Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On Thu, Jul 30, 2020 at 11:09 PM Luben Tuikov wrote:
> On 2020-07-29 9:49 a.m., Alex Deucher wrote:
> > On Wed, Jul 29, 2020 at 4:11 AM Christian König
> > wrote:
> >>
> >> Am 28.07.20 um 21:29 schrieb Peilin Ye:
> >>> Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> >>> amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
> >>> when `size` is greater than 356.
> >>>
> >>> In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
> >>> unfortunately does not initialize that 4-byte hole. Fix it by using
> >>> memset() instead.
> >>>
> >>> Cc: sta...@vger.kernel.org
> >>> Fixes: c193fa91b918 ("drm/amdgpu: information leak in
> >>> amdgpu_info_ioctl()")
> >>> Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> >>> Suggested-by: Dan Carpenter
> >>> Signed-off-by: Peilin Ye
> >>
> >> Reviewed-by: Christian König
> >>
> >> I can't count how many of those we have fixed over the years.
> >>
> >> At some point we should probably document that using "= {}" or "= { 0 }"
> >> in the kernel is a really bad idea and should be avoided.
> >
> > Moreover, it seems like different compilers seem to behave relatively
> > differently with these and we often get reports of warnings with these
> > on clang. When in doubt, memset.
>
> There are quite a few of those under drivers/gpu/drm, for "amd/", "scheduler/"
> drm*.c files,
>
> $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ | wc
> -l
> 374
> $_
>
> Out of which only 16 are of the non-ISO C variety, "= {}",
>
> $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
> 16
> $_
That is an unrelated issue, those were introduced to deal with older compilers
that do not accept '{0}' as an initializer for an aggregate whose
first member is
another aggregate. Generally speaking, '= { }' is better to use in the
kernel than
'= { 0 }' because all supported compilers interpret that the same way for all
structures.
Arnd
___
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
Am 31.07.20 um 09:10 schrieb Greg Kroah-Hartman:
On Fri, Jul 31, 2020 at 08:57:53AM +0200, Christian König wrote:
Am 31.07.20 um 08:53 schrieb Greg Kroah-Hartman:
On Thu, Jul 30, 2020 at 05:09:07PM -0400, Luben Tuikov wrote:
On 2020-07-29 9:49 a.m., Alex Deucher wrote:
On Wed, Jul 29, 2020 at 4:11 AM Christian König
wrote:
Am 28.07.20 um 21:29 schrieb Peilin Ye:
Compiler leaves a 4-byte hole near the end of `dev_info`, causing
amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
when `size` is greater than 356.
In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
unfortunately does not initialize that 4-byte hole. Fix it by using
memset() instead.
Cc: sta...@vger.kernel.org
Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
Suggested-by: Dan Carpenter
Signed-off-by: Peilin Ye
Reviewed-by: Christian König
I can't count how many of those we have fixed over the years.
At some point we should probably document that using "= {}" or "= { 0 }"
in the kernel is a really bad idea and should be avoided.
Moreover, it seems like different compilers seem to behave relatively
differently with these and we often get reports of warnings with these
on clang. When in doubt, memset.
There are quite a few of those under drivers/gpu/drm, for "amd/", "scheduler/"
drm*.c files,
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex "./scheduler/.*\.c" \)
-exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ | wc -l
374
$_
Out of which only 16 are of the non-ISO C variety, "= {}",
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex "./scheduler/.*\.c" \)
-exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
16
$_
Perhaps the latter are the more pressing ones, since it is a C++ initializer
and not a ISO C one.
It only matters when we care copying the data to userspace, if it all
stays in the kernel, all is fine.
Well only as long as you don't try to compute a CRC32, MD5 or any
fingerprint for a hash from the bytes from the structure.
Then it fails horrible and you wonder why the code doesn't works as
expected.
True, but the number of times I have ever needed to do that to a
structure for a driver is, um, never...
If a structure ever needs to have that happen to it, I would sure hope
the developer was aware of padding fields, otherwise, well, someone
needs to take away their C language certification :)
Well it is very likely that stack allocated structures have the same
values in the padding bytes most of the time. So the problem is very
subtle and hard to detect.
We've seen enough problems with that over the last ~10 years that I'm
clearly in favor of adding something to checkpatch.pl to spill out a
warning if "= { }" is used for zero initialization.
Alternatively some of the people who know gcc/clang better than I do
could come up with a warning that you shouldn't cast a structure with
uninitialized padding to void* or u8*.
