On Wed, 30 Jun 2021 at 14:42, Jiri Olsa <jolsa@xxxxxxxxxx> wrote: > > On Wed, Jun 30, 2021 at 12:34:58PM +0200, Brendan Jackman wrote: > > On Tue, 29 Jun 2021 at 23:09, Jiri Olsa <jolsa@xxxxxxxxxx> wrote: > > > > > > On Tue, Jun 29, 2021 at 06:41:24PM +0200, Jiri Olsa wrote: > > > > On Tue, Jun 29, 2021 at 06:25:33PM +0200, Brendan Jackman wrote: > > > > > On Tue, 29 Jun 2021 at 18:04, Jiri Olsa <jolsa@xxxxxxxxxx> wrote: > > > > > > On Tue, Jun 29, 2021 at 04:10:12PM +0200, Jiri Olsa wrote: > > > > > > > On Mon, Jun 28, 2021 at 11:21:42AM +0200, Brendan Jackman wrote: > > > > > > > > > > atomics in .imm). Any idea if this test was ever passing on PowerPC? > > > > > > > > > > > > > > > > > > > > > > hum, I guess not.. will check > > > > > > > > > > > > nope, it locks up the same: > > > > > > > > > > Do you mean it locks up at commit 91c960b0056 too? > > > > Sorry I was being stupid here - the test didn't exist at this commit > > > > > > I tried this one: > > > > 37086bfdc737 bpf: Propagate stack bounds to registers in atomics w/ BPF_FETCH > > > > > > > > I will check also 91c960b0056, but I think it's the new test issue > > > > So yeah hard to say whether this was broken on PowerPC all along. How > > hard is it for me to get set up to reproduce the failure? Is there a > > rootfs I can download, and some instructions for running a PowerPC > > QEMU VM? If so if you can also share your config and I'll take a look. > > > > If it's not as simple as that, I'll stare at the code for a while and > > see if anything jumps out. > > > > I have latest fedora ppc server and compile/install latest bpf-next tree > I think it will be reproduced also on vm, I attached my config OK, getting set up to boot a PowerPC QEMU isn't practical here unless someone's got commands I can copy-paste (suspect it will need .config hacking too). Looks like you need to build a proper bootloader, and boot an installer disk. Looked at the code for a bit but nothing jumped out. It seems like the verifier is seeing a BPF_ADD | BPF_FETCH, which means it doesn't detect an infinite loop, but then we lose the BPF_FETCH flag somewhere between do_check in verifier.c and bpf_jit_build_body in bpf_jit_comp64.c. That would explain why we don't get the "eBPF filter atomic op code %02x (@%d) unsupported", and would also explain the lockup because a normal atomic add without fetch would leave BPF R1 unchanged. We should be able to confirm that theory by disassembling the JITted code that gets hexdumped by bpf_jit_dump when bpf_jit_enable is set to 2... at least for PowerPC 32-bit... maybe you could paste those lines into the 64-bit version too? Here's some notes I made for disassembling the hexdump on x86, I guess you'd just need to change the objdump flags: -- - Enable console JIT output: ```shell echo 2 > /proc/sys/net/core/bpf_jit_enable ``` - Load & run the program of interest. - Copy the hex code from the kernel console to `/tmp/jit.txt`. Here's what a short program looks like. This includes a line of context - don't paste the `flen=` line. ``` [ 79.381020] flen=8 proglen=54 pass=4 image=000000001af6f390 from=test_verifier pid=258 [ 79.389568] JIT code: 00000000: 0f 1f 44 00 00 66 90 55 48 89 e5 48 81 ec 08 00 [ 79.397411] JIT code: 00000010: 00 00 48 c7 45 f8 64 00 00 00 bf 04 00 00 00 48 [ 79.405965] JIT code: 00000020: f7 df f0 48 29 7d f8 8b 45 f8 48 83 f8 60 74 02 [ 79.414719] JIT code: 00000030: c9 c3 31 c0 eb fa ``` - This incantation will split out and decode the hex, then disassemble the result: ```shell cat /tmp/jit.txt | cut -d: -f2- | xxd -r >/tmp/obj && objdump -D -b binary -m i386:x86-64 /tmp/obj ``` -- Sandipan, Naveen, do you know of anything in the PowerPC code that might be leading us to drop the BPF_FETCH flag from the atomic instruction in tools/testing/selftests/bpf/verifier/atomic_bounds.c?
