http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c2 --- Comment #2 from Mark Rohrbacher --- Created attachment 852563 --> http://bugzilla.opensuse.org/attachment.cgi?id=852563&action=edit dmesg.txt (2) -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c4 --- Comment #4 from Mark Rohrbacher --- Created attachment 852565 --> http://bugzilla.opensuse.org/attachment.cgi?id=852565&action=edit dmesg.txt (4) -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c6 --- Comment #6 from Mark Rohrbacher --- Created attachment 852569 --> http://bugzilla.opensuse.org/attachment.cgi?id=852569&action=edit dmesg.txt (macvlan) The problem is not limited to veth, it also occurs on macvlan interfaces. -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c8 --- Comment #8 from Mark Rohrbacher --- Created attachment 852615 --> http://bugzilla.opensuse.org/attachment.cgi?id=852615&action=edit partial output from crash "foreach bt" Attached a cropped output from the "foreach bt" command from "crash", showing the backtrace of the crashing veth_poll+0x337 (I suppose this is veth.c:580, bpf_prog_run_xdp) and the backtrace of the installation of the xdp program. It seems that the call to bpf_prog_run_xdp finally leads to an half-patched bpf_dispatcher_xdp_func: crash> dis bpf_dispatcher_xdp_func 0xffffffff99594790 : int3 0xffffffff99594791 : imul $0xffffffc6,0x76e92729(%rax),%ebx 0xffffffff99594798 : rex.RX add %r12b,0xf(%rsi) 0xffffffff9959479c : (bad) 0xffffffff9959479d : add %r8b,(%rax) Not 100% sure about the following, but: crash> p bp_patching bp_patching = $2 = { vec = 0xffffb0d965ff7488, nr_entries = 1 } crash> p bp_patching.vec[0] $4 = { detour = 0x0, addr = 0xffffffff99594790, len = 5, opcode = "\351k\240)'" } in combination with the line regs->ip = (unsigned long) tp->detour; from the poke_int3_handler could be the problem (note that the text_poke_bp call from __bpf_arch_text_poke supplies "NULL" as detour handler). I re-tested with a recent Tumbleweed kernel (5.13.13-1-default) on a desktop machine with 8 cores (Intel Xeon W-2123) and couldn't reproduce the problem there. The text_poke_bp_patch / poke_int3_handler combination is completely different here and seems to be safe. -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c9 --- Comment #9 from Mark Rohrbacher --- Reverting the commit 28924f7b0963da7e4ce098baa067c3d36dbf49ff could solve the problem: bpf: Simplify __bpf_arch_text_poke poke type handling (bsc#1154353). suse-commit: e9229505be158934fd6d02eb5003ad7c3da74663 -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1190561 http://bugzilla.opensuse.org/show_bug.cgi?id=1190561#c14 --- Comment #14 from Mark Rohrbacher --- 5.3.18-1 looks older than the official kernel, which was 5.3.18-59.19. Was that intentional? Anyway, the new kernel version is now loading/unloading xdp programs for more than one hour. With the last kernel version, the crash occurred after approx. 20mins, so this looks pretty good. Nevertheless, I'll leave it running through the weekend. Thanks for your effort. -- You are receiving this mail because: You are on the CC list for the bug.