I mean KASAN is already doing a great job detecting that kind of stuff,
but for this you still need to hit the offending code path.
Thanks,
Christian.
thanks,
greg k-h
___
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On Fri, Jul 31, 2020 at 08:57:53AM +0200, Christian König wrote:
> Am 31.07.20 um 08:53 schrieb Greg Kroah-Hartman:
> > On Thu, Jul 30, 2020 at 05:09:07PM -0400, Luben Tuikov wrote:
> > > On 2020-07-29 9:49 a.m., Alex Deucher wrote:
> > > > On Wed, Jul 29, 2020 at 4:11 AM Christian König
> > > > wrote:
> > > > > Am 28.07.20 um 21:29 schrieb Peilin Ye:
> > > > > > Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> > > > > > amdgpu_info_ioctl() to copy uninitialized kernel stack memory to
> > > > > > userspace
> > > > > > when `size` is greater than 356.
> > > > > >
> > > > > > In 2015 we tried to fix this issue by doing `= {};` on `dev_info`,
> > > > > > which
> > > > > > unfortunately does not initialize that 4-byte hole. Fix it by using
> > > > > > memset() instead.
> > > > > >
> > > > > > Cc: sta...@vger.kernel.org
> > > > > > Fixes: c193fa91b918 ("drm/amdgpu: information leak in
> > > > > > amdgpu_info_ioctl()")
> > > > > > Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> > > > > > Suggested-by: Dan Carpenter
> > > > > > Signed-off-by: Peilin Ye
> > > > > Reviewed-by: Christian König
> > > > >
> > > > > I can't count how many of those we have fixed over the years.
> > > > >
> > > > > At some point we should probably document that using "= {}" or "= { 0
> > > > > }"
> > > > > in the kernel is a really bad idea and should be avoided.
> > > > Moreover, it seems like different compilers seem to behave relatively
> > > > differently with these and we often get reports of warnings with these
> > > > on clang. When in doubt, memset.
> > > There are quite a few of those under drivers/gpu/drm, for "amd/",
> > > "scheduler/"
> > > drm*.c files,
> > >
> > > $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> > > "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ |
> > > wc -l
> > > 374
> > > $_
> > >
> > > Out of which only 16 are of the non-ISO C variety, "= {}",
> > >
> > > $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> > > "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
> > > 16
> > > $_
> > >
> > > Perhaps the latter are the more pressing ones, since it is a C++
> > > initializer and not a ISO C one.
> > It only matters when we care copying the data to userspace, if it all
> > stays in the kernel, all is fine.
>
> Well only as long as you don't try to compute a CRC32, MD5 or any
> fingerprint for a hash from the bytes from the structure.
>
> Then it fails horrible and you wonder why the code doesn't works as
> expected.
True, but the number of times I have ever needed to do that to a
structure for a driver is, um, never...
If a structure ever needs to have that happen to it, I would sure hope
the developer was aware of padding fields, otherwise, well, someone
needs to take away their C language certification :)
thanks,
greg k-h
___
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On Thu, Jul 30, 2020 at 05:09:07PM -0400, Luben Tuikov wrote:
> On 2020-07-29 9:49 a.m., Alex Deucher wrote:
> > On Wed, Jul 29, 2020 at 4:11 AM Christian König
> > wrote:
> >>
> >> Am 28.07.20 um 21:29 schrieb Peilin Ye:
> >>> Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> >>> amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
> >>> when `size` is greater than 356.
> >>>
> >>> In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
> >>> unfortunately does not initialize that 4-byte hole. Fix it by using
> >>> memset() instead.
> >>>
> >>> Cc: sta...@vger.kernel.org
> >>> Fixes: c193fa91b918 ("drm/amdgpu: information leak in
> >>> amdgpu_info_ioctl()")
> >>> Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> >>> Suggested-by: Dan Carpenter
> >>> Signed-off-by: Peilin Ye
> >>
> >> Reviewed-by: Christian König
> >>
> >> I can't count how many of those we have fixed over the years.
> >>
> >> At some point we should probably document that using "= {}" or "= { 0 }"
> >> in the kernel is a really bad idea and should be avoided.
> >
> > Moreover, it seems like different compilers seem to behave relatively
> > differently with these and we often get reports of warnings with these
> > on clang. When in doubt, memset.
>
> There are quite a few of those under drivers/gpu/drm, for "amd/", "scheduler/"
> drm*.c files,
>
> $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ | wc
> -l
> 374
> $_
>
> Out of which only 16 are of the non-ISO C variety, "= {}",
>
> $find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
> "./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
> 16
> $_
>
> Perhaps the latter are the more pressing ones, since it is a C++ initializer
> and not a ISO C one.
It only matters when we care copying the data to userspace, if it all
stays in the kernel, all is fine.
thanks,
greg k-h
___
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
Am 31.07.20 um 08:53 schrieb Greg Kroah-Hartman:
On Thu, Jul 30, 2020 at 05:09:07PM -0400, Luben Tuikov wrote:
On 2020-07-29 9:49 a.m., Alex Deucher wrote:
On Wed, Jul 29, 2020 at 4:11 AM Christian König
wrote:
Am 28.07.20 um 21:29 schrieb Peilin Ye:
Compiler leaves a 4-byte hole near the end of `dev_info`, causing
amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
when `size` is greater than 356.
In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
unfortunately does not initialize that 4-byte hole. Fix it by using
memset() instead.
Cc: sta...@vger.kernel.org
Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
Suggested-by: Dan Carpenter
Signed-off-by: Peilin Ye
Reviewed-by: Christian König
I can't count how many of those we have fixed over the years.
At some point we should probably document that using "= {}" or "= { 0 }"
in the kernel is a really bad idea and should be avoided.
Moreover, it seems like different compilers seem to behave relatively
differently with these and we often get reports of warnings with these
on clang. When in doubt, memset.
There are quite a few of those under drivers/gpu/drm, for "amd/", "scheduler/"
drm*.c files,
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex "./scheduler/.*\.c" \)
-exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ | wc -l
374
$_
Out of which only 16 are of the non-ISO C variety, "= {}",
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex "./scheduler/.*\.c" \)
-exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
16
$_
Perhaps the latter are the more pressing ones, since it is a C++ initializer
and not a ISO C one.
It only matters when we care copying the data to userspace, if it all
stays in the kernel, all is fine.
Well only as long as you don't try to compute a CRC32, MD5 or any
fingerprint for a hash from the bytes from the structure.
Then it fails horrible and you wonder why the code doesn't works as
expected.
Regards,
Christian.
thanks,
greg k-h
___
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On 2020-07-29 9:49 a.m., Alex Deucher wrote:
> On Wed, Jul 29, 2020 at 4:11 AM Christian König
> wrote:
>>
>> Am 28.07.20 um 21:29 schrieb Peilin Ye:
>>> Compiler leaves a 4-byte hole near the end of `dev_info`, causing
>>> amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
>>> when `size` is greater than 356.
>>>
>>> In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
>>> unfortunately does not initialize that 4-byte hole. Fix it by using
>>> memset() instead.
>>>
>>> Cc: sta...@vger.kernel.org
>>> Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
>>> Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
>>> Suggested-by: Dan Carpenter
>>> Signed-off-by: Peilin Ye
>>
>> Reviewed-by: Christian König
>>
>> I can't count how many of those we have fixed over the years.
>>
>> At some point we should probably document that using "= {}" or "= { 0 }"
>> in the kernel is a really bad idea and should be avoided.
>
> Moreover, it seems like different compilers seem to behave relatively
> differently with these and we often get reports of warnings with these
> on clang. When in doubt, memset.
There are quite a few of those under drivers/gpu/drm, for "amd/", "scheduler/"
drm*.c files,
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
"./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *(|NULL|0) *}" \{\} \+ | wc -l
374
$_
Out of which only 16 are of the non-ISO C variety, "= {}",
$find . \( -regex "./drm.*\.c" -or -regex "./amd/.*\.c" -or -regex
"./scheduler/.*\.c" \) -exec egrep -n -- " *= *{ *}" \{\} \+ | wc -l
16
$_
Perhaps the latter are the more pressing ones, since it is a C++ initializer
and not a ISO C one.
Regards,
Luben
>
> Alex
>
>>
>> Thanks,
>> Christian.
>>
>>> ---
>>> $ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
>>> struct drm_amdgpu_info_device {
>>> __u32 device_id;/* 0 4 */
>>> __u32 chip_rev; /* 4 4 */
>>> __u32 external_rev; /* 8 4 */
>>> __u32 pci_rev; /*12 4 */
>>> __u32 family; /*16 4 */
>>> __u32 num_shader_engines; /*20 4 */
>>> __u32 num_shader_arrays_per_engine; /*24
>>> 4 */
>>> __u32 gpu_counter_freq; /*28 4 */
>>> __u64 max_engine_clock; /*32 8 */
>>> __u64 max_memory_clock; /*40 8 */
>>> __u32 cu_active_number; /*48 4 */
>>> __u32 cu_ao_mask; /*52 4 */
>>> __u32 cu_bitmap[4][4]; /*5664 */
>>> /* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
>>> __u32 enabled_rb_pipes_mask; /* 120 4 */
>>> __u32 num_rb_pipes; /* 124 4 */
>>> /* --- cacheline 2 boundary (128 bytes) --- */
>>> __u32 num_hw_gfx_contexts; /* 128 4 */
>>> __u32 _pad; /* 132 4 */
>>> __u64 ids_flags;/* 136 8 */
>>> __u64 virtual_address_offset; /* 144 8 */
>>> __u64 virtual_address_max; /* 152 8 */
>>> __u32 virtual_address_alignment; /* 160 4
>>> */
>>> __u32 pte_fragment_size;/* 164 4 */
>>> __u32 gart_page_size; /* 168 4 */
>>> __u32 ce_ram_size; /* 172 4 */
>>> __u32 vram_type;/* 176 4 */
>>> __u32 vram_bit_width; /* 180 4 */
>>> __u32 vce_harvest_config; /* 184 4 */
>>> __u32 gc_double_offchip_lds_buf; /* 188 4
>>> */
>>> /* --- cacheline 3 boundary (192 bytes) --- */
>>> __u64 prim_buf_gpu_addr;/* 192 8 */
>>> __u64 pos_buf_gpu_addr; /* 200 8 */
>>> __u64 cntl_sb_buf_gpu_addr; /* 208 8 */
>>> __u64 param_buf_gpu_addr; /* 216 8 */
>>> __u32 prim_buf_size;/* 224 4 */
>>> __u32 pos_buf_size; /* 228 4 */
>>> __u32 cntl_sb_buf_size; /* 232 4 */
>>> __u32 param_buf_size; /* 236 4 */
>>> __u32 wave_front_size; /* 240 4 */
>>>
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
Applied. Thanks!
Alex
On Wed, Jul 29, 2020 at 4:11 AM Christian König
wrote:
>
> Am 28.07.20 um 21:29 schrieb Peilin Ye:
> > Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> > amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
> > when `size` is greater than 356.
> >
> > In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
> > unfortunately does not initialize that 4-byte hole. Fix it by using
> > memset() instead.
> >
> > Cc: sta...@vger.kernel.org
> > Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
> > Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> > Suggested-by: Dan Carpenter
> > Signed-off-by: Peilin Ye
>
> Reviewed-by: Christian König
>
> I can't count how many of those we have fixed over the years.
>
> At some point we should probably document that using "= {}" or "= { 0 }"
> in the kernel is a really bad idea and should be avoided.
>
> Thanks,
> Christian.
>
> > ---
> > $ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
> > struct drm_amdgpu_info_device {
> > __u32 device_id;/* 0 4 */
> > __u32 chip_rev; /* 4 4 */
> > __u32 external_rev; /* 8 4 */
> > __u32 pci_rev; /*12 4 */
> > __u32 family; /*16 4 */
> > __u32 num_shader_engines; /*20 4 */
> > __u32 num_shader_arrays_per_engine; /*24
> > 4 */
> > __u32 gpu_counter_freq; /*28 4 */
> > __u64 max_engine_clock; /*32 8 */
> > __u64 max_memory_clock; /*40 8 */
> > __u32 cu_active_number; /*48 4 */
> > __u32 cu_ao_mask; /*52 4 */
> > __u32 cu_bitmap[4][4]; /*5664 */
> > /* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
> > __u32 enabled_rb_pipes_mask; /* 120 4 */
> > __u32 num_rb_pipes; /* 124 4 */
> > /* --- cacheline 2 boundary (128 bytes) --- */
> > __u32 num_hw_gfx_contexts; /* 128 4 */
> > __u32 _pad; /* 132 4 */
> > __u64 ids_flags;/* 136 8 */
> > __u64 virtual_address_offset; /* 144 8 */
> > __u64 virtual_address_max; /* 152 8 */
> > __u32 virtual_address_alignment; /* 160 4
> > */
> > __u32 pte_fragment_size;/* 164 4 */
> > __u32 gart_page_size; /* 168 4 */
> > __u32 ce_ram_size; /* 172 4 */
> > __u32 vram_type;/* 176 4 */
