[RFC 06/23] mm/memblock: Add memblock early memory allocation apis

[santosh.shilimkar@ti.com (Santosh Shilimkar)]

On Sunday 13 October 2013 01:56 PM, Tejun Heo wrote:
> Hello,
> 
> On Sat, Oct 12, 2013 at 05:58:49PM -0400, Santosh Shilimkar wrote:
>> Introduce memblock early memory allocation APIs which allow to support
>> LPAE extension on 32 bits archs. More over, this is the next step
> 

[..]

>> +/* FIXME: Move to memblock.h at a point where we remove nobootmem.c */
>> +void *memblock_early_alloc_try_nid_nopanic(int nid, phys_addr_t size,
>> +		phys_addr_t align, phys_addr_t from, phys_addr_t max_addr);
>> +void *memblock_early_alloc_try_nid(int nid, phys_addr_t size,
>> +		phys_addr_t align, phys_addr_t from, phys_addr_t max_addr);
> 
> Wouldn't it make more sense to put @nid at the end.  @size is the main
> parameter here and it gets confusing with _alloc_node() interface as
> the positions of paramters change.  Plus, kmalloc_node() puts @node at
> the end too.
> 
Ok. Will make @nid as a last parameter.

>> +void __memblock_free_early(phys_addr_t base, phys_addr_t size);
>> +void __memblock_free_late(phys_addr_t base, phys_addr_t size);
> 
> Would it be possible to drop "early"?  It's redundant and makes the
> function names unnecessarily long.  When memblock is enabled, these
> are basically doing about the same thing as memblock_alloc() and
> friends, right?  Wouldn't it make more sense to define these as
> memblock_alloc_XXX()?
> 
A small a difference w.r.t existing memblock_alloc() vs these new
exports returns virtual mapped memory pointers. Actually I started
with memblock_alloc_xxx() but then memblock already exports memblock_alloc_xx()
returning physical memory pointer. So just wanted to make these interfaces
distinct and added "early". But I agree with you that the 'early' can
be dropped. Will fix it.

>> +#define memblock_early_alloc(x) \
>> +	memblock_early_alloc_try_nid(MAX_NUMNODES, x, SMP_CACHE_BYTES, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
>> +#define memblock_early_alloc_align(x, align) \
>> +	memblock_early_alloc_try_nid(MAX_NUMNODES, x, align, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
>> +#define memblock_early_alloc_nopanic(x) \
>> +	memblock_early_alloc_try_nid_nopanic(MAX_NUMNODES, x, SMP_CACHE_BYTES, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
>> +#define memblock_early_alloc_pages(x) \
>> +	memblock_early_alloc_try_nid(MAX_NUMNODES, x, PAGE_SIZE, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
>> +#define memblock_early_alloc_pages_nopanic(x) \
>> +	memblock_early_alloc_try_nid_nopanic(MAX_NUMNODES, x, PAGE_SIZE, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
> 
> I always felt a bit weird about _pages() interface.  It says pages but
> takes bytes in size.  Maybe we're better off just converting the
> current _pages users to _alloc_align()?
> 
I thought the pages interfaces are more for asking the memory
allocations which are page aligned. So yes, we could convert
these users to make use of align interfaces.


>> +#define memblock_early_alloc_node(nid, x) \
>> +	memblock_early_alloc_try_nid(nid, x, SMP_CACHE_BYTES, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
>> +#define memblock_early_alloc_node_nopanic(nid, x) \
>> +	memblock_early_alloc_try_nid_nopanic(nid, x, SMP_CACHE_BYTES, \
>> +			BOOTMEM_LOW_LIMIT, BOOTMEM_ALLOC_ACCESSIBLE)
> 
> Ditto as above.  Maybe @nid can be moved to the end?
>
ok
 
>> +static void * __init _memblock_early_alloc_try_nid_nopanic(int nid,
>> +				phys_addr_t size, phys_addr_t align,
>> +				phys_addr_t from, phys_addr_t max_addr)
>> +{
>> +	phys_addr_t alloc;
>> +	void *ptr;
>> +
>> +	if (WARN_ON_ONCE(slab_is_available())) {
>> +		if (nid == MAX_NUMNODES)
> 
> Shouldn't we be using NUMA_NO_NODE?
> 
>> +			return kzalloc(size, GFP_NOWAIT);
>> +		else
>> +			return kzalloc_node(size, GFP_NOWAIT, nid);
> 
> And kzalloc_node() understands NUMA_NO_NODE.
> 
Will try this out.

>> +	}
>> +
>> +	if (WARN_ON(!align))
>> +		align = __alignof__(long long);
> 
> Wouldn't SMP_CACHE_BYTES make more sense?  Also, I'm not sure we
> actually want WARN on it.  Interpreting 0 as "default align" isn't
> that weird.
> 
Will drop that WARN and use SMP_CACHE_BYTES as a default.


>> +	/* align @size to avoid excessive fragmentation on reserved array */
>> +	size = round_up(size, align);
>> +
>> +again:
>> +	alloc = memblock_find_in_range_node(from, max_addr, size, align, nid);
>> +	if (alloc)
>> +		goto done;
>> +
>> +	if (nid != MAX_NUMNODES) {
>> +		alloc =
>> +			memblock_find_in_range_node(from, max_addr, size,
>> +						    align, MAX_NUMNODES);
>> +		if (alloc)
>> +			goto done;
>> +	}
>> +
>> +	if (from) {
>> +		from = 0;
>> +		goto again;
>> +	} else {
>> +		goto error;
>> +	}
>> +
>> +done:
>> +	memblock_reserve(alloc, size);
>> +	ptr = phys_to_virt(alloc);
>> +	memset(ptr, 0, size);
> 
> What if the address is high?  Don't we need kmapping here?
>
The current nobootmem code actually don't handle the high
addresses since the max memory is limited by memblock.current_limit
which is max_low_pfn. So I am assuming we don't need to support
it. __alloc_bootmem_node_high() interface underneath uses
__alloc_memory_core_early() and we tried to keep the same
functionality in new code.
 
>> +
>> +	/*
>> +	 * The min_count is set to 0 so that bootmem allocated blocks
>> +	 * are never reported as leaks.
>> +	 */
>> +	kmemleak_alloc(ptr, size, 0, 0);
>> +
>> +	return ptr;
>> +
>> +error:
>> +	return NULL;
>> +}
>> +
>> +void * __init memblock_early_alloc_try_nid_nopanic(int nid,
>> +				phys_addr_t size, phys_addr_t align,
>> +				phys_addr_t from, phys_addr_t max_addr)
>> +{
>> +	memblock_dbg("%s: %llu bytes align=0x%llx nid=%d from=0x%llx max_addr=0x%llx %pF\n",
>> +			__func__, (u64)size, (u64)align, nid, (u64)from,
>> +			(u64)max_addr, (void *)_RET_IP_);
>> +	return _memblock_early_alloc_try_nid_nopanic(nid, size,
>> +						align, from, max_addr);
> 
> Do we need the extra level of wrapping?  Just implement
> alloc_try_nid_nopanic() here and make the panicky version call it?
> 
It was useful to have caller information (_RET_IP_) for debug. But
it can be dropped if you insist.

Regards,
Santosh