Re: [PATCH 2/4] virtual block device driver (ramzswap)

[Pekka Enberg <penberg@cs.helsinki.fi>]

On Tue, Sep 15, 2009 at 11:21 AM, Nitin Gupta <ngupta@vflare.org> wrote:
> I don't want to ponder too much about this point now. If you all are okay
> with keeping this function buried in driver, I will do so. I'm almost tired
> maintaining this compcache thing outside of mainline.

Yup, whatever makes most sense to you.

>> Then make ramzswap depend on !CONFIG_ARM. In any case, CONFIG_ARM bits
>> really don't belong into drivers/block.
>
> ARM is an extremely important user of compcache -- Its currently being
> tested (unofficially) on Android, Nokia etc.

That's not a technical argument for keeping CONFIG_ARM in the driver.

>>>>> +
>>>>> +       trace_mark(ramzswap_lock_wait, "ramzswap_lock_wait");
>>>>> +       mutex_lock(&rzs->lock);
>>>>> +       trace_mark(ramzswap_lock_acquired, "ramzswap_lock_acquired");
>>>>
>>>> Hmm? What's this? I don't think you should be doing ad hoc
>>>> trace_mark() in driver code.
>>>
>>> This is not ad hoc. It is to see contention over this lock which I believe is a
>>> major bottleneck even on dual-cores. I need to keep this to measure improvements
>>> as I gradually make this locking more fine grained (using per-cpu buffer etc).
>>
>> It is ad hoc. Talk to the ftrace folks how to do it properly. I'd keep
>> those bits out-of-tree until the issue is resolved, really.
>
> /me is speechless.

That's fine, I CC'd the ftrace folks. Hopefully they'll be able to help you.

>
>>>>> +       rzs->compress_buffer = kzalloc(2 * PAGE_SIZE, GFP_KERNEL);
>>>>
>>>> Use alloc_pages(__GFP_ZERO) here?
>>>
>>> alloc pages then map them (i.e. vmalloc). What did we gain? With
>>> vmalloc, pages might
>>> not be physically contiguous which might hurt performance as
>>> compressor runs over this buffer.
>>>
>>> So, use kzalloc().
>>
>> I don't know what you're talking about. kzalloc() calls
>> __get_free_pages() directly for your allocation. You probably should
>> use that directly.
>
> What is wrong with kzalloc? I'm wholly totally stumped.
> I respect your time reviewing the code but this really goes over my head.
> We can continue arguing about get_pages vs kzalloc but I doubt if we will
> gain anything out of it.

The slab allocator needs metadata for the allocation so you're wasting
memory. If you really want *two pages*, why don't you simply use the
page allocator for that?

Btw, Nitin, why are you targeting drivers/block and not
drivers/staging at this point? It seems obvious enough that there are
still some issues that need to be ironed out (like the CONFIG_ARM
thing) so submitting the driver for inclusion in drivers/staging and
fixing it up there incrementally would likely save you from a lot of
trouble. Greg, does ramzswap sound like something that you'd be
willing to take?

                        Pekka