On Mon, Feb 11, 2013 at 04:10:46PM +0100, Andreas Färber wrote: > Am 11.02.2013 15:57, schrieb Stefan Hajnoczi: > > On Mon, Feb 11, 2013 at 3:34 PM, Peter Maydell <peter.mayd...@linaro.org> > > wrote: > >> On 11 February 2013 14:19, Andreas Färber <afaer...@suse.de> wrote: > >>> Am 11.02.2013 15:01, schrieb Markus Armbruster: > >>>> Kevin Wolf <kw...@redhat.com> writes: > >>>> > >>>>> Am 11.02.2013 14:27, schrieb Stefan Hajnoczi: > >>>>>> I think we need to side-track this patch email to figure out what to > >>>>>> use: > >>>>>> > >>>>>> fprintf(stderr) - some warnings/errors use this > >>>>>> error_report() - goes to the monitor, if possible, otherwise stderr > >>>>> > >>>>> These look wrong to me. > >>>> > >>>> "Wrong" is a bit strong, in particular since there's ample precedence > >>>> for these uses. > >> > >> Certainly for fprintf() I would say "deprecated" wherever > >> we have a better API available. > >> > >>>>>> qemu_log_*() - goes to the qemu log, seems a little TCG-centric > >>>>> > >>>>> I would suggest either this or just trace points. (And by the way, it's > >>>>> a pity that -d is so TCG-centric, it's been more than once the reason > >>>>> why I disabled KVM when debugging a guest... Having at least -d int > >>>>> would be so useful.) > >>>> > >>>> Tracepoints don't really fit when we want to report the guest does > >>>> something we don't handle. Users deserve fair warning then, don't they? > >>>> > >>>> Could qemu_log() & friends be made fit for general use? What's missing? > >>> > >>> Blue already did some work to make it more usable, and I believe Peter > >>> adopted LOG_UNIMPL for ARM devices in place of hw_error() > >> > >> Yes; in particular where we have classes of error message which the > >> user may wish to enable or disable (of which "QEMU doesn't implement > >> this" and "the guest just did something that's probably a guest bug" > >> are two common ones) qemu_log_mask(LOG_*, ...) is the preferred > >> API for devices IMHO. So I think Herve's patch is entirely the > >> right thing. > > > > qemu_log_mask() can replace fprintf() but it needs to default to > > stderr and a reasonable default mask. It should be used for > > non-monitor command errors and warnings. > > > > Having said that, I think we should use hw_error() or fprintf() in > > this patch until qemu_log_mask() replaces them. > > Nack. Like I just pointed out, hw_error() is a fatal abort and not a > replacement for logging. I don't mind which of the non-fatal logging > methods you choose, but guest-triggerable aborting is a semantic change > that we should not choose just because someone doesn't like logging API > or defaults. Especially not for 1.4.
You are right. Stefan
