On 23 March 2017 at 09:32, Srinivas Ramana [off-list ref] wrote:

On 03/22/2017 07:15 PM, Srinivas Ramana wrote:

quoted

On 03/22/2017 06:10 PM, Will Deacon wrote:

quoted

On Wed, Mar 22, 2017 at 12:16:24PM +0000, Ard Biesheuvel wrote:

quoted

On 22 March 2017 at 11:38, Srinivas Ramana [off-list ref]
wrote:

quoted

From: Neeraj Upadhyay <redacted>

If kernel image extends across alignment boundary, existing
code increases the KASLR offset by size of kernel image. The
offset is masked after resizing. There are cases, where after
masking, we may still have kernel image extending across
boundary. This eventually results in only 2MB block getting
mapped while creating the page tables. This results in data aborts
while accessing unmapped regions during second relocation (with
kaslr offset) in __primary_switch. To fix this problem, round up the
kernel image size, by swapper block size, before adding it for
correction.

For example consider below case, where kernel image still crosses
1GB alignment boundary, after masking the offset, which is fixed
by rounding up kernel image size.

SWAPPER_TABLE_SHIFT = 30
Swapper using section maps with section size 2MB.
CONFIG_PGTABLE_LEVELS = 3
VA_BITS = 39

_text  : 0xffffff8008080000
_end   : 0xffffff800aa1b000
offset : 0x1f35600000
mask = ((1UL << (VA_BITS - 2)) - 1) & ~(SZ_2M - 1)

(_text + offset) >> SWAPPER_TABLE_SHIFT = 0x3fffffe7c
(_end + offset) >> SWAPPER_TABLE_SHIFT  = 0x3fffffe7d

offset after existing correction (before mask) = 0x1f37f9b000
(_text + offset) >> SWAPPER_TABLE_SHIFT = 0x3fffffe7d
(_end + offset) >> SWAPPER_TABLE_SHIFT  = 0x3fffffe7d

offset (after mask) = 0x1f37e00000
(_text + offset) >> SWAPPER_TABLE_SHIFT = 0x3fffffe7c
(_end + offset) >> SWAPPER_TABLE_SHIFT  = 0x3fffffe7d

new offset w/ rounding up = 0x1f38000000
(_text + offset) >> SWAPPER_TABLE_SHIFT = 0x3fffffe7d
(_end + offset) >> SWAPPER_TABLE_SHIFT  = 0x3fffffe7d

Fixes: f80fb3a3d508 ("arm64: add support for kernel ASLR")
Signed-off-by: Neeraj Upadhyay <redacted>
Signed-off-by: Srinivas Ramana <redacted>

Reviewed-by: Ard Biesheuvel <redacted>

... and thanks for the excellent commit log message!

Thanks both. I've picked this up as a fix.

Will

Thanks Ard and Will for the review and picking this patch.
can we also CC: [off-list ref] ?

Thanks,
-- Srinivas R

Sorry, there is a checkpatch error in the last patch. I will submit v3
after fixing the checkpatch error.

I wouldn't worry about that. Will has already queued the patch.