Re: [PATCH] staging: zcache: fix crash on cpu remove

[Dave Hansen <dave@linux.vnet.ibm.com>]

On Thu, 2011-10-06 at 14:28 -0500, Seth Jennings wrote:
> In the case that a cpu is taken offline before zcache_do_preload() is
> ever called on the cpu, the per-cpu zcache_preloads structure will
> be uninitialized.  In the CPU_DEAD case for zcache_cpu_notifier(),
> kp->obj is not checked before calling kmem_cache_free() on it.
> If it is NULL, a crash results.
> 
> This patch ensures that both kp->obj and kp->page are not NULL before
> calling the respective free functions. In practice, just checking
> one or the other should be sufficient since they are assigned together
> in zcache_do_preload(), but I check both for safety.

zcache_objnode_alloc() and zcache_obj_alloc() seem to do the same thing:
dereference 'kp' without checking it for NULL.  That's fine, but it
requires that preemption be off between zcache_do_preload() and calling
those functions.  That seems to be the case today, but it worries me.

I can see another user getting added that does zcache_obj_alloc() but
forgot to disable preemption.  

We at least need a BUG_ON(!in_atomic()) near the dereferences.  The
enable/disable paths out of zcache_do_preload() look OK to me, but
they're not trivial.

Was there another purpose behind the preempt_disabling()?  We have a
number of spots on the kernel that do preloads but *don't* protect the
percpu buffers in the same way.  The radix-tree code, for instance, uses
preempt_disable() to provide the guarantee that an allocation does not
occur, but its data structures are safe to access without preempt.

IOW, with the radix code, a missed preempt_disable() costs you an
-ENOMEM.  With zcache it costs you an oops.

-- Dave