lkml.org 
[lkml]   [2010]   [Aug]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 04/10] Use percpu buffers
Hi,

On 08/10/2010 10:35 AM, Pekka Enberg wrote:
> On 10.8.2010 7.47, Nitin Gupta wrote:
>> On 08/10/2010 12:27 AM, Pekka Enberg wrote:
>>> On Mon, Aug 9, 2010 at 8:26 PM, Nitin Gupta<ngupta@vflare.org> wrote:
>>>> @@ -303,38 +307,41 @@ static int zram_write(struct zram *zram, struct bio *bio)
>>>> zram_test_flag(zram, index, ZRAM_ZERO))
>>>> zram_free_page(zram, index);
>>>>
>>>> - mutex_lock(&zram->lock);
>>>> + preempt_disable();
>>>> + zbuffer = __get_cpu_var(compress_buffer);
>>>> + zworkmem = __get_cpu_var(compress_workmem);
>>>> + if (unlikely(!zbuffer || !zworkmem)) {
>>>> + preempt_enable();
>>>> + goto out;
>>>> + }
>>> The per-CPU buffer thing with this preempt_disable() trickery looks
>>> overkill to me. Most block device drivers seem to use mempool_alloc()
>>> for this sort of thing. Is there some reason you can't use that here?
>>>
>> Other block drivers are allocating relatively small structs using
>> mempool_alloc(). However, in case of zram, these buffers are quite
>> large (compress_workmem is 64K!). So, allocating them on every write
>> would probably be much slower than using a pre-allocated per-cpu buffer.
> The mempool API is precisely for that - using pre-allocated buffers instead of allocating every time. The preempt_disable() games make the code complex and have the downside of higher scheduling latencies so why not give mempools a try?
>

mempool_alloc() first calls alloc_fn with ~(__GFP_WAIT | __GFP_IO)
and *then* falls down to pre-allocated buffers. So, it will always
be slower than directly using pre-allocated buffers as is done
currently.

One trick we can use is to have alloc_fn such that it always returns
failure with ~__GFP_WAIT and do actual allocation otherwise. But still
it seems like unnecessary cost.

Thanks,
Nitin





\
 
 \ /
  Last update: 2010-08-10 07:35    [W:0.073 / U:0.620 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site