Re: [RFC 0/2] guarantee natural alignment for kmalloc()

[Vlastimil Babka]

On 4/7/19 10:00 AM, Christoph Hellwig wrote:
> On Fri, Apr 05, 2019 at 07:11:17PM +0200, Vlastimil Babka wrote:
>> On 3/22/19 6:52 PM, Christopher Lameter wrote:
>> > On Thu, 21 Mar 2019, Vlastimil Babka wrote:
>> > 
>> >> That however doesn't work well for the xfs/IO case where block sizes are
>> >> not known in advance:
>> >>
>> >> https://lore.kernel.org/linux-fsdevel/20190225040904.5557-1-ming.lei@xxxxxxxxxx/T/#ec3a292c358d05a6b29cc4a9ce3ae6b2faf31a23f
>> > 
>> > I thought we agreed to use custom slab caches for that?
>> 
>> Hm maybe I missed something but my impression was that xfs/IO folks would have
>> to create lots of them for various sizes not known in advance, and that it
>> wasn't practical and would welcome if kmalloc just guaranteed the alignment.
>> But so far they haven't chimed in here in this thread, so I guess I'm wrong.
> 
> Yes, in XFS we might have quite a few.  Never mind all the other
> block level consumers that might have similar reasonable expectations
> but haven't triggered the problematic drivers yet.

What about a LSF session/BoF to sort this out, then? Would need to have people
from all three MM+FS+IO groups, I suppose.