Calling free_pages on part of the memory returned by get_free_pages?

Calling free_pages on part of the memory returned by get_free_pages?

[Timur Tabi]

According to LDD3, if I call get_free_pages() to allocate X bytes, I have to
free all of those pages with free_pages().  The VM internals are a little bit
over my head, but I looked at the code and I didn't see why that is a requirement.

For example, let's say I want to allocated 6MB of physically-contiguous memory.
 If I call x = get_free_pages(11) to get 8MB.  What happens if I then do
"free_pages(x + 6 * 1024 * 1024, 9)"?

I remember doing this on the 2.4 kernel, and it never gave me any problems.

-- 
Timur Tabi
Linux kernel developer at Freescale

Re: Calling free_pages on part of the memory returned by get_free_pages?

[Andi Kleen]

Timur Tabi <timur@freescale.com> writes:

> According to LDD3, if I call get_free_pages() to allocate X bytes, I have to
> free all of those pages with free_pages().  The VM internals are a little bit
> over my head, but I looked at the code and I didn't see why that is a requirement.
>
> For example, let's say I want to allocated 6MB of physically-contiguous memory.
>  If I call x = get_free_pages(11) to get 8MB.  What happens if I then do
> "free_pages(x + 6 * 1024 * 1024, 9)"?
>
> I remember doing this on the 2.4 kernel, and it never gave me any problems.

It is ok, as long as you don't use compound pages (__GFP_COMP) and call
split_page() to fix up the reference counts.

Also you do this to save memory right? The large system hash code does it too
and I used to do it in some 2.4 change with an alloc_pages_exact()
which never made it into 2.6.

If it's reasonably common we should re-add alloc/get_pages_exact() helper to 
make this pattern clear and easier to use.

-Andi

Re: Calling free_pages on part of the memory returned by get_free_pages?

[Timur Tabi]

Andi Kleen wrote:

> It is ok, as long as you don't use compound pages (__GFP_COMP) and call
> split_page() to fix up the reference counts.

Thanks.  I've never even heard of compound pages, so that's not a problem.  I'll
look up split_page().

> Also you do this to save memory right? 

Yes, I expect to allocate a 5MB chunk of memory, which means I'll be wasting 3MB
of physically-contiguous memory on an embedded system.

> The large system hash code does it too

Ah, I see that in alloc_large_system_hash().  Thanks

> If it's reasonably common we should re-add alloc/get_pages_exact() helper to 
> make this pattern clear and easier to use.

I'd be in favor of that!  I was planning on doing something like that anyway.

-- 
Timur Tabi
Linux kernel developer at Freescale

Re: Calling free_pages on part of the memory returned by get_free_pages?

[Timur Tabi]

Andi Kleen wrote:
> Timur Tabi <timur@freescale.com> writes:
> 
>> According to LDD3, if I call get_free_pages() to allocate X bytes, I have to
>> free all of those pages with free_pages().  The VM internals are a little bit
>> over my head, but I looked at the code and I didn't see why that is a requirement.
>>
>> For example, let's say I want to allocated 6MB of physically-contiguous memory.
>>  If I call x = get_free_pages(11) to get 8MB.  What happens if I then do
>> "free_pages(x + 6 * 1024 * 1024, 9)"?
>>
>> I remember doing this on the 2.4 kernel, and it never gave me any problems.
> 
> It is ok, as long as you don't use compound pages (__GFP_COMP) and call
> split_page() to fix up the reference counts.
> 
> Also you do this to save memory right? The large system hash code does it too
> and I used to do it in some 2.4 change with an alloc_pages_exact()
> which never made it into 2.6.
> 
> If it's reasonably common we should re-add alloc/get_pages_exact() helper to 
> make this pattern clear and easier to use.

Follow-up question:

Say I allocate 8MB with __get_free_pages() and then use multiple free_pages() /
split_page() calls to free the last 3MB.  Will the de-allocated blocks be merged
into larger-order chunks, if they're contiguous?  That is, if I repeat this process:

1. Allocate 8MB
2. Free 3MB
3. Free 5MB

over and over again, will I totally fragment memory to the point that 8MB
allocates will never work again?

-- 
Timur Tabi
Linux kernel developer at Freescale