Re: [linux-pm] Freezer: Don't count threads waiting for frozenfilesystems.

[Alan Stern]

On Wed, 29 Oct 2008, Miklos Szeredi wrote:

> Not all callbacks.  I don't know what the current model is but AFAIR
> it should be something like this:
> 
>   1) call drivers to prepare for suspend (allocate space, etc)
>   2) stop all driver activity (plug queues, disable interrupts, etc)
>   3) call drivers to actually save state and power down
>   4) suspend
> 
> The part we are concerned is stopping driver activity.  It could be
> done with a mutex, or it could be done by freezing tasks.  Adding a
> mutex or other mechanism is the one I most like, but it's probably the
> biggest work, so lets look at how to fix the freezing:

Not only is adding a mutex the biggest amount of work, it has has the 
largest impact.  Every I/O pathway would have to acquire the 
appropriate mutex.  That's a significant additional load on the system.

> Currently the criteria for freezing is that userspace task has to exit
> kernelspace, and kernel task has to hit a specific "freeze point".
> This causes problems where we want to freeze tasks which are "stuck"
> inside filesystems or other non-driver parts of the kernel.  We can
> fix this two ways:
> 
>   a) mark additional places to freeze for userspace tasks as
>   well. This is the direction Nigel seems to be taking.
> 
>   b) or instead we could allow freezing anywhere in uninterruptible
>   sleep, _except_ where explicity marked.
> 
> Which is easier?  I don't know.  But I very storgly feel that marking
> un-freezable places instead of marking freezable places is a much
> cleaner solution.  It only affects parts of the kernel which have
> something to do with suspend, instead of affecting parts of the kernel
> which have absolutely nothing to do with suspend.

The problem with unrestricted freezing shows up when you freeze tasks
that hold a mutex or other sort of lock.  If this mutex is needed later
on for suspending a device then the suspend will hang, because a frozen
task can't release any mutexes.

I suppose you could try to categorize mutexes as "freezable" and 
"non-freezable".  Ones used by device drives would generally be 
non-freezable, whereas others (such as those used by VFS) would be 
freezable.  Still, it would be pretty difficult.  Among other things, 
it would be necessary to verify that a task holding a non-freezable 
mutex never tries to acquire a freezable mutex.

Alan Stern

--
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/