> > __u32 vram_bit_width; /* 180 4 */
> > __u32 vce_harvest_config; /* 184 4 */
> > __u32 gc_double_offchip_lds_buf; /* 188 4
> > */
> > /* --- cacheline 3 boundary (192 bytes) --- */
> > __u64 prim_buf_gpu_addr;/* 192 8 */
> > __u64 pos_buf_gpu_addr; /* 200 8 */
> > __u64 cntl_sb_buf_gpu_addr; /* 208 8 */
> > __u64 param_buf_gpu_addr; /* 216 8 */
> > __u32 prim_buf_size;/* 224 4 */
> > __u32 pos_buf_size; /* 228 4 */
> > __u32 cntl_sb_buf_size; /* 232 4 */
> > __u32 param_buf_size; /* 236 4 */
> > __u32 wave_front_size; /* 240 4 */
> > __u32 num_shader_visible_vgprs; /* 244 4 */
> > __u32 num_cu_per_sh;/* 248 4 */
> > __u32 num_tcc_blocks; /* 252 4 */
> > /* --- cacheline 4 boundary (256 bytes) --- */
> > __u32 gs_vgt_table_depth; /* 256 4 */
> > __u32 gs_prim_buffer_depth; /* 260 4 */
> > __u32 max_gs_waves_per_vgt; /* 264 4 */
> > __u32 _pad1;/* 268 4 */
> > __u32 cu_ao_bitmap[4][4]; /* 27264 */
> > /* --- cacheline 5 boundary (320 bytes) was 16 bytes ago --- */
> > __u64 high_va_offset; /*
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On Wed, Jul 29, 2020 at 4:11 AM Christian König
wrote:
>
> Am 28.07.20 um 21:29 schrieb Peilin Ye:
> > Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> > amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
> > when `size` is greater than 356.
> >
> > In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
> > unfortunately does not initialize that 4-byte hole. Fix it by using
> > memset() instead.
> >
> > Cc: sta...@vger.kernel.org
> > Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
> > Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> > Suggested-by: Dan Carpenter
> > Signed-off-by: Peilin Ye
>
> Reviewed-by: Christian König
>
> I can't count how many of those we have fixed over the years.
>
> At some point we should probably document that using "= {}" or "= { 0 }"
> in the kernel is a really bad idea and should be avoided.
Moreover, it seems like different compilers seem to behave relatively
differently with these and we often get reports of warnings with these
on clang. When in doubt, memset.
Alex
>
> Thanks,
> Christian.
>
> > ---
> > $ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
> > struct drm_amdgpu_info_device {
> > __u32 device_id;/* 0 4 */
> > __u32 chip_rev; /* 4 4 */
> > __u32 external_rev; /* 8 4 */
> > __u32 pci_rev; /*12 4 */
> > __u32 family; /*16 4 */
> > __u32 num_shader_engines; /*20 4 */
> > __u32 num_shader_arrays_per_engine; /*24
> > 4 */
> > __u32 gpu_counter_freq; /*28 4 */
> > __u64 max_engine_clock; /*32 8 */
> > __u64 max_memory_clock; /*40 8 */
> > __u32 cu_active_number; /*48 4 */
> > __u32 cu_ao_mask; /*52 4 */
> > __u32 cu_bitmap[4][4]; /*5664 */
> > /* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
> > __u32 enabled_rb_pipes_mask; /* 120 4 */
> > __u32 num_rb_pipes; /* 124 4 */
> > /* --- cacheline 2 boundary (128 bytes) --- */
> > __u32 num_hw_gfx_contexts; /* 128 4 */
> > __u32 _pad; /* 132 4 */
> > __u64 ids_flags;/* 136 8 */
> > __u64 virtual_address_offset; /* 144 8 */
> > __u64 virtual_address_max; /* 152 8 */
> > __u32 virtual_address_alignment; /* 160 4
> > */
> > __u32 pte_fragment_size;/* 164 4 */
> > __u32 gart_page_size; /* 168 4 */
> > __u32 ce_ram_size; /* 172 4 */
> > __u32 vram_type;/* 176 4 */
> > __u32 vram_bit_width; /* 180 4 */
> > __u32 vce_harvest_config; /* 184 4 */
> > __u32 gc_double_offchip_lds_buf; /* 188 4
> > */
> > /* --- cacheline 3 boundary (192 bytes) --- */
> > __u64 prim_buf_gpu_addr;/* 192 8 */
> > __u64 pos_buf_gpu_addr; /* 200 8 */
> > __u64 cntl_sb_buf_gpu_addr; /* 208 8 */
> > __u64 param_buf_gpu_addr; /* 216 8 */
> > __u32 prim_buf_size;/* 224 4 */
> > __u32 pos_buf_size; /* 228 4 */
> > __u32 cntl_sb_buf_size; /* 232 4 */
> > __u32 param_buf_size; /* 236 4 */
> > __u32 wave_front_size; /* 240 4 */
> > __u32 num_shader_visible_vgprs; /* 244 4 */
> > __u32 num_cu_per_sh;/* 248 4 */
> > __u32 num_tcc_blocks; /* 252 4 */
> > /* --- cacheline 4 boundary (256 bytes) --- */
> > __u32 gs_vgt_table_depth; /* 256 4 */
> > __u32 gs_prim_buffer_depth; /* 260 4 */
> > __u32 max_gs_waves_per_vgt; /* 264 4 */
> > __u32 _pad1;/* 268 4 */
> > __u32
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
On Wed, Jul 29, 2020 at 10:11 AM Christian König
wrote:
>
> Am 28.07.20 um 21:29 schrieb Peilin Ye:
> > Compiler leaves a 4-byte hole near the end of `dev_info`, causing
> > amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
> > when `size` is greater than 356.
> >
> > In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
> > unfortunately does not initialize that 4-byte hole. Fix it by using
> > memset() instead.
> >
> > Cc: sta...@vger.kernel.org
> > Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
> > Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
> > Suggested-by: Dan Carpenter
> > Signed-off-by: Peilin Ye
>
> Reviewed-by: Christian König
>
> I can't count how many of those we have fixed over the years.
>
> At some point we should probably document that using "= {}" or "= { 0 }"
> in the kernel is a really bad idea and should be avoided.
I think the rule is also "don't create uapi structs with holes in
them", but we've also fumbled that one quite a few times :-/
-Daniel
>
> Thanks,
> Christian.
>
> > ---
> > $ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
> > struct drm_amdgpu_info_device {
> > __u32 device_id;/* 0 4 */
> > __u32 chip_rev; /* 4 4 */
> > __u32 external_rev; /* 8 4 */
> > __u32 pci_rev; /*12 4 */
> > __u32 family; /*16 4 */
> > __u32 num_shader_engines; /*20 4 */
> > __u32 num_shader_arrays_per_engine; /*24
> > 4 */
> > __u32 gpu_counter_freq; /*28 4 */
> > __u64 max_engine_clock; /*32 8 */
> > __u64 max_memory_clock; /*40 8 */
> > __u32 cu_active_number; /*48 4 */
> > __u32 cu_ao_mask; /*52 4 */
> > __u32 cu_bitmap[4][4]; /*5664 */
> > /* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
> > __u32 enabled_rb_pipes_mask; /* 120 4 */
> > __u32 num_rb_pipes; /* 124 4 */
> > /* --- cacheline 2 boundary (128 bytes) --- */
> > __u32 num_hw_gfx_contexts; /* 128 4 */
> > __u32 _pad; /* 132 4 */
> > __u64 ids_flags;/* 136 8 */
> > __u64 virtual_address_offset; /* 144 8 */
> > __u64 virtual_address_max; /* 152 8 */
> > __u32 virtual_address_alignment; /* 160 4
> > */
> > __u32 pte_fragment_size;/* 164 4 */
> > __u32 gart_page_size; /* 168 4 */
> > __u32 ce_ram_size; /* 172 4 */
> > __u32 vram_type;/* 176 4 */
> > __u32 vram_bit_width; /* 180 4 */
> > __u32 vce_harvest_config; /* 184 4 */
> > __u32 gc_double_offchip_lds_buf; /* 188 4
> > */
> > /* --- cacheline 3 boundary (192 bytes) --- */
> > __u64 prim_buf_gpu_addr;/* 192 8 */
> > __u64 pos_buf_gpu_addr; /* 200 8 */
> > __u64 cntl_sb_buf_gpu_addr; /* 208 8 */
> > __u64 param_buf_gpu_addr; /* 216 8 */
> > __u32 prim_buf_size;/* 224 4 */
> > __u32 pos_buf_size; /* 228 4 */
> > __u32 cntl_sb_buf_size; /* 232 4 */
> > __u32 param_buf_size; /* 236 4 */
> > __u32 wave_front_size; /* 240 4 */
> > __u32 num_shader_visible_vgprs; /* 244 4 */
> > __u32 num_cu_per_sh;/* 248 4 */
> > __u32 num_tcc_blocks; /* 252 4 */
> > /* --- cacheline 4 boundary (256 bytes) --- */
> > __u32 gs_vgt_table_depth; /* 256 4 */
> > __u32 gs_prim_buffer_depth; /* 260 4 */
> > __u32 max_gs_waves_per_vgt; /* 264 4 */
> > __u32 _pad1;/* 268 4 */
> > __u32 cu_ao_bitmap[4][4]; /* 27264 */
> > /* ---
Re: [Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
Am 28.07.20 um 21:29 schrieb Peilin Ye:
Compiler leaves a 4-byte hole near the end of `dev_info`, causing
amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
when `size` is greater than 356.
