lkml.org 
[lkml]   [2017]   [Dec]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] iommu/vt-d: Fix shift overflow in qi_flush_dev_iotlb
On Wed, Dec 13, 2017 at 11:31:02AM -0600, Hook, Gary wrote:
> On 12/13/2017 11:15 AM, Alex Williamson wrote:
> > On Wed, 13 Dec 2017 10:41:47 -0600
> > "Hook, Gary" <ghook@amd.com> wrote:
> >
> > > On 12/13/2017 9:58 AM, Alex Williamson wrote:
> > > > On Wed, 13 Dec 2017 15:13:55 +0800
> > > > Peter Xu <peterx@redhat.com> wrote:
> > > > > On Tue, Dec 12, 2017 at 03:43:08PM -0700, Alex Williamson wrote:
> > > > >
> > > > > [...]
> > > > > > diff --git a/drivers/iommu/dmar.c b/drivers/iommu/dmar.c
> > > > > > index 9a7ffd13c7f0..87888b102057 100644
> > > > > > --- a/drivers/iommu/dmar.c
> > > > > > +++ b/drivers/iommu/dmar.c
> > > > > > @@ -1345,7 +1345,9 @@ void qi_flush_dev_iotlb(struct intel_iommu *iommu, u16 sid, u16 qdep,
> > > > > > struct qi_desc desc;
> > > > > > if (mask) {
> > > > > > - BUG_ON(addr & ((1 << (VTD_PAGE_SHIFT + mask)) - 1));
> > > > > > + BUG_ON((mask > MAX_AGAW_PFN_WIDTH) ||
> > > > > > + ((mask == MAX_AGAW_PFN_WIDTH) && addr) ||
> > > > > > + (addr & ((1 << (VTD_PAGE_SHIFT + mask)) - 1)));
> > > > >
> > > > > Could it work if we just use 1ULL instead of 1 here? Thanks,
> > > >
> > > > In either case we're talking about shifting off the end of the
> > > > variable, which I understand to be undefined. Right? Thanks,
> > >
> > > How so? Bits fall off the left (MSB) end, zeroes fill in the right (LSB)
> > > end. I believe that behavior is pretty set.
> >
> > Maybe I'm relying too much on stackoverflow, but:
> >
> > https://stackoverflow.com/questions/11270492/what-does-the-c-standard-say-about-bitshifting-more-bits-than-the-width-of-type
>
> No, probably not. I don't have my copy of c99 handy, so can't check it. But
> it is beyond me why any compiler implementation would choose to use a rotate
> instead of a shift... probably a performance issue.
>
> So, yeah, when you have silly parameters, you get what you get.
>
> I'll stick to my suggestion. Which seems unambiguous... but I could be
> wrong.

Hi, Alex, Hook,

I did a quick test:

xz-mi:tmp $ cat a.c
#include <stdio.h>

void main(void)
{
unsigned long long val = 0x8000000000000001ULL;
int shift;

printf("origin: 0x%llx\n", val);
shift = 1; printf("shl 1: 0x%llx\n", val << shift);
shift = 62; printf("shl 62: 0x%llx\n", val << shift);
shift = 63; printf("shl 63: 0x%llx\n", val << shift);
shift = 64; printf("shl 64: 0x%llx\n", val << shift);
shift = 65; printf("shl 65: 0x%llx\n", val << shift);
}
xz-mi:tmp $ gcc -std=c99 a.c
xz-mi:tmp $ ./a.out
origin: 0x8000000000000001
shl 1: 0x2
shl 62: 0x4000000000000000
shl 63: 0x8000000000000000
shl 64: 0x8000000000000001
shl 65: 0x2
xz-mi:tmp $ objdump -d a.out | grep -A20 "<main>"
00000000004004d7 <main>:
4004d7: 55 push %rbp
4004d8: 48 89 e5 mov %rsp,%rbp
4004db: 48 83 ec 10 sub $0x10,%rsp
4004df: 48 b8 01 00 00 00 00 movabs $0x8000000000000001,%rax
4004e6: 00 00 80
4004e9: 48 89 45 f8 mov %rax,-0x8(%rbp)
4004ed: 48 8b 45 f8 mov -0x8(%rbp),%rax
4004f1: 48 89 c6 mov %rax,%rsi
4004f4: bf 60 06 40 00 mov $0x400660,%edi
4004f9: b8 00 00 00 00 mov $0x0,%eax
4004fe: e8 ed fe ff ff callq 4003f0 <printf@plt>
400503: c7 45 f4 01 00 00 00 movl $0x1,-0xc(%rbp)
40050a: 8b 45 f4 mov -0xc(%rbp),%eax
40050d: 48 8b 55 f8 mov -0x8(%rbp),%rdx
400511: 89 c1 mov %eax,%ecx
400513: 48 d3 e2 shl %cl,%rdx
400516: 48 89 d0 mov %rdx,%rax
400519: 48 89 c6 mov %rax,%rsi
40051c: bf 70 06 40 00 mov $0x400670,%edi
400521: b8 00 00 00 00 mov $0x0,%eax

So it seems not really a rotation operation, but it looks more like
convering a "shifting N" into "shifting N%64" when N>=64. So now I
agree with Alex's change. Thanks all.

--
Peter Xu

\
 
 \ /
  Last update: 2017-12-15 08:31    [W:0.045 / U:0.120 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site