OK6410A Development Board (VIII) 40 linux-5.11 OK6410A buddy alloc and free

The third stage is to establish a buddy

Buddy's lifespan

mm_init->mem_init returns - no end point

Buddy managed memory size

The memory managed by buddy is determined by memblock

The portion of memblock.memory that does not include memblock.reserved

memblock can reserve memory through memblock_alloc or memblock_reserve

How to use buddy

alloc

alloc_pages/alloc_page // Returns struct page

get_zeroed_page // Returns the virtual address

__get_free_pages/__get_free_page // Returns the virtual address

get_dma_pages

__get_free_pages

struct page *page = alloc_pages

alloc_pages_node

__alloc_pages_node

__alloc_pages

__alloc_pages_nodemask

struct page * page = get_page_from_freelist

return (unsigned long) page_address(page);

__alloc_pages_nodemask details

__alloc_pages_nodemask

  prepare_alloc_context //1. Prepare parameters

  get_page_from_freelist //2. According to the parameters obtained by prepare_alloc_context, the fast path attempts to allocate memory 

  __alloc_pages_slowpath //3. If the fast path fails to allocate memory, the slow path attempts to allocate memory

For details, please visit https://zhuanlan.zhihu.com/p/258921453

highly recommended

prepare_alloc_context // Set three parameters of get_page_from_freelist function

1. struct alloc_context *ac

ac->migratetype = gfp_migratetype(gfp_mask); // Set the required migration type, related to zone->free_area[order].free_list[migratetype]

2. gfp_t *alloc_mask

*alloc_mask |= __GFP_HARDWALL;

3. unsigned int *alloc_flags

*alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);

get_page_from_freelist 

for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx, ac->nodemask) { // Traverse the zones according to the priority of ZONE_HIGHMEM ZONE_NORMAL ZONE_DMA

mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); // Get the current watermark

if (!zone_watermark_fast(zone, order, mark, ac->highest_zoneidx, alloc_flags, gfp_mask)) { 

// Compare with the watermark FLAG set by alloc_pages

// Entering the if function body at this time indicates that the current watermark <= the watermark FLAG set by alloc_pages, indicating that no more physical pages can be allocated

if (alloc_flags & ALLOC_NO_WATERMARKS) goto try_this_zone;

// Indicates the watermark set by alloc_pages FLAG means ignoring the watermark, and directly allocates pages to the try_this_zone label

}

try_this_zone:

page = rmqueue(ac->preferred_zoneref->zone, zone, order, gfp_mask, alloc_flags, ac->migratetype);

if (likely(order == 0)) rmqueue_pcplist(preferred_zone, zone, gfp_flags, migratetype, alloc_flags);

// Check whether the linked list maintained by per_cpu_pages (pcp) can satisfy the memory request

else __rmqueue(zone, order, migratetype, alloc_flags);

// Apply from the partner system, the query list is

// zone->free_area[order].free_list[migratetype]

__rmqueue_smallest(zone, order, migratetype);

area = &(zone->free_area[current_order]);

page = get_page_from_free_area(area, migratetype);

del_page_from_free_list(page, zone, current_order);

}

__alloc_pages_slowpath

Various mechanisms for invoking memory shortage

...relaim

... compact

... oom killer

... kswapd

get_page_from_freelist(gfp_mask, order, alloc_flags, ac);

free

free_pages/free_page

__free_pages/__free_page

free_pages(unsigned long addr, unsigned int order)

__free_pages(virt_to_page((void *)addr), order);

free_the_page(page, order)

__free_pages_ok

free_one_page

__free_one_page

add_to_free_list(page, zone, order, migratetype);

other

The essence of buddy

The buddy system only maintains free blocks, and the allocated blocks do not belong to the buddy system.

After calling free_pages on some blocks, they will be put into the buddy system.

---

Disassembly & Merge

The essence of buddy is to split and merge according to 2^n

Buddy Problems and Solutions

// Is there any virtual memory problem?

1. Physical memory fragmentation problem

Buddy Merge

Timing: free_pages

method:

Threads: None

Compact

Timing: alloc

method:

Threads: 1 thread kcompactd0

2. Insufficient physical memory

Reclaim

Timing: when alloc memory is scarce/cycle/system sleep

method:

Threads: 2 threads kswapd0 oom_reaper

3. Unable to apply for consecutive physical pages

CMA contiguous memory allocator

Timing: when allocating continuous memory

method: 

A section of memory is reserved for the driver to use, but when the driver is not in use, the CMA area can be allocated to the user process for use as anonymous memory or page cache.

When the driver needs to use it, the memory occupied by the process is recycled or migrated to free up the reserved memory previously occupied for use by the driver.

Threads: None

miscellaneous

---

5. If __GFP_DMA is set, pages can only be obtained from the ZONE_DMA memory management area

6. If __GFP_HIGH is not set, get pages from the ZONE_NORMAL ZONE_DMA memory management area in turn

7. If __GFP_HIGH is set, get pages from ZONE_HIGHMEM ZONE_NORMAL ZONE_DMA memory management areas in turn

---

Page frames are unlimited and can store any data

x86 hardware has restrictions on page frames, resulting in the addition of ZONE (ZONE page frame classification)

There is no zone, or there is only one zone : ZONE_NORMAL 16MB - 896MB

1.DMA can only address the first 16MB of RAM : Added ZONE_DMA 0MB - 16MB

2.32bit cannot address all physical memory : Added ZONE_HIGHMEM 896M - 

What about arm32???

Page frames are unlimited and can store any data

The arm32 hardware has restrictions on page frames, resulting in the addition of ZONE (ZONE page frame classification)

There is no zone, or there is only one zone : ZONE_NORMAL 0MB - 760MB

1.32bit cannot address all physical memory : Added ZONE_HIGHMEM 760MB -  

---

status quo

CONFIG_ZONE_DMA CONFIG_ZONE_DMA32 CONFIG_HIGHMEM is not configured

There are only two zones: ZONE_NORMAL ZONE_MOVABLE

---

For arm32, when do you need to configure CONFIG_HIGHMEM?

200MB does not require CONFIG_HIGHMEM

1GB needs to be configured because the kernel space is only 1G, and the memblock reservation is less than 1G, so it cannot be fully mapped.

---

S3C6410 DMA

6410 has DMA

Source and destination are both available in system bus/peripheral bus (4 cases)

It uses ARM's IP PL080

4 DMAs, 8 channels per DMA

No limit on memory