Re: [Xen-devel] [RFC PATCH 0/4] Add missing default labels to switch statements

[Oleksandr Andrushchenko <andr2000@xxxxxxxxx>]

On 2/25/19 1:10 PM, Julien Grall wrote:
> Hi,
>
> On 25/02/2019 10:06, Oleksandr Andrushchenko wrote:
> > On 2/23/19 1:41 AM, Andrew Cooper wrote:
> > > On 22/02/2019 23:22, Julien Grall wrote:
> > > > Hi,
> > > >
> > > > On 22/02/2019 22:34, Andrew Cooper wrote:
> > > > > On 22/02/2019 22:11, Julien Grall wrote:
> > > > > > Hi Stefano,
> > > > > >
> > > > > > On 22/02/2019 21:58, Stefano Stabellini wrote:
> > > > > > > On Fri, 22 Feb 2019, Andrew Cooper wrote:
> > > > > > > > On 22/02/2019 21:00, Stefano Stabellini wrote:
> > > > > > > > > On Fri, 22 Feb 2019, Julien Grall wrote:
> > > > > > > > > > > > > BTW, I checked the series with -Wswitch-default:
> > > > > > > > > > > > > -Wswitch-default
> > > > > > > > > > > > > Warn whenever a switch statement does not have a default 
> > > > > > > > > > > > > case.
> > > > > > > > > > > > > > Furthermore, using BUG() is a pretty bad idea in switch.
> > > > > > > > > > > > > It is and not only in the switch. The reason I put BUG is 
> > > > > > > > > > > > > that I tried
> > > > > > > > > > > > > to follow
> > > > > > > > > > > > > the existing "error handling" at those places.
> > > > > > > > > > > > It is not because BUG() is been used today in some places 
> > > > > > > > > > > > that we need to
> > > > > > > > > > > > continue to spread it.
> > > > > > > > > > > >
> > > > > > > > > > > > > Use of BUG() itself is another topic which will also need 
> > > > > > > > > > > > > to be
> > > > > > > > > > > > > addressed
> > > > > > > > > > > > So we should not add more of them...
> > > > > > > > > > > Again, I see this as a dedicated change. So, in the current 
> > > > > > > > > > > series I think
> > > > > > > > > > > it is
> > > > > > > > > > > acceptable to use the existing way of error handling if any 
> > > > > > > > > > > at all.
> > > > > > > > > > That's not how it works in upstream. If you know some 
> > > > > > > > > > constructs are wrong, it
> > > > > > > > > > is best to try to address partially the problem directly then 
> > > > > > > > > > having so you
> > > > > > > > > > reduce the amounts of change afterwards.
> > > > > > > > > >
> > > > > > > > > > So please try to not introduce more BUG() in the code base.
> > > > > > > > > Hi Oleksandr, Julien,
> > > > > > > > >
> > > > > > > > > Julien's right that we should not introduce any more BUG()s. 
> > > > > > > > > In fact,
> > > > > > > > > each of them makes the code less safe, not more safe! The 
> > > > > > > > > purpose of
> > > > > > > > > MISRAC 16.4 is "defensive programming": write the code in a 
> > > > > > > > > way that is
> > > > > > > > > more (not less!) resilient to failure.
> > > > > > > > >
> > > > > > > > > So, I think it is a good idea to introduce a default label 
> > > > > > > > > because it
> > > > > > > > > can help us spot unexpected issues. Instead of calling BUG() 
> > > > > > > > > in the
> > > > > > > > > default handler, which is detrimental, we should return an 
> > > > > > > > > error when
> > > > > > > > > possible, or just print a warning.
> > > > > > > > domain_crash() is almost always better than BUG(). It is very 
> > > > > > > > obvious
> > > > > > > > if it gets hit, and wont crash Xen.
> > > > > > > That's a good suggestion.
> > > > > > >
> > > > > > >
> > > > > > > > > As 16.4 clearly state, even a simple comment would be enough 
> > > > > > > > > to address
> > > > > > > > > the rule. We just need to explain why a default label is not 
> > > > > > > > > needed.
> > > > > > > > > Such as:
> > > > > > > > >
> > > > > > > > >      default:
> > > > > > > > >      /* unreachable because blah and blah */
> > > > > > > > What a simple comment doesn't do is avoid breaking -Wswitch.
> > > > > > > I don't know how to reconcile 16.4 with -Wswitch. One could 
> > > > > > > argue that
> > > > > > > -Wswitch could be a good way to address 16.4, but then we 
> > > > > > > introduce a
> > > > > > > compiler specific requirement. Typically gcc is not the compiler of
> > > > > > > choice for these environments, unfortunately forcing gcc is not an
> > > > > > > option.
> > > > > > Well, you could build with GCC and then build with your custom
> > > > > > compiler... But, GCC is pretty much the only choice for Xen on 
> > > > > > Arm today
> > > > > > as we don't build with clang and I pretty doubt we can build with 
> > > > > > compcert.
> > > > > So the suggestion I had was to have an overall CONFIG_MISRA which 
> > > > > we can
> > > > > hide some of this nonsense behind, and then
> > > > >
> > > > > #ifdef CONFIG_MISRA
> > > > > #define MISRA_BLE_DEFAULT default:
> > > > > #else
> > > > > #define MISRA_BLE_DEFAULT
> > > > > #endif
> > > > This is pretty disgusting :). But then, it makes the code is bit more
> > > > obscure. So how that rule is making Xen more safe? Furthermore, one
> > > > default may not rule them all. So aren't we just adding code to make
> > > > MISRA happy at the risk of introducing more bug?
> > > Any switch statement where the default isn't empty
> > > (assert/bug/domain_crash/print) should probably be a regular default:
> > >
> > > I very much doubt MISRA intended for people to make themselves 
> > > compliant
> > > by adding "default: break;" everywhere, but lethargy is a powerful
> > > driving force.
> > >
> > > An alternative, and substantially more ugly would be to have the else
> > > case be "if ( 0 )" or similar, so you do take out the full next block,
> > > but that requires people to member braces.
> > >
> > > Even more ugly would be to take (x) and construct the block manually,
> > > but then you get at minimum a set of brackets where you would expect to
> > > see braces in normal C, and this alone is confusing to read.
> > >
> > > > > So when you disable CONFIG_MISRA, your compiler starts being able to
> > > > > help you again.
> > > > >
> > > > > TBH, it would also be nice to hide the SYMBOL nonsense behind, so 
> > > > > we can
> > > > > continue doing it the efficient way for ~100% of the time.
> > > > My main concern is we only looked at 2 MISRA rules and we already need
> > > > some nonsense code to spread everywhere. How much more are we going to
> > > > get like that?
> > (Un)fortunately many more...
> Do you mind providing a list of MISRA C rule that will require 
> tree-wide changes?
I am not quite sure anyone else or I can do this at the moment
> > > MISRA scanning is just like Coverity scanning.  The item classes with
> > > the highest seen-count tend to be the least interesting and 
> > > relevant, as
> > > they tend to be the ones where we are systematically different.
> > >
> > > It is the one-off flagged issues which tend to be important.
> > I do agree. There are many more more serious issues on the list.
> > The "default" rule is just one to try and seemed very straight forward
> > to fix (I put aside BUG() issue - this was expected to be discussed)
> It feels to me we should try to address serious issues first and then 
> look at more controversial changes afterwards.
While I do agree in common with this approach the reason for bringing in
some "controversial changes" was our hope to understand community's feeling
about that in general and try to work out an accepted way for such changes.

> Cheers,

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel