Skip to contentSkip to conversation

virtio: Always reset vhost devices

PATCH In review Hanna Czenczek 2 participants 4 messages Jul 23, 2024 – Aug 1, 2024

lore
Share & tools
lore
#0 lore raw

v2 cover virtio: Always reset vhost devices

Hi,

As explained in patch 2 (the main one) of this series, we currently don’t issue the RESET_DEVICE command to vhost back-ends, even though we fully intend to do so.

The problem is that sending this command is gated behind a vhost_started check, but at that point (during the device reset process), the device is actually stopped, and so vhost_started is false. We still want to send RESET_DEVICE there, so patch 2 removes the vhost_started condition.

This means that we need to be able to call VirtioDeviceClass.get_vhost() when vhost_started is false. For most .get_vhost() implementations, that’s perfectly fine; but three of them (crypto, gpu, net) dereference some pointers, so if any of them is NULL, we have to explicitly return NULL in those implementations. That’s what patch 1 is for.

This time, I’ve run make check with ubsan; I can confirm that v1 generated errors for vhost-net, but with patch 1 added, it’s clean. I’ve also run the CI pipeline:
https://gitlab.com/hreitz/qemu/-/pipelines/1384524949
Specifically, clang-system passed:
https://gitlab.com/hreitz/qemu/-/jobs/7406688234
The only failure is msys2-64bit, which timed out (re-tried repeatedly), but judging from https://gitlab.com/qemu-project/qemu/-/pipelines, I think that’s expected.

v2: Added patch 1, left patch 2 unchanged.

Hanna Czenczek (2):
virtio: Allow .get_vhost() without vhost_started
virtio: Always reset vhost devices

 include/hw/virtio/virtio.h  |  1 +
 hw/display/vhost-user-gpu.c |  2 +-
 hw/net/virtio-net.c         | 19 +++++++++++++++++--
 hw/virtio/virtio-crypto.c   | 18 +++++++++++++++---
 hw/virtio/virtio.c          |  8 ++++++--
 5 files changed, 40 insertions(+), 8 deletions(-)
Signature · 1 line
2.45.2
#1 lore raw

PATCH v2 1/2 virtio: Allow .get_vhost() without vhost_started

Historically, .get_vhost() was probably only called when vdev->vhost_started is true. However, we now decidedly want to call it also when vhost_started is false, specifically so we can issue a reset to the vhost back-end while device operation is stopped.

Some .get_vhost() implementations dereference some pointers (or return offsets from them) that are probably guaranteed to be non-NULL when vhost_started is true, but not necessarily otherwise. This patch makes all such implementations check all such pointers, returning NULL if any is NULL.

---
 include/hw/virtio/virtio.h  |  1 +
 hw/display/vhost-user-gpu.c |  2 +-
 hw/net/virtio-net.c         | 19 +++++++++++++++++--
 hw/virtio/virtio-crypto.c   | 18 +++++++++++++++---
 4 files changed, 34 insertions(+), 6 deletions(-)
include/hw/virtio/virtio.h +1 -0
index 7512afbc84..ff29088232 100644
@@ -221,6 +221,7 @@ struct VirtioDeviceClass {
     int (*post_load)(VirtIODevice *vdev);
     const VMStateDescription *vmsd;
     bool (*primary_unplug_pending)(void *opaque);
+    /* May be called even when vdev->vhost_started is false */
     struct vhost_dev *(*get_vhost)(VirtIODevice *vdev);
     void (*toggle_device_iotlb)(VirtIODevice *vdev);
 };
hw/display/vhost-user-gpu.c +1 -1
index 63c64ddde6..5056ffd618 100644
@@ -642,7 +642,7 @@ vhost_user_gpu_device_realize(DeviceState *qdev, Error **errp)
 static struct vhost_dev *vhost_user_gpu_get_vhost(VirtIODevice *vdev)
 {
     VhostUserGPU *g = VHOST_USER_GPU(vdev);
-    return &g->vhost->dev;
+    return g->vhost ? &g->vhost->dev : NULL;
 }
 
 static Property vhost_user_gpu_properties[] = {
hw/net/virtio-net.c +17 -2
index 8f30972708..874ed9c032 100644
@@ -3890,8 +3890,23 @@ static bool dev_unplug_pending(void *opaque)
 static struct vhost_dev *virtio_net_get_vhost(VirtIODevice *vdev)
 {
     VirtIONet *n = VIRTIO_NET(vdev);
-    NetClientState *nc = qemu_get_queue(n->nic);
-    struct vhost_net *net = get_vhost_net(nc->peer);
+    NetClientState *nc;
+    struct vhost_net *net;
+
+    if (!n->nic) {
+        return NULL;
+    }
+
+    nc = qemu_get_queue(n->nic);
+    if (!nc) {
+        return NULL;
+    }
+
+    net = get_vhost_net(nc->peer);
+    if (!net) {
+        return NULL;
+    }
+
     return &net->dev;
 }
 
