On 08/21/2012 12:03 AM, Christoph Lameter wrote:
> V11->V12
> - Rediff against current slab/next from Pekka
> - Drop label name change patch
> 
> V10->V11
> - Fix issues pointed out by Joonsoo and Glauber
> - Simplify Slab bootstrap further
> 
> V9->V10
> - Memory leak was a false alarm
> - Resequence patches to make it easier
>   to apply.
> - Do more boot sequence consolidation in slab/slub.
>   [We could still do much more like common kmalloc
>   handling]
> - Fixes suggested by David and Glauber
> 
> V8->V9:
> - Fix numerous things pointed out by Glauber.
> - Cleanup the way error handling works in the
>   common kmem_cache_create() function.
> - General cleanup by breaking things up
>   into multiple patches were necessary.
> 
> V7->V8:
> - Do not use kfree for kmem_cache in slub.
> - Add more patches up to a common
>   scheme for object alignment.
> 
> V6->V7:
> - Omit pieces that were merged for 3.6
> - Fix issues pointed out by Glauber.
> - Include the patches up to the point at which
>   the slab name handling is unified
> 
> V5->V6:
> - Patches against Pekka's for-next tree.
> - Go slow and cut down to just patches that are safe
>   (there will likely be some churn already due to the
>   mutex unification between slabs)
> - More to come next week when I have more time (
>   took me almost the whole week to catch up after
>   being gone for awhile).
> 
> V4->V5
> - Rediff against current upstream + Pekka's cleanup branch.
> 
> V3->V4:
> - Do not use the COMMON macro anymore.
> - Fixup various issues
> - No general sysfs support yet due to lockdep issues with
>   keys in kmalloc'ed memory.
> 
> V2->V3:
> - Incorporate more feedback from Joonsoo Kim and Glauber Costa
> - And a couple more patches to deal with slab duping and move
>   more code to slab_common.c
> 
> V1->V2:
> - Incorporate glommers feedback.
> - Add 2 more patches dealing with common code in kmem_cache_destroy
> 
> This is a series of patches that extracts common functionality from
> slab allocators into a common code base. The intend is to standardize
> as much as possible of the allocator behavior while keeping the
> distinctive features of each allocator which are mostly due to their
> storage format and serialization approaches.
> 
> This patchset makes a beginning by extracting common functionality in
> kmem_cache_create() and kmem_cache_destroy(). However, there are
> numerous other areas where such work could be beneficial:
> 
> 1. Extract the sysfs support from SLUB and make it common. That way
>    all allocators have a common sysfs API and are handleable in the same
>    way regardless of the allocator chose.
> 
> 2. Extract the error reporting and checking from SLUB and make
>    it available for all allocators. This means that all allocators
>    will gain the resiliency and error handling capabilties.
> 
> 3. Extract the memory hotplug and cpu hotplug handling. It seems that
>    SLAB may be more sophisticated here. Having common code here will
>    make it easier to maintain the special code.
> 
> 4. Extract the aliasing capability of SLUB. This will enable fast
>    slab creation without creating too many additional slab caches.
>    The arrays of caches of varying sizes in numerous subsystems
>    do not cause the creation of numerous slab caches. Storage
>    density is increased and the cache footprint is reduced.
> 
> Ultimately it is to be hoped that the special code for each allocator
> shrinks to a mininum. This will also make it easier to make modification
> to allocators.
> 
> In the far future one could envision that the current allocators will
> just become storage algorithms that can be chosen based on the need of
> the subsystem. F.e.
> 
> Cpu cache dependend performance		= Bonwick allocator (SLAB)
> Minimal cycle count and cache footprint	= SLUB
> Maximum storage density			= K&R allocator (SLOB)
> 
> 
With the whole series applied, I get a bug (dmesg attached). Allocator
is SLUB, with all the debugging options ontop.

Triggered by executing the test routine "mybug()" after the kernel is
fully functional, and then issuing "cat /proc/slabinfo".

This is the test case I was using before, but this time it all works
immediately after the caches are destructed and recreated - so it's
better. But transversing the list triggers it.

Code I used is below:

void mybug(void)
{
        struct kmem_cache *c1;

        c1 = KMEM_CACHE(st1,
                SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);

        printk("c1: %p\n", c1);
        kmem_cache_destroy(c1);

        c1 = KMEM_CACHE(st1,
                SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
        printk("c1 again: %p\n", c1);

        kmem_cache_destroy(c1);
}

I tried to bisect it to the precise point, but I couldn't. The series is
not runtime bisectable. Since it is such a fragile code, having all
patches to at least boot would be of great help. (I'll post the output
in reply to the relevant patch)