Re: [PATCH] arm64/cpufeature: Move BUG_ON() inside get_arm64_ftr_reg()

[Anshuman Khandual]

On 05/21/2020 10:29 PM, Catalin Marinas wrote:
> On Thu, May 21, 2020 at 05:22:15PM +0100, Will Deacon wrote:
>> On Thu, May 21, 2020 at 08:45:38AM +0530, Anshuman Khandual wrote:
>>> On 05/20/2020 11:09 PM, Will Deacon wrote:
>>>> On Wed, May 20, 2020 at 04:47:11PM +0100, Catalin Marinas wrote:
>>>>> On Wed, May 20, 2020 at 01:20:13PM +0100, Will Deacon wrote:
>>>>>> On Wed, May 20, 2020 at 06:52:54AM +0530, Anshuman Khandual wrote:
>>>>>>> There is no way to proceed when requested register could not be searched in
>>>>>>> arm64_ftr_reg[]. Requesting for a non present register would be an error as
>>>>>>> well. Hence lets just BUG_ON() when the search fails in get_arm64_ftr_reg()
>>>>>>> rather than checking for return value and doing the same in some individual
>>>>>>> callers.
>>>>>>>
>>>>>>> But there are some callers that dont BUG_ON() upon search failure. It adds
>>>>>>> an argument 'failsafe' that provides required switch between callers based
>>>>>>> on whether they could proceed or not.
>>>>>>>
>>>>>>> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
>>>>>>> Cc: Will Deacon <will@xxxxxxxxxx>
>>>>>>> Cc: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
>>>>>>> Cc: Mark Brown <broonie@xxxxxxxxxx>
>>>>>>> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
>>>>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx
>>>>>>>
>>>>>>> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
>>>>>>> ---
>>>>>>> Applies on next-20200518 that has recent cpufeature changes from Will.
>>>>>>>
>>>>>>>  arch/arm64/kernel/cpufeature.c | 26 +++++++++++++-------------
>>>>>>>  1 file changed, 13 insertions(+), 13 deletions(-)
>>>>>>>
>>>>>>> diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
>>>>>>> index bc5048f152c1..62767cc540c3 100644
>>>>>>> --- a/arch/arm64/kernel/cpufeature.c
>>>>>>> +++ b/arch/arm64/kernel/cpufeature.c
>>>>>>> @@ -557,7 +557,7 @@ static int search_cmp_ftr_reg(const void *id, const void *regp)
>>>>>>>   *         - NULL on failure. It is upto the caller to decide
>>>>>>>   *	     the impact of a failure.
>>>>>>>   */
>>>>>>> -static struct arm64_ftr_reg *get_arm64_ftr_reg(u32 sys_id)
>>>>>>> +static struct arm64_ftr_reg *get_arm64_ftr_reg(u32 sys_id, bool failsafe)
>>>>>>
>>>>>> Generally, I'm not a big fan of boolean arguments because they are really
>>>>>> opaque at the callsite. It also seems bogus to me that we don't trust the
>>>>>> caller to pass a valid sys_id, but we trust it to get "failsafe" right,
>>>>>> which seems to mean "I promise to check the result isn't NULL before
>>>>>> dereferencing it."
>>>>>>
>>>>>> So I don't see how this patch improves anything. I'd actually be more
>>>>>> inclined to stick a WARN() in get_arm64_ftr_reg() when it returns NULL and
>>>>>> have the callers handle NULL by returning early, getting rid of all the
>>>>>> BUG_ONs in here. Sure, the system might end up in a funny state, but we
>>>>>> WARN()d about it and tried to keep going (and Linus has some strong opinions
>>>>>> on this too).
>>>>>
>>>>> Such WARN can be triggered by the user via emulate_sys_reg(), so we
>>>>> can't really have it in get_arm64_ftr_reg() without a 'failsafe' option.
>>>>
>>>> Ah yes, that would be bad. In which case, I don't think the existing code
>>>> should change.
>>>
>>> The existing code has BUG_ON() in three different callers doing exactly the
>>> same thing that can easily be taken care in get_arm64_ftr_reg() itself. As
>>> mentioned before an enum variable (as preferred - over a bool) can still
>>> preserve the existing behavior for emulate_sys_reg().
>>>
>>> IMHO these are very good reasons for us to change the code which will make
>>> it cleaner while also removing three redundant BUG_ON() instances. Hence I
>>> will request you to please reconsider this proposal.
>>
>> Hmm, then how about trying my proposal with the WARN_ON(), but having a
>> get_arm64_ftr_reg_nowarn() variant for the user emulation case?
> 
> That works for me, get_arm64_ftr_reg() would be a wrapper over the
> _nowarn function with the added WARN_ON.

Sure, will do.

> 
> read_sanitised_ftr_reg() would need to return something though. Would
> all 0s be ok? I think it works as long as we don't have negative CPUID
> fields.

Just trying to understand. If get_arm64_ftr_reg() returns NULL, then
read_sanitised_ftr_reg() should also return 0 for that non existent
register (already been warned in get_arm64_ftr_reg).

@@ -961,8 +972,8 @@ u64 read_sanitised_ftr_reg(u32 id)
 {
        struct arm64_ftr_reg *regp = get_arm64_ftr_reg(id);

-       /* We shouldn't get a request for an unsupported register */
-       BUG_ON(!regp);
+       if (!regp)
+               return 0;
        return regp->sys_val;
 }