hw/virtio/virtio-crypto.c +15 -3
index bbe8aa4b99..3296b7f219 100644
@@ -1247,9 +1247,21 @@ static bool virtio_crypto_guest_notifier_pending(VirtIODevice *vdev, int idx)
 static struct vhost_dev *virtio_crypto_get_vhost(VirtIODevice *vdev)
 {
     VirtIOCrypto *vcrypto = VIRTIO_CRYPTO(vdev);
-    CryptoDevBackend *b = vcrypto->cryptodev;
-    CryptoDevBackendClient *cc = b->conf.peers.ccs[0];
-    CryptoDevBackendVhost *vhost_crypto = cryptodev_get_vhost(cc, b, 0);
+    CryptoDevBackend *b;
+    CryptoDevBackendClient *cc;
+    CryptoDevBackendVhost *vhost_crypto;
+
+    b = vcrypto->cryptodev;
+    if (!b) {
+        return NULL;
+    }
+
+    cc = b->conf.peers.ccs[0];
+    vhost_crypto = cryptodev_get_vhost(cc, b, 0);
+    if (!vhost_crypto) {
+        return NULL;
+    }
+
     return &vhost_crypto->dev;
 }
 
Signature · 1 line
2.45.2
Reviewed-by on v2 1/2
#2 lore raw

Re: [PATCH v2 1/2] virtio: Allow .get_vhost() without vhost_started

Quoted 11 lines · Hanna Czenczek

Historically, .get_vhost() was probably only called when vdev->vhost_started is true. However, we now decidedly want to call it also when vhost_started is false, specifically so we can issue a reset to the vhost back-end while device operation is stopped.

Some .get_vhost() implementations dereference some pointers (or return offsets from them) that are probably guaranteed to be non-NULL when vhost_started is true, but not necessarily otherwise. This patch makes all such implementations check all such pointers, returning NULL if any is NULL.

Quoted 7 lines
---
 include/hw/virtio/virtio.h  |  1 +
 hw/display/vhost-user-gpu.c |  2 +-
 hw/net/virtio-net.c         | 19 +++++++++++++++++--
 hw/virtio/virtio-crypto.c   | 18 +++++++++++++++---
 4 files changed, 34 insertions(+), 6 deletions(-)
#3 lore raw

PATCH v2 2/2 virtio: Always reset vhost devices

In reply to Hanna Czenczek

Requiring vhost_started to be true for resetting vhost devices in virtio_reset() seems like the wrong condition: Most importantly, the preceding virtio_set_status(vdev, 0) call will (for vhost devices) end up in vhost_dev_stop() (through vhost devices' .set_status implementations), setting vdev->vhost_started = false. Therefore, the gated vhost_reset_device() call is unreachable.

vhost_started is not documented, so it is hard to say what exactly it is supposed to mean, but judging from the fact that vhost_dev_start() sets it and vhost_dev_stop() clears it, it seems like it indicates whether there is a vhost back-end, and whether that back-end is currently running and processing virtio requests.

Making a reset conditional on whether the vhost back-end is processing virtio requests seems wrong; in fact, it is probably better to reset it only when it is not currently processing requests, which is exactly the current order of operations in virtio_reset(): First, the back-end is stopped through virtio_set_status(vdev, 0), then we want to send a reset.

Therefore, we should drop the vhost_started condition, but in its stead we then have to verify that we can indeed send a reset to this vhost device, by not just checking k->get_vhost != NULL (introduced by commit 95e1019a4a9), but also that the vhost back-end is connected (hdev = k->get_vhost(); hdev != NULL && hdev->vhost_ops != NULL).

---
 hw/virtio/virtio.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
hw/virtio/virtio.c +6 -2
index 583a224163..35dfc01074 100644
@@ -2150,8 +2150,12 @@ void virtio_reset(void *opaque)
         vdev->device_endian = virtio_default_endian();
     }
 
-    if (vdev->vhost_started && k->get_vhost) {
-        vhost_reset_device(k->get_vhost(vdev));
+    if (k->get_vhost) {
+        struct vhost_dev *hdev = k->get_vhost(vdev);
+        /* Only reset when vhost back-end is connected */
+        if (hdev && hdev->vhost_ops) {
+            vhost_reset_device(hdev);
+        }
     }
 
     if (k->reset) {
Signature · 1 line
2.45.2