2014-10-21 18:15:33

by Sasha Levin

Subject: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

hstate_sizelog() would shift left an int rather than long, triggering
undefined behaviour and passing an incorrect value when the requested
page size was more than 4GB, thus breaking >4GB pages.

Signed-off-by: Sasha Levin <[email protected]>
---
include/linux/hugetlb.h | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
index 65e12a2..57e0dfd 100644
--- a/include/linux/hugetlb.h
+++ b/include/linux/hugetlb.h
@@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
{
if (!page_size_log)
return &default_hstate;
- return size_to_hstate(1 << page_size_log);
+
+ return size_to_hstate(1UL << page_size_log);
}

static inline struct hstate *hstate_vma(struct vm_area_struct *vma)
--
1.7.10.4

2014-10-22 05:42:51

by Andrey Ryabinin

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

On 10/21/2014 10:15 PM, Sasha Levin wrote:
> hstate_sizelog() would shift left an int rather than long, triggering
> undefined behaviour and passing an incorrect value when the requested
> page size was more than 4GB, thus breaking >4GB pages.

>
> Signed-off-by: Sasha Levin <[email protected]>
> ---
> include/linux/hugetlb.h | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> index 65e12a2..57e0dfd 100644
> --- a/include/linux/hugetlb.h
> +++ b/include/linux/hugetlb.h
> @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
> {
> if (!page_size_log)
> return &default_hstate;
> - return size_to_hstate(1 << page_size_log);
> +
> + return size_to_hstate(1UL << page_size_log);

That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.


> }
>
> static inline struct hstate *hstate_vma(struct vm_area_struct *vma)
>

2014-10-22 18:44:39

by Andrew Morton

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

On Wed, 22 Oct 2014 09:42:46 +0400 Andrey Ryabinin <[email protected]> wrote:

> On 10/21/2014 10:15 PM, Sasha Levin wrote:
> > hstate_sizelog() would shift left an int rather than long, triggering
> > undefined behaviour and passing an incorrect value when the requested
> > page size was more than 4GB, thus breaking >4GB pages.
>
> >
> > Signed-off-by: Sasha Levin <[email protected]>
> > ---
> > include/linux/hugetlb.h | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> > index 65e12a2..57e0dfd 100644
> > --- a/include/linux/hugetlb.h
> > +++ b/include/linux/hugetlb.h
> > @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
> > {
> > if (!page_size_log)
> > return &default_hstate;
> > - return size_to_hstate(1 << page_size_log);
> > +
> > + return size_to_hstate(1UL << page_size_log);
>
> That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.
>

But

struct hstate *size_to_hstate(unsigned long size)

2014-10-22 18:50:41

by Sasha Levin

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

On 10/22/2014 02:44 PM, Andrew Morton wrote:
> On Wed, 22 Oct 2014 09:42:46 +0400 Andrey Ryabinin <[email protected]> wrote:
>
>> > On 10/21/2014 10:15 PM, Sasha Levin wrote:
>>> > > hstate_sizelog() would shift left an int rather than long, triggering
>>> > > undefined behaviour and passing an incorrect value when the requested
>>> > > page size was more than 4GB, thus breaking >4GB pages.
>> >
>>> > >
>>> > > Signed-off-by: Sasha Levin <[email protected]>
>>> > > ---
>>> > > include/linux/hugetlb.h | 3 ++-
>>> > > 1 file changed, 2 insertions(+), 1 deletion(-)
>>> > >
>>> > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
>>> > > index 65e12a2..57e0dfd 100644
>>> > > --- a/include/linux/hugetlb.h
>>> > > +++ b/include/linux/hugetlb.h
>>> > > @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
>>> > > {
>>> > > if (!page_size_log)
>>> > > return &default_hstate;
>>> > > - return size_to_hstate(1 << page_size_log);
>>> > > +
>>> > > + return size_to_hstate(1UL << page_size_log);
>> >
>> > That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.
>> >
> But
>
> struct hstate *size_to_hstate(unsigned long size)

True, but "(1 << page_size_log)" produces an integer rather than long because "1"
is an int and not long.

#include <stdio.h>

int main(void)
{
unsigned long a, b;

a = 1 << 32;
b = 1UL << 32;

printf("a: %lu b: %lu\n", a, b);
}


$ ./a.out
a: 0 b: 4294967296


With the patch, size_to_hstate() gets the unsigned long it expects.


