Re: [Qemu-devel] [PATCH 03/15] coroutine-ucontext: reduce stack size to 64kB

[Peter Lieven]

Am 28.06.2016 um 16:20 schrieb Dr. David Alan Gilbert:
> * Peter Lieven (address@hidden) wrote:
> > Am 28.06.2016 um 13:35 schrieb Dr. David Alan Gilbert:
> > > * Peter Lieven (address@hidden) wrote:
> > > > Am 28.06.2016 um 12:57 schrieb Dr. David Alan Gilbert:
> > > > > * Paolo Bonzini (address@hidden) wrote:
> > > > > > On 28/06/2016 11:01, Peter Lieven wrote:
> > > > > > > evaluation with the recently introduced maximum stack size monitoring revealed
> > > > > > > that the actual used stack size was never above 4kB so allocating 1MB stack
> > > > > > > for each coroutine is a lot of wasted memory. So reduce the stack size to
> > > > > > > 64kB which should still give enough head room.
> > > > > > If we make the stack this much smaller, there is a non-zero chance of
> > > > > > smashing it.  You must add a guard page if you do this (actually more
> > > > > > than one because QEMU will happily have stack frames as big as 16 KB).
> > > > > > The stack counts for RSS but it's not actually allocated memory, so why
> > > > > > does it matter?
> > > > > I think I'd be interested in seeing the /proc/.../smaps before and after this
> > > > > change to see if anything is visible and if we can see the difference
> > > > > in rss etc.
> > > > Can you advise what in smaps should be especially looked at.
> > > >
> > > > As for RSS I can report hat the long term usage is significantly lower.
> > > > I had the strange observation that when the VM is running for some minutes
> > > > the RSS suddenly increases to the whole stack size.
> > > You can see the Rss of each mapping; if you knew where your stacks were
> > > it would be easy to see if it was the stacks that were Rss and if
> > > there was anything else odd about them.
> > > If you set hte mapping as growsdown then you can see the area that has a 'gd'
> > > in it's VmFlags.
> > Would you expect to see each 1MB allocation in smaps or is it possible that
> > the kernel merges some mappings to bigger ones?
> >
> > And more importantly if the regions are merged Paolos comment about we
> > do not need a guard page would not be true because a coroutine stack could
> > grow into annother coroutines stack. Looking at the commit from Linus it
> > would also be good to have that guard page not having the gd flag.
> Hmm I'm not sure; one for Paolo.

My fault. The second mmap call with the pointer to the stack must carry
the MAP_FIXED flag.

Peter