Re: [PATCH] mm: Make failslab writable again

[Vlastimil Babka]

On 9/20/22 12:21, Alexander Atanasov wrote:
> On 20.09.22 12:29, Vlastimil Babka wrote:
>> On 9/20/22 11:17, Alexander Atanasov wrote:
>>> Hello,
>>>
>>> On 20.09.22 11:42, Vlastimil Babka wrote:
>>>>> +static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
>>>>> +                size_t length)
>>>>> +{
>>>>> +    if (s->refcount > 1)
>>>>> +        return -EINVAL;
>>>>> +
>>>>> +    s->flags &= ~SLAB_FAILSLAB;
>>>>> +    if (buf[0] == '1')
>>>>> +        s->flags |= SLAB_FAILSLAB;
>>>>
>>>> Could we at least use a temporary variable to set up the final value and
>>>> then do a WRITE_ONCE() to s->flags, so the compiler is not allowed to do
>>>> some funky stuff? Assuming this is really the only place where we modify
>>>> s->flags during runtime, so we can't miss other updates due to RMW.
>>>
>>> Since it is set or clear - instead of temporary variable and potentially two
>>> writes and RMW issues i would suggest this:
>>> +    if (buf[0] == '1')
>>> +        s->flags |= SLAB_FAILSLAB;
>>> +       else
>>> +        s->flags &= ~SLAB_FAILSLAB;
>>
>> This way also has RMW issues, and also the compiler is allowed to
>> temporarily modify s->flags any way it likes; with WRITE_ONCE() it can't.
> 
> Okay, so the safest way is this?
> 
> if (buf[0] == '1')
>     WRITE_ONCE(s->flags, READ_ONCE(s->flags) | SLAB_FAILSLAB);
> else
>     WRITE_ONCE(s->flags, READ_ONCE(s->flags) & ~SLAB_FAILSLAB);

Yeah, that would work. Given we are the only writer, we shouldn't even need
a READ_ONCE.

> It got me thinking how many places would break if the compiler
> starts to temporariliy modify the flags - i hope it never does.

That's likely true as well. But the macros have been introduced for this
purpose AFAIK.