[PATCH V2 4/4] riscv: mm: Optimize TASK_SIZE definition

Guo Ren guoren at kernel.org
Sat Dec 23 17:24:00 PST 2023


On Sat, Dec 23, 2023 at 6:31 PM David Laight <David.Laight at aculab.com> wrote:
>
> From: Guo Ren
> > Sent: 23 December 2023 02:53
> >
> > On Fri, Dec 22, 2023 at 7:52 PM David Laight <David.Laight at aculab.com> wrote:
> > >
> > > From: Guo Ren
> > > > Sent: 22 December 2023 11:25
> > > ...
> > > > > > +#define TASK_SIZE    (is_compat_task() ? \
> > > > > >                        TASK_SIZE_32 : TASK_SIZE_64)
> > > > I would remove is_compat_task() in the next version because your patch
> > > > contains that.
> > >
> > > Does TASK_SIZE get used in access_ok() ?
> > > If so the repeated expansion of that 'mess' will slow things down.
> > >
> > > OTOH access_ok(ptr, len) can just check (ptr | (ptr + len)) < 0)
> > > and rely on the page faults for everything else.
> > Or do you want to discuss the bad side effect of is_compat_task()?
> >
> > Yes, test_thread_flag(TIF_32BIT) would slow down access_ok(). But if
> > we use TASK_SIZE_MAX, VA_BITS still needs pgtable_l5_enabled,
> > pgtable_l4_enabled detectation for riscv.
> >
> > It's not only for compat mode, but also Sv39, Sv48, Sv57. All treat
> > TASK_SIZE_MAX as 0x8000000000000000, right? Then:
> > access_ok(ptr, len) can just check (ptr | (ptr + len)) < 0)
> >
> > It's another feature and does not relate to compat mode.
>
> Compat mode just makes it worse...
It's hard to observe.

>
> One possibility would be to save the task's max user address
> in the task structure itself - that would save all the conditionals
> at a cost of an extra value in the task structure.
It would still cause memory load operation, although it is $tp->xxx.
If we want to gain observability benefits, "just check (ptr | (ptr +
len)) < 0)" is better.

>
> There is also the question of whether a normally 64-bit task
> can actually make the compat mmap() system call?
No.

> On x86 that is certainly possible (IIRC wine does it), x86
> userspace can flip between 32bit and 63bit mode without a
> system call.
RISC-V can't do that because it needs sstatux.uxl=32/64, which can
only be modified in S-mode.


>
>         David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)



-- 
Best Regards
 Guo Ren



More information about the linux-riscv mailing list