per-cpu stats in block device: overkill?

per-cpu stats in block device: overkill?

[Rusty Russell]

Hi Jens, Tejun, Jerome,

  I've been auditing alloc_per_cpu users, and got to genhd.  The code is fairly complex, but I can't help wondering if per-cpu counters are overkill.  After all, we have a single queue lock.

  The reason I care is that I'm changing alloc_per_cpu to use the static per-cpu area: at 40/80 bytes (32/64 bit) per stat, we'd be restricted to a few hundred disks unless the percpu area is enlarged (in current patches, a cmdline param).  Or, I can change genhd to use big_percpu_alloc which will use the current inefficient dynamic per-cpu system until we get dynamic per-cpu regions (if ever).

Cheers,
Rusty.

Re: per-cpu stats in block device: overkill?

[Tejun Heo]

Hello, Rusty.

Rusty Russell wrote:
> Hi Jens, Tejun, Jerome,
> 
> I've been auditing alloc_per_cpu users, and got to genhd.  The code
> is fairly complex, but I can't help wondering if per-cpu counters
> are overkill.  After all, we have a single queue lock.

Yeah, maybe.

> The reason I care is that I'm changing alloc_per_cpu to use the
> static per-cpu area: at 40/80 bytes (32/64 bit) per stat, we'd be
> restricted to a few hundred disks unless the percpu area is enlarged
> (in current patches, a cmdline param).  Or, I can change genhd to
> use big_percpu_alloc which will use the current inefficient dynamic
> per-cpu system until we get dynamic per-cpu regions (if ever).

I'm working on local counter (local_t) allocator which is used to
replace percpu allocation in percpu_counter and used as basis for
percpu_ref which replaces module ref counting and will be used to
simplify block/char lifetime rules.

The local counter allocator allocates per-cpu pages and the space
overhead is minimal.  If per-cpu stats in genhd is necessary, I think
converting it to percpu local counter allocation should do it.

BTW, why make percpu area static?

Thanks.

-- 
tejun

Re: per-cpu stats in block device: overkill?

[Rusty Russell]

On Monday 22 December 2008 12:25:54 Tejun Heo wrote:
> I'm working on local counter (local_t) allocator which is used to
> replace percpu allocation in percpu_counter and used as basis for
> percpu_ref which replaces module ref counting and will be used to
> simplify block/char lifetime rules.

Hi Tejun,

   Interesting!  Thanks to Christoph's dynamic percpu efforts, I've
been revising efforts to make alloc_percpu use the same efficient
mechanism that static percpu vars use.  We actually have this code
already, tucked away in module.c.

   This work is basically complete; the step I started this morning
is to remove the per_cpu__ prefix hackery from the per-cpu ops (in favour
of sparse annotations).  This leads to cpu_local_inc et. al. being usable
for alloc_percpu-created percpu vars, not just static ones.

> The local counter allocator allocates per-cpu pages and the space
> overhead is minimal.  If per-cpu stats in genhd is necessary, I think
> converting it to percpu local counter allocation should do it.

   Interesting; an allyesconfig boot uses 194 per-cpu allocs from
lib/percpu_counter.c at the moment.  The module.c allocator is fairly space
efficient: 4 bytes per "block" (ie. each allocation or hole) but slow,
which I figure is OK.  Packing is good though.

> BTW, why make percpu area static?

   Good question.  Archs use a simple offset for per-cpu areas: some hold
this in a register (eg. %fs for x86-32).  This means that the layout must be
"congruent" (ie. have the same inter-cpu spacing) if we allocate a new
per-cpu area (hard for non-NUMA).

  For 5 years I waited for this to be fixed, and avoided exposing the per-cpu
core, and the alloc_percpu stuff was a standin implementation.  But Christoph
L. showed that even with the size limit, there are numerous places which want
small per-cpu allocations which are optimally accessed, so I restarted work.
See Message-Id: <20081117132630.33F09DDDF5@ozlabs.org> "[PATCH 1/7] Improve
alloc_percpu: make the per cpu reserve configurable and larger." and thread.

  In addition, Mathieu and I have been discussing local_t: it's wandered
off its original purpose and we're debating what to do about it.  See
Message-ID: <alpine.DEB.1.10.0812150823370.18692@gandalf.stny.rr.com>
"local_add_return" and thread.

I look forward to always-cogent your thoughts on these issues!
Rusty.