Re: [PATCH] include/linux/slab.h: new KFREE() macro.

[Randy Dunlap <randy.dunlap@oracle.com>]

On Tue, 9 Jan 2007 11:02:35 -0800 (PST) Amit Choudhary wrote:

> 
> --- Valdis.Kletnieks@vt.edu wrote:
> 
> > On Mon, 08 Jan 2007 01:06:12 PST, Amit Choudhary said:
> > > I do not see how a double free can result in _logical_wrong_behaviour_ of the program and the
> > > program keeps on running (like an incoming packet being dropped because of double free).
> > Double
> > > free will _only_and_only_ result in system crash that can be solved by setting 'x' to NULL.
> > 
> > The problem is that very rarely is there a second free() with no intervening
> > use - what actually *happens* usually is:
> > 
> > 1) You alloc the memory
> > 2) You use the memory
> > 3) You take a reference on the memory, so you know where it is.
> > 4) You free the memory
> > 5) You use the memory via the reference you took in (3)
> > 6) You free it again - at which point you finally know for sure that
> > everything in step 5 was doing a fandango on core....
> > 
> 
> Correct. And doing kfree(x); x=NULL; is not hiding that. These issues can still be debugged by
> using the slab debugging options. One other benefit of doing this is that if someone tries to
> access the same memory again using the variable 'x', then he will get an immediate crash. And the
> problem can be solved immediately, without using the slab debugging options. I do not yet
> understand how doing this hides the bugs, obfuscates the code, etc. because I haven't seen an
> example yet, but only blanket statements.
> 
> But now I know better, since I haven't heard anything in support of this case, I have concluded
> that doing kfree(x); x=NULL; is _not_needed_ in the "linux kernel". I hope that no one does it in
> the future. And since people vehemently opposed this, I think its better to add another item on
> the kernel janitor's list to remove all the (x=NULL) statements where people are doing "kfree(x);
> x=NULL".

No thanks.  If a driver author wants to maintain driver state
that way, it's OK, but that doesn't make it a global requirement.

---
~Randy