In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
unfortunately does not initialize that 4-byte hole. Fix it by using
memset() instead.
Cc: sta...@vger.kernel.org
Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
Suggested-by: Dan Carpenter
Signed-off-by: Peilin Ye
Reviewed-by: Christian König
I can't count how many of those we have fixed over the years.
At some point we should probably document that using "= {}" or "= { 0 }"
in the kernel is a really bad idea and should be avoided.
Thanks,
Christian.
---
$ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
struct drm_amdgpu_info_device {
__u32 device_id;/* 0 4 */
__u32 chip_rev; /* 4 4 */
__u32 external_rev; /* 8 4 */
__u32 pci_rev; /*12 4 */
__u32 family; /*16 4 */
__u32 num_shader_engines; /*20 4 */
__u32 num_shader_arrays_per_engine; /*24 4
*/
__u32 gpu_counter_freq; /*28 4 */
__u64 max_engine_clock; /*32 8 */
__u64 max_memory_clock; /*40 8 */
__u32 cu_active_number; /*48 4 */
__u32 cu_ao_mask; /*52 4 */
__u32 cu_bitmap[4][4]; /*5664 */
/* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
__u32 enabled_rb_pipes_mask; /* 120 4 */
__u32 num_rb_pipes; /* 124 4 */
/* --- cacheline 2 boundary (128 bytes) --- */
__u32 num_hw_gfx_contexts; /* 128 4 */
__u32 _pad; /* 132 4 */
__u64 ids_flags;/* 136 8 */
__u64 virtual_address_offset; /* 144 8 */
__u64 virtual_address_max; /* 152 8 */
__u32 virtual_address_alignment; /* 160 4 */
__u32 pte_fragment_size;/* 164 4 */
__u32 gart_page_size; /* 168 4 */
__u32 ce_ram_size; /* 172 4 */
__u32 vram_type;/* 176 4 */
__u32 vram_bit_width; /* 180 4 */
__u32 vce_harvest_config; /* 184 4 */
__u32 gc_double_offchip_lds_buf; /* 188 4 */
/* --- cacheline 3 boundary (192 bytes) --- */
__u64 prim_buf_gpu_addr;/* 192 8 */
__u64 pos_buf_gpu_addr; /* 200 8 */
__u64 cntl_sb_buf_gpu_addr; /* 208 8 */
__u64 param_buf_gpu_addr; /* 216 8 */
__u32 prim_buf_size;/* 224 4 */
__u32 pos_buf_size; /* 228 4 */
__u32 cntl_sb_buf_size; /* 232 4 */
__u32 param_buf_size; /* 236 4 */
__u32 wave_front_size; /* 240 4 */
__u32 num_shader_visible_vgprs; /* 244 4 */
__u32 num_cu_per_sh;/* 248 4 */
__u32 num_tcc_blocks; /* 252 4 */
/* --- cacheline 4 boundary (256 bytes) --- */
__u32 gs_vgt_table_depth; /* 256 4 */
__u32 gs_prim_buffer_depth; /* 260 4 */
__u32 max_gs_waves_per_vgt; /* 264 4 */
__u32 _pad1;/* 268 4 */
__u32 cu_ao_bitmap[4][4]; /* 27264 */
/* --- cacheline 5 boundary (320 bytes) was 16 bytes ago --- */
__u64 high_va_offset; /* 336 8 */
__u64 high_va_max; /* 344 8 */
__u32 pa_sc_tile_steering_override; /* 352 4
*/
/* XXX 4 bytes hole, try to pack */
__u64 tcc_disabled_mask;
[Linux-kernel-mentees] [PATCH] drm/amdgpu: Prevent kernel-infoleak in amdgpu_info_ioctl()
Compiler leaves a 4-byte hole near the end of `dev_info`, causing
amdgpu_info_ioctl() to copy uninitialized kernel stack memory to userspace
when `size` is greater than 356.
