Re: [PATCH] mm/hugetlb: allow crashkernel to reuse hugetlb CMA memory

[Kiryl Shutsemau]

On Wed, Nov 04, 2026 at 06:27:27AM -0800, Usama Arif wrote:
> Use a separate suffix to make the exclusion of existing HugeTLB
> allocations from the vmcore an explicit choice.

The crashkernel=size,cma documentation already says that pages allocated
from the range are not in the vmcore and that it should not be used if
dumping of userspace memory is intended. HugeTLB pages are userspace
memory. I don't see what a new suffix makes explicit that ",cma" does
not already.

Why not let crashkernel=size,cma take hugetlb CMA first and reserve a
dedicated area only for what hugetlb CMA does not cover? With
hugetlb_cma=6G crashkernel=2G,cma there would be no extra CMA at all.
With hugetlb_cma=2G crashkernel=4G,cma the crash kernel would reuse the
2G and a dedicated 2G area would be reserved for the rest. Either way,
nobody has to learn a new option.

The catch is ordering. reserve_crashkernel_cma() runs from setup_arch(),
but hugetlb_cma_reserve() runs later, from mm_core_init_early(). The arch
code could record the size and leave the reservation to generic code
right after hugetlb_cma_reserve(). That is still memblock time, so the
remainder can be declared there. An early_initcall is too late for that.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov