These patches try to test the fix made in commit e2f7fc0ac695 ("bpf:
fix undefined behavior in narrow load handling"). The problem existed
in the generated BPF bytecode that was doing a 32bit narrow read of a
64bit field, so to test it the code would need to be executed.
Currently the only such field exists in BPF_PROG_TYPE_PERF_EVENT,
which was not supported by bpf_prog_test_run().
I'm sending these patches to bpf-next now as they introduce a new
feature. But maybe some of those patches could go to the bpf branch?
There is a bit of yak shaving to do for the test to be run:
1. Print why the program could not be run (patch 1).
2. Some fixes for errno clobbering (patches 2 and 3).
3. Using bpf_prog_test_run_xattr, so I can pass ctx_in stuff too
(patch 4).
4. Adding ctx stuff and data out size override to struct bpf_test, and
use them for the perf event tests (patches 5 and 6).
5. Some tools headers syncing (patches 7 and 8).
6. Split out some useful functions for bpf_prog_test_run
implementation out of the net/bpf/test_run.c (patch 9)
7. Implement bpf_prog_test_run for perf event programs and test it
(patches 10 and 11).
The last point is where I'm least sure how things should be done
properly:
1. There is a bunch of stuff to prepare for the
bpf_perf_prog_read_value to work, and that stuff is very hacky. I
would welcome some hints about how to set up the perf_event and
perf_sample_data structs in a way that is a bit more future-proof
than just setting some fields in a specific way, so some other code
won't use some other fields (like setting event.oncpu to -1 to
avoid event.pmu to be used for reading the value of the event).
2. The tests try to see if the test run for perf event sets up the
context properly, so they verify the struct pt_regs contents. They
way it is currently written Works For Me, but surely it won't work
on other arch. So what would be the way forward? Just put the test
case inside #ifdef __x86_64__?
3. Another thing in tests - I'm trying to make sure that the
bpf_perf_prog_read_value helper works as it seems to be the only
bpf perf helper that takes the ctx as a parameter. Is that enough
or I should test other helpers too?
About the test itself - I'm not sure if it will work on a big endian
machine. I think it should, but I don't have anything handy here to
verify it.
Krzesimir Nowak (12):
selftests/bpf: Print a message when tester could not run a program
selftests/bpf: Avoid a clobbering of errno
selftests/bpf: Avoid another case of errno clobbering
selftests/bpf: Use bpf_prog_test_run_xattr
selftests/bpf: Allow passing more information to BPF prog test run
selftests/bpf: Make sure that preexisting tests for perf event work
tools headers: Adopt compiletime_assert from kernel sources
tools headers: Sync struct bpf_perf_event_data
bpf: Split out some helper functions
bpf: Implement bpf_prog_test_run for perf event programs
selftests/bpf: Add tests for bpf_prog_test_run for perf events progs
selftests/bpf: Test correctness of narrow 32bit read on 64bit field
include/linux/bpf.h | 28 ++
kernel/bpf/Makefile | 1 +
kernel/bpf/test_run.c | 212 ++++++++++++++
kernel/trace/bpf_trace.c | 60 ++++
net/bpf/test_run.c | 263 +++++-------------
tools/include/linux/compiler.h | 28 ++
tools/include/uapi/linux/bpf_perf_event.h | 1 +
tools/testing/selftests/bpf/test_verifier.c | 197 ++++++++++++-
.../selftests/bpf/verifier/perf_event_run.c | 96 +++++++
.../bpf/verifier/perf_event_sample_period.c | 4 +
.../testing/selftests/bpf/verifier/var_off.c | 21 ++
11 files changed, 700 insertions(+), 211 deletions(-)
create mode 100644 kernel/bpf/test_run.c
create mode 100644 tools/testing/selftests/bpf/verifier/perf_event_run.c
--
2.20.1