Messages in this thread |  | | | Date | Wed, 4 Feb 2009 23:25:43 +0100 | | From | Ingo Molnar <> | | Subject | Re: [git pull] timer fix |
| |
* Linus Torvalds <torvalds@linux-foundation.org> wrote:
> On Wed, 4 Feb 2009, Ingo Molnar wrote: > > > > Pavel Emelyanov (1): > > x86: fix hpet timer reinit for x86_64 > > > > > > arch/x86/kernel/hpet.c | 2 +- > > 1 files changed, 1 insertions(+), 1 deletions(-) > > > > diff --git a/arch/x86/kernel/hpet.c b/arch/x86/kernel/hpet.c > > index 64d5ad0..ec319d1 100644 > > --- a/arch/x86/kernel/hpet.c > > +++ b/arch/x86/kernel/hpet.c > > @@ -1075,7 +1075,7 @@ static void hpet_rtc_timer_reinit(void) > > hpet_t1_cmp += delta; > > hpet_writel(hpet_t1_cmp, HPET_T1_CMP); > > lost_ints++; > > - } while ((long)(hpet_readl(HPET_COUNTER) - hpet_t1_cmp) > 0); > > + } while ((long)(hpet_readl(HPET_COUNTER) - (u32)hpet_t1_cmp) > 0); > > This is bordering on not being correct.
yeah, i had to look twice. The only reason i left it that way was because i couldnt reproduce the problem and hpet is hellishly fragile and this patch was tested so i chickened out.
OTOH that fragility is partly because such constructs have piled up so you very much have a valid point ...
We'll clean this up. I've already added the clean 32-bit casts - which also has another advantage: it does not actually trust the hw to always return 32-bit values - it explicitly cuts to 32 bits and does signed arithmetics on that. Will also do the helper function cleanup to abstract the counter arithmetics away.
> In particular, think about when HPET_COUNTER or hpet_t1_cmp overflows in > 32 bits, and what you want to happen. If you do the subtract add test in > 64 bits, it will simply do the wrong thing. Think what happens if > hpet_t1_cmp is actually _larger_ than HPET_COUNTER, but overflowed in 32 > bits, and you're now looking at: > > (long) (0xffffffff - 0x00000001) > > which is actually > 0, so the thing will continue to loop INCORRECTLY. It > should have stopped (and _would_ have stopped on 32-bit x86).
yeah, allowing that to happen is just wrong.
Ingo
|  |