On Thu, 26 Nov 2020 at 09:01, Björn Töpel bjorn.topel@intel.com wrote:
On 2020-11-26 07:44, Yonghong Song wrote:
[...]
What other configures I am missing?
BTW, I cherry-picked the following pick from bpf tree in this experiment. commit e7f4a5919bf66e530e08ff352d9b78ed89574e6b (HEAD -> xsk) Author: Björn Töpel bjorn.topel@intel.com Date: Mon Nov 23 18:56:00 2020 +0100
net, xsk: Avoid taking multiple skbuff referencesHmm, I'm getting an oops, unless I cherry-pick:
36ccdf85829a ("net, xsk: Avoid taking multiple skbuff references")
*AND*
537cf4e3cc2f ("xsk: Fix umem cleanup bug at socket destruct")
from bpf/master.
Same as Bjorn's findings ^^^, additionally applying the second patch 537cf4e3cc2f [PASS] all tests for me
PREREQUISITES: [ PASS ] SKB NOPOLL: [ PASS ] SKB POLL: [ PASS ] DRV NOPOLL: [ PASS ] DRV POLL: [ PASS ] SKB SOCKET TEARDOWN: [ PASS ] DRV SOCKET TEARDOWN: [ PASS ] SKB BIDIRECTIONAL SOCKETS: [ PASS ] DRV BIDIRECTIONAL SOCKETS: [ PASS ]
With the first patch alone, as soon as we enter DRV/Native NOPOLL mode kernel panics, whereas in your case NOPOLL tests were falling with packets being *lost* as per seqnum mismatch.
Can you please test this out with both patches and let us know?
Can I just run test_xsk.sh at tools/testing/selftests/bpf/ directory? This will be easier than the above for bpf developers. If it does not work, I would like to recommend to make it work.
yes test_xsk.shis self contained, will update the instructions in there with v4.
Thanks, /Weqaar
Björn