Re: [PATCH] memblock: fix memblock_estimated_nr_free_pages() for soft-reserved memory

[Akinobu Mita]

2025年11月5日(水) 2:18 Mike Rapoport <rppt@xxxxxxxxxx>:
>
> (added Dan Williams)
>
> Hi,
>
> On Tue, Nov 04, 2025 at 09:39:21AM +0900, Akinobu Mita wrote:
> > memblock_estimated_nr_free_pages() returns the difference between the total
> > size of the "memory" memblock type and the "reserved" memblock type.
> >
> > The "soft-reserved" memory regions are added to the "reserved" memblock
> > type, but not to the "memory" memblock type. Therefore,
>
> @Dan, do we really need to memblock_reserve() the E820_TYPE_SOFT_RESERVED
> ranges?
> Quick scan didn't show anything that requires this, but I could easily miss
> something.
>
> > memblock_estimated_nr_free_pages() may return a smaller value than
> > expected, or if it underflows, an extremely large value.
> >
> > /proc/sys/kernel/threads-max is determined by the value of
> > memblock_estimated_nr_free_pages().  This issue was discovered on machines
> > with CXL memory because kernel.threads-max was either smaller than expected
> > or extremely large for the installed DRAM size.
> >
> > This fixes the issue by improving the accuracy of
> > memblock_estimated_nr_free_pages() by subtracting only the overlapping size
> > of regions with "memory" and "reserved" memblock types.
> >
> > Signed-off-by: Akinobu Mita <akinobu.mita@xxxxxxxxx>
> > ---
> >  mm/memblock.c | 33 ++++++++++++++++++++++++++++++++-
> >  1 file changed, 32 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/memblock.c b/mm/memblock.c
> > index e23e16618e9b..af014fa10a44 100644
> > --- a/mm/memblock.c
> > +++ b/mm/memblock.c
>
> ...
>
> > @@ -1826,7 +1842,22 @@ phys_addr_t __init_memblock memblock_reserved_kern_size(phys_addr_t limit, int n
> >   */
> >  unsigned long __init memblock_estimated_nr_free_pages(void)
> >  {
> > -     return PHYS_PFN(memblock_phys_mem_size() - memblock_reserved_size());
>
> We have memblock_reserved_kern_size() that tells how much memory was
> reserved from the actual RAM. Replacing memblock_reserved_size() with
> memblock_reserved_kern_size() will omit "soft-reserved" ranges.

Replacing memblock_reserved_size() with memblock_reserved_kern_size(
MEMBLOCK_ALLOC_ANYWHERE, NUMA_NO_NODE) also fixed the problem. Thank you.