Date: Thu, 2 Apr 2015 18:45:27 +0200
From: Greg Kurz <gkurz@linux.vnet.ibm.com>
To: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>, linux-api@vger.kernel.org,
        linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        virtualization@lists.linux-foundation.org
Subject: Re: [PATCH v2 7/7] vhost: feature to set the vring endianness
Message-ID: <20150402184527.0bc68efe@bahia.local>
In-Reply-To: <20150402142046.GA23121@redhat.com>
References: <20150402131637.24676.60700.stgit@bahia.lab.toulouse-stg.fr.ibm.com>
	<20150402131713.24676.9924.stgit@bahia.lab.toulouse-stg.fr.ibm.com>
	<20150402142046.GA23121@redhat.com>
Organization: IBM
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 9138
Lines: 276

On Thu, 2 Apr 2015 16:20:46 +0200
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Thu, Apr 02, 2015 at 03:17:13PM +0200, Greg Kurz wrote:
> > This patch brings cross-endian support to vhost when used to implement
> > legacy virtio devices. Since it is a relatively rare situation, the feature
> > availability is controlled by a kernel config option (not set by default).
> > 
> > If cross-endian support is compiled in, vhost abvertises a new feature
> > to be negotiated with userspace. If userspace acknowledges the feature,
> > it can inform vhost about the endianness to use with a new ioctl.
> > 
> > This feature is mutually exclusive with virtio 1.0. Even if the vhost device
> > advertises virtio 1.0 and legacy cross-endian features, it cannot receive
> > aknowledgement for both at the same time.
> > 
> > Hot paths are being preserved from any penalty when the config option is
> > disabled or when virtio 1.0 is being used.
> > 
> > Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
> > ---
> >  drivers/vhost/Kconfig      |   10 ++++++++++
> >  drivers/vhost/net.c        |    5 +++++
> >  drivers/vhost/scsi.c       |    4 ++++
> >  drivers/vhost/test.c       |    4 ++++
> >  drivers/vhost/vhost.c      |   19 +++++++++++++++++++
> >  drivers/vhost/vhost.h      |   13 ++++++++++++-
> >  include/uapi/linux/vhost.h |   10 ++++++++++
> >  7 files changed, 64 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/vhost/Kconfig b/drivers/vhost/Kconfig
> > index 017a1e8..5bb8da9 100644
> > --- a/drivers/vhost/Kconfig
> > +++ b/drivers/vhost/Kconfig
> > @@ -32,3 +32,13 @@ config VHOST
> >  	---help---
> >  	  This option is selected by any driver which needs to access
> >  	  the core of vhost.
> > +
> > +config VHOST_SET_ENDIAN_LEGACY
> > +	bool "Cross-endian support for host kernel accelerator"
> > +	default n
> > +	---help---
> > +	  This option allows vhost to support guests with a different byte
> > +	  ordering
> 
> from host
> 
> >. It is disabled by default since it adds overhead and it
> > +	  is only needed by a few platforms (powerpc and arm).
> > +
> > +	  If unsure, say "n".
> 
> "N"
> 
> > diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> > index 2bbfc25..5274a44 100644
> > --- a/drivers/vhost/net.c
> > +++ b/drivers/vhost/net.c
> > @@ -1011,6 +1011,11 @@ static int vhost_net_set_features(struct vhost_net *n, u64 features)
> >  		vhost_hlen = 0;
> >  		sock_hlen = hdr_len;
> >  	}
> > +
> > +	if (features & ((1ULL << VHOST_F_SET_ENDIAN_LEGACY) |
> > +			(1ULL << VIRTIO_F_VERSION_1)))
> > +		return -EINVAL;
> > +
> >  	mutex_lock(&n->dev.mutex);
> >  	if ((features & (1 << VHOST_F_LOG_ALL)) &&
> >  	    !vhost_log_access_ok(&n->dev)) {
> > diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> > index 71df240..b53e9c2 100644
> > --- a/drivers/vhost/scsi.c
> > +++ b/drivers/vhost/scsi.c
> > @@ -1544,6 +1544,10 @@ static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
> >  	if (features & ~VHOST_SCSI_FEATURES)
> >  		return -EOPNOTSUPP;
> >  
> > +	if (features & ((1ULL << VHOST_F_SET_ENDIAN_LEGACY) |
> > +			(1ULL << VIRTIO_F_VERSION_1)))
> > +		return -EINVAL;
> > +
> >  	mutex_lock(&vs->dev.mutex);
> >  	if ((features & (1 << VHOST_F_LOG_ALL)) &&
> >  	    !vhost_log_access_ok(&vs->dev)) {
> > diff --git a/drivers/vhost/test.c b/drivers/vhost/test.c
> > index d9c501e..cfefdad 100644
> > --- a/drivers/vhost/test.c
> > +++ b/drivers/vhost/test.c
> > @@ -243,6 +243,10 @@ static int vhost_test_set_features(struct vhost_test *n, u64 features)
> >  {
> >  	struct vhost_virtqueue *vq;
> >  
> > +	if (features & ((1ULL << VHOST_F_SET_ENDIAN_LEGACY) |
> > +			(1ULL << VIRTIO_F_VERSION_1)))
> > +		return -EINVAL;
> > +
> >  	mutex_lock(&n->dev.mutex);
> >  	if ((features & (1 << VHOST_F_LOG_ALL)) &&
> >  	    !vhost_log_access_ok(&n->dev)) {
> 
> Above seems to prevent users from specifying either
> VIRTIO_F_VERSION_1 or VHOST_F_SET_ENDIAN_LEGACY.
> Does it actually work?
> 

