Re: [Qemu-devel] [Qemu-block] Limiting coroutine stack usage

[Daniel P . Berrangé]

On Thu, Feb 22, 2018 at 01:02:05PM +0100, Peter Lieven wrote:
> Am 22.02.2018 um 13:00 schrieb Daniel P. Berrangé:
> > On Thu, Feb 22, 2018 at 12:51:58PM +0100, Peter Lieven wrote:
> >> Am 22.02.2018 um 12:40 schrieb Daniel P. Berrangé:
> >>> On Thu, Feb 22, 2018 at 12:32:04PM +0100, Kevin Wolf wrote:
> >>>> Am 22.02.2018 um 12:01 hat Peter Lieven geschrieben:
> >>>>> Am 22.02.2018 um 11:57 schrieb Kevin Wolf:
> >>>>>> Am 20.02.2018 um 22:54 hat Paolo Bonzini geschrieben:
> >>>>>>> On 20/02/2018 18:04, Peter Lieven wrote:
> >>>>>>>> Hi,
> >>>>>>>>
> >>>>>>>> I remember we discussed a long time ago to limit the stack usage of 
> >>>>>>>> all
> >>>>>>>> functions that are executed in a coroutine
> >>>>>>>> context to a very low value to be able to safely limit the coroutine
> >>>>>>>> stack size as well.
> >>>>>>> IIRC the only issue was that hw/ide/atapi.c has mutual recursion 
> >>>>>>> between
> >>>>>>> ide_atapi_cmd_reply_end -> ide_transfer_start -> ahci_start_transfer 
> >>>>>>> ->
> >>>>>>> ide_atapi_cmd_reply_end.
> >>>>>>>
> >>>>>>> But perhaps it's not an issue, somebody needs to audit the code.
> >>>>>> I think John intended to get rid of the recursion sometime, but I doubt
> >>>>>> he has had the time so far.
> >>>>> Apart from this is is possible to define special cflags in the
> >>>>> Makefile.objs just for a subdirectory? I have patches ready to make
> >>>>> the block layer files and other coroutine users compile with
> >>>>> -Wstack-size=2048. But I do not want to specify each file separately.
> >>>> Our Makefiles have lines like this:
> >>>>
> >>>>     iscsi.o-cflags     := $(LIBISCSI_CFLAGS)
> >>>>
> >>>> I don't think there is a direct mechanism to apply cflags to a whole
> >>>> directory or just to block-obj-y/block-obj-m, but just looping over them
> >>>> could work. I'm not a Makefile expert at all, but after some toying with
> >>>> a simple example, something like this might work:
> >>>>
> >>>>     $(foreach x,$(block-obj-y),$(eval $x-cflags += -Wstack-size=2048))
> >>> You'll need it for anything block layer depends on too - so that's much
> >>> of util/, crypto/ and io/ directories at least.
> >>>
> >>> So perhaps it would be shorter if we do the opposite - set 
> >>> -Wstack-size=2048
> >>> globally for everything in QEMU, and then override -Wstack-size=$BIGGER
> >>> for the (hopefully) few sources that have a larger stack need ?
> >> I tried that already. 2048 is a strong limit for many functions.
> >> It breaks already as soon as some buffer has a size of PATH_MAX, but
> >> thats handleable. But there are some structs around that are very large.
> > There are surprisingly few "char [PATH_MAX]" variables left in QEMU - we
> > should have a final push to eliminate them regardless.
> >
> >> Generally, it would be a good idea to have a global limit, of course.
> > We could at least put a limit on that matches the current worst case to
> > prevent it getting worse than it already is.
> 
> That would be a good idea, yes.
> 
> How would you handle the override for a smaller -Wstack-usage ?

If you have multiple -Wstack-size=$XXX  flags to GCC, I expect the last
one wins. So just need to double check that the per-object file CFLAGS
occur after the global CFLAS in the compiler args


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|