Patchwork [2/2] virtio-blk: disable write cache if not negotiated

login
register
mail settings
Submitter Paolo Bonzini
Date July 3, 2012, 1:20 p.m.
Message ID <1341321642-24598-3-git-send-email-pbonzini@redhat.com>
Download mbox | patch
Permalink /patch/168810/
State New
Headers show

Comments

Paolo Bonzini - July 3, 2012, 1:20 p.m.
If the guest does not support flushes, we should run in writethrough mode.
The setting is temporary until the next reset, so that for example the
BIOS will run in writethrough mode while Linux will run with a writeback
cache.

VIRTIO_BLK_F_FLUSH has been introduced in Linux 2.6.32 (in 2009) and
was backported to RHEL/CentOS 5.6 (in 2010).  The Windows drivers have
two bugs, which I reported on the Red Hat Bugzilla as bugs 837321 and
837324.  With these patches they will suffer a performance hit but
gain correctness.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 hw/virtio-blk.c |   14 ++++++++++++++
 1 file changed, 14 insertions(+)
Kevin Wolf - July 3, 2012, 1:49 p.m.
Am 03.07.2012 15:20, schrieb Paolo Bonzini:
> If the guest does not support flushes, we should run in writethrough mode.
> The setting is temporary until the next reset, so that for example the
> BIOS will run in writethrough mode while Linux will run with a writeback
> cache.
> 
> VIRTIO_BLK_F_FLUSH has been introduced in Linux 2.6.32 (in 2009) and
> was backported to RHEL/CentOS 5.6 (in 2010).  The Windows drivers have
> two bugs, which I reported on the Red Hat Bugzilla as bugs 837321 and
> 837324.  With these patches they will suffer a performance hit but
> gain correctness.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

I generally like the idea for a default, but doesn't this override even
an explicit cache=writeback? Are we sure that we want this?

Kevin
Paolo Bonzini - July 3, 2012, 1:51 p.m.
Il 03/07/2012 15:49, Kevin Wolf ha scritto:
>> If the guest does not support flushes, we should run in writethrough mode.
>> > The setting is temporary until the next reset, so that for example the
>> > BIOS will run in writethrough mode while Linux will run with a writeback
>> > cache.
>> > 
>> > VIRTIO_BLK_F_FLUSH has been introduced in Linux 2.6.32 (in 2009) and
>> > was backported to RHEL/CentOS 5.6 (in 2010).  The Windows drivers have
>> > two bugs, which I reported on the Red Hat Bugzilla as bugs 837321 and
>> > 837324.  With these patches they will suffer a performance hit but
>> > gain correctness.
>> > 
>> > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> I generally like the idea for a default, but doesn't this override even
> an explicit cache=writeback?

Yes.  It doesn't override cache=unsafe though.

> Are we sure that we want this?

The idea is that this change will overcome Anthony's objections to
switching the default to writeback...

Paolo
Kevin Wolf - July 4, 2012, 10:16 a.m.
Am 03.07.2012 15:51, schrieb Paolo Bonzini:
> Il 03/07/2012 15:49, Kevin Wolf ha scritto:
>>> If the guest does not support flushes, we should run in writethrough mode.
>>>> The setting is temporary until the next reset, so that for example the
>>>> BIOS will run in writethrough mode while Linux will run with a writeback
>>>> cache.
>>>>
>>>> VIRTIO_BLK_F_FLUSH has been introduced in Linux 2.6.32 (in 2009) and
>>>> was backported to RHEL/CentOS 5.6 (in 2010).  The Windows drivers have
>>>> two bugs, which I reported on the Red Hat Bugzilla as bugs 837321 and
>>>> 837324.  With these patches they will suffer a performance hit but
>>>> gain correctness.
>>>>
>>>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>> I generally like the idea for a default, but doesn't this override even
>> an explicit cache=writeback?
> 
> Yes.  It doesn't override cache=unsafe though.

When the guest doesn't support flushes, cache=writeback is equivalent to
cache=unsafe, so if you want the old behaviour back you can switch to
cache=unsafe without additional risks.

We don't have a cache=directunsafe, though, so if you want to get the
old behaviour of cache=none back, you're out of luck. Not sure how
acceptable this is.

Irrespective of this concern I've come to the conclusion that I agree
and we actually must enforce this for non-unsafe mode, and not doing it
is a bug.

Kevin
Paolo Bonzini - July 4, 2012, 12:21 p.m.
Il 04/07/2012 12:16, Kevin Wolf ha scritto:
>> > Yes.  It doesn't override cache=unsafe though.
> When the guest doesn't support flushes, cache=writeback is equivalent to
> cache=unsafe, so if you want the old behaviour back you can switch to
> cache=unsafe without additional risks.
> 
> We don't have a cache=directunsafe, though, so if you want to get the
> old behaviour of cache=none back, you're out of luck. Not sure how
> acceptable this is.

If we want to fix this, let's take the occasion to split the parameters
into cache=on/off (well, we have that already), flush=on/off, and a
device-side wce=on/off.

> Irrespective of this concern I've come to the conclusion that I agree
> and we actually must enforce this for non-unsafe mode, and not doing it
> is a bug.

Thanks!  Is that an Acked-by/Reviewed-by? :)

Paolo
Kevin Wolf - July 4, 2012, 12:50 p.m.
Am 04.07.2012 14:21, schrieb Paolo Bonzini:
> Il 04/07/2012 12:16, Kevin Wolf ha scritto:
>>>> Yes.  It doesn't override cache=unsafe though.
>> When the guest doesn't support flushes, cache=writeback is equivalent to
>> cache=unsafe, so if you want the old behaviour back you can switch to
>> cache=unsafe without additional risks.
>>
>> We don't have a cache=directunsafe, though, so if you want to get the
>> old behaviour of cache=none back, you're out of luck. Not sure how
>> acceptable this is.
> 
> If we want to fix this, let's take the occasion to split the parameters
> into cache=on/off (well, we have that already), flush=on/off, and a
> device-side wce=on/off.

You're volunteering? Great! ;-)

>> Irrespective of this concern I've come to the conclusion that I agree
>> and we actually must enforce this for non-unsafe mode, and not doing it
>> is a bug.
> 
> Thanks!  Is that an Acked-by/Reviewed-by? :)

Before merging the patches (or actually this patch, I think patch 1 is
fairly independent), I'd like to hear more opinions on whether we need
the cache parameter split first. But as far as the hardware is
concerned, sure, take it as an Acked-by and go forward with the spec and
kernel side of things.

Kevin
Paolo Bonzini - July 4, 2012, 1:20 p.m.
Il 04/07/2012 14:50, Kevin Wolf ha scritto:
> Before merging the patches (or actually this patch, I think patch 1 is
> fairly independent),

Yes, it is.

> I'd like to hear more opinions on whether we need
> the cache parameter split first

Ok, let's discuss it next week on the KVM/QEMU call.  Getting the cache
parameter split before 1.2 is going to be hard, but who knows.

> But as far as the hardware is
> concerned, sure, take it as an Acked-by and go forward with the spec and
> kernel side of things.

Rusty already committed everything, so it's too late! :)  (Sorry for the
tabs vs. spaces BTW).

Paolo

Patch

diff --git a/hw/virtio-blk.c b/hw/virtio-blk.c
index 280f96d..500e026 100644
--- a/hw/virtio-blk.c
+++ b/hw/virtio-blk.c
@@ -543,6 +543,19 @@  static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
     return features;
 }
 
+static void virtio_blk_set_status(VirtIODevice *vdev, uint8_t status)
+{
+    VirtIOBlock *s = to_virtio_blk(vdev);
+    uint32_t features;
+
+    if (!(status & VIRTIO_CONFIG_S_DRIVER_OK)) {
+        return;
+    }
+
+    features = vdev->guest_features;
+    bdrv_set_enable_write_cache(s->bs, !!(features & (1 << VIRTIO_BLK_F_WCE)));
+}
+
 static void virtio_blk_save(QEMUFile *f, void *opaque)
 {
     VirtIOBlock *s = opaque;
@@ -628,6 +641,7 @@  VirtIODevice *virtio_blk_init(DeviceState *dev, VirtIOBlkConf *blk)
     s->vdev.get_config = virtio_blk_update_config;
     s->vdev.set_config = virtio_blk_set_config;
     s->vdev.get_features = virtio_blk_get_features;
+    s->vdev.set_status = virtio_blk_set_status;
     s->vdev.reset = virtio_blk_reset;
     s->bs = blk->conf.bs;
     s->conf = &blk->conf;