On 11/28/23 11:33, Larysa Zaremba wrote:
On Tue, Nov 28, 2023 at 11:26:59AM +0100, Jesper Dangaard Brouer wrote:
On 11/27/23 19:32, Larysa Zaremba wrote:
Currently, there is no reason for data_meta to be limited to 32 bytes.
Loosen this limitation and make maximum meta size 252.
First I though you made a type here with 252 bytes, but then I remembered
the 4 byte alignment.
I think commit message should elaborate on why 252 bytes.
You are right, will do.
I'm looking at the code to see if metadata can be extended into area
used by xdp_frame, such that a BPF-prog get direct memory access to this
(which should not be allowed). The bpf_xdp_adjust_meta() helper does
handle this (as you/Olek also mentioned in commit msg). A driver could
set data_meta such that they overlap, but I guess we will categorize
this as a driver bug.
The XDP headroom have evolved to become dynamic (commonly 192 or 256
bytes). Thus, we cannot statically reduce metalen with sizeof(struct
xdp_frame). The maximum meta size 252, could be achieved (and valid) if
a driver chooses to have more XDP headroom e.g. 288 bytes. So, I guess
it is correct to say maximum valid meta size is 252 bytes, but can be
limited and reduced further by drivers chosen XDP headroom and memory
reserved by struct xdp_frame (limited in bpf_xdp_adjust_meta).
Also, modify the selftest, so test_xdp_context_error does not complain
about the unexpected success.
v2->v3:
* Fix main patch author
* Add selftests path
v1->v2:
* replace 'typeof(metalen)' with the actual type
Aleksander Lobakin (1):
net, xdp: allow metadata > 32
Larysa Zaremba (1):
selftests/bpf: increase invalid metadata size
include/linux/skbuff.h | 13 ++++++++-----
include/net/xdp.h | 7 ++++++-
.../selftests/bpf/prog_tests/xdp_context_test_run.c | 4 ++--
3 files changed, 16 insertions(+), 8 deletions(-)
