Re: [Qemu-devel] [PATCH v5 6/6] KVM: s390: Wire up ioeventfd.

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Mar 05, 2013 at 09:13:08AM +0100, Cornelia Huck wrote:
> On Mon, 4 Mar 2013 21:07:41 -0300
> Marcelo Tosatti <mtosatti@redhat.com> wrote:
> 
> > On Thu, Feb 28, 2013 at 12:33:21PM +0100, Cornelia Huck wrote:
> > > Enable ioeventfd support on s390 and hook up diagnose 500 virtio-ccw
> > > notifications.
> > > 
> > > Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
> > > ---
> > >  arch/s390/kvm/Kconfig    |  1 +
> > >  arch/s390/kvm/Makefile   |  2 +-
> > >  arch/s390/kvm/diag.c     | 26 ++++++++++++++++++++++++++
> > >  arch/s390/kvm/kvm-s390.c |  1 +
> > >  4 files changed, 29 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/s390/kvm/Kconfig b/arch/s390/kvm/Kconfig
> > > index b58dd86..3c43e30 100644
> > > --- a/arch/s390/kvm/Kconfig
> > > +++ b/arch/s390/kvm/Kconfig
> > > @@ -22,6 +22,7 @@ config KVM
> > >  	select PREEMPT_NOTIFIERS
> > >  	select ANON_INODES
> > >  	select HAVE_KVM_CPU_RELAX_INTERCEPT
> > > +	select HAVE_KVM_EVENTFD
> > >  	---help---
> > >  	  Support hosting paravirtualized guest machines using the SIE
> > >  	  virtualization capability on the mainframe. This should work
> > > diff --git a/arch/s390/kvm/Makefile b/arch/s390/kvm/Makefile
> > > index 3975722..8fe9d65 100644
> > > --- a/arch/s390/kvm/Makefile
> > > +++ b/arch/s390/kvm/Makefile
> > > @@ -6,7 +6,7 @@
> > >  # it under the terms of the GNU General Public License (version 2 only)
> > >  # as published by the Free Software Foundation.
> > >  
> > > -common-objs = $(addprefix ../../../virt/kvm/, kvm_main.o)
> > > +common-objs = $(addprefix ../../../virt/kvm/, kvm_main.o eventfd.o)
> > >  
> > >  ccflags-y := -Ivirt/kvm -Iarch/s390/kvm
> > >  
> > > diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
> > > index a390687..1c01a99 100644
> > > --- a/arch/s390/kvm/diag.c
> > > +++ b/arch/s390/kvm/diag.c
> > > @@ -13,6 +13,7 @@
> > >  
> > >  #include <linux/kvm.h>
> > >  #include <linux/kvm_host.h>
> > > +#include <asm/virtio-ccw.h>
> > >  #include "kvm-s390.h"
> > >  #include "trace.h"
> > >  #include "trace-s390.h"
> > > @@ -104,6 +105,29 @@ static int __diag_ipl_functions(struct kvm_vcpu *vcpu)
> > >  	return -EREMOTE;
> > >  }
> > >  
> > > +static int __diag_virtio_hypercall(struct kvm_vcpu *vcpu)
> > > +{
> > > +	int ret, idx;
> > > +
> > > +	/* No virtio-ccw notification? Get out quickly. */
> > > +	if (!vcpu->kvm->arch.css_support ||
> > > +	    (vcpu->run->s.regs.gprs[1] != KVM_S390_VIRTIO_CCW_NOTIFY))
> > > +		return -EOPNOTSUPP;
> > > +
> > > +	idx = srcu_read_lock(&vcpu->kvm->srcu);
> > > +	/*
> > > +	 * The layout is as follows:
> > > +	 * - gpr 2 contains the subchannel id (passed as addr)
> > > +	 * - gpr 3 contains the virtqueue index (passed as datamatch)
> > > +	 */
> > > +	ret = kvm_io_bus_write(vcpu->kvm, KVM_VIRTIO_CCW_NOTIFY_BUS,
> > > +				vcpu->run->s.regs.gprs[2],
> > > +				8, &vcpu->run->s.regs.gprs[3]);
> > > +	srcu_read_unlock(&vcpu->kvm->srcu, idx);
> > > +	/* kvm_io_bus_write returns -EOPNOTSUPP if it found no match. */
> > > +	return ret < 0 ? ret : 0;
> > > +}
> > > +
> > 
> > What about the cookie?
> 
> This is the not-performance-optimized version which went through our
> testing and which I'd like to see applied. I also have a not-yet-tested
> patch that adds cookie support (see below), but I'd like to postpone
> this to 3.10.

I agree it's best to delay this optimization, focus on basic
functionality now. As guest with cookie support is
compatible with host without said support (all we do
really is reserve a gpr), there won't be issues adding it later.
Some comments on the draft below.


> commit 8fce4a8a3478252f7ee2eb74d2732673f5911120
> Author: Cornelia Huck <cornelia.huck@de.ibm.com>
> Date:   Tue Feb 26 17:05:01 2013 +0100
> 
>     virtio-ccw: ioeventfd: use cookies
>     
>     Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
> 
> diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
> index 1c01a99..bd2a013 100644
> --- a/arch/s390/kvm/diag.c
> +++ b/arch/s390/kvm/diag.c
> @@ -108,22 +108,29 @@ static int __diag_ipl_functions(struct kvm_vcpu *vcpu)
>  static int __diag_virtio_hypercall(struct kvm_vcpu *vcpu)
>  {
>  	int ret, idx;
> +	long cookie;
>  
>  	/* No virtio-ccw notification? Get out quickly. */
>  	if (!vcpu->kvm->arch.css_support ||
>  	    (vcpu->run->s.regs.gprs[1] != KVM_S390_VIRTIO_CCW_NOTIFY))
>  		return -EOPNOTSUPP;
>  
> +	cookie = vcpu->run->s.regs.gprs[4];

Hmm signed so this can become negative.

>  	idx = srcu_read_lock(&vcpu->kvm->srcu);
>  	/*
>  	 * The layout is as follows:
>  	 * - gpr 2 contains the subchannel id (passed as addr)
>  	 * - gpr 3 contains the virtqueue index (passed as datamatch)
> +	 * - gpr 4 contains the index on the bus (optionally)
>  	 */
> -	ret = kvm_io_bus_write(vcpu->kvm, KVM_VIRTIO_CCW_NOTIFY_BUS,
> -				vcpu->run->s.regs.gprs[2],
> -				8, &vcpu->run->s.regs.gprs[3]);
> +	ret = kvm_io_bus_write_cookie(vcpu->kvm, KVM_VIRTIO_CCW_NOTIFY_BUS,
> +				      vcpu->run->s.regs.gprs[2],
> +				      8, &vcpu->run->s.regs.gprs[3],
> +				      &cookie);
>  	srcu_read_unlock(&vcpu->kvm->srcu, idx);
> +	if (!ret)
> +		/* Return cookie in gpr 2. */
> +		vcpu->run->s.regs.gprs[2] = cookie;

We should also set gprs[2] to something sane if exiting to userspace
so avoid adding overhead for that case.
Probably # of devices + 1 or a negative value.
Are there 32 bit guests? If yes can they pass a negative value?

>  	/* kvm_io_bus_write returns -EOPNOTSUPP if it found no match. */
>  	return ret < 0 ? ret : 0;
>  }
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index 206247f..2f65a3a 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -154,6 +154,8 @@ enum kvm_bus {
>  
>  int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  		     int len, const void *val);
> +int kvm_io_bus_write_cookie(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
> +			    int len, const void *val, long *cookie);
>  int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, int len,
>  		    void *val);
>  int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 7c188a3..b77ea5f 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -2710,6 +2710,51 @@ int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  	return -EOPNOTSUPP;
>  }
>  
> +/* kvm_io_bus_write_cookie - called under kvm->slots_lock */
> +int kvm_io_bus_write_cookie(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
> +			    int len, const void *val, long *cookie)
> +{
> +	int idx, ret = -EOPNOTSUPP;
> +	struct kvm_io_bus *bus;
> +	struct kvm_io_range range;
> +
> +	range = (struct kvm_io_range) {
> +		.addr = addr,
> +		.len = len,
> +	};
> +
> +	if (!cookie)
> +		return -EINVAL;
> +

Hmm does this ever trigger?

> +	bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
> +
> +	/* First try the device referenced by *cookie. */
> +	if (kvm_io_bus_sort_cmp(&range, &bus->range[*cookie]) == 0)

We need to range-check the cookie.

> +		if (!kvm_iodevice_write(bus->range[*cookie].dev, addr, len,
> +					val))
> +			return 0;
> +
> +	/*
> +	 * *cookie contained garbage; fall back to search and return the
> +	 * correct value in *cookie.
> +	 */
> +	idx = kvm_io_bus_get_first_dev(bus, addr, len);
> +	if (idx < 0)
> +		goto out;
> +
> +	while (idx < bus->dev_count &&
> +		kvm_io_bus_sort_cmp(&range, &bus->range[idx]) == 0) {
> +		if (!kvm_iodevice_write(bus->range[idx].dev, addr, len, val)) {
> +			ret = 0;
> +			goto out;
> +		}
> +		idx++;
> +	}
> +out:
> +	*cookie = idx;
> +	return ret;
> +}
> +
>  /* kvm_io_bus_read - called under kvm->slots_lock */
>  int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
>  		    int len, void *val)