Re: [PATCH] drm/radeon/evergreen_cs: fix int overflow errors in cs track offsets
From: Christian König
Date: Wed Jul 31 2024 - 02:57:14 EST
Am 30.07.24 um 19:36 schrieb Nikita Zhandarovich:
On 7/29/24 11:12, Christian König wrote:
Am 29.07.24 um 20:04 schrieb Christian König:
Am 29.07.24 um 19:26 schrieb Nikita Zhandarovich:
Hi,
On 7/29/24 02:23, Christian König wrote:
Am 26.07.24 um 14:52 schrieb Alex Deucher:
On Fri, Jul 26, 2024 at 3:05 AM Christian König
<christian.koenig@xxxxxxx> wrote:
Am 25.07.24 um 20:09 schrieb Nikita Zhandarovich:
Several cs track offsets (such as 'track->db_s_read_offset')
either are initialized with or plainly take big enough values that,
once shifted 8 bits left, may be hit with integer overflow if the
resulting values end up going over u32 limit.
Some debug prints take this into account (see according
dev_warn() in
evergreen_cs_track_validate_stencil()), even if the actual
calculated value assigned to local 'offset' variable is missing
similar proper expansion.
Mitigate the problem by casting the type of right operands to the
wider type of corresponding left ones in all such cases.
Found by Linux Verification Center (linuxtesting.org) with static
analysis tool SVACE.
Fixes: 285484e2d55e ("drm/radeon: add support for evergreen/ni
tiling informations v11")
Cc: stable@xxxxxxxxxxxxxxx
Well first of all the long cast doesn't makes the value 64bit, it
depends on the architecture.
Then IIRC the underlying hw can only handle a 32bit address space so
having the offset as long is incorrect to begin with.
Evergreen chips support a 36 bit internal address space and NI and
newer support a 40 bit one, so this is applicable.
In that case I strongly suggest that we replace the unsigned long with
u64 or otherwise we get different behavior on 32 and 64bit machines.
Regards,
Christian.
To be clear, I'll prepare v2 patch that changes 'offset' to u64 as well
as the cast of 'track->db_z_read_offset' (and the likes) to u64 too.
On the other note, should I also include casting to wider type of the
expression surf.layer_size * mslice (example down below) in
evergreen_cs_track_validate_cb() and other similar functions? I can't
properly gauge if the result will definitively fit into u32, maybe it
makes sense to expand it as well?
The integer overflows caused by shifts are irrelevant and doesn't need
any fixing in the first place.
Wait a second.
Thinking more about it the integer overflows are actually necessary
because that is exactly what happens in the hardware as well.
If you don't overflow those shifts you actually create a security
problem because the HW the might access at a different offset then you
calculated here.
We need to use something like a mask or use lower_32_bits() here.
Christian,
My apologies, I may be getting a bit confused here.
If integer overflows caused by shifts are predictable and constitute
normal behavior in this case, and there is no need to "fix" them, does
it still make sense to use any mitigations at all, i.e. masks or macros?
Well you stumbled over that somehow, so some automated checker things
that this is a bad idea.
Leaving these shifts to u32 variables as they are now will achieve the
same result as, for example, doing something along the lines of:
offset = lower_32_bits((u64)track->cb_color_bo_offset[id] << 8);
which seems clunky and unnecessary, even if it suppresses some static
analyzer triggers (and that seems overboard).
Or am I missing something obvious here?
No, it's just about suppressing the static checker warnings.
I'm also not 100% sure how that old hw works. Alex mentioned that it is
using 36bits internally.
So it could be that we need to switch to using u64 here. I need to
double check that with Alex.
But using unsigned long is certainly incorrect cause we then get
different behavior based on the CPU architecture.
Thanks for pointing this out,
Christian.
Regards,
Christian.
The point is rather that we need to avoid multiplication overflows and
the security problems which come with those.
441 }
442
443 offset += surf.layer_size * mslice;
In other words that here needs to be validated correctly.
Agreed, I think either casting right operand to u64 (once 'offset' is
also changed from unsigned long to u64) or using mul_u32_u32() here and
in other places should suffice.
Regards,
Christian.
444 if (offset > radeon_bo_size(track->cb_color_bo[id])) {
445 /* old ddx are broken they allocate bo with w*h*bpp
Regards,
Nikita
Alex
And finally that is absolutely not material for stable.
Regards,
Christian.
