For platforms on powerpc architecture with a default page size greater
than 4096, there was an inconsistency in fragment size calculation.
This caused the BPF selftest xdp_adjust_tail/xdp_adjust_frags_tail_grow
to fail on powerpc.
The issue occurred because the fragment buffer size in
bpf_prog_test_run_xdp() was set to 4096, while the actual data size in
the fragment within the shared skb was checked against PAGE_SIZE
(65536 on powerpc) in min_t, causing it to exceed 4096 and be set
accordingly. This discrepancy led to an overflow when
bpf_xdp_frags_increase_tail() checked for tailroom, as skb_frag_size(frag)
could be greater than rxq->frag_size (when PAGE_SIZE > 4096).
This change fixes:
1. test_run by getting the correct arch dependent PAGE_SIZE.
2. selftest by caculating tailroom and getting correct PAGE_SIZE.
Changes:
v1 -> v2:
* Address comments from Alexander
* Use dynamic page size, cacheline size and size of
struct skb_shared_info to calculate parameters.
* Fixed both test_run and selftest.
v1: https://lore.kernel.org/all/20250122183720.1411176-1-skb99@xxxxxxxxxxxxx/
Saket Kumar Bhaskar (2):
bpf, test_run: Replace hardcoded page size with dynamic PAGE_SIZE in
test_run
selftests/bpf: Refactor xdp_adjust_tail selftest with dynamic sizing
.../bpf/prog_tests/xdp_adjust_tail.c | 160 +++++++++++++-----
.../bpf/progs/test_xdp_adjust_tail_grow.c | 41 +++--
2 files changed, 149 insertions(+), 52 deletions(-)
--
2.43.5
