Re: [PATCH v2] module: check if memory leak by module.

[Michal Hocko]

On Fri 31-03-17 10:05:29, Joel Fernandes wrote:
> On Fri, Mar 31, 2017 at 1:00 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > On Thu 30-03-17 23:49:52, Joel Fernandes wrote:
> >> Hi Michal,
> >>
> >> On Wed, Mar 29, 2017 at 3:43 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >> > On Wed 29-03-17 09:23:32, Vaneet Narang wrote:
> >> >> Hi,
> >> >>
> >> >> >> Hmm, how can you track _all_ vmalloc allocations done on behalf of the
> >> >> >> module? It is quite some time since I've checked kernel/module.c but
> >> >> >> from my vague understading your check is basically only about statically
> >> >> >> vmalloced areas by module loader. Is that correct? If yes then is this
> >> >> >> actually useful? Were there any bugs in the loader code recently? What
> >> >> >> led you to prepare this patch? All this should be part of the changelog!
> >> >>
> >> >> First of all there is no issue in kernel/module.c. This patch add functionality
> >> >> to detect scenario where some kernel module does some memory allocation but gets
> >> >> unloaded without doing vfree. For example
> >> >> static int kernel_init(void)
> >> >> {
> >> >>         char * ptr = vmalloc(400 * 1024);
> >> >>         return 0;
> >> >> }
> >> >
> >> > How can you track that allocation back to the module? Does this patch
> >> > actually works at all? Also why would be vmalloc more important than
> >> > kmalloc allocations?
> >>
> >> Doesn't the patch use caller's (in this case, the module is the
> >> caller) text address for tracking this? vma->vm->caller should track
> >> the caller doing the allocation?
> >
> > Not really. First of all it will be vmalloc() to be tracked in the above
> > the example because vmalloc is not inlined. And secondly even if the
> 
> vmalloc is not inlined, but __built_in_address(0) will return the
> *return address* of vmalloc:
> 
> >From https://gcc.gnu.org/onlinedocs/gcc-4.6.0/gcc/Return-Address.html :
> "The level argument is number of frames to scan up the call stack. A
> value of 0 yields the return address of the current function"

yes, sorry, I meant to say s@vmalloc is not @__vmalloc_node_flags is not@
I can see some arguments to make __vmalloc_node_flags inline to make
/proc/vmallocinfo output more useful but...

> > caller of the vmalloc was tracked then it would be hopelessly
> > insufficient because you would get coverage of the _direct_ module usage
> > of vmalloc rather than anything that the module triggered and that is
> > outside of the module. Which means any library function etc...
> 
> Yes true, but I think the check is for direct allocations, done by the
> module, not indirect ones... it may not be a catch-all issues type of
> deal but is still IMO a good check since we already have
> va->vm->caller available.

I disagree. We have a full featured kmemleak to catch all potential
leaks. This code is IMHO not worth it.

-- 
Michal Hocko
SUSE Labs