You might need vmalloc_sync_all()

You might need vmalloc_sync_all()

[Andi Kleen]

Since someone mentioned again.

Right now only x86 implements vmalloc_sync_all(). But iff:

- Your module mappings might require a page fault calling C 
code in kernel
- You got a pagefault notifier at the beginning of the page fault
handler

you risk a endless pagefault recursion with modular pagefault
notifiers. vmalloc_sync_all should sync all the available mappings 
on notifier register so that the C pf handler is not called for module 
faults.

Apparently at least ARM has this problem.

If the kernel module fault is only assembler code without notifier
(e.g. to reload a TLB) it is probably not needed.

-Andi

Re: You might need vmalloc_sync_all()

[Russell King]

On Tue, May 01, 2007 at 06:26:55AM +0200, Andi Kleen wrote:
> 
> Since someone mentioned again.
> 
> Right now only x86 implements vmalloc_sync_all(). But iff:
> 
> - Your module mappings might require a page fault calling C 
> code in kernel

Yes, just like x86.

> - You got a pagefault notifier at the beginning of the page fault
> handler

What's this?

> Apparently at least ARM has this problem.

Is there a bug report?

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: You might need vmalloc_sync_all()

[Andi Kleen]

> What's this?

Hmm, when I look now at your source I don't see one either. Maybe
the reporter had a strangely patched kernel.

Anyways, point stands -- for on demand module mappings you likely
want vmalloc_sync_all to avoid nested faults. If you think you 
can handle the nested faults safely then ignore it @)

> 
> > Apparently at least ARM has this problem.
> 
> Is there a bug report?

Forwarded it.

-Andi

Re: You might need vmalloc_sync_all()

[Russell King]

On Tue, May 01, 2007 at 01:37:33PM +0200, Andi Kleen wrote:
> > What's this?
> 
> Hmm, when I look now at your source I don't see one either. Maybe
> the reporter had a strangely patched kernel.
> 
> Anyways, point stands -- for on demand module mappings you likely
> want vmalloc_sync_all to avoid nested faults. If you think you 
> can handle the nested faults safely then ignore it @)
> 
> > 
> > > Apparently at least ARM has this problem.
> > 
> > Is there a bug report?
> 
> Forwarded it.

For the record here, Andi forwarded the report about kprobes on ARM
causing recursive page faults.  In essence, my opinion is that the
kprobes hook is probably incorrectly placed - probably far too early
in the data exception processing.

Since the data exception path is used for all sorts of things (missing
page table entries, permission, ECC errors, hardware errors, alignment
faults, etc) putting a call designed to intercept page faults early
will pick up all this other noise.

If it's placed in the same way as i386 then there shouldn't be an issue.

Since I haven't seen the code which adds this hook to ARM, I can only
speculate at the time being.  So I feel that until the hook comes up
as a candidate for merging, there's no point in fixing a problem in
mainline which doesn't yet exist there.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of: