Re: [RFC PATCH 05/12] [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit

[Cyril Bur <cyrilbur@gmail.com>]

On Tue, 2018-02-20 at 13:50 +1100, Michael Neuling wrote:
> On Tue, 2018-02-20 at 11:22 +1100, Cyril Bur wrote:
> 
> 
> The comment from the cover sheet should be here
> 
> > ---
> >  arch/powerpc/include/asm/exception-64s.h | 25 +++++++++++++++++++++
> >  arch/powerpc/kernel/entry_64.S           |  5 +++++
> >  arch/powerpc/kernel/process.c            | 37 ++++++++++++++++++++++++++++----
> >  3 files changed, 63 insertions(+), 4 deletions(-)
> > 
> > diff --git a/arch/powerpc/include/asm/exception-64s.h b/arch/powerpc/include/asm/exception-64s.h
> > index 471b2274fbeb..f904f19a9ec2 100644
> > --- a/arch/powerpc/include/asm/exception-64s.h
> > +++ b/arch/powerpc/include/asm/exception-64s.h
> > @@ -35,6 +35,7 @@
> >   * implementations as possible.
> >   */
> >  #include <asm/head-64.h>
> > +#include <asm/tm.h>
> >  
> >  /* PACA save area offsets (exgen, exmc, etc) */
> >  #define EX_R9		0
> > @@ -127,6 +128,26 @@
> >  	hrfid;								\
> >  	b	hrfi_flush_fallback
> >  
> > +#ifdef CONFIG_PPC_TRANSACTIONAL_MEM
> > +#define TM_KERNEL_ENTRY		                                        \
> > +	ld	r3,_MSR(r1);			                        \
> > +	/* Probably don't need to check if coming from user/kernel */	\
> > +	/* If TM is suspended or active then we must have come from*/	\
> > +	/* userspace */							\
> > +	andi.	r0,r3,MSR_PR;						\
> > +	beq	1f;							\
> > +	rldicl. r3,r3,(64-MSR_TS_LG),(64-2); /* SUSPENDED or ACTIVE*/   \
> > +	beql+	1f;                   	/* Not SUSPENDED or ACTIVE */   \
> > +	bl	save_nvgprs;						\
> > +	RECONCILE_IRQ_STATE(r10,r11);					\
> > +	li	r3,TM_CAUSE_MISC;					\
> > +	bl	tm_reclaim_current;	/* uint8 cause		   */	\
> > +1:
> > +
> > +#else /* CONFIG_PPC_TRANSACTIONAL_MEM */
> > +#define TM_KERNEL_ENTRY
> > +#endif /* CONFIG_PPC_TRANSACTIONAL_MEM */
> > +
> >  #ifdef CONFIG_RELOCATABLE
> >  #define __EXCEPTION_RELON_PROLOG_PSERIES_1(label, h)			\
> >  	mfspr	r11,SPRN_##h##SRR0;	/* save SRR0 */			\
> > @@ -675,6 +696,9 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)
> >  	EXCEPTION_PROLOG_COMMON(trap, area);			\
> >  	/* Volatile regs are potentially clobbered here */	\
> >  	additions;						\
> > +	/* This is going to need to go somewhere else as well */\
> > +	/* See comment in tm_recheckpoint()		      */\
> > +	TM_KERNEL_ENTRY;					\
> >  	addi	r3,r1,STACK_FRAME_OVERHEAD;			\
> >  	bl	hdlr;						\
> >  	b	ret
> > @@ -689,6 +713,7 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)
> >  	EXCEPTION_PROLOG_COMMON_3(trap);			\
> >  	/* Volatile regs are potentially clobbered here */	\
> >  	additions;						\
> > +	TM_KERNEL_ENTRY;					\
> >  	addi	r3,r1,STACK_FRAME_OVERHEAD;			\
> >  	bl	hdlr
> >  
> > diff --git a/arch/powerpc/kernel/entry_64.S b/arch/powerpc/kernel/entry_64.S
> > index 2cb5109a7ea3..107c15c6f48b 100644
> > --- a/arch/powerpc/kernel/entry_64.S
> > +++ b/arch/powerpc/kernel/entry_64.S
> > @@ -126,6 +126,11 @@ BEGIN_FW_FTR_SECTION
> >  33:
> >  END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR)
> >  #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE && CONFIG_PPC_SPLPAR */
> > +	TM_KERNEL_ENTRY
> > +	REST_GPR(0,r1)
> > +	REST_4GPRS(3,r1)
> > +	REST_2GPRS(7,r1)
> > +	addi	r9,r1,STACK_FRAME_OVERHEAD
> 
> Why are we doing these restores here now?

