Re: [PATCH v3 2/6] futex2: Implement vectorized wait

[Gabriel Krisman Bertazi]

André Almeida <andrealmeid@xxxxxxxxxxxxx> writes:

>>> +/**
>>> + * struct futex_waitv - A waiter for vectorized wait
>>> + * @val:	Expected value at uaddr
>>> + * @uaddr:	User address to wait on
>>> + * @flags:	Flags for this waiter
>>> + * @__reserved:	Reserved member to preserve data alignment. Should be 0.
>>> + */
>>> +struct futex_waitv {
>>> +	__u64 val;
>>> +	__u64 uaddr;
>>> +	__u32 flags;
>>> +	__u32 __reserved;
>>> +};
>> 
>> why force uaddr  to be __u64, even for 32-bit?  uaddr could be a (void*) for
>> all we care, no?  Also, by adding a reserved field, you are wasting 32
>> bits even on 32-bit architectures.
>> 
>
> We do that to make the structure layout compatible with both entry
> points, remove the need for special cast and duplicated code, as
> suggested by Thomas and Arnd:
>
> https://lore.kernel.org/lkml/87v94310gm.ffs@tglx/
>
> https://lore.kernel.org/lkml/CAK8P3a0MO1qJLRkCH8KrZ3+=L66KOsMRmcbrUvYdMoKykdKoyQ@xxxxxxxxxxxxxx/

I find this weird.  I'm not even juts talking about compat, but even on
native 32-bit. But also, 32 applications on 64, which is a big use
case for games.

The structure is mandating a 64 bit uaddr field and has an unnecessary
pad.  You are wasting 20% of the space, which is gonna be elements of a
vector coming from user space.  Worst case, you are doing copy_from_user
of an extra 1k bytes in the critical path of futex_waitv for no good
reason.

Also, if I understand correctly, Arnd suggestion, at least, was to have
two parser functions and a single syscall entry point, that would do the
translation:

if (in_compat_syscall())
   futex_parse_waitv_compat(futexv, waiters, nr_futexes);
else
   futex_parse_waitv(futexv, waiters, nr_futexes);

-- 
Gabriel Krisman Bertazi