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

[Stefan Hajnoczi]

On Wed, Mar 07, 2018 at 09:36:38PM +0100, Peter Lieven wrote:
> Am 06.03.2018 um 12:51 schrieb Stefan Hajnoczi:
> > On Tue, Feb 20, 2018 at 06:04:02PM +0100, Peter Lieven wrote:
> >> 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.
> >>
> >> I checked through all functions in block/, migration/ and nbd/ and there 
> >> are only very few larger or unbound stack
> >> allocations that can easily be fixed.
> >>
> >> Now my question: Is there an easy way to add a cflag like 
> >> -Wstack-usage=2048 to all objects in a given directory only?
> >> I tried to add a llimit to the whole project, but fixing this will be a 
> >> larger task.
> > 2KB is fine for QEMU code but actual coroutine stack sizes will have to
> > be at least 8KB, I guess, in order for third-party libraries to work
> > (e.g. curl, rbd).  PATH_MAX is 4KB on Linux.
> >
> > Nested event loops in QEMU code can also result in deep call stacks.
> > This happens when aio_poll() invokes an fd handler or BH that also
> > invokes aio_poll().
> 
> The plan was to limit the stack usage only as a compiler option. I would 
> leave the coroutine stack size at 1MB
> for now until we have a way to identify the worst case usage.

I'm not sure we'll be able to confidently set a small stack size, but a
compile-time check doesn't hurt.

Maybe someday we'll switch to stackless coroutines.  Clang now has
coroutines support and maybe gcc will get it too (for C++20).  If the
compiler transforms the code then coroutine state can be kept in a
smaller data area that is not a call stack.  A single per-thread call
stack is used to run any coroutine and only during coroutine execution.

Stefan

signature.asc
Description: PGP signature