The syscall handler expects the syscall params to still be in their
respective regs.

> 
> >  
> >  	/*
> >  	 * A syscall should always be called with interrupts enabled
> > diff --git a/arch/powerpc/kernel/process.c b/arch/powerpc/kernel/process.c
> > index 77dc6d8288eb..ea75da0fd506 100644
> > --- a/arch/powerpc/kernel/process.c
> > +++ b/arch/powerpc/kernel/process.c
> > @@ -951,6 +951,23 @@ void tm_recheckpoint(struct thread_struct *thread)
> >  	if (!(thread->regs->msr & MSR_TM))
> >  		return;
> >  
> > +	/*
> > +	 * This is 'that' comment.
> 
> I think I'm in the loop here but I don't actually know what this means. 
> 
> Senior Mikey moment or Crazy Cyril comments? I'll let the peanut gallery decide.
> 

Oh quite possibly crazy Cyril comment that will have to be...
normalised. I should actually delete this and see if that's still the
case.

> > +	 *
> > +	 * If we get where with tm suspended or active then something
> 
> s/where/here/
> 
> > +	 * has gone wrong. I've added this now as a proof of concept.
> > +	 *
> > +	 * The problem I'm seeing without it is an attempt to
> > +	 * recheckpoint a CPU without a previous reclaim.
> > +	 *
> > +	 * I'm probably missed an exception entry with the
> > +	 * TM_KERNEL_ENTRY macro. Should be easy enough to find.
> > +	 */
> > +	if (MSR_TM_ACTIVE(mfmsr()))
> > +		return;
> 
> I don't really get this.  Wouldn't this test apply now?
> 
> > +
> > +	tm_enable();
> 
> Why did we add this?
> 

Ah yes that was a cleanup I noticed along the way and clearly forgot to
finish.

At the moment there's a bunch of tm_enable()s either before calling
functions like tm_recheckpoint() or tm_reclaim_current() or inside
helpers (tm_reclaim_current() for example again). I feel like callers
shouldn't have to worry, it should be up to the function actually doing
the TM work to enable it.

> > +
> >  	/* We really can't be interrupted here as the TEXASR registers can't
> >  	 * change and later in the trecheckpoint code, we have a userspace R1.
> >  	 * So let's hard disable over this region.
> > @@ -1009,6 +1026,13 @@ static inline void tm_recheckpoint_new_task(struct task_struct *new)
> >  static inline void __switch_to_tm(struct task_struct *prev,
> >  		struct task_struct *new)
> >  {
> > +	/*
> > +	 * So, with the rework none of this code should not be needed.
> > +	 * I've left in the reclaim for now. This *should* save us
> > +	 * from any mistake in the new code. Also the
> > +	 * enabling/disabling logic of MSR_TM really should be
> > +	 * refactored into a common way with MSR_{FP,VEC,VSX}
> > +	 */
> >  	if (cpu_has_feature(CPU_FTR_TM)) {
> >  		if (tm_enabled(prev) || tm_enabled(new))
> >  			tm_enable();
> > @@ -1016,11 +1040,14 @@ static inline void __switch_to_tm(struct task_struct *prev,
> >  		if (tm_enabled(prev)) {
> >  			prev->thread.load_tm++;
> >  			tm_reclaim_task(prev);
> > -			if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0)
> > -				prev->thread.regs->msr &= ~MSR_TM;
> > +			/*
> > +			 * The disabling logic may be confused don't
> > +			 * disable for now
> > +			 *
> > +			 * if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0)
> > +			 *	prev->thread.regs->msr &= ~MSR_TM;
> > +			 */
> 
> Why are you doing this when you just remove all this code in the next patch?

The next 3 or so patches will need squashing into this one before
merging.

> >  		}
> > -
> > -		tm_recheckpoint_new_task(new);
> >  	}
> >  }
> >  
> > @@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs)
> >  	msr_diff = current->thread.ckpt_regs.msr & ~regs->msr;
> >  	msr_diff &= MSR_FP | MSR_VEC | MSR_VSX;
> >  
> > +	tm_recheckpoint(&current->thread);
> > +
> 
> So why do we do tm_recheckpoint at all? Shouldn't most of the tm_blah code go
> away in process.c after all this?
> 

I'm not sure I follow, we need to recheckpoint because we're going back
to userspace? Or would you rather calling the tm.S code directly from
the exception return path?

Yes, I hope we'll be able to have a fairly big cleanup commit of tm_
code in process.c at the end of this series.

> >  	/* Ensure that restore_math() will restore */
> >  	if (msr_diff & MSR_FP)
> >  		current->thread.load_fp = 1;