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

Re: [Xen-devel] [PATCH 5/5] xen: clean up VPF flags macros



On 09/28/2015 09:52 AM, Jan Beulich wrote:
On 28.09.15 at 09:29, <JGross@xxxxxxxx> wrote:
On 09/28/2015 08:22 AM, Jan Beulich wrote:
On 28.09.15 at 07:23, <JGross@xxxxxxxx> wrote:
On 09/25/2015 05:42 PM, Jan Beulich wrote:
On 25.09.15 at 13:54, <JGross@xxxxxxxx> wrote:
Per-VCPU pause flags in sched.h are defined as bit positions and as
values derived from the bit defines. There is only one user of a value
which can be easily converted to use a bit number as well.

I'm not convinced:

--- a/xen/common/domctl.c
+++ b/xen/common/domctl.c
@@ -172,7 +172,7 @@ void getdomaininfo(struct domain *d, struct
xen_domctl_getdomaininfo *info)
            info->max_vcpu_id = v->vcpu_id;
            if ( !test_bit(_VPF_down, &v->pause_flags) )
            {
-            if ( !(v->pause_flags & VPF_blocked) )
+            if ( !test_bit(_VPF_blocked, &v->pause_flags) )

test_bit() is quite a bit more complex an operation than a simple &,
and with (on x86) even constant_test_bit() involving a cast to
a pointer to volatile I'm afraid we can't even hope that compilers
would produce identical code for both in cases like this one (as that
casts limits freedom of the compiler). IOW I'd rather see other
test_bit(_VPF_...) uses converted the inverse way (which as a nice
but minor side effect would yield slightly smaller source code).

What about introducing __test_bit() being a variant which can be
reordered by omitting the volatile modifier? I think this would have
the same effect.

I'm not convinced it always would - the inline function is still more
complex than the plain operation.

Depends on the way it is done. What about:

#define __test_bit(nr, addr) ({                         \
      if ( bitop_bad_size(addr) ) __bitop_bad_size();     \
      (__builtin_constant_p(nr) ?                         \
       !!(*(addr) & ((typeof)(*(addr))1 << (nr))) :       \
       __variable_test_bit((nr),(addr)));                 \
})

But that's not correct - addr may point to wider than a single entry
array, irrespective of whether nr is a compile time constant.

It would even be possible to drop the test for bitop_bad_size(addr) in
the constant case.

In which case 1 << nr may reference a bit beyond the type
of *addr.

Hmm, yes, you are right, of course.

It could be fixed, however.

The question is: does it make sense to follow this path any longer,
or would you reject it even in case of correctness? I wouldn't mind
either way, I just don't want to waste time (mine and yours).

And we could still get rid of many double definitions
of the same bit. Even if the mask definition of a bit is not error prone
by relying on the definition of the bit position, it makes it harder to
find all users of this bit.

Why so? Just omit the leading underscore when grep-ing, and you'll
find all instances (less preprocessor token concatenation, but that's
orthogonal).

I do use grep for this purpose occasionally, but I prefer tools like
cscope. BTW: IMO using grep the way you are suggesting here is annoying
for cases where the search string is contained in other items.

There may be cases, yes, but surely not this one: How likely is it for
some other variable name to include, say, VPF_blocked?

I think we both agree that [_]VPF_blocked poses no problem using grep.

It's more a matter of taste and what people are used to. If someone
is using cscope for that purpose on a regular basis he will either have
to search for two different variables or he will have to use another
tool, possibly after seeing the definition of VPF_blocked depending on
_VPF_blocked. Both cases will be annoying as an extra action is required
for him.


Juergen

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.