Re: [PATCH] Fix two potential mem leaks in MPT Fusion (mpt_attach())

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 1 Aug 2007 21:03:50 -0600 Matthew Wilcox <matthew@wil.cx> wrote:

> On Wed, Aug 01, 2007 at 05:26:53PM -0700, Andrew Morton wrote:
> > Why on earth is that using GFP_ATOMIC?  This function later goes on to
> > create procfs files and such things.
> 
> Seems fairly common in driver initialisation code.  I removed three
> instances of this in the advansys driver.

hrm.  People reach for GFP_ATOMIC so often that it becomes a habit, I guess.

It makes one wonder how much that lovely fault-injection framework is being
used.


> > y'know, we could have a debug option which will spit warnings if someone
> > does a !__GFP_WAIT allocation while !in_atomic() (only works if
> > CONFIG_PREEMPT).  
> > 
> > But please, make it depend on !CONFIG_AKPM.  I shudder to think about all
> > the stuff it would pick up.
> 
> Seems like you'd get a lot of false positives.

There would be a few.  mempool does a non-__GFP_WAIT allocation
deliberately, for example (I still think that's fishy btw).

But I don't expect there would be a large number of falsies.  We could add
a __GFP_I_REALLY_MEANT_ATOMIC flag to shut those up.

>  How about a call:
> 
> slab_warn_about_atomic_allocs();
> 
> right before calling the initcalls, and then
> 
> slab_stop_warning_about_atomic_allocs();
> 
> after calling them?  That should give people a lot to chew on for a few
> months.  Obviously, you would need to not warn about allocations from
> interrupt context, as you say above.

Could.  But GFP_ATOMIC at initcall-time really isn't a problem (except that
it can probably also happen at modprobe-time).

What is the major concern is needlessly atomic allocations at regular
runtime.