Arggg no it doesn't *of course*... I've added these bogus checks lately
and didn't re-test :(

> > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > index 2ee2826..60a0f32 100644
> > --- a/drivers/vhost/vhost.c
> > +++ b/drivers/vhost/vhost.c
> > @@ -199,6 +199,7 @@ static void vhost_vq_reset(struct vhost_dev *dev,
> >  	vq->call = NULL;
> >  	vq->log_ctx = NULL;
> >  	vq->memory = NULL;
> > +	vq->legacy_big_endian = false;
> >  }
> >  
> >  static int vhost_worker(void *data)
> > @@ -806,6 +807,24 @@ long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
> >  		} else
> >  			filep = eventfp;
> >  		break;
> > +#ifdef CONFIG_VHOST_SET_ENDIAN_LEGACY
> > +	case VHOST_SET_VRING_ENDIAN_LEGACY:
> > +	{
> > +		struct vhost_vring_endian e;
> > +
> > +		if (!vhost_has_feature(vq, VHOST_F_SET_ENDIAN_LEGACY)) {
> > +			r = -EINVAL;
> > +			break;
> > +		}
> > +
> > +		if (copy_from_user(&e, argp, sizeof(e))) {
> > +			r = -EFAULT;
> > +			break;
> > +		}
> > +		vq->legacy_big_endian = e.is_big_endian;
> > +		break;
> > +	}
> > +#endif
> >  	default:
> >  		r = -ENOIOCTLCMD;
> >  	}
> > diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
> > index 4e9a186..d50881d 100644
> > --- a/drivers/vhost/vhost.h
> > +++ b/drivers/vhost/vhost.h
> > @@ -106,6 +106,9 @@ struct vhost_virtqueue {
> >  	/* Log write descriptors */
> >  	void __user *log_base;
> >  	struct vhost_log *log;
> > +
> > +	/* We need to know the device endianness with legacy virtio. */
> > +	bool legacy_big_endian;
> >  };
> >  
> >  struct vhost_dev {
> > @@ -165,7 +168,11 @@ enum {
> >  	VHOST_FEATURES = (1ULL << VIRTIO_F_NOTIFY_ON_EMPTY) |
> >  			 (1ULL << VIRTIO_RING_F_INDIRECT_DESC) |
> >  			 (1ULL << VIRTIO_RING_F_EVENT_IDX) |
> > -			 (1ULL << VHOST_F_LOG_ALL),
> > +			 (1ULL << VHOST_F_LOG_ALL) |
> > +#ifdef CONFIG_VHOST_SET_ENDIAN_LEGACY
> > +			 (1ULL << VHOST_F_SET_ENDIAN_LEGACY) |
> > +#endif
> > +			 0ULL,
> >  };
> >  
> >  static inline bool vhost_has_feature(struct vhost_virtqueue *vq, int bit)
> > @@ -177,6 +184,10 @@ static inline bool vhost_is_little_endian(struct vhost_virtqueue *vq)
> >  {
> >  	if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
> >  		return true;
> > +#ifdef CONFIG_VHOST_SET_ENDIAN_LEGACY
> > +	if (vhost_has_feature(vq, VHOST_F_SET_ENDIAN_LEGACY))
> > +		return !vq->legacy_big_endian;
> 
> why do we need to check the feature bit?
> How about simple
> 	return !vq->legacy_big_endian;
> here?

