Re: [PATCH v5 11/15] KVM: s390: restore IAM in get_ipm() when IPM is clean

[Michael Mueller]

On 07.01.19 00:32, Halil Pasic wrote:
> On Wed, 19 Dec 2018 20:17:52 +0100
> Michael Mueller <mimu@xxxxxxxxxxxxx> wrote:
>
> > The patch adds the parameter irq_flags and allows to
> > restore the Interruption Alert Mask (IAM) in the GISA
> > atomically while guaranteeing the IPM is clean.
> >
> > The function returns the IPM of the GISA. If the returned
> > value is 0x00 and the IRQ_FLAG_IAM was set, the IAM has
> > been restored.
> >
> > New irq flag:
> >    IRQ_FLAG_IAM: When set, the IAM is restored if no ISC bit
> >    		is set in the IPM, i.e. no new airqs are
> > 		pending. The test and restore operations
> > 		are done atomically.
> >
> > Signed-off-by: Michael Mueller <mimu@xxxxxxxxxxxxx>
> > ---
> >   arch/s390/kvm/interrupt.c | 34 +++++++++++++++++++++++++++++-----
> >   1 file changed, 29 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/s390/kvm/interrupt.c b/arch/s390/kvm/interrupt.c
> > index 1cc3ad2e6c7e..8307717e3caf 100644
> > --- a/arch/s390/kvm/interrupt.c
> > +++ b/arch/s390/kvm/interrupt.c
> > @@ -37,6 +37,7 @@
> >   #define IRQ_FLAG_LOCAL    0x8000 /* include local interruption pending mask */
> >   #define IRQ_FLAG_FLOATING 0x4000 /* include float interruption pending mask */
> >   #define IRQ_FLAG_GISA     0x2000 /* include GISA interruption pending mask */
> > +#define IRQ_FLAG_IAM      0x0080 /* when set try to restore IAM */
> >   
> >   #define IRQ_MASK_ALL      (IRQ_FLAG_LOCAL | IRQ_FLAG_FLOATING | IRQ_FLAG_GISA)
> >   #define IRQ_MASK_NO_GISA  (IRQ_MASK_ALL & ~IRQ_FLAG_GISA)
> > @@ -253,9 +254,32 @@ static inline void set_ipm_gisc(struct kvm_s390_gisa *gisa, u32 gisc)
> >   	set_bit_inv(IPM_BIT_OFFSET + gisc, (unsigned long *) gisa);
> >   }
> >   
> > -static inline u8 get_ipm(struct kvm_s390_gisa *gisa)
> > +static inline u8 get_ipm(struct kvm_s390_gisa *gisa, u16 irq_flags)
> >   {
> > -	return READ_ONCE(gisa->ipm);
> > +	u64 word, _word;
> > +	u8 ipm;
> > +
> > +	if (!(irq_flags & IRQ_FLAG_IAM))
> > +		return READ_ONCE(gisa->ipm);
> > +
> > +	do {
> > +		word = READ_ONCE(gisa->u64.word[0]);
> > +		ipm = word >> 24;
> > +		/* If the GISA is in the alert list, return the IPM. */
> > +		if ((u64)gisa != word >> 32)
> > +			return ipm;
> > +		/* If the IPM is dirty, return the IPM. */
> > +		if (ipm)
> > +			return ipm;
> > +		/*
> > +		 * Try to restore the IAM or loop, if the IPM is dirty
> > +		 * again or the GISA has been inserted into the alert list.
> > +		 */
> > +		_word = (word & ~0xffUL) |
> > +			container_of(gisa, struct sie_page2, gisa)->kvm->arch.iam;
> > +	} while (cmpxchg(&gisa->u64.word[0], word, _word) != _word);
>                                                               ^^
>
> Shouldn't this rather be:
> 	} while (cmpxchg(&gisa->u64.word[0], word, _word) != word);

Yes, the comparison needs to be done against the old value of
the memory region.

> Regards,
> Halil
> > +
> > +	return 0;
> >   }