[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [PATCH] dynamic debug patch



On Wed, Nov 05, 2008 at 04:42:21PM +0000, Daniel P. Berrange wrote:
> On Wed, Nov 05, 2008 at 04:26:15PM +0100, Daniel Veillard wrote:
> >   First version of a small patch allowing to gather debug informations
> > from a running libvirt daemon. Basically one can send signal USR2
> > to the daemon, the daemon will from that point dump its current internal
> > state and all verbose debug output to a file /tmp/libvirt_debug_xxxx.
> > When sending back signal USR2 the file is closed and can be looked at
> > for analysis, this allow to save extensive debug informations from a
> > running daemon.
> 
> Hardcoding a file in /tmp is not very good practice - it'll certainly
> not play nicely with SELinux. We should make the logfile name be 

  I don't see why SELinux should object a binary open and write a file
into /tmp/ . Also that's a path that a binary linking to libvirt
but running under as non-root can still use.
  /tmp/libvirt_debug_xxxx is of course opened with mkstemp !

> configurable via /etc/libvirt/libvirtd.conf. 

  Which is libvirtd specific. Seems I failed to carry my suggestion
that the thing should work work any app linked to libvirt shared lib
even if my initial patch didn't do this.

> > and similar for main internal data structures and drivers
> > The patch is just a work in progress but trying to get early feedback.
> > Maybe this could be used to get remote introspection capabilities too
> > but ATM I'm more looking at providing an easy way to get debug
> > informations on a libvirt program.
> 
> I think use virBufferPtr is rather too cumbersome for the calling
> program. The internal API for logging currently just takes
> 
>   void qemudLog(int priority, const char *fmt, ...)

  which again is tied to QEmu

> Either we can pass variable args out to the app, or format the error
> message and then pass it out. The calling app can write this straight
> to a file, send to syslog, or anything else it wants, without having
> the intermediate step of potentially  uneccessary buffering.  I think

  I don't think the buffering cost is serious, and that would allow the
internal debug routines to be used in different ways. We should not tie
those entry points to a specific output processing IMHO.

> it'll also be desirable to include a 'category' string to allow 
> applications to filter the messages it uses. 

  yes filtering would be nice.

> In parallel with debugging / logging of functional aspects of the 
> daemon / library, it is also desirable to get the ability to record
> a log of operations invoked on objects. eg, I want to record all
> operation invoked against a particular driver, so as a support 
> person I can ask a bug-reporter to send me the log of the operations
> they did leading upto the problem.

  yes that's the goal.

> As an application using libvirt API, I'd expect to have ability to
> register a callback to receive debug info, and APIs to control the
> logging level
> 
> 
>    enum {
>       VIR_LOG_DEBUG,
>       VIR_LOG_INFO,
>       VIR_LOG_WARN,
>       VIR_LOG_ERROR,
>    };
> 
>    typedef void (*virLogHandler)(const char *category,
>                                  int priority,
>                                  const char *msg);
> 
>    virLogAddHandler(virLogHandler callback);
> 
>    virLogSetDefaultPriority(int priority);
>    virLogSetPriority(const char *category,
>                      int priority);
> 

  Okay, that's an API, expecting the app will use it. What i want is
allow debug without the application being modified which is rather
different.

> And a semi-public internal API (ie for use by libvirt.so and libvirtd)
> 
>     virLogMessage(const char *category, int priority, const char *fmt, ....);

  That's fine but i would rather structure object dump based on
something more neutral, which is why i think the dump to a buffer
is more generally useful. that could be reused in various ways.

