linux-kernel - Re: [PATCH v6 9/9] drm/amd/display: Mark dc_fixpt_from

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <bc35fddc-498a-cc58-b6a1-f5234a4f6d0d@loongson.cn>
Date: Sun, 22 Dec 2024 12:27:47 +0800
From: Tiezhu Yang <yangtiezhu@...ngson.cn>
To: Xi Ruoyao <xry111@...111.site>, Nathan Chancellor <nathan@...nel.org>,
 Peter Zijlstra <peterz@...radead.org>
Cc: Alex Deucher <alexdeucher@...il.com>, Josh Poimboeuf
 <jpoimboe@...nel.org>, Huacai Chen <chenhuacai@...nel.org>,
 loongarch@...ts.linux.dev, amd-gfx@...ts.freedesktop.org,
 linux-kernel@...r.kernel.org, Nick Desaulniers <ndesaulniers@...gle.com>,
 llvm@...ts.linux.dev
Subject: Re: [PATCH v6 9/9] drm/amd/display: Mark dc_fixpt_from_fraction()
 noinline

On 12/21/2024 03:40 PM, Xi Ruoyao wrote:
> On Fri, 2024-12-20 at 15:34 -0700, Nathan Chancellor wrote:
>>> Now, the thing is, these ASSERT()s are checking for divide-by-zero, I
>>> suspect clang figured that out and invokes UB on us and just stops
>>> code-gen.
>>
>> Yeah, I think your analysis is spot on, as this was introduced by a
>> change in clang from a few months ago according to my bisect:
>>
>> https://github.com/llvm/llvm-project/commit/37932643abab699e8bb1def08b7eb4eae7ff1448
>>
>> Since the ASSERT does not do anything to prevent the divide by zero (it
>> just flags it with WARN_ON) and the rest of the code doesn't either, I
>> assume that the codegen stops as soon as it encounters the unreachable
>> that change created from the path where divide by zero would occur via
>>
>>   dc_fixpt_recip() ->
>>     dc_fixpt_from_fraction() ->
>>       complete_integer_division_u64() ->
>>         div64_u64_rem()
>>
>> Shouldn't callers of division functions harden them against dividing by
>> zero?
>
> Yes I think it'd be the correct solution.

Thank you all. Do you mean like this?

--- >8 ---

diff --git a/drivers/gpu/drm/amd/display/dc/basics/fixpt31_32.c 
b/drivers/gpu/drm/amd/display/dc/basics/fixpt31_32.c
index 88d3f9d7dd55..848d8e67304a 100644
--- a/drivers/gpu/drm/amd/display/dc/basics/fixpt31_32.c
+++ b/drivers/gpu/drm/amd/display/dc/basics/fixpt31_32.c
@@ -79,11 +79,13 @@ struct fixed31_32 dc_fixpt_from_fraction(long long 
numerator, long long denomina
         unsigned long long arg2_value = arg2_negative ? -denominator : 
denominator;

         unsigned long long remainder;
+       unsigned long long res_value;

         /* determine integer part */

-       unsigned long long res_value = complete_integer_division_u64(
-               arg1_value, arg2_value, &remainder);
+       ASSERT(arg2_value);
+
+       res_value = complete_integer_division_u64(arg1_value, 
arg2_value, &remainder);

         ASSERT(res_value <= LONG_MAX);

@@ -214,8 +216,6 @@ struct fixed31_32 dc_fixpt_recip(struct fixed31_32 arg)
          * Good idea to use Newton's method
          */

-       ASSERT(arg.value);
-
         return dc_fixpt_from_fraction(
                 dc_fixpt_one.value,
                 arg.value);

With the above changes, there is no "falls through" objtool warning
compiled with both clang 19 and the latest mainline clang 20.

If you are OK with it, I will send a separate formal patch to handle
this issue after doing some more testing.

Thanks,
Tiezhu