Re: [PATCH 01/10] vhost-user-blk: reconnect on any error during realize

From: Kevin Wolf <kwolf@redhat.com>
To: "Roman Kagan" <rvkagan@yandex-team.ru>,
	qemu-devel@nongnu.org, qemu-block@nongnu.org,
	"Michael S. Tsirkin" <mst@redhat.com>,
	"Hanna Reitz" <hreitz@redhat.com>,
	yc-core@yandex-team.ru,
	"Marc-André Lureau" <marcandre.lureau@redhat.com>,
	"Paolo Bonzini" <pbonzini@redhat.com>
Subject: Re: [PATCH 01/10] vhost-user-blk: reconnect on any error during realize
Date: Fri, 12 Nov 2021 12:37:59 +0100	[thread overview]
Message-ID: <YY5SF0uPkoEvX9qM@redhat.com> (raw)
In-Reply-To: <YY4aGok6e8Z6BRQu@rvkaganb.lan>

Am 12.11.2021 um 08:39 hat Roman Kagan geschrieben:
> On Thu, Nov 11, 2021 at 06:52:30PM +0100, Kevin Wolf wrote:
> > Am 11.11.2021 um 16:33 hat Roman Kagan geschrieben:
> > > vhost-user-blk realize only attempts to reconnect if the previous
> > > connection attempt failed on "a problem with the connection and not an
> > > error related to the content (which would fail again the same way in the
> > > next attempt)".
> > > 
> > > However this distinction is very subtle, and may be inadvertently broken
> > > if the code changes somewhere deep down the stack and a new error gets
> > > propagated up to here.
> > > 
> > > OTOH now that the number of reconnection attempts is limited it seems
> > > harmless to try reconnecting on any error.
> > > 
> > > So relax the condition of whether to retry connecting to check for any
> > > error.
> > > 
> > > This patch amends a527e312b5 "vhost-user-blk: Implement reconnection
> > > during realize".
> > > 
> > > Signed-off-by: Roman Kagan <rvkagan@yandex-team.ru>
> > 
> > It results in less than perfect error messages. With a modified export
> > that just crashes qemu-storage-daemon during get_features, I get:
> > 
> > qemu-system-x86_64: -device vhost-user-blk-pci,chardev=c: Failed to read msg header. Read 0 instead of 12. Original request 1.
> > qemu-system-x86_64: -device vhost-user-blk-pci,chardev=c: Reconnecting after error: vhost_backend_init failed: Protocol error
> > qemu-system-x86_64: -device vhost-user-blk-pci,chardev=c: Reconnecting after error: Failed to connect to '/tmp/vsock': Connection refused
> > qemu-system-x86_64: -device vhost-user-blk-pci,chardev=c: Reconnecting after error: Failed to connect to '/tmp/vsock': Connection refused
> > qemu-system-x86_64: -device vhost-user-blk-pci,chardev=c: Failed to connect to '/tmp/vsock': Connection refused
> 
> This patch doesn't change any error messages.  Which ones specifically
> became less than perfect as a result of this patch?

But it adds error messages (for each retry), which are different from
the first error message. As I said this is not the end of the world, but
maybe a bit more confusing.

> > I guess this might be tolerable. On the other hand, the patch doesn't
> > really fix anything either, but just gets rid of possible subtleties.
> 
> The remaining patches in the series make other errors beside -EPROTO
> propagate up to this point, and some (most) of them are retryable.  This
> was the reason to include this patch at the beginning of the series (I
> guess I should've mentioned that in the patch log).

I see. I hadn't looked at the rest of the series yet because I ran out
of time, but now that I'm skimming them, I see quite a few places that
use non-EPROTO, but I wonder which of them actually should be
reconnected. So far all I saw were presumably persistent errors where a
retry won't help. Can you give me some examples?

Kevin