Re: [PATCH] mm: Support memblock alloc on the exact node for sparse_buffer_init()

[Mike Rapoport <rppt@linux.ibm.com>]

On Thu, Sep 19, 2019 at 03:14:22PM +0800, Yunfeng Ye wrote:
> 
> 
> On 2019/9/19 12:47, Mike Rapoport wrote:
> > Hi,
> > 
> > On Wed, Sep 18, 2019 at 12:22:29PM +0800, Yunfeng Ye wrote:
> >> Currently, when memblock_find_in_range_node() fail on the exact node, it
> >> will use %NUMA_NO_NODE to find memblock from other nodes. At present,
> >> the work is good, but when the large memory is insufficient and the
> >> small memory is enough, we want to allocate the small memory of this
> >> node first, and do not need to allocate large memory from other nodes.
> >>
> >> In sparse_buffer_init(), it will prepare large chunks of memory for page
> >> structure. The page management structure requires a lot of memory, but
> >> if the node does not have enough memory, it can be converted to a small
> >> memory allocation without having to allocate it from other nodes.
> >>
> >> Add %MEMBLOCK_ALLOC_EXACT_NODE flag for this situation. Normally, the
> >> behavior is the same with %MEMBLOCK_ALLOC_ACCESSIBLE, only that it will
> >> not allocate from other nodes when a single node fails to allocate.
> >>
> >> If large contiguous block memory allocated fail in sparse_buffer_init(),
> >> it will allocates small block memmory section by section later.
> > 
> > Did you see the sparse_buffer_init() actually falling back to allocate from a
> > different node? If a node does not have enough memory to hold it's own
> > memory map, filling only it with parts of the memory map will not make such
> > node usable.
> >  
> Normally, it won't happen that sparse_buffer_init() falling back from a different
> node, because page structure size is 64 bytes per 4KB of memory, no more than 2%
> of total available memory. But in the special cases, for eaxmple, memory address
> is isolated by BIOS when memory failure, split the total memory many pieces,
> although we have enough memory, but no large contiguous block memory in one node.
> sparse_buffer_init() needs large contiguous block memory to be alloc in one time,
> 
> Eg, the size of memory is 1TB, sparse_buffer_init() need 1TB * 64/4096 = 16GB, but
> we have 100 blocks memory which every block only have 10GB, although total memory
> have almost 100*10GB=1TB, but no contiguous 16GB block.
 
An explanation that a node memory may become highly fragmented should be a
part of the changelog.

> Before commit 2a3cb8baef71 ("mm/sparse: delete old sparse_init and enable new one"),
> we have %CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER config to meeting this situation,
> after that, it fall back to allocate memory from other nodes, so have the performance
> impact by remote numa access.
> 
> commit 85c77f791390 ("mm/sparse: add new sparse_init_nid() and sparse_init()") wrote
> that:
>     "
>     sparse_init_nid(), which only
>     operates within one memory node, and thus allocates memory either in large
>     contiguous block or allocates section by section
>     "
> it means that allocates section by section is a normal choice too, so I think add
> %MEMBLOCK_ALLOC_EXACT_NODE is also a choice for this situation. Most cases,
> sparse_buffer_init() works good and not allocated from other nodes at present.

I'd prefer to see memblock_alloc_exact_nid_raw() wrapper for
memblock_find_in_range_node() rather than using a flag.
 
> thanks.
> Yunfeng Ye
> 
> >> Signed-off-by: Yunfeng Ye <yeyunfeng@huawei.com>
> >> ---
> >>  include/linux/memblock.h | 1 +
> >>  mm/memblock.c            | 3 ++-
> >>  mm/sparse.c              | 2 +-
> >>  3 files changed, 4 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/include/linux/memblock.h b/include/linux/memblock.h
> >> index f491690..9a81d9c 100644
> >> --- a/include/linux/memblock.h
> >> +++ b/include/linux/memblock.h
> >> @@ -339,6 +339,7 @@ static inline int memblock_get_region_node(const struct memblock_region *r)
> >>  #define MEMBLOCK_ALLOC_ANYWHERE	(~(phys_addr_t)0)
> >>  #define MEMBLOCK_ALLOC_ACCESSIBLE	0
> >>  #define MEMBLOCK_ALLOC_KASAN		1
> >> +#define MEMBLOCK_ALLOC_EXACT_NODE	2
> >>
> >>  /* We are using top down, so it is safe to use 0 here */
> >>  #define MEMBLOCK_LOW_LIMIT 0
> >> diff --git a/mm/memblock.c b/mm/memblock.c
> >> index 7d4f61a..dbd52c3c 100644
> >> --- a/mm/memblock.c
> >> +++ b/mm/memblock.c
> >> @@ -277,6 +277,7 @@ static phys_addr_t __init_memblock memblock_find_in_range_node(phys_addr_t size,
> >>
> >>  	/* pump up @end */
> >>  	if (end == MEMBLOCK_ALLOC_ACCESSIBLE ||
> >> +	    end == MEMBLOCK_ALLOC_EXACT_NODE ||
> >>  	    end == MEMBLOCK_ALLOC_KASAN)
> >>  		end = memblock.current_limit;
> >>
> >> @@ -1365,7 +1366,7 @@ static phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
> >>  	if (found && !memblock_reserve(found, size))
> >>  		goto done;
> >>
> >> -	if (nid != NUMA_NO_NODE) {
> >> +	if (end != MEMBLOCK_ALLOC_EXACT_NODE && nid != NUMA_NO_NODE) {
> >>  		found = memblock_find_in_range_node(size, align, start,
> >>  						    end, NUMA_NO_NODE,
> >>  						    flags);
> >> diff --git a/mm/sparse.c b/mm/sparse.c
> >> index 72f010d..828db46 100644
> >> --- a/mm/sparse.c
> >> +++ b/mm/sparse.c
> >> @@ -477,7 +477,7 @@ static void __init sparse_buffer_init(unsigned long size, int nid)
> >>  	sparsemap_buf =
> >>  		memblock_alloc_try_nid_raw(size, PAGE_SIZE,
> >>  						addr,
> >> -						MEMBLOCK_ALLOC_ACCESSIBLE, nid);
> >> +						MEMBLOCK_ALLOC_EXACT_NODE, nid);
> >>  	sparsemap_buf_end = sparsemap_buf + size;
> >>  }
> >>
> >> -- 
> >> 2.7.4.huawei.3
> >>
> >>
> > 
> 

-- 
Sincerely yours,
Mike.