <ankor2023@xxxxxxxxx
> <mailto:ankor2023@xxxxxxxxx>>:
>
> Hello Alex,
>
> Thanks for the analysis.
> Squidonlyallowsalphanumericcharactersinthe servicename,
> soitrefusesto usethe originalservice namein the -n option (-n
> squid.user.service).
> I added-n squiduser option to the ExecStart string of the second
> instance.
> Now it looks good:
>
> # the first instance
> lsof -p 3746105 | grep shm
> squid 3746105 root mem REG 0,23 525572
> 1213614834 /dev/shm/squid-cf__queues.shm
> squid 3746105 root mem REG 0,23 136
> 1213614835 /dev/shm/squid-cf__readers.shm
> squid 3746105 root mem REG 0,23 8
> 1213614833 /dev/shm/squid-cf__metadata.shm
> squid 3746105 root 7u REG 0,23 8
> 1213614833 /dev/shm/squid-cf__metadata.shm
> squid 3746105 root 8u REG 0,23 525572
> 1213614834 /dev/shm/squid-cf__queues.shm
> squid 3746105 root 9u REG 0,23 136
> 1213614835 /dev/shm/squid-cf__readers.shm
>
> # the second instance
> lsof -p 3685356 | grep shm
> squid.use 3685356 root mem REG 0,23 2093368
> 1212704041 /dev/shm/squiduser-tls_session_cache.shm
> squid.use 3685356 root mem REG 0,23 525572
> 1212704039 /dev/shm/squiduser-cf__queues.shm
> squid.use 3685356 root mem REG 0,23 136
> 1212704040 /dev/shm/squiduser-cf__readers.shm
> squid.use 3685356 root mem REG 0,23 8
> 1212704038 /dev/shm/squiduser-cf__metadata.shm
> squid.use 3685356 root 6u REG 0,23 8
> 1212704038 /dev/shm/squiduser-cf__metadata.shm
> squid.use 3685356 root 7u REG 0,23 525572
> 1212704039 /dev/shm/squiduser-cf__queues.shm
> squid.use 3685356 root 8u REG 0,23 136
> 1212704040 /dev/shm/squiduser-cf__readers.shm
> squid.use 3685356 root 9u REG 0,23 2093368
> 1212704041 /dev/shm/squiduser-tls_session_cache.shm
>
> Kind regards,
> Ankor.
>
>
>
>
>
> пт, 7 мар. 2025 г. в 17:48, Alex Rousskov
> <rousskov@xxxxxxxxxxxxxxxxxxxxxxx
> <mailto:rousskov@xxxxxxxxxxxxxxxxxxxxxxx>>:
>
> On 2025-03-07 06:50, Andrey K wrote:
>
> > Squid Cache: Version 6.13
> > Service Name: squid
>
> > Squid Cache: Version 6.10
> > Service Name: squid
>
> > # the first instance
> > 1318 DEL ... 30205 /dev/shm/squid-cf__queues.shm
> > 1318 DEL ... 30206 /dev/shm/squid-cf__readers.shm
> > 1318 DEL ... 30204 /dev/shm/squid-cf__metadata.shm
> > 1318 8u ... 8 30204 /dev/shm/squid-cf__metadata.shm
> (deleted)
> > 1318 9u ... 525572 30205 /dev/shm/squid-cf__queues.shm
> (deleted)
> > 1318 10u ... 136 30206 /dev/shm/squid-cf__readers.shm
> (deleted)
> >
> > # the second instance
> > 1514 mem ... 2093368 31497
> /dev/shm/squid-tls_session_cache.shm
> > 1514 mem ... 525572 31495 /dev/shm/squid-cf__queues.shm
> > 1514 mem ... 136 31496 /dev/shm/squid-cf__readers.shm
> > 1514 mem ... 8 31494 /dev/shm/squid-cf__metadata.shm
> > 1514 6u ... 8 31494 /dev/shm/squid-cf__metadata.shm
> > 1514 7u ... 525572 31495 /dev/shm/squid-cf__queues.shm
> > 1514 8u ... 136 31496 /dev/shm/squid-cf__readers.shm
> > 1514 9u ... 2093368 31497
> /dev/shm/squid-tls_session_cache.shm
>
> As suspected, these two Squid instances use the same shared memory
> segments (e.g., /dev/shm/squid-cf*). Such shared use violates
> critical
> code assumptions and results in undefined behavior.
