In alloc_flex_gd(), when flexbg_size is large, kmalloc size would
overflow and flex_gd->groups would point to a buffer smaller than
expected, causing OOB accesses when it is used.
Note that in ext4_resize_fs(), flexbg_size is calculated using
sbi->s_log_groups_per_flex, which is read from the disk and only bounded
to [1, 31]. The patch returns NULL for too large flexbg_size.
Signed-off-by: Haogang Chen <[email protected]>
---
fs/ext4/resize.c | 2 ++
1 files changed, 2 insertions(+), 0 deletions(-)
diff --git a/fs/ext4/resize.c b/fs/ext4/resize.c
index f9d948f..8601f4b 100644
--- a/fs/ext4/resize.c
+++ b/fs/ext4/resize.c
@@ -161,6 +161,8 @@ static struct ext4_new_flex_group_data *alloc_flex_gd(unsigned long flexbg_size)
if (flex_gd == NULL)
goto out3;
+ if (flexbg_size >= UINT_MAX / sizeof(struct ext4_new_flex_group_data))
+ goto out2;
flex_gd->count = flexbg_size;
flex_gd->groups = kmalloc(sizeof(struct ext4_new_group_data) *
--
1.7.5.4
On 2/20/12 4:41 PM, Haogang Chen wrote:
> In alloc_flex_gd(), when flexbg_size is large, kmalloc size would
> overflow and flex_gd->groups would point to a buffer smaller than
> expected, causing OOB accesses when it is used.
>
> Note that in ext4_resize_fs(), flexbg_size is calculated using
> sbi->s_log_groups_per_flex, which is read from the disk and only bounded
> to [1, 31]. The patch returns NULL for too large flexbg_size
Hm this raises a few questions I think.
On the one hand, making sure the kmalloc arg doesn't overflow here is
certainly a good thing and probably the right thing to do in the short term.
So I guess:
Reviewed-by: Eric Sandeen <[email protected]>
for that, to close the hole.
But the types are a mess; alloc_flex_gd() takes an unsigned long;
it's passed an int, and assigns to flex_gd->count, an ext4_group_t
(which is an unsigned int). They should probably all be ext4_group_t
for consistency.
But that's not the worst of it...
Doesn't this also mean that a valid s_log_groups_per_flex (i.e. 31)
will fail in this resize code? That would be an unexpected outcome.
2^31 groups per flex is a little crazy, but still technically valid
according to the limits in the code.
So really, trying to allocate an array of all possible groups-per-flex
in the resize code is probably a really bad idea to start with, and the
resize code has got serious problems if kmalloc(UINT_MAX-1) is expected
to work...
-Eric
> Signed-off-by: Haogang Chen <[email protected]>
> ---
> fs/ext4/resize.c | 2 ++
> 1 files changed, 2 insertions(+), 0 deletions(-)
>
> diff --git a/fs/ext4/resize.c b/fs/ext4/resize.c
> index f9d948f..8601f4b 100644
> --- a/fs/ext4/resize.c
> +++ b/fs/ext4/resize.c
> @@ -161,6 +161,8 @@ static struct ext4_new_flex_group_data *alloc_flex_gd(unsigned long flexbg_size)
> if (flex_gd == NULL)
> goto out3;
>
> + if (flexbg_size >= UINT_MAX / sizeof(struct ext4_new_flex_group_data))
> + goto out2;
> flex_gd->count = flexbg_size;
>
> flex_gd->groups = kmalloc(sizeof(struct ext4_new_group_data) *
On Feb 20, 2012, at 6:47 PM, Eric Sandeen wrote:
> Hm this raises a few questions I think.
>
> On the one hand, making sure the kmalloc arg doesn't overflow here is
> certainly a good thing and probably the right thing to do in the short term.
>
> So I guess:
>
> Reviewed-by: Eric Sandeen <[email protected]>
>
> for that, to close the hole.
Another possibility is to wait for knalloc/kmalloc_array in the -mm
tree, which is basically the non-zeroing version of kcalloc that
performs overflow checking.
> Doesn't this also mean that a valid s_log_groups_per_flex (i.e. 31)
> will fail in this resize code? That would be an unexpected outcome.
> 2^31 groups per flex is a little crazy, but still technically valid
> according to the limits in the code.
Or we could limit s_log_groups_per_flex/groups_per_flex to a
reasonable upper bound in ext4_fill_flex_info(), right?
- xi
On 02/21/2012 07:55 AM, Xi Wang wrote:
> On Feb 20, 2012, at 6:47 PM, Eric Sandeen wrote:
>> Hm this raises a few questions I think.
>>
>> On the one hand, making sure the kmalloc arg doesn't overflow here is
>> certainly a good thing and probably the right thing to do in the short term.
>>
>> So I guess:
>>
>> Reviewed-by: Eric Sandeen <[email protected]>
>>
>> for that, to close the hole.
>
> Another possibility is to wait for knalloc/kmalloc_array in the -mm
> tree, which is basically the non-zeroing version of kcalloc that
> performs overflow checking.
>
>> Doesn't this also mean that a valid s_log_groups_per_flex (i.e. 31)
>> will fail in this resize code? That would be an unexpected outcome.
>> 2^31 groups per flex is a little crazy, but still technically valid
>> according to the limits in the code.
>
> Or we could limit s_log_groups_per_flex/groups_per_flex to a
> reasonable upper bound in ext4_fill_flex_info(), right?
Depends on the "flex_bg" design intent, I guess.
I don't know if the 2^31 was an intended design limit, or just a
mathematical limit that based on container sizes etc...
I'd have to look at the resize code more carefully but I can't imagine
that it's imperative to allocate this stuff all at once.
-Eric
> - xi
>
On 2012-02-21, at 9:36 AM, Eric Sandeen wrote:
> On 02/21/2012 07:55 AM, Xi Wang wrote:
>> On Feb 20, 2012, at 6:47 PM, Eric Sandeen wrote:
>>> Hm this raises a few questions I think.
>>>
>>> On the one hand, making sure the kmalloc arg doesn't overflow here is
>>> certainly a good thing and probably the right thing to do in the short term.
>>>
>>> So I guess:
>>>
>>> Reviewed-by: Eric Sandeen <[email protected]>
>>>
>>> for that, to close the hole.
>>
>> Another possibility is to wait for knalloc/kmalloc_array in the -mm
>> tree, which is basically the non-zeroing version of kcalloc that
>> performs overflow checking.
>>
>>> Doesn't this also mean that a valid s_log_groups_per_flex (i.e. 31)
>>> will fail in this resize code? That would be an unexpected outcome.
>>> 2^31 groups per flex is a little crazy, but still technically valid
>>> according to the limits in the code.
>>
>> Or we could limit s_log_groups_per_flex/groups_per_flex to a
>> reasonable upper bound in ext4_fill_flex_info(), right?
>
> Depends on the "flex_bg" design intent, I guess.
>
> I don't know if the 2^31 was an intended design limit, or just a
> mathematical limit that based on container sizes etc...
>
> I'd have to look at the resize code more carefully but I can't imagine
> that it's imperative to allocate this stuff all at once.
We previously tried to use a large flex_bg size to put all metadata into a
single group so it could easily be allocated on a separate SSD device, but
that didn't work very well. Once the number of bitmaps in group 0 is more
than the number of free blocks in that group (below 16k groups, due to group
descriptors) then they need to overflow into group 1 and collide with the
group descriptors there. Then mke2fs chokes, AFAIR.
It may be different with bigalloc, since the number of blocks in a group can
be very large, I haven't tried that.
In any case, I don't think anyone expects vmalloc(2^32 * struct size) to work,
but I wouldn't sweat fixing this until there is some real reason to do so.
Cheers, Andreas
On Mon, Feb 20, 2012 at 05:41:24PM -0500, Haogang Chen wrote:
> In alloc_flex_gd(), when flexbg_size is large, kmalloc size would
> overflow and flex_gd->groups would point to a buffer smaller than
> expected, causing OOB accesses when it is used.
>
> Note that in ext4_resize_fs(), flexbg_size is calculated using
> sbi->s_log_groups_per_flex, which is read from the disk and only bounded
> to [1, 31]. The patch returns NULL for too large flexbg_size.
>
> Signed-off-by: Haogang Chen <[email protected]>
Thanks, applied. Apologies for missing this during the last cycle.
- Ted