> Currently we toggle between using stderr, and syslog according to
> whether the --deamon flag is passed. This isn't particularly
> great. We should be configurable independantly of thise.
> 
> I'd expect /etc/libvirt/libvirtd.conf  to allow you to set 
> 
>    # Choice of 'null', 'syslog', 'stderr', 'file'
>    log.backend = "syslog"
> 
>    # If 'file' was chosen, then also allow
>    log.file = "/var/log/libvirtd.log"
> 
>    # or if 'syslog' was chosen, then allow admin to set the
>    # syslog facility name for openlog() call.
>    log.syslog = "libvirtd"
> 
>    # One of VIR_LOG_XXX constants
>    log.priority = "WARN"
> 
>    # And per category overrides
>    log.category.events = "DEBUG"
>    log.category.mdns = "INFO"
> 
> We already have SIGUSR1 hooked up to make it re-read the domain XML
> files, so we could extend this to also re-read the master config
> file. So if someone wanted to temporarily debug the system they

  Hum, that's libvirtd specific.

> could just edit 'log.backend' from 'null' to 'file', and send the
> 'SIGUSR1'.  Or if they already have it logging to a file by default
> and just wanted to increase verbosity, they'd edit 'log.priority'
> and send SIGUSR1.

  It's still quite a more complex operation than just sending the
signal. It also means reloading the preferences, it's not that neutral

> At any time, they can also send SIGUSR2 as per your patch, and get
> a dump listing all active sockets, sent to whatever log.backend
> they currently have configured - either syslog, or file, or whateve
> else we implement.

  Yes I still think it fits slightly different uses. The idea of being
able to select the parts doing logs is clearly needed in some ways
just right now all the event loop logging can be quite verbose.

> This comprehensive logging solution is rather alot of work, so for
> immediate use I'd suggest just adding the config for 'log.file' and
> 'log.backend' choice, and ignore the per-category debug levels for 
> 0.5.0 release. 
> 
> > Also I'm unclear, do we really want to have all the debug strings
> > internationalized with _() , that's more work for localization team
> > and it's unclear this would benefit 'end users'.
> 
> I think debug messages are potentially important not to localize. They
> are not typically intended for end users, but for developers & vendor
> support people, where having everything in english may make it easier
> to search for similar problems. 
> 
> >  
> > +static FILE *error_out = NULL;
> > +static FILE *info_out = NULL;
> > +static int debug_activated = 0;
> > +
> > +#ifdef ENABLE_DEBUG
> > +
> > +/*
> > + * Signal entry point on USR2 we can't do anything at that point except
> > + * log the signal and have qemudToggleDebug called when back into the
> > + * main loop.
> > + */
> > +static void sig_debug(int sig, siginfo_t * siginfo, void* context) {
> > +    sig_handler(sig, siginfo, context);
> > +}
> > +
> > +/*
> > + * Debug the state of a client
> > + */
> > +static void qemudDumpDebugClient(struct qemud_client *client) {
> > +    if (client->magic != QEMUD_CLIENT_MAGIC) {
> > +        qemudLog(QEMUD_DEBUG,
> > +                 _("\nQEmud client: invalid magic %X, skipping\n"),
> > +                 (unsigned int) client->magic);
> > +        return;
> > +    }
> > +    qemudLog(QEMUD_DEBUG, _("QEmud client: fd %d readonly %d mode %d"),
> > +             client->fd, client->readonly, client->mode);
> > +
> > +    /* TODO: more complete dump of state, especially the connection */
> > +}
> > +
> > +/*
> > + * Debug the state of a server
> > + */
> > +static void qemudDumpDebugServer(struct qemud_server *server) {
> > +    struct qemud_client *client;
> > +
> > +    if (server == NULL) {
> > +        qemudLog(QEMUD_DEBUG, "%s",
> > +            _("QEmud server: NULL pointer\n"));
> > +        return;
> > +    }
> > +    qemudLog(QEMUD_DEBUG,
> > +        _("QEmud server: listening on %d sockets, %d clients\n"),
> > +             server->nsockets, server->nclients);
> > +
> > +    client = server->clients;
> > +
> > +    while (client != NULL) {
> > +        qemudDumpDebugClient(client);
> > +        client = client->next;
> > +    }
> > +}
> > +
> > +/*
> > + * Toggle the debug status on/off, on on create a new temporary
> > + * debug file and start saving the output there
> > + * It is called when the signal USR2 is received.
> > + */
> > +static void qemudToggleDebug(struct qemud_server *server) {
> > +    if (debug_activated == 0) {
> > +        char path[50] = "/tmp/libvirt_debug_XXXXXX";
> 
> This path needs to be configurable from the /etc/libvirt/libvirtd.conf
> file, and default to /var/log to be selinux compliant.

  drwxr-xr-x 18 root root
  /var/log doesn't fit my use cases.

