There are two issues with bpf_prog_pack:
(1) On NUMA systems, bpf_prog_pack need to be bigger
(PMD_SIZE * num_online_nodes) to use huge pages.
(2) If the system doesn't support huge pages (nohugevmalloc in cmdline),
allocating PMD_SIZE for bpf_prog_pack is a waste.
Address these issues with flexible bpf_prog_pack_size().
Song Liu (2):
vmalloc: expose vmap_allow_huge via get_vmap_allow_huge()
bpf: flexible size for bpf_prog_pack
include/linux/vmalloc.h | 1 +
kernel/bpf/core.c | 47 +++++++++++++++++++++++------------------
mm/vmalloc.c | 5 +++++
3 files changed, 33 insertions(+), 20 deletions(-)
--
2.30.2
