1 Feb 2006 22:41
Re: [RFC][PATCH -mm][Experimental] swsusp: freeze userspace processes first
Nigel Cunningham <nigel <at> suspend2.net>
2006-02-01 21:41:43 GMT
2006-02-01 21:41:43 GMT
Hi.
On Wednesday 01 February 2006 22:49, Pavel Machek wrote:
> Hi!
>
> > > > This is an experimantal patch aimed at the "unable to freeze
> > > > processes under load" problem.
> > > >
> > > > On my box the 2.6.16-rc1-mm4 kernel with this patch applied survives
> > > > the "dd if=/dev/hda of=/dev/null" test.
> > > >
> > > > Please have a look.
> > >
> > > It makes it better (well, I used my own, simpler variant, but that
> > > should not matter; patch is attached). I now can't reproduce hangs
> > > with simple stress testing, but running kernel make alongside that
> > > makes it hang sometimes. Example of non-frozen gcc:
> > >
> > > gcc D EEE06A70 0 1750 1749 1751
> > > (NOTLB)
> > > df85df38 00000046 bf878130 eee06a70 00004111 eee06a70 eee06a70
> > > 003d0900
> > > 00000000 c0137cf5 df85c000 00000000 c058ada2 c012503e ef2c915c
> > > ef2c9030
> > > c1c0b480 7c3b8500 003d0927 df85c000 00000a98 7c3b8500 003d0927
> > > c0770800
> > > Call Trace:
> > > [<c0137cf5>] attach_pid+0x25/0xb0
> > > [<c058ada2>] _write_unlock_irq+0x12/0x30
> > > [<c012503e>] copy_process+0xe5e/0x11b0
> > > [<c0588f74>] wait_for_completion+0x94/0xd0
> > > [<c0121690>] default_wake_function+0x0/0x10
> > > [<c01254d9>] do_fork+0x149/0x210
> > > [<c0101218>] sys_vfork+0x28/0x30
> > > [<c0103231>] syscall_call+0x7/0xb
> > >
> > > ...maybe solving this would solve journalling problems, too? It is
> > > similar AFAICT.
> >
> > What exactly is the journalling problem?
>
> Hangs by freezing everything at same time only happen with journalling
> filesystems; there kjournald needs to be running if we want user
> threads to be stoppable.
Ah. Right. That should be fixed by doing the kernelspace threads after the
usespace ones (as I believe you're now doing?). I have seen XFS still
submitting I/O after the sys_sync is finished (it apparently treats sys_sync
as a weak and useless indication that it should think about considering
flushing a buffer or two). That's why I'm now using bdev freezing instead of
sys_sync.
> > > <at> <at> -87,7 +87,6 <at> <at> static int prepare_processes(void)
> > > int error;
> > >
> > > pm_prepare_console();
> > > - sys_sync();
> > > disable_nonboot_cpus();
> > >
> > > if (freeze_processes()) {
> >
> > That will help speed up freezing, but it won't help the integrity of your
> > data if you don't resume.
>
> See the patch better; it is now done between freezing userspace and
> kernel threads.
Ah. So I see. Sorry.
> > > /*
> > > * Timeout for stopping processes
> > > */
> > > -#define TIMEOUT (6 * HZ)
> > > +#define TIMEOUT (60 * HZ)
> >
> > You're kidding, right?
>
> sync takes long time... and 6 seconds were not enough to deliver
> signals on highly-loaded ext2.
I'm using a timeout per step. Perhaps you could try that approach? 1 minute is
an awfully long time to wait if you do hang.
Regards,
Nigel
--
See our web page for Howtos, FAQs, the Wiki and mailing list info.
http://www.suspend2.net IRC: #suspend2 on Freenode
Hi.
On Wednesday 01 February 2006 22:49, Pavel Machek wrote:
> Hi!
>
> > > > This is an experimantal patch aimed at the "unable to freeze
> > > > processes under load" problem.
> > > >
> > > > On my box the 2.6.16-rc1-mm4 kernel with this patch applied survives
> > > > the "dd if=/dev/hda of=/dev/null" test.
> > > >
> > > > Please have a look.
> > >
> > > It makes it better (well, I used my own, simpler variant, but that
> > > should not matter; patch is attached). I now can't reproduce hangs
> > > with simple stress testing, but running kernel make alongside that
> > > makes it hang sometimes. Example of non-frozen gcc:
> > >
> > > gcc D EEE06A70 0 1750 1749 1751
> > > (NOTLB)
> > > df85df38 00000046 bf878130 eee06a70 00004111 eee06a70 eee06a70
> > > 003d0900
> > > 00000000 c0137cf5 df85c000 00000000 c058ada2 c012503e ef2c915c
> > > ef2c9030
> > > c1c0b480 7c3b8500 003d0927 df85c000 00000a98 7c3b8500 003d0927
> > > c0770800
> > > Call Trace:
> > > [<c0137cf5>] attach_pid+0x25/0xb0
> > > [<c058ada2>] _write_unlock_irq+0x12/0x30
> > > [<c012503e>] copy_process+0xe5e/0x11b0
> > > [<c0588f74>] wait_for_completion+0x94/0xd0
> > > [<c0121690>] default_wake_function+0x0/0x10
> > > [<c01254d9>] do_fork+0x149/0x210
> > > [<c0101218>] sys_vfork+0x28/0x30
> > > [<c0103231>] syscall_call+0x7/0xb
> > >
> > > ...maybe solving this would solve journalling problems, too? It is
> > > similar AFAICT.
> >
> > What exactly is the journalling problem?
>
> Hangs by freezing everything at same time only happen with journalling
> filesystems; there kjournald needs to be running if we want user
> threads to be stoppable.
Ah. Right. That should be fixed by doing the kernelspace threads after the
usespace ones (as I believe you're now doing?). I have seen XFS still
submitting I/O after the sys_sync is finished (it apparently treats sys_sync
as a weak and useless indication that it should think about considering
flushing a buffer or two). That's why I'm now using bdev freezing instead of
sys_sync.
> > > <at> <at> -87,7 +87,6 <at> <at> static int prepare_processes(void)
> > > int error;
> > >
> > > pm_prepare_console();
> > > - sys_sync();
> > > disable_nonboot_cpus();
> > >
> > > if (freeze_processes()) {
> >
> > That will help speed up freezing, but it won't help the integrity of your
> > data if you don't resume.
>
> See the patch better; it is now done between freezing userspace and
> kernel threads.
Ah. So I see. Sorry.
> > > /*
> > > * Timeout for stopping processes
> > > */
> > > -#define TIMEOUT (6 * HZ)
> > > +#define TIMEOUT (60 * HZ)
> >
> > You're kidding, right?
>
> sync takes long time... and 6 seconds were not enough to deliver
> signals on highly-loaded ext2.
I'm using a timeout per step. Perhaps you could try that approach? 1 minute is
an awfully long time to wait if you do hang.
Regards,
Nigel
--
--
See our web page for Howtos, FAQs, the Wiki and mailing list info.
http://www.suspend2.net IRC: #suspend2 on Freenode
RSS Feed