Messages in this thread |  | | | Date | Wed, 30 Sep 2009 09:14:02 -0700 (PDT) | | From | Linus Torvalds <> | | Subject | Re: Linux 2.6.32-rc1 |
| |
On Wed, 30 Sep 2009, Eric Dumazet wrote: > > lock; cmpxchg8b (%%esi) > > gives 4 bytes opcode : f0 0f c7 0e > Because alternative (call cmpxchg8b_emu) uses 5 bytes, a nop will be added. > > Choosing ".byte 0xf0, 0x0f, 0xc7, 0x4e, 0x00" aka "lock cmpxchg8b 0x0(%esi)" is a litle bit better ?
And if you want to be really clever, you would want to handle the non-SMP case too, where you have just "cmpxchgb (%%esi)" (three bytes) without the lock prefix.
However, at this point I think Arjan's patch is already way superior to what we have now (feel free to take a look at what we generate on 32-bit without PAE today - just have a barf-bag handy), so all I'd really want is a few "tested-by"s to make me feel happier about it, and a few more people looking at the emulation routine to all say "ok, looks sane, ACK".
And at that point we can then either make "cmpxchg()" just do the 8-byte case natively, or just take your patch to change sched_clock.c to now use the no-longer-entirely-disgusting cmpxchg64().
Ingo - I suspect both those patches should just go through you. You do both x86 and scheduler, so I'll happily pull the end result.
Linus
|  |