Re: [PATCH net-next 0/4] net: ipa: kill IPA_VALIDATION

[Leon Romanovsky]

On Tue, Jul 27, 2021 at 07:34:41AM -0500, Alex Elder wrote:
> On 7/27/21 6:16 AM, Leon Romanovsky wrote:
> > On Mon, Jul 26, 2021 at 12:40:06PM -0500, Alex Elder wrote:
> >> A few months ago I proposed cleaning up some code that validates
> >> certain things conditionally, arguing that doing so once is enough,
> >> thus doing so always should not be necessary.
> >>   https://lore.kernel.org/netdev/20210320141729.1956732-1-elder@xxxxxxxxxx/
> >> Leon Romanovsky felt strongly that this was a mistake, and in the
> >> end I agreed to change my plans.
> > 
> > <...>
> > 
> >> The second patch fixes a bug that wasn't normally exposed because of
> >> the conditional compilation (a reason Leon was right about this).
> > 
> > Thanks Alex,
> > 
> > If you want another anti pattern that is very popular in netdev, the following pattern is
> > wrong by definition :):
> > if (WARN_ON(...))
> >   return ...
> 
> I understand this reasoning.
> 
> I had it return an error if the WARN_ON() condition was true in cases
> where the function returned a value and callers already handled errors.
> I looked back at the patch and here is one of those cases:
> 
> gsi_channel_trans_alloc()
> - If too many TREs are requested we do not want to allocate them
>   from the pool, or it will cause further breakage.  By returning
>   early, no transaction will be filled or committed, and an error
>   message will (often) be reported, which will indicate the source
>   of the error.  If any error occurs during initialization, we fail
>   that whole process and everything should be cleaned up.  So in
>   this case at least, returning if this ever occurred is better
>   than allowing control to continue into the function.
> 
> In any case I take your point.  I will now add to my task list
> a review of these spots.  I'd like to be sure an error message
> *is* reported at an appropriate level up the chain of callers so
> I can always identify the culprit in the a WARN_ON() fires (even
> though it should never
>  happen).  And in each case I'll evaluate
> whether returning is better than not.

You can, but users don't :). So if it is valid but error flow, that
needs user awareness, simply print something to the dmesg with *_err()
prints.


BTW, I'm trying to untangle some of the flows in net/core/devlink.c
and such if(WARN()) pattern is even harmful, because it is very hard to
understand when that error is rare/non-exist/real.

Thanks

> 
> Thanks.
> 
> 					-Alex
> 
> > The WARN_*() macros are intended catch impossible flows, something that
> > shouldn't exist. The idea that printed stack to dmesg and return to the
> > caller will fix the situation is a very naive one. That stack already
> > says that something very wrong in the system.
> > 
> > If such flow can be valid use "if(...) return ..", if not use plain
> > WARN_ON(...).
> > 
> > Thanks
> > 
>