Signed-off-by: Nikita Zhandarovich <n.zhandarovich@xxxxxxxxxx>
---
P.S. While I am not certain that track->cb_color_bo_offset[id]
actually ends up taking values high enough to cause an overflow,
nonetheless I thought it prudent to cast it to ulong as well.
drivers/gpu/drm/radeon/evergreen_cs.c | 18 +++++++++---------
1 file changed, 9 insertions(+), 9 deletions(-)
diff --git a/drivers/gpu/drm/radeon/evergreen_cs.c
b/drivers/gpu/drm/radeon/evergreen_cs.c
index 1fe6e0d883c7..d734d221e2da 100644
--- a/drivers/gpu/drm/radeon/evergreen_cs.c
+++ b/drivers/gpu/drm/radeon/evergreen_cs.c
@@ -433,7 +433,7 @@ static int evergreen_cs_track_validate_cb(struct
radeon_cs_parser *p, unsigned i
return r;
}
- offset = track->cb_color_bo_offset[id] << 8;
+ offset = (unsigned long)track->cb_color_bo_offset[id] << 8;
if (offset & (surf.base_align - 1)) {
dev_warn(p->dev, "%s:%d cb[%d] bo base %ld not
aligned with %ld\n",
__func__, __LINE__, id, offset,
surf.base_align);
@@ -455,7 +455,7 @@ static int evergreen_cs_track_validate_cb(struct
radeon_cs_parser *p, unsigned i
min = surf.nby - 8;
}
bsize =
radeon_bo_size(track->cb_color_bo[id]);
- tmp = track->cb_color_bo_offset[id] << 8;
+ tmp = (unsigned
long)track->cb_color_bo_offset[id] << 8;
for (nby = surf.nby; nby > min; nby--) {
size = nby * surf.nbx * surf.bpe *
surf.nsamples;
if ((tmp + size * mslice) <=
bsize) {
@@ -476,10 +476,10 @@ static int
evergreen_cs_track_validate_cb(struct radeon_cs_parser *p,
unsigned i
}
}
dev_warn(p->dev, "%s:%d cb[%d] bo too small (layer
size %d, "
- "offset %d, max layer %d, bo size %ld, slice
%d)\n",
+ "offset %ld, max layer %d, bo size %ld, slice
%d)\n",
__func__, __LINE__, id, surf.layer_size,
- track->cb_color_bo_offset[id] << 8, mslice,
- radeon_bo_size(track->cb_color_bo[id]), slice);
+ (unsigned long)track->cb_color_bo_offset[id]
<< 8,
+ mslice,
radeon_bo_size(track->cb_color_bo[id]), slice);
dev_warn(p->dev, "%s:%d problematic surf: (%d %d)
(%d
%d %d %d %d %d %d)\n",
__func__, __LINE__, surf.nbx, surf.nby,
surf.mode, surf.bpe, surf.nsamples,
@@ -608,7 +608,7 @@ static int
evergreen_cs_track_validate_stencil(struct radeon_cs_parser *p)
return r;
}
- offset = track->db_s_read_offset << 8;
+ offset = (unsigned long)track->db_s_read_offset << 8;
if (offset & (surf.base_align - 1)) {
dev_warn(p->dev, "%s:%d stencil read bo base %ld not
aligned with %ld\n",
__func__, __LINE__, offset,
surf.base_align);
@@ -627,7 +627,7 @@ static int
evergreen_cs_track_validate_stencil(struct radeon_cs_parser *p)
return -EINVAL;
}
- offset = track->db_s_write_offset << 8;
+ offset = (unsigned long)track->db_s_write_offset << 8;
if (offset & (surf.base_align - 1)) {
dev_warn(p->dev, "%s:%d stencil write bo base %ld
not
aligned with %ld\n",
__func__, __LINE__, offset,
surf.base_align);
@@ -706,7 +706,7 @@ static int
evergreen_cs_track_validate_depth(struct radeon_cs_parser *p)
return r;
}
- offset = track->db_z_read_offset << 8;
+ offset = (unsigned long)track->db_z_read_offset << 8;
if (offset & (surf.base_align - 1)) {
dev_warn(p->dev, "%s:%d stencil read bo base %ld not
aligned with %ld\n",
__func__, __LINE__, offset,
surf.base_align);
@@ -722,7 +722,7 @@ static int
evergreen_cs_track_validate_depth(struct radeon_cs_parser *p)
return -EINVAL;
}
- offset = track->db_z_write_offset << 8;
+ offset = (unsigned long)track->db_z_write_offset << 8;
if (offset & (surf.base_align - 1)) {
dev_warn(p->dev, "%s:%d stencil write bo base %ld
not
aligned with %ld\n",
__func__, __LINE__, offset,
surf.base_align);