Re: [PATCH 2/2] MIPS: memset.S: Fix return of __clear_user from Lpartial_fixup

From: Matt Redfearn
Date: Tue Apr 17 2018 - 09:21:20 EST


Hi James,

On 16/04/18 23:13, James Hogan wrote:
On Thu, Mar 29, 2018 at 10:28:24AM +0100, Matt Redfearn wrote:
The __clear_user function is defined to return the number of bytes that
could not be cleared. From the underlying memset / bzero implementation
this means setting register a2 to that number on return. Currently if a
page fault is triggered within the memset_partial block, the value
loaded into a2 on return is meaningless.

The label .Lpartial_fixup\@ is jumped to on page fault. Currently it
masks the remaining count of bytes (a2) with STORMASK, meaning that the
least significant 2 (32bit) or 3 (64bit) bits of the remaining count are
always clear.

Are you sure about that. It seems to do that *to ensure those bits are
set correctly*...

Secondly, .Lpartial_fixup\@ expects t1 to contain the end address of the
copy. This is set up by the initial block:
PTR_ADDU t1, a0 /* end address */
However, the .Lmemset_partial\@ block then reuses register t1 to
calculate a jump through a block of word copies. This leaves it no
longer containing the end address of the copy operation if a page fault
occurs, and the remaining bytes calculation is incorrect.

Fix these issues by removing the and of a2 with STORMASK, and replace t1
with register t2 in the .Lmemset_partial\@ block.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Matt Redfearn <matt.redfearn@xxxxxxxx>
---

arch/mips/lib/memset.S | 9 ++++-----
1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/arch/mips/lib/memset.S b/arch/mips/lib/memset.S
index 90bcdf1224ee..3257dca58cad 100644
--- a/arch/mips/lib/memset.S
+++ b/arch/mips/lib/memset.S
@@ -161,19 +161,19 @@
.Lmemset_partial\@:
R10KCBARRIER(0(ra))
- PTR_LA t1, 2f /* where to start */
+ PTR_LA t2, 2f /* where to start */
#ifdef CONFIG_CPU_MICROMIPS
LONG_SRL t7, t0, 1

Hmm, on microMIPS t7 isn't on the clobber list for __bzero, and nor is
t8...

#endif
#if LONGSIZE == 4
- PTR_SUBU t1, FILLPTRG
+ PTR_SUBU t2, FILLPTRG
#else
.set noat
LONG_SRL AT, FILLPTRG, 1
- PTR_SUBU t1, AT
+ PTR_SUBU t2, AT
.set at
#endif
- jr t1
+ jr t2
PTR_ADDU a0, t0 /* dest ptr */

^^^ note this...

.set push
@@ -250,7 +250,6 @@
.Lpartial_fixup\@:
PTR_L t0, TI_TASK($28)
- andi a2, STORMASK

... this isn't right.

If I read correctly, t1 (after the above change stops clobbering it) is
the end of the full 64-byte blocks, i.e. the start address of the final
partial block.


The .Lfwd_fixup calculation (for full blocks) appears to be:

a2 = ((len & 0x3f) + start_of_partial) - badvaddr

which is spot on. (len & 0x3f) is the partial block and remaining bytes
that haven't been set yet, add start_of_partial to get end of the full
range, subtract bad address to find how much didn't copy.


The calculation for .Lpartial_fixup however appears to (currently) do:

a2 = ((len & STORMASK) + start_of_partial) - badvaddr

Which might make sense if start_of_partial (t1) was replaced with
end_of_partial, which does seem to be calculated as noted above, and put
in a0 ready for the final few bytes to be set.

LONG_L t0, THREAD_BUADDR(t0)
LONG_ADDU a2, t1

^^ So I think either it needs to just s/t1/a0/ here and not bother
preserving t1 above (smaller change and probably the original intent),
or preserve t1 and mask 0x3f instead of STORMASK like .Lfwd_fixup does
(which would work but seems needlessly complicated to me).

Does that make any sense or have I misunderstood some subtlety?

Thanks for taking the time to work this through - you're right, changing t1 to a0 in the fault handler does give the right result and is much less invasive. Updated patch incoming :-)

Thanks,
Matt



Cheers
James

jr ra
--
2.7.4