>
>
> > Maybe I'm not experiencing any difficulties because I have
> caching turned off on
> > both instances?
>
> Well, you _are_ experiencing at least one difficulty -- the
> assertion
> that started this email thread. If you have fully disabled
> caching, the
> difficulties you experience should not include cache corruption.
> However, it looks like at least one of the two instances does
> cache TLS
> sessions (in /dev/shm/squid-tls_session_cache.shm). If both
> instances do
> that, then all bets are off!
>
>
> FWIW, related future Squid improvements may include:
>
> * Detecting such shared memory segments clashes; refusing to start.
> * Disabling shared memory use when caching is completely disabled.
>
> Quality pull requests welcome.
>
>
> Cheers,
>
> Alex.
>
>
> > чт, 6 мар. 2025 г. в 17:11, Alex Rousskov:
> >
> > On 2025-03-06 08:59, Amos Jeffries wrote:
> > > On 6/03/25 19:17, Andrey K wrote:
> > >> Hello,
> > >>
> > >> I have a similar configuration: two SMP squids
> running on the
> > same OEL
> > >> host.
> > >>
> > >> They were built with different configurations: with
> different
> > >> installation path prefixes and different names of
> binary files:
> > squid
> > >> and squid.user and they listen to different ports.
> > >> They are launched from two different
> services:squid.service and
> > >> squid.user.service with the service Type=forking:
> > >>
> > >> ExecStart=/usr/sbin/squid -sYC
> > >> ExecStart=/sbin/squid.user -f
> /etc/squid.user/squid.conf
> > >>
> > >> I have not experienced any troubles with this
> configuration yet.
> > >>
> > >> /> Please be aware that "squid -n ..." is a
> REQUIREMENT for running/
> > >> /multiple Squid instances on the same machine
> regardless of what
> > features
> > >> are used./
> > >>
> > >> Could you please tell me if I should use the -n
> option in the
> > >> ExecStart strings?
> > >> The arguments of the options should be the service names?
> > >>
> > >> ExecStart=/usr/sbin/squid -sYC -n squid.service
> > >> ExecStart=/sbin/squid.user -f
> /etc/squid.user/squid.conf -n
> > >> squid.user.service
> > >>
> > > Yes you should. The different ./configure options has
> helped you
> > avoid
> > > major issues, but some may still appear.
> >
> > I agree. Moreover, I do not understand how your two SMP
> Squids could
> > work correctly without distinct service names because (on
> OEL) I would
> > expect them to share the same shared memory segments
> (which they must
> > not do to remain distinct instances).
> >
> > What is your Squid version? Can you tell how your Squids
> name their
> > shared memory segment "files"? For example, on some Linux
> OSes, those
> > segments could be in /var/run/shm/ with names like
> > squid-tr_map_anchors.shm and squid-tr_spaces.shm.
> >
> >
> > Thank you,
> >
> > Alex.
> >
> > _______________________________________________
> > squid-users mailing list
> > squid-users@xxxxxxxxxxxxxxxxxxxxx
> <mailto:squid-users@xxxxxxxxxxxxxxxxxxxxx>
> > <mailto:squid-users@xxxxxxxxxxxxxxxxxxxxx
> <mailto:squid-users@xxxxxxxxxxxxxxxxxxxxx>>
> > https://lists.squid-cache.org/listinfo/squid-users
> <https://lists.squid-cache.org/listinfo/squid-users>
> > <https://lists.squid-cache.org/listinfo/squid-users
> <https://lists.squid-cache.org/listinfo/squid-users>>
> >
>
_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
https://lists.squid-cache.org/listinfo/squid-users
_______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx https://lists.squid-cache.org/listinfo/squid-users