In 2015 we tried to fix this issue by doing `= {};` on `dev_info`, which
unfortunately does not initialize that 4-byte hole. Fix it by using
memset() instead.
Cc: sta...@vger.kernel.org
Fixes: c193fa91b918 ("drm/amdgpu: information leak in amdgpu_info_ioctl()")
Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
Suggested-by: Dan Carpenter
Signed-off-by: Peilin Ye
---
$ pahole -C "drm_amdgpu_info_device" drivers/gpu/drm/amd/amdgpu/amdgpu_kms.o
struct drm_amdgpu_info_device {
__u32 device_id;/* 0 4 */
__u32 chip_rev; /* 4 4 */
__u32 external_rev; /* 8 4 */
__u32 pci_rev; /*12 4 */
__u32 family; /*16 4 */
__u32 num_shader_engines; /*20 4 */
__u32 num_shader_arrays_per_engine; /*24 4
*/
__u32 gpu_counter_freq; /*28 4 */
__u64 max_engine_clock; /*32 8 */
__u64 max_memory_clock; /*40 8 */
__u32 cu_active_number; /*48 4 */
__u32 cu_ao_mask; /*52 4 */
__u32 cu_bitmap[4][4]; /*5664 */
/* --- cacheline 1 boundary (64 bytes) was 56 bytes ago --- */
__u32 enabled_rb_pipes_mask; /* 120 4 */
__u32 num_rb_pipes; /* 124 4 */
/* --- cacheline 2 boundary (128 bytes) --- */
__u32 num_hw_gfx_contexts; /* 128 4 */
__u32 _pad; /* 132 4 */
__u64 ids_flags;/* 136 8 */
__u64 virtual_address_offset; /* 144 8 */
__u64 virtual_address_max; /* 152 8 */
__u32 virtual_address_alignment; /* 160 4 */
__u32 pte_fragment_size;/* 164 4 */
__u32 gart_page_size; /* 168 4 */
__u32 ce_ram_size; /* 172 4 */
__u32 vram_type;/* 176 4 */
__u32 vram_bit_width; /* 180 4 */
__u32 vce_harvest_config; /* 184 4 */
__u32 gc_double_offchip_lds_buf; /* 188 4 */
/* --- cacheline 3 boundary (192 bytes) --- */
__u64 prim_buf_gpu_addr;/* 192 8 */
__u64 pos_buf_gpu_addr; /* 200 8 */
__u64 cntl_sb_buf_gpu_addr; /* 208 8 */
__u64 param_buf_gpu_addr; /* 216 8 */
__u32 prim_buf_size;/* 224 4 */
__u32 pos_buf_size; /* 228 4 */
__u32 cntl_sb_buf_size; /* 232 4 */
__u32 param_buf_size; /* 236 4 */
__u32 wave_front_size; /* 240 4 */
__u32 num_shader_visible_vgprs; /* 244 4 */
__u32 num_cu_per_sh;/* 248 4 */
__u32 num_tcc_blocks; /* 252 4 */
/* --- cacheline 4 boundary (256 bytes) --- */
__u32 gs_vgt_table_depth; /* 256 4 */
__u32 gs_prim_buffer_depth; /* 260 4 */
__u32 max_gs_waves_per_vgt; /* 264 4 */
__u32 _pad1;/* 268 4 */
__u32 cu_ao_bitmap[4][4]; /* 27264 */
/* --- cacheline 5 boundary (320 bytes) was 16 bytes ago --- */
__u64 high_va_offset; /* 336 8 */
__u64 high_va_max; /* 344 8 */
__u32 pa_sc_tile_steering_override; /* 352 4
*/
/* XXX 4 bytes hole, try to pack */
__u64 tcc_disabled_mask;/* 360 8 */
/* size: 368, cachelines: 6, members: 49 */
/* sum members: 364, holes: 1, sum holes: 4 */
/* last cacheline: 48 bytes */
};
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git
