Re: [PATCH v2] libxenstore: prefer using the character device

[Jonathan Creekmore <jonathan.creekmore@gmail.com>]

Ian Jackson <Ian.Jackson@eu.citrix.com> writes:

> Wei Liu writes ("Re: [Xen-devel] [PATCH v2] libxenstore: prefer 
> using the character device"): 
>> On Thu, Aug 27, 2015 at 09:04:38AM -0500, Jonathan Creekmore 
>> wrote: 
>> > With the addition of FMODE_ATOMIC_POS in the Linux 3.14 
>> > kernel, concurrent blocking file accesses to a single open 
>> > file descriptor can cause a deadlock trying to grab the file 
>> > position lock. If a watch has been set up, causing a 
>> > read_thread to blocking read on the file descriptor, then 
>> > future writes that would cause the background read to 
>> > complete will block waiting on the file position lock before 
>> > they can execute. This race condition only occurs when 
>> > libxenstore is accessing the xenstore daemon through the 
>> > /proc/xen/xenbus file and not through the unix domain socket, 
>> > which is the case when the xenstore daemon is running as a 
>> > stub domain or when oxenstored is passed 
>> > --disable-socket. Accessing the daemon from the true 
>> > character device also does not exhibit this problem.   On 
>> > Linux, prefer using the character device file over the proc 
>> > file if the character device exists. 
> 
> I confess I still see this as working around a kernel bug.  Only 
> this time we are switching from a buggy to non-buggy kernel 
> interface. 
> 
> Why don't we have the kernel provide only non-buggy interfaces ?

I was just trying to implement what David suggested; CC'ing him on 
this as well.
 
> 
>> > diff --git a/tools/xenstore/xs_lib.c 
>> > b/tools/xenstore/xs_lib.c index af4f75a..0c7744e 100644 --- 
>> > a/tools/xenstore/xs_lib.c +++ b/tools/xenstore/xs_lib.c @@ 
>> > -81,6 +81,8 @@ const char *xs_domain_dev(void) 
>> >  #if defined(__RUMPUSER_XEN__) || defined(__RUMPRUN__) return 
>> >  "/dev/xen/xenbus"; #elif defined(__linux__) 
>> > +	if (access("/dev/xen/xenbus", F_OK) == 0) + 
>> > return "/dev/xen/xenbus"; 
> 
> Also, previously xs_domain_dev was a function which simply 
> returned a static value.  I feel vaguely uneasy at putting this 
> kind of autodetection logic here. 

Not entirely; the existing code queried an environment variable first
and, only if that was not set, did it return a static value. I added the
autodetection logic to fall back in the case where, for some reason,
/dev/xen/xenbus did not exist. Handling that kind of a fallback at a
higher layer would be a larger refactor to probe for the existence of
the character device first.