On Sun, 17 Jul 2016 11:25:43 +0800
Simon Guo <simon...@linux.vnet.ibm.com> wrote:
> Hi Cyril,
> On Wed, Jun 08, 2016 at 02:00:34PM +1000, Cyril Bur wrote:
> > @@ -917,24 +907,10 @@ static inline void tm_recheckpoint_new_task(struct
> > task_struct *new)
> > "(new->msr 0x%lx, new->origmsr 0x%lx)\n",
> > new->pid, new->thread.regs->msr, msr);
> >
> > - /* This loads the checkpointed FP/VEC state, if used */
> > tm_recheckpoint(&new->thread, msr);
> >
> > - /* This loads the speculative FP/VEC state, if used */
> > - if (msr & MSR_FP) {
> > - do_load_up_transact_fpu(&new->thread);
> > - new->thread.regs->msr |=
> > - (MSR_FP | new->thread.fpexc_mode);
> > - }
> > -#ifdef CONFIG_ALTIVEC
> > - if (msr & MSR_VEC) {
> > - do_load_up_transact_altivec(&new->thread);
> > - new->thread.regs->msr |= MSR_VEC;
> > - }
> > -#endif
> > - /* We may as well turn on VSX too since all the state is restored now */
> > - if (msr & MSR_VSX)
> > - new->thread.regs->msr |= MSR_VSX;
> > + /* Won't restore math get called later? */
> > + restore_math(new->thread.regs);
>
> I have some question for the "restore_math" in tm_recheckpoint_new_task().
>
Please ask!
> Per my understanding, now after tm_recheckpoint, the fp_state content
> is obsolete. However restore_math() will try to restore FP/Vec/VSX state
> from fp_state, (orginally it is right, since fp_state was the valid
> checkpointed state and consistent with FP register ). Should we remove
> the restore_math() here?
>
The aim of this patch is to ensure that pt_regs, fp_state and vr_state always
hold a threads 'live' registers. So, after a recheckpoint fp_state is where the
the state should be. tm_reclaim_thread() does a save_all() before doing the
reclaim.
This means that the call to restore_math() is a replacement for all deleted
lines above it.
I added it here because I'd prefer to be safe but I left that comment in
because I suspect restore_math() will be called later and we can get away with
not calling it here.
> And, should the thread's MSR now set FP bit in tm_recheckpoint(), to
> indicate that FP register content is "fresh" in contrast to thread.fp_state?
>
I'm not sure what you mean by 'fresh'. You do highlight that we'll have to be
sure that the MSR bits are off (so that restore_math() doesn't assume the
registers are already loaded) which makes me think that tm_reclaim_thread()
should be doing a giveup_all(), I'll fix that.
I hope that helps,
Cyril
> Please correct me if I am wrong.
>
> Thanks,
> - Simon
>
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
