Re: [PATCH V3 1/2] MIPS: Loongson-3: Enable COP2 usage in kernel

[Thomas Bogendoerfer <tsbogend@alpha.franken.de>]

On Wed, Sep 02, 2020 at 02:54:10PM +0800, Huacai Chen wrote:
> Hi, Thomas,
> 
> On Wed, Aug 26, 2020 at 8:48 PM Thomas Bogendoerfer
> <tsbogend@alpha.franken.de> wrote:
> >
> > On Fri, Aug 14, 2020 at 05:44:18PM +0800, Huacai Chen wrote:
> > > On Tue, Aug 11, 2020 at 8:08 PM Thomas Bogendoerfer
> > > <tsbogend@alpha.franken.de> wrote:
> > > > this comes with it's own memcpy/memset and stuff, I don't see a reason why
> > > > COP2 needs to be enabled there,
> > > gslq/gssq can also be generated by toolchains.
> >
> > I don't want to introduce every single CPU optimization bits into such
> > a closed first stage loader. So please use $(filter-out) in
> > arch/mips/boot/compressed/Makefile to disable creation of 16byte load/stores.
> >
> > > > which is correct for all user space process, otherwise the whole
> > > > cop2 exception thing wouldn't work. And if cop2 exception handling
> > > > has been run it's set in THREAD_STATUS.
> > > >
> > > THREAD_STATUS means thread_struct.cp0_status, which is the cp0_status
> > > when a process runs in kernel-space. KSTK_STATUS (what you have seen
> > > in copy_thread_tls() below) means cp0_status in a process's kernel
> > > stack, which saves the cp0_status when a process runs in user-space.
> > > Whether COP2 exception can work depends on that KSTK_STATUS (but not
> > > THREAD_STATUS) should not contain CU2 at the first time. So, whether
> > > or not THREAD_STATUS contains CU2, it won't break COP2 handling.
> >
> > so why don't we fix the the in-kernel cp0_status instead ?
> >
> > How about this ?
> >
> > diff --git a/arch/mips/kernel/process.c b/arch/mips/kernel/process.c
> > index 90b869297893..26fb77a8d406 100644
> > --- a/arch/mips/kernel/process.c
> > +++ b/arch/mips/kernel/process.c
> > @@ -133,6 +133,7 @@ int copy_thread_tls(unsigned long clone_flags, unsigned long usp,
> >         /*  Put the stack after the struct pt_regs.  */
> >         childksp = (unsigned long) childregs;
> >         p->thread.cp0_status = read_c0_status() & ~(ST0_CU2|ST0_CU1);
> > +       p->thread.cp0_status |= ST0_KERNEL_CUMASK;
> >         if (unlikely(p->flags & PF_KTHREAD)) {
> >                 /* kernel thread */
> >                 unsigned long status = p->thread.cp0_status;
> I tried this way but it doesn't work, the reason is that the resume
> routine in r4k_switch.S save the current hardware status into
> THREAD_STATUS, but CU2 in hardware is cleared in its caller (i.e.,
> switch_to).

so let's fix it there:

diff --git a/arch/mips/include/asm/switch_to.h b/arch/mips/include/asm/switch_to.h
index 0b0a93bf83cd..a4374b4cb88f 100644
--- a/arch/mips/include/asm/switch_to.h
+++ b/arch/mips/include/asm/switch_to.h
@@ -117,6 +117,8 @@ do {                                                                        \
                __restore_dsp(next);                                    \
        }                                                               \
        if (cop2_present) {                                             \
+               u32 status = read_c0_status();                          \
+                                                                       \
                set_c0_status(ST0_CU2);                                 \
                if ((KSTK_STATUS(prev) & ST0_CU2)) {                    \
                        if (cop2_lazy_restore)                          \
@@ -127,7 +129,7 @@ do {                                                                        \
                    !cop2_lazy_restore) {                               \
                        cop2_restore(next);                             \
                }                                                       \
-               clear_c0_status(ST0_CU2);                               \
+               write_c0_status(status);                                \
        }                                                               \
        __clear_r5_hw_ll_bit();                                         \
        __clear_software_ll_bit();                                      \

BTW. if we come up to a final solution, this change should be a seperate
patch. And the change in process.c probably, too.

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]