Messages in this thread | | | From | Andy Lutomirski <> | Date | Wed, 7 May 2014 08:27:09 -0700 | Subject | Re: [ABOMINATION] x86: Fast interrupt return to userspace |
| |
On Wed, May 7, 2014 at 4:14 AM, Ingo Molnar <mingo@kernel.org> wrote: > > * Andy Lutomirski <luto@amacapital.net> wrote: > >> > Whatever. I got enough profile data to say that it seems to have >> > cut 'iret' overhead by at least two thirds. So it may not *work*, >> > but from a "hey look, some random numbers" standpoint it is worth >> > playing with. >> >> :) >> >> Is there actual interest in turning something like this into a real >> patch? It would almost certainly have to default off and no one >> sane would ever use it except for special-purpose machines. > > The macro speedup looks rather impressive, and we've done ugly things > for far smaller speedups. > > But I don't think it should be a 'special mode'. It either is made to > work unconditionally and can be a prime speedup to be proud of in a > politely disgusted fashion, or we don't want the complexity (and > future bitrot) of some special switch.
The problem is that we'll break anything that expects to be able to use more than the specified 128-byte redzone. Doing that is currently safe as long as no signals are delivered.
I could imagine something like Go blowing up badly.
We might be able to get away with automatically disabling this thing if sigaltstack is enabled.
--Andy
| |