[Tarantool-patches] [PATCH v2 2/2] fiber: exit with panic if we unable to revert guard page

Cyrill Gorcunov gorcunov at gmail.com
Tue Feb 4 19:08:00 MSK 2020


On Tue, Feb 04, 2020 at 06:47:42PM +0300, Konstantin Osipov wrote:
> > --- a/src/lib/core/fiber.c
> > +++ b/src/lib/core/fiber.c
> > @@ -1041,13 +1041,17 @@ fiber_stack_destroy(struct fiber *fiber, struct slab_cache *slabc)
> >  			 * to setup the original protection back in
> >  			 * background.
> >  			 *
> > +			 * For now lets exit with panic: if mprotect
> > +			 * failed we must not allow to reuse such slab
> > +			 * with PROT_NONE'ed page somewhere inside.
> > +			 *
> 
> somewhere inside its stack area would be more clear.

Thanks!

> >  			 * Note that in case if we're called from
> >  			 * fiber_stack_create() the @mprotect_flags is
> >  			 * the same as the slab been created with, so
> >  			 * calling mprotect for VMA with same flags
> >  			 * won't fail.
> >  			 */
> > -			diag_log();
> > +			panic_syserror("fiber: Can't put guard page to slab");
> 
> While the patch itself is LGTM, we need to nail down the cause of
> the failure even at the cost of crash, I suspect what we're getting
> here is ENOMEM from the kernel. I suspect we have too many
> mprotect regions and the kernel runs out of some internal
> resources for them.

Yes, seems we might get out of VMA when there are too many fibers
and system doesn't have enough memory to continue splitting.

> I think adding better diagnostics could help us identify the
> issue: the failed address, its slab, the total number of fibers (and by
> induction mprotected pages). 
> 
> I have also discussed the issue with @xemul, and he suggests that 
> wrong slab alignment could be causing this.

True, and we need investigate it. Once we have the "fix" merged
in I'll file a bug to investigate this ideas. Thanks a huge, Kostya!

> 
> Finally, we could try to clear mprotect() first, and if it fails,
> avoid destroying the fiber and keep it cached for a while more.
> We could retry destroying it when kernel has more memory.

Yes. I put FIXME for such intelligent exit (I thought about
stack slabs only not the complete fibers though).

> 
> These are of course ideas for follow ups.


More information about the Tarantool-patches mailing list