On Mon 03-07-17 09:30:35, Linus Torvalds wrote: > On Mon, Jul 3, 2017 at 8:49 AM, Michal Hocko <mho...@kernel.org> wrote: > > > > If you think this is worth pursuing in upstream, just let me know and I > > can polish it, add a patch for the man page and other things. > > Hmm. This doesn't look bad, except the bprm games there really look annoying. > > Also, I'm wondering whether this should be per-thread - conceptually > "expand_stack()" really is a thread thing. All callers are using > "current", although it's not always obvious. > > So I'm wondering if a slightly larger patch that simply made the > "limit" be an _argument_ to expand_stack() would clean up both of > these issues. The execve() use would simply pass in the stack limit, > and the fault users would pass in "current->expand_stack_limit". > > Again, I'm not sure how many people really use multiple GROW_DOWN > stacks for threading, but it's conceptually the right thing to do, so > I think conceptually this should be per-thread. And the fact that it > might clean up the execve() thing makes me think it's the right thing > to do. > > What do you think?
I am not sure about the per-thread vs. per mm part. If for nothing else, MAP_GROWSDOWN can be something else than the main thread stack which can be modified by all threads and then the semantic would be quite surprising if different threads had a different idea about the expansion. No? But an additional argument to expand_stack would surely clean things up a bit and will get rid of the ugly bprm part as well. I will think about it some more and then post the patch to linux-api to have a larger audience. Thanks! -- Michal Hocko SUSE Labs