Thanks,
Sasha

2014-10-22 19:04:13

by Andrey Ryabinin

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

2014-10-22 22:44 GMT+04:00 Andrew Morton <[email protected]>:
> On Wed, 22 Oct 2014 09:42:46 +0400 Andrey Ryabinin <[email protected]> wrote:
>
>> On 10/21/2014 10:15 PM, Sasha Levin wrote:
>> > hstate_sizelog() would shift left an int rather than long, triggering
>> > undefined behaviour and passing an incorrect value when the requested
>> > page size was more than 4GB, thus breaking >4GB pages.
>>
>> >
>> > Signed-off-by: Sasha Levin <[email protected]>
>> > ---
>> > include/linux/hugetlb.h | 3 ++-
>> > 1 file changed, 2 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
>> > index 65e12a2..57e0dfd 100644
>> > --- a/include/linux/hugetlb.h
>> > +++ b/include/linux/hugetlb.h
>> > @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
>> > {
>> > if (!page_size_log)
>> > return &default_hstate;
>> > - return size_to_hstate(1 << page_size_log);
>> > +
>> > + return size_to_hstate(1UL << page_size_log);
>>
>> That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.
>>
>
> But
>
> struct hstate *size_to_hstate(unsigned long size)
>

What's wrong? On 32 bits, if page_size_log >= 32, then (unsingned
long)(1ULL << page_size_log) will be truncated to 0. I guess it's ok.
size_to_hstate will just return NULL in that case.


--
Best regards,
Andrey Ryabinin

2014-10-22 20:13:10

by Andrew Morton

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

On Wed, 22 Oct 2014 14:50:22 -0400 Sasha Levin <[email protected]> wrote:

> On 10/22/2014 02:44 PM, Andrew Morton wrote:
> > On Wed, 22 Oct 2014 09:42:46 +0400 Andrey Ryabinin <[email protected]> wrote:
> >
> >> > On 10/21/2014 10:15 PM, Sasha Levin wrote:
> >>> > > hstate_sizelog() would shift left an int rather than long, triggering
> >>> > > undefined behaviour and passing an incorrect value when the requested
> >>> > > page size was more than 4GB, thus breaking >4GB pages.
> >> >
> >>> > >
> >>> > > Signed-off-by: Sasha Levin <[email protected]>
> >>> > > ---
> >>> > > include/linux/hugetlb.h | 3 ++-
> >>> > > 1 file changed, 2 insertions(+), 1 deletion(-)
> >>> > >
> >>> > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> >>> > > index 65e12a2..57e0dfd 100644
> >>> > > --- a/include/linux/hugetlb.h
> >>> > > +++ b/include/linux/hugetlb.h
> >>> > > @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
> >>> > > {
> >>> > > if (!page_size_log)
> >>> > > return &default_hstate;
> >>> > > - return size_to_hstate(1 << page_size_log);
> >>> > > +
> >>> > > + return size_to_hstate(1UL << page_size_log);
> >> >
> >> > That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.
> >> >
> > But
> >
> > struct hstate *size_to_hstate(unsigned long size)
>
> True, but "(1 << page_size_log)" produces an integer rather than long because "1"
> is an int and not long.

My point is that there's no point in using 1ULL because
size_to_hstate() will truncate it anyway.

2014-10-22 20:26:56

by Andrey Ryabinin

Subject: Re: [PATCH] mm, hugetlb: correct bit shift in hstate_sizelog

2014-10-23 0:13 GMT+04:00 Andrew Morton <[email protected]>:
> On Wed, 22 Oct 2014 14:50:22 -0400 Sasha Levin <[email protected]> wrote:
>
>> On 10/22/2014 02:44 PM, Andrew Morton wrote:
>> > On Wed, 22 Oct 2014 09:42:46 +0400 Andrey Ryabinin <[email protected]> wrote:
>> >
>> >> > On 10/21/2014 10:15 PM, Sasha Levin wrote:
>> >>> > > hstate_sizelog() would shift left an int rather than long, triggering
>> >>> > > undefined behaviour and passing an incorrect value when the requested
>> >>> > > page size was more than 4GB, thus breaking >4GB pages.
>> >> >
>> >>> > >
>> >>> > > Signed-off-by: Sasha Levin <[email protected]>
>> >>> > > ---
>> >>> > > include/linux/hugetlb.h | 3 ++-
>> >>> > > 1 file changed, 2 insertions(+), 1 deletion(-)
>> >>> > >
>> >>> > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
>> >>> > > index 65e12a2..57e0dfd 100644
>> >>> > > --- a/include/linux/hugetlb.h
>> >>> > > +++ b/include/linux/hugetlb.h
>> >>> > > @@ -312,7 +312,8 @@ static inline struct hstate *hstate_sizelog(int page_size_log)
>> >>> > > {
>> >>> > > if (!page_size_log)
>> >>> > > return &default_hstate;
>> >>> > > - return size_to_hstate(1 << page_size_log);
>> >>> > > +
>> >>> > > + return size_to_hstate(1UL << page_size_log);
>> >> >
>> >> > That still could be undefined on 32-bits. Either use 1ULL or reduce SHM_HUGE_MASK on 32bits.
>> >> >
>> > But
>> >
>> > struct hstate *size_to_hstate(unsigned long size)
>>
>> True, but "(1 << page_size_log)" produces an integer rather than long because "1"
>> is an int and not long.
>
> My point is that there's no point in using 1ULL because
> size_to_hstate() will truncate it anyway.
>

There is a point to use 1ULL
On 32-bit with size >= 32
(1UL << size) - undefined, so size_to_hstate() will truncate it to
undefined as well. E.g. It definitely won't be zero on x86.
While (1ULL << size) - is defined and size_to_hstate() will truncate it to zero.


--
Best regards,
Andrey Ryabinin