Re: [Qemu-devel] [PATCH v2] block/curl: Don't lose original error when a connection fails.

[Kevin Wolf <kwolf@redhat.com>]

Am 03.07.2015 um 14:35 hat Markus Armbruster geschrieben:
> "Richard W.M. Jones" <rjones@redhat.com> writes:
> 
> > Currently if qemu is connected to a curl source (eg. web server), and
> > the web server fails / times out / dies, you always see a bogus EIO
> > "Input/output error".
> >
> > For example, choose a large file located on any local webserver which
> > you control:
> >
> >   $ qemu-img convert -p http://example.com/large.iso /tmp/test
> >
> > Once it starts copying the file, stop the webserver and you will see
> > qemu-img fail with:
> >
> >   qemu-img: error while reading sector 61440: Input/output error
> >
> > This patch does two things: Firstly print the actual error from curl
> > so it doesn't get lost.  Secondly, change EIO to EPROTO.  EPROTO is a
> > POSIX.1 compatible errno which more accurately reflects that there was
> > a protocol error, rather than some kind of hardware failure.
> >
> > After this patch is applied, the error changes to:
> >
> >   $ qemu-img convert -p http://example.com/large.iso /tmp/test
> >   qemu-img: curl: transfer closed with 469989 bytes remaining to read
> >   qemu-img: error while reading sector 16384: Protocol error
> >
> > Signed-off-by: Richard W.M. Jones <rjones@redhat.com>
> > Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  block/curl.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/block/curl.c b/block/curl.c
> > index 3a2b63e..2fd7c06 100644
> > --- a/block/curl.c
> > +++ b/block/curl.c
> > @@ -22,6 +22,7 @@
> >   * THE SOFTWARE.
> >   */
> >  #include "qemu-common.h"
> > +#include "qemu/error-report.h"
> >  #include "block/block_int.h"
> >  #include "qapi/qmp/qbool.h"
> >  #include "qapi/qmp/qstring.h"
> > @@ -298,6 +299,12 @@ static void curl_multi_check_completion(BDRVCURLState *s)
> >              /* ACBs for successful messages get completed in curl_read_cb */
> >              if (msg->data.result != CURLE_OK) {
> >                  int i;
> > +
> > +                /* Don't lose the original error message from curl, since
> > +                 * it contains extra data.
> > +                 */
> > +                error_report("curl: %s", state->errmsg);
> > +
> >                  for (i = 0; i < CURL_NUM_ACB; i++) {
> >                      CURLAIOCB *acb = state->acb[i];
> >  
> 
> Printing an error message, then returning an error code is problematic.
> 
> It works when the caller is going to print its own error message to the
> same destination.  Callee produces a specific error message devoid of
> context, caller produces an unspecific one with hopefully more context.
> Better than just one of them.  Worse than a single specific error with
> context, but that can't be done with just a "return errno code"
> interface.
> 
> It's kind of wrong when the caller reports its own error somewhere else,
> e.g. to a monitor.  Still, when barfing extra info to stderr is the best
> we can do, it's better than nothing.
> 
> It's more wrong when the caller handles the error quietly.  I guess
> that's never the case here, but I can't be sure without a lot more
> sleuthing.  Perhaps Kevin or Stefan can judge this immediately.

I'm not worried too much about requests made by the monitor or during
startup. I don't like the error_report() there, but having a more
specific error message on stderr is better than having nothing.

The case that bothers me more is guest requests. Depending on the
werror/rerror settings, this may allow the guest to flood the log file
with curl error messages.

> > @@ -305,7 +312,7 @@ static void curl_multi_check_completion(BDRVCURLState *s)
> >                          continue;
> >                      }
> >  
> > -                    acb->common.cb(acb->common.opaque, -EIO);
> > +                    acb->common.cb(acb->common.opaque, -EPROTO);
> >                      qemu_aio_unref(acb);
> >                      state->acb[i] = NULL;
> >                  }
> 
> To understand impact exactly, we'd need to figure out where the changed
> error code gets consumed.  However, I don't expect consumers to check
> the actual error code.  A quick grep for comparisons with EIO or -EIO
> finds nothing related to block I/O, except for nbd_trip() checking the
> value of nbd_co_receive_request(), and that's unrelated.

Yes, I wouldn't expect any problems caused by this change.

Kevin