Re: [PATCH 2.6.19 5/5] fs: freeze_bdev with semaphore not mutex

On Friday, 10 November 2006 00:32, Pavel Machek wrote:
Hi!

On Fri, Nov 10, 2006 at 12:11:46AM +0100, Pavel Machek wrote:
? Not sure if I quite understand, but if dm breaks sync... something
is teribly wrong with dm. And we do simple sys_sync()... so I do not
think we have a problem.

If you want to handle arbitrary kernel state, you might have a device-mapper
device somewhere lower down the stack of devices that is queueing any I/O
that reaches it. So anything waiting for I/O completion will wait until
the dm process that suspended that device has finished whatever it is doing
- and that might be a quick thing carried out by a userspace lvm tool, or
a long thing carried out by an administrator using dmsetup.

I'm guessing you need a way of detecting such state lower down the stack
then optionally either aborting the operation telling the user it can't be
done at present; waiting for however long it takes (perhaps for ever if
the admin disappeared); or more probably skipping those devices on a
'best endeavours' basis.

Okay, so you claim that sys_sync can stall, waiting for administator?

In such case we can simply do one sys_sync() before we start freezing
userspace... or just more the only sys_sync() there. That way, admin
has chance to unlock his system.

Well, this is a different story.

My point is that if we call sys_sync() _anyway_ before calling
freeze_filesystems(), then freeze_filesystems() is _safe_ (either the
sys_sync() blocks, or it doesn't in which case freeze_filesystems() won't
block either).

This means, however, that we can leave the patch as is (well, with the minor
fix I have already posted), for now, because it doesn't make things worse a
bit, but:
(a) it prevents xfs from being corrupted and
(b) it prevents journaling filesystems in general from replaying journals
after a failing resume.

Still, there is a problem with the possibility of potential lock-up - either
with the bdevs-freezing patch or without it - due to a suspended dm device
down the stack and solving that is a _separate_ issue.

Now if we use the userland suspend, there's no problem at all, I think,
because s2disk calls sync() before it goes to suspend_system(), so the
admin will have a chance to unclock the system and everything is fine and
dandy (although it should be documented somewhere, IMHO).

However, if the built-in swsusp is used, then well ... <looks> ... we can put
a call to sys_sync() before prepare_processes() in pm_suspend_disk().

Greetings,
Rafael


--
You never change things by fighting the existing reality.
R. Buckminster Fuller
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Relevant Pages

  • Re: pthread cond signal vs broadcast
    ... But assuming more than one thread can be waiting ... So choosing signal over broadcast is more of an optimazation. ... if I have a stack and waiters are waiting for a non-empty ...
    (comp.programming.threads)
  • RE: Bluetooth suspend resume
    ... hardware notification that the device is gone away. ... communication with the software stack was needed. ... | I have a very annoying problem with bluetooth, ... | What can I do to solve this problem (API calls before suspend ...
    (microsoft.public.windowsce.platbuilder)
  • sniffing out a hung disk subsystem, Re: 5.0.6 grinds to a complete halt
    ... > disk-susbsystem hang would reflect in how ose 5.0.6 behaved. ... > like a process that I can hopefully identify as disk IO, ... is waiting on. ... [shows the stack of a process that isn't hung] ...
    (comp.unix.sco.misc)
  • Re: pthread cond signal vs broadcast
    ... If more than one thread could be waiting on the condition, ... when the predicate actually changes. ... if I have a stack and waiters are waiting for a non-empty ...
    (comp.programming.threads)
  • Re: Opinions, please..
    ... "I would say you're waiting for him to bust out. ... about to eat a BB and SB for 20 percent of his stack. ... Calculating implied odds and watching the blinds is how you need to ...
    (rec.gambling.poker)