Re: [PATCH v2] mm: kmemleak: Use mempool allocations for kmemleak objects

[Catalin Marinas]

On Thu, Aug 01, 2019 at 08:41:53AM +0200, Michal Hocko wrote:
> On Wed 31-07-19 16:44:50, Catalin Marinas wrote:
> > On Tue, Jul 30, 2019 at 01:02:15PM -0700, Andrew Morton wrote:
> > > On Sat, 27 Jul 2019 14:23:33 +0100 Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> > > > Add mempool allocations for struct kmemleak_object and
> > > > kmemleak_scan_area as slightly more resilient than kmem_cache_alloc()
> > > > under memory pressure. Additionally, mask out all the gfp flags passed
> > > > to kmemleak other than GFP_KERNEL|GFP_ATOMIC.
> > > > 
> > > > A boot-time tuning parameter (kmemleak.mempool) is added to allow a
> > > > different minimum pool size (defaulting to NR_CPUS * 4).
> > > 
> > > btw, the checkpatch warnings are valid:
> > > 
> > > WARNING: usage of NR_CPUS is often wrong - consider using cpu_possible(), num_possible_cpus(), for_each_possible_cpu(), etc
> > > #70: FILE: mm/kmemleak.c:197:
> > > +static int min_object_pool = NR_CPUS * 4;
> > > 
> > > WARNING: usage of NR_CPUS is often wrong - consider using cpu_possible(), num_possible_cpus(), for_each_possible_cpu(), etc
> > > #71: FILE: mm/kmemleak.c:198:
> > > +static int min_scan_area_pool = NR_CPUS * 1;
> > > 
> > > There can be situations where NR_CPUS is much larger than
> > > num_possible_cpus().  Can we initialize these tunables within
> > > kmemleak_init()?
> > 
> > We could and, at least on arm64, cpu_possible_mask is already
> > initialised at that point. However, that's a totally made up number. I
> > think we would better go for a Kconfig option (defaulting to, say, 1024)
> > similar to the CONFIG_DEBUG_KMEMLEAK_EARLY_LOG_SIZE and we grow it if
> > people report better values in the future.
> 
> If you really want/need to make this configurable then the command line
> parameter makes more sense - think of distribution kernel users for
> example.

I doubt you'd have pre-built distribution kernels with kmemleak enabled.

> But I am still not sure why this is really needed. The initial
> size is a "made up" number of course. There is no good estimation to
> make (without a crystal ball). The value might be increased based on
> real life usage.

We had a similar situation with the early log buffer (before slab is
initialised), initially 400 which was good enough for my needs (embedded
systems) but others had entirely different requirements. A configurable
(cmdline, Kconfig) option would make it easier for people to change,
especially if coupled with a meaningful suggestion in dmesg.

Another option is to use the early log as an emergency pool after
initialisation instead of freeing it (it's currently __initdata) and
drop the mempool idea. I may give this a go, at least we only have a
single Kconfig option.

-- 
Catalin