This is a runtime feature. Even for powerpc, cross-endian won't be the
most common scenario. Userspace may have cleared the bit if it doesn't
need/want it.

> All you need to do is set legacy_big_endian to
> !virtio_legacy_is_little_endian() on reset.
> Maybe rename to legacy_is_little_endian so you don't
> need to reverse the logic.
> 

Do you mean vhost doesn't need userspace to provide the endianness
to be used ? Please elaborate on the meaning of "reset" here... I
am confused.

> > +#endif
> >  	return virtio_legacy_is_little_endian();
> >  }
> >  
> > diff --git a/include/uapi/linux/vhost.h b/include/uapi/linux/vhost.h
> > index bb6a5b4..09d2a48 100644
> > --- a/include/uapi/linux/vhost.h
> > +++ b/include/uapi/linux/vhost.h
> > @@ -47,6 +47,11 @@ struct vhost_vring_addr {
> >  	__u64 log_guest_addr;
> >  };
> >  
> > +struct vhost_vring_endian {
> > +	unsigned int index;
> > +	bool is_big_endian;
> 
> bool in uapi is a bad idea.
> Generally, I think you can use vhost_vring_state here.
> 

Ok to turn is_big_endian into an int but why hijacking vhost_vring_state ?

> > +};
> > +
> >  struct vhost_memory_region {
> >  	__u64 guest_phys_addr;
> >  	__u64 memory_size; /* bytes */
> > @@ -103,6 +108,9 @@ struct vhost_memory {
> >  /* Get accessor: reads index, writes value in num */
> >  #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
> >  
> > +/* Set endianness for the ring (legacy virtio only) */
> > +#define VHOST_SET_VRING_ENDIAN_LEGACY _IOW(VHOST_VIRTIO, 0x13, struct vhost_vring_endian)
> > +
> >  /* The following ioctls use eventfd file descriptors to signal and poll
> >   * for events. */
> >  
> 
> You also need a GET ioctl, as a matter of policy.
> 

Ok. Will add.

> > @@ -126,6 +134,8 @@ struct vhost_memory {
> >  #define VHOST_F_LOG_ALL 26
> >  /* vhost-net should add virtio_net_hdr for RX, and strip for TX packets. */
> >  #define VHOST_NET_F_VIRTIO_NET_HDR 27
> > +/* the vring endianness can be explicitly set (legacy virtio only). */
> > +#define VHOST_F_SET_ENDIAN_LEGACY 28
> >  
> >  /* VHOST_SCSI specific definitions */
> 
> 
> VHOST_F_SET_ENDIAN_LEGACY doesn't seem too useful.
> Is this so userspace can detect kernel configuration?

Yes. Also, it is quite easy to "un-ack" the feature when
it is not wanted.

> I think probing VHOST_SET_VRING_ENDIAN_LEGACY should
> be sufficient for this.
> 

I'll get rid of the feature.

Thanks for your time Michael.

--
Greg

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/