[PATCH v4 02/14] memblock: add MEMBLOCK_RSRV_KERN flag

[Wei Yang]

On Tue, Feb 25, 2025 at 09:46:28AM +0200, Mike Rapoport wrote:
>On Mon, Feb 24, 2025 at 01:31:31AM +0000, Wei Yang wrote:
>> On Wed, Feb 19, 2025 at 09:24:31AM +0200, Mike Rapoport wrote:
>> >Hi,
>> >
>> >On Tue, Feb 18, 2025 at 03:50:04PM +0000, Wei Yang wrote:
>> >> On Thu, Feb 06, 2025 at 03:27:42PM +0200, Mike Rapoport wrote:
>> >> >From: "Mike Rapoport (Microsoft)" <rppt at kernel.org>
>> >> >
>> >> >to denote areas that were reserved for kernel use either directly with
>> >> >memblock_reserve_kern() or via memblock allocations.
>> >> >
>> >> >Signed-off-by: Mike Rapoport (Microsoft) <rppt at kernel.org>
>> >> >---
>> >> > include/linux/memblock.h | 16 +++++++++++++++-
>> >> > mm/memblock.c            | 32 ++++++++++++++++++++++++--------
>> >> > 2 files changed, 39 insertions(+), 9 deletions(-)
>> >> >
>> >> >diff --git a/include/linux/memblock.h b/include/linux/memblock.h
>> >> >index e79eb6ac516f..65e274550f5d 100644
>> >> >--- a/include/linux/memblock.h
>> >> >+++ b/include/linux/memblock.h
>> >> >@@ -50,6 +50,7 @@ enum memblock_flags {
>> >> > 	MEMBLOCK_NOMAP		= 0x4,	/* don't add to kernel direct mapping */
>> >> > 	MEMBLOCK_DRIVER_MANAGED = 0x8,	/* always detected via a driver */
>> >> > 	MEMBLOCK_RSRV_NOINIT	= 0x10,	/* don't initialize struct pages */
>> >> >+	MEMBLOCK_RSRV_KERN	= 0x20,	/* memory reserved for kernel use */
>> >> 
>> >> Above memblock_flags, there are comments on explaining those flags.
>> >> 
>> >> Seems we miss it for MEMBLOCK_RSRV_KERN.
>> >
>> >Right, thanks!
>> > 
>> >> > 
>> >> > #ifdef CONFIG_HAVE_MEMBLOCK_PHYS_MAP
>> >> >@@ -1459,14 +1460,14 @@ phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
>> >> > again:
>> >> > 	found = memblock_find_in_range_node(size, align, start, end, nid,
>> >> > 					    flags);
>> >> >-	if (found && !memblock_reserve(found, size))
>> >> >+	if (found && !__memblock_reserve(found, size, nid, MEMBLOCK_RSRV_KERN))
>> >> 
>> >> Maybe we could use memblock_reserve_kern() directly. If my understanding is
>> >> correct, the reserved region's nid is not used.
>> >
>> >We use nid of reserved regions in reserve_bootmem_region() (commit
>> >61167ad5fecd ("mm: pass nid to reserve_bootmem_region()")) but KHO needs to
>> >know the distribution of reserved memory among the nodes before
>> >memmap_init_reserved_pages().
>> > 
>> >> BTW, one question here. How we handle concurrent memblock allocation? If two
>> >> threads find the same available range and do the reservation, it seems to be a
>> >> problem to me. Or I missed something?
>> >
>> >memblock allocations end before smp_init(), there is no possible concurrency.
>> > 
>> 
>> Thanks, I still have one question here.
>> 
>> Below is a simplified call flow.
>> 
>>     mm_core_init()
>>         mem_init()
>>             memblock_free_all()
>>                 free_low_memory_core_early()
>>                     memmap_init_reserved_pages()
>>                         memblock_set_node(..., memblock.reserved, )   --- (1)
>>                     __free_memory_core()
>>         kmem_cache_init()
>>             slab_state = UP;                                          --- (2)
>> 
>> And memblock_allloc_range_nid() is not supposed to be called after
>> slab_is_available(). Even someone do dose it, it will get memory from slab
>> instead of reserve region in memblock.
>> 
>> From the above call flow and background, there are three cases when
>> memblock_alloc_range_nid() would be called:
>> 
>>   * If it is called before (1), memblock.reserved's nid would be adjusted correctly.
>>   * If it is called after (2), we don't touch memblock.reserved.
>>   * If it happens between (1) and (2), it looks would break the consistency of
>>     nid information in memblock.reserved. Because when we use
>>     memblock_reserve_kern(), NUMA_NO_NODE would be stored in region.
>> 
>> So my question is if the third case happens, would it introduce a bug? If it
>> won't happen, seems we don't need to specify the nid here?
>
>We don't really care about proper assignment of nodes between (1) and (2)
>from one side and the third case does not happen on the other side. Nothing
>should call membloc_alloc() after memblock_free_all(). 
>

My point is if no one would call memblock_alloc() after memblock_free_all(),
which set nid in memblock.reserved properly, it seems not necessary to do
__memblock_reserve() with exact nid during memblock_alloc()? 

As you did __memblock_reserve(found, size, nid, MEMBLOCK_RSRV_KERN) in this
patch.

>But it's easy to make the window between (1) and (2) disappear by replacing
>checks for slab_is_available() in memblock with a variable local to
>memblock.
> 
>> -- 
>> Wei Yang
>> Help you, Help me
>
>-- 
>Sincerely yours,
>Mike.

-- 
Wei Yang
Help you, Help me