On 05/30/2012 08:56 AM, Jiri Denemark wrote:
On Wed, May 30, 2012 at 08:48:22 -0600, Eric Blake wrote:
> On 05/30/2012 08:34 AM, Jiri Denemark wrote:
>> ---
>> src/util/virfile.c | 10 ++++++++++
>> 1 file changed, 10 insertions(+)
>>
>> diff --git a/src/util/virfile.c b/src/util/virfile.c
>> index db3d737..bc7f2c9 100644
>> --- a/src/util/virfile.c
>> +++ b/src/util/virfile.c
>> @@ -35,6 +35,7 @@
>> #include "configmake.h"
>> #include "memory.h"
>> #include "virterror_internal.h"
>> +#include "logging.h"
>>
>> #define VIR_FROM_THIS VIR_FROM_NONE
>> #define virFileError(code, ...) \
>> @@ -51,6 +52,15 @@ int virFileClose(int *fdptr, bool preserve_errno)
>> if (preserve_errno)
>> saved_errno = errno;
>> rc = close(*fdptr);
>> + if (rc < 0) {
>> + if (errno != EBADF) {
>> + char ebuf[1024] ATTRIBUTE_UNUSED;
>> + VIR_DEBUG("Failed to close fd %d: %d",
>> + *fdptr, virStrerror(errno, ebuf, sizeof(ebuf)));
>
> I think we _also_ need to log EBADF errors, those are usually the sign
> of developer bugs, as you are probably guilty of a double-close race.
> In fact, I think EBADF should be logged with higher severity than
> VIR_DEBUG, as we want to see that right away.
I explicitly didn't want to log them since you would see a lot of errors when
after fork we are closing all available file descriptors except those we
explicitly want to leave open... I'll add EBADF logging and handle those
mass-closes separately.
Oh, good point. EBADF in the parent is almost always a bug (and if the
race had gone the other way, you'd have a double-close). EBADF in the
child is an annoyance - we know we are single-threaded and therefore
there is no double-close race, and it is more efficient to try and waste
a close() than it is to do an fcntl() filter to see if the fd is valid
in order to decide whether to close.
But that means that the _only_ place that should be ignoring EBADF on a
mass-close is in the post-fork() closing of a child; so maybe the
solution is:
change the signature of virFileClose() to take an extra parameter to say
whether to ignore EBADF,
change VIR_FORCE_CLOSE to pass false to the new parameter,
and change command.c's mass-close to use virFileClose() directly, with
the parameter set to true
--
Eric Blake eblake(a)redhat.com +1-919-301-3266
Libvirt virtualization library
http://libvirt.org