> > +/*
> > + * Set up the debugging environment
> > + */
> > +static void qemudInitDebug(void) {
> > +#ifdef ENABLE_DEBUG
> > +    struct sigaction oldact;
> > +    struct sigaction sig_action;
> > +
> > +    /*
> > +     * if there is already an handler, leave it as is to
> > +     * avoid disturbing the application's behaviour
> > +     */
> > +    if (sigaction (SIGUSR2, NULL, &oldact) == 0) {
> > +        if (oldact.sa_handler == NULL && oldact.sa_sigaction == NULL) {
> > +            sig_action.sa_sigaction = sig_debug;
> > +            sigaction(SIGUSR2, &sig_action, NULL);
> > +        }
> > +    }
> 
> Since this is inside the daemon, we already know that no other 
> SIGUSR2 handler is present, so can ignore that check.

  I intend to have similar code in the client shared lib.

> > +#ifdef ENABLE_DEBUG
> > +    case SIGUSR2:
> > +        qemudToggleDebug(server);
> > +        break;
> > +#endif
> 
> IMHO, toggle of log level should be done with SIGUSR1 and libvirtd.conf
> reload, and SIGUSR2 should just be doing the 'state dump' of active
> sockets, clients & other interesting info, to whatever logging target
> is active, be it syslog or a file, or stderr.

  Again only works in the daemon, not in a client linked to the lib
and I would prefer to have one single way to expose debugging.

> >  #ifdef ENABLE_DEBUG
> >          case QEMUD_DEBUG:
> > -            if (verbose) {
> > -                vprintf(fmt, args);
> > -                fputc('\n', stdout);
> > +            if ((verbose) || (debug_activated)) {
> > +                vfprintf(info_out, fmt, args);
> > +                fputc('\n', info_out);
> >              }
> >              break;
> 
> 
> This qemudLog() function is really flawed - if toggles between the use
> of syslog() vs stderr, based on whether you pass --daemon flag. It needs
> to be configurable by the config file, independantly of --daemon. As
> this patch stands, nothing will ever be logged because 99% of libvirtd
> instances run with --daemon flag set, and thus are using the syslog()
> code branch

  You missed the debug_activated flags. I tested this.

> > QEmud server: listening on 2 sockets, 1 clients
> > 
> > QEmud client: fd 10 readonly 0 mode 0
> > EVENT: Calculate expiry of 1 timers
> > EVENT: Timeout at 0 due in -1 ms
> > EVENT: Poll on 5 handles 0x24d91f0 timeout -1
> > EVENT: Poll got 1 event
> > EVENT: Dispatch 10 1 0x24c1610
> > EVENT: Remove handle 10
> > EVENT: mark delete 4
> > EVENT: Add handle 10 13 0x4087d0 0x24c1610
> > EVENT: Remove handle 10
> > EVENT: mark delete 5
> > EVENT: Add handle 10 14 0x4087d0 0x24c1610
> > EVENT: Calculate expiry of 1 timers
> 
> 
> THis pile of (mostly useless) EVENT log messages is why we need to
> ultimately have 'category' filters per file :-)

  Agreed some filtering is really needed, but that could be done
as post processing on the log untill we have structured logging in
place.

  thanks !

Daniel

-- 
Daniel Veillard      | libxml Gnome XML XSLT toolkit  http://xmlsoft.org/
daniel veillard com  | Rpmfind RPM search engine http://rpmfind.net/
http://veillard.com/ | virtualization library  http://libvirt.org/


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]