Re:Re: [PATCH] mm/page_alloc: fix the potential memory waste in page_frag_alloc_align

[王伟]

At 2023-10-24 00:17:41, "Matthew Wilcox" <willy@xxxxxxxxxxxxx> wrote:
>On Mon, Oct 23, 2023 at 11:42:16PM +0800, wang wei wrote:
>> First step, allocating a memory fragment with size 1KB bytes uses
>> page_frag_alloc_align.  It will allocate PAGE_FRAG_CACHE_MAX_SIZE
>> bytes by __page_frag_cache_refill, store the pointer at nc->va and
>> then return the last 1KB memory fragment which address is nc->va
>> + PAGE_FRAG_CACHE_MAX_SIZE - 1KB. The remaining PAGE_FRAG_CACHE_MAX_SIZE
>> - 1KB bytes of memory can Meet future memory requests.
>> 
>> Second step, if the caller requests a memory fragment with size
>> more then PAGE_FRAG_CACHE_MAX_SIZE bytes,  page_frag_alloc_align,
>> it will also allocate PAGE_FRAG_CACHE_MAX_SIZE bytes by
>> __page_frag_cache_refill,  store the pointer at nc->va, and
>> return NULL.  this behavior makes the rest of
>> PAGE_FRAG_CACHE_MAX_SIZE - 1KB bytes memory at First step are
>> wasted(allocate from buddy system but not used).
>
>We could do this, but have you ever seen this happen, or are you
>just reading code and looking for problems?  If the latter, I think
>you've misunderstood how this allocator is normally used.

I am indeed reading the kernel mm-subsystem code recently. And I did 
two experiments in a ARM32 platform when i read page_frag_alloc_align code. 
In this platform, PAGE_FRAG_CACHE_MAX_SIZE = 32767(32K).

First experiment code:

static int __init test_ko_init(void)
{
	struct page_frag_cache nc;
	void *data1, *data2;

	data1 = __alloc_page_frag(&nc, 33 * 1024, NULL);
	printk("alloc 33KB data1 = %x, nc.va = %x \n", data1, nc.va);
	data2 = __alloc_page_frag(&nc, 33 * 1024, NULL);
	printk("alloc 33KB data2 = %x, nc.va = %x \n", data2, nc.va);
	return 0;
}

output:

[   19.939263] alloc 33KB data1 = 0, nc.va = c3930000
[   19.943790] alloc 33KB data2 = 0, nc.va = c3930000

In this experiment, i allocate two 33KB-size memory.  
__alloc_page_frag returns NULL indicates the memory request could not be meet.

Second experiment code:

static int __init test_ko_init(void)
{
	struct page_frag_cache nc;
	void *data1, *data2;

	data1 = __alloc_page_frag(&nc, 1 * 1024, NULL);
	printk("alloc 1KB data1 = %x, nc.va = %x \n", data1, nc.va);
	data2 = __alloc_page_frag(&nc, 33 * 1024, NULL);
	printk("alloc 33KB data2 = %x, nc.va = %x \n", data2, nc.va);
	return 0;
}

output:

[   15.906566] alloc 1KB data1 = c3937c00, nc.va = c3930000
[   15.911632] alloc 33KB data2 = 0, nc.va = c3938000

In this experiment, i allocate 1KB-size and 33KB-size memory.  
__alloc_page_frag returns NULL when i allocate 33KB-size memory.
But, nc.va is new pointer to a 32KB memory fragment. So the memory 
between c3930000 and c3937bff is unused. this what i said memory waste.

Based on the two experiments above, I suspect this will happen on 
other platforms as well. 

I am new commer for kernel code, and I am not very understand 
lost of kernel functions' use. Thanks for your advice.