lkml.org 
[lkml]   [2005]   [Dec]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC] genalloc != generic DEVICE memory allocator
Hi Jes,

Jes Sorensen wrote:
>>>>>>"Andrey" == Andrey Volkov <avolkov@varma-el.com> writes:
>
>
> Andrey> Hello Jes and all I try to use your allocator (gen_pool_xxx),
> Andrey> idea of which is a cute nice thing. But current implementation
> Andrey> of it is inappropriate for a _device_ (aka onchip, like
> Andrey> framebuffer) memory allocation, by next reasons:
>
> Andrey,
>
> Keep in mind that genalloc was meant to be simple for basic memory
> allocations. It was never meant to be an over complex super high
> performance allocation mechanism.
>
> Andrey> 1) Device memory is expensive resource by access time and/or
> Andrey> size cost. So we couldn't use (usually) this memory for the
> Andrey> free blocks lists.
>
> This really is irrelevant, the space is only used within the object
> when it's on the free list. Ie. if all memory is handed out there's
> no space used for this purpose.

I point out 2 reasons: ACCESS TIME was first :), let take very
widespread case: PCI device with some onboard memory and any
N GHz proc. - result may be terrible: each access to device mem (which
usually uncached) will slowed down this super fast proc to 33 MHZ, i.e
same as we made busy-wait with disabled interrupts after each read/write...

I possible awry when use 'control structures' in 2), I've in view
allocator's control structures (size/next etc), not device specific
control structs.

>
> Andrey> 3) Obvious (IMHO) workflow of mem. allocator
> Andrey> look like: - at startup time, driver allocate some big
> Andrey> (almost) static mem. chunk(s) for a control/data structures.
> Andrey> - during work of the device, driver allocate many small
> Andrey> mem. blocks with almost identical size. such behavior lead to
> Andrey> degeneration of buddy method and transform it to the
> Andrey> first/best fit method (with long seek by the free node list).
>
> This is only really valid for network devices, and even then it's not
> quite so. For things like uncached allocations your observation is
> completely off.

Could you give me some examples? Possible I overlooked something
significant.

>
> For the case of more traditional devices, the control structures will
> be allocated from one end of the block, the rest will be used for
> packet descriptors which will be going in and out of the memory pool
> on a regular basis.

This was main reason why I try to modify genalloc: I needed in
generic allocator for both short-live strictly aligned blocks and
long-live blocks with restriction by size.

> In most normal cases these will all be of the same
> size and it doesn't matter where in the memory space they were
> allocated.

And thats also why I consider that 'buddy' is not appropriate to be
'generic' (most cases == generic, isn't is :)?): when you're allocate
mainly same sized blocks, 'buddy' degraded to the first-fit.

Possible solution I see in mixed first-fit with lazy coalescent for
short lived blocks and first-fit with immediately coalescent for
long-lived blocks. But, again, I may overlook something significant.
And, certainly, I could overlooked someone else allocator implementation
in some driver.

>
> Andrey> 4) The simple binary buddy method is far away from perfect for
> Andrey> a device due to a big internal fragmentation. Especially for a
> Andrey> network/mfd devices, for which, size of allocated data very
> Andrey> often is not a power of 2.
>
snip
>
> Andrey> I start to modify your code to satisfy above demands, but
> Andrey> firstly I wish to know your, or somebody else, opinion.
>
> I honestly don't think the majority of your demands are valid.
> genalloc was meant to be simple, not an ultra fast at any random
> block size allocator. So far I don't see any reason for changing to
> the allocation algorithm into anything much more complex - doesn't
> mean there couldn't be a reason for doing so, but I don't think you
> have described any so far.
I disagree here, generic couldn't be very simple and slow, because in
this case simply no one will be use it, and hence we'll get today's
picture: reimplemented allocators in many drivers.

>
> You mentioned frame buffers, but what is the kernel supposed to do
> with those allocation wise? If you have a frame buffer console, the
> memory is allocated once and handed to the frame buffer driver.
> Ie. you don't need a ton of on demand allocations for that and for
> X, the memory management is handled in the X server, not by the
> kernel.

For video-only device this is true, but if device is a multifunctional,
which is frequent case in embedded systems, then kernel must control of
device memory allocation. Currently, however, even video cards for
desktops become more and more multifunctional (VIVO/audio etc.).

>
> The only thing I think would make sense to implement is to allow it to
> use indirect descriptor blocks for the memory it manages. This is not
> because it's wrong to use the memory for the free list, as it will
> only be used for this when the chunk is not in use, but because access
> to certain types of memory isn't always valid through normal direct
> access. Ie. if one used descriptor blocks residing in normal
> GFP_KERNEL memory, it would be possible to use the allocator to manage
> memory sitting on the other side of a PCI bus.
I describe above, why we couldn't/wouldn't use onboard memory for
allocator specific data.

Pantelis, Am I answered to your question (...what are you trying to
do...) too?

--
Regards
Andrey Volkov
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-12-22 19:24    [from the cache]
©2003-2011 Jasper Spaans