> On Jan 10, 2023, at 9:33 AM, Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 1/5/23 1:26 AM, tong@xxxxxxxxxxxxx wrote:
>> diff --git a/tools/testing/selftests/bpf/prog_tests/htab_deadlock.c b/tools/testing/selftests/bpf/prog_tests/htab_deadlock.c
>> new file mode 100644
>> index 000000000000..137dce8f1346
>> --- /dev/null
>> +++ b/tools/testing/selftests/bpf/prog_tests/htab_deadlock.c
>> @@ -0,0 +1,75 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/* Copyright (c) 2022 DiDi Global Inc. */
>> +#define _GNU_SOURCE
>> +#include <pthread.h>
>> +#include <sched.h>
>> +#include <test_progs.h>
>> +
>> +#include "htab_deadlock.skel.h"
>> +
>> +static int perf_event_open(void)
>> +{
>> + struct perf_event_attr attr = {0};
>> + int pfd;
>> +
>> + /* create perf event on CPU 0 */
>> + attr.size = sizeof(attr);
>> + attr.type = PERF_TYPE_HARDWARE;
>> + attr.config = PERF_COUNT_HW_CPU_CYCLES;
>> + attr.freq = 1;
>> + attr.sample_freq = 1000;
>> + pfd = syscall(__NR_perf_event_open, &attr, -1, 0, -1, PERF_FLAG_FD_CLOEXEC);
>> +
>> + return pfd >= 0 ? pfd : -errno;
>> +}
>> +
>> +void test_htab_deadlock(void)
>> +{
>> + unsigned int val = 0, key = 20;
>> + struct bpf_link *link = NULL;
>> + struct htab_deadlock *skel;
>> + int err, i, pfd;
>> + cpu_set_t cpus;
>> +
>> + skel = htab_deadlock__open_and_load();
>> + if (!ASSERT_OK_PTR(skel, "skel_open_and_load"))
>> + return;
>> +
>> + err = htab_deadlock__attach(skel);
>> + if (!ASSERT_OK(err, "skel_attach"))
>> + goto clean_skel;
>> +
>> + /* NMI events. */
>> + pfd = perf_event_open();
>> + if (pfd < 0) {
>> + if (pfd == -ENOENT || pfd == -EOPNOTSUPP) {
>> + printf("%s:SKIP:no PERF_COUNT_HW_CPU_CYCLES\n", __func__);
>> + test__skip();
>
> This test is a SKIP in bpf CI, so it won't be useful.
> https://github.com/kernel-patches/bpf/actions/runs/3858084722/jobs/6579470256#step:6:5198
>
> Is there other way to test it or do you know what may be missing in vmtest.sh? Not sure if the cloud setup in CI blocks HW_CPU_CYCLES. If it is, I also don't know a good way (Cc: Manu).
Hi
Other test cases using PERF_COUNT_HW_CPU_CYCLES were skipped too. For example,
send_signal
find_vma
get_stackid_cannot_attach
>
>> + goto clean_skel;
>> + }
>> + if (!ASSERT_GE(pfd, 0, "perf_event_open"))
>> + goto clean_skel;
>> + }
>> +
>> + link = bpf_program__attach_perf_event(skel->progs.bpf_empty, pfd);
>> + if (!ASSERT_OK_PTR(link, "attach_perf_event"))
>> + goto clean_pfd;
>> +
>> + /* Pinned on CPU 0 */
>> + CPU_ZERO(&cpus);
>> + CPU_SET(0, &cpus);
>> + pthread_setaffinity_np(pthread_self(), sizeof(cpus), &cpus);
>> +
>> + /* update bpf map concurrently on CPU0 in NMI and Task context.
>> + * there should be no kernel deadlock.
>> + */
>> + for (i = 0; i < 100000; i++)
>> + bpf_map_update_elem(bpf_map__fd(skel->maps.htab),
>> + &key, &val, BPF_ANY);
>> +
>> + bpf_link__destroy(link);
>> +clean_pfd:
>> + close(pfd);
>> +clean_skel:
>> + htab_deadlock__destroy(skel);
>> +}
>> diff --git a/tools/testing/selftests/bpf/progs/htab_deadlock.c b/tools/testing/selftests/bpf/progs/htab_deadlock.c
>> new file mode 100644
>> index 000000000000..dacd003b1ccb
>> --- /dev/null
>> +++ b/tools/testing/selftests/bpf/progs/htab_deadlock.c
>> @@ -0,0 +1,30 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/* Copyright (c) 2022 DiDi Global Inc. */
>> +#include <linux/bpf.h>
>> +#include <bpf/bpf_helpers.h>
>> +#include <bpf/bpf_tracing.h>
>> +
>> +char _license[] SEC("license") = "GPL";
>> +
>> +struct {
>> + __uint(type, BPF_MAP_TYPE_HASH);
>> + __uint(max_entries, 2);
>> + __uint(map_flags, BPF_F_ZERO_SEED);
>> + __type(key, unsigned int);
>> + __type(value, unsigned int);
>> +} htab SEC(".maps");
>> +
>> +SEC("fentry/perf_event_overflow")
>> +int bpf_nmi_handle(struct pt_regs *regs)
>> +{
>> + unsigned int val = 0, key = 4;
>> +
>> + bpf_map_update_elem(&htab, &key, &val, BPF_ANY);
>
> I ran it in my qemu setup which does not skip the test. I got this splat though:
This is a false alarm, not deadlock(this patch fix deadlock, only). I fix waring in other patch, please review
https://patchwork.kernel.org/project/netdevbpf/patch/20230105112749.38421-1-tong@xxxxxxxxxxxxx/
>
> [ 42.990306] ================================
> [ 42.990307] WARNING: inconsistent lock state
> [ 42.990310] 6.2.0-rc2-00304-gaf88a1bb9967 #409 Tainted: G O
> [ 42.990313] --------------------------------
> [ 42.990315] inconsistent {INITIAL USE} -> {IN-NMI} usage.
> [ 42.990317] test_progs/1546 [HC1[1]:SC0[0]:HE0:SE1] takes:
> [ 42.990322] ffff888101245768 (&htab->lockdep_key){....}-{2:2}, at: htab_map_update_elem+0x1e7/0x810
> [ 42.990340] {INITIAL USE} state was registered at:
> [ 42.990341] lock_acquire+0x1e6/0x530
> [ 42.990351] _raw_spin_lock_irqsave+0xb8/0x100
> [ 42.990362] htab_map_update_elem+0x1e7/0x810
> [ 42.990365] bpf_map_update_value+0x40d/0x4f0
> [ 42.990371] map_update_elem+0x423/0x580
> [ 42.990375] __sys_bpf+0x54e/0x670
> [ 42.990377] __x64_sys_bpf+0x7c/0x90
> [ 42.990382] do_syscall_64+0x43/0x90
> [ 42.990387] entry_SYSCALL_64_after_hwframe+0x72/0xdc
>
> Please check.
>
>> + return 0;
>> +}
>> +
>> +SEC("perf_event")
>> +int bpf_empty(struct pt_regs *regs)
>> +{
>
> btw, from a quick look at __perf_event_overflow, I suspect doing the bpf_map_update_elem() here instead of the fentry/perf_event_overflow above can also reproduce the patch 1 issue?
No
bpf_overflow_handler will check the bpf_prog_active, if syscall increase it, bpf_overflow_handler will skip the bpf prog.
Fentry will not check the bpf_prog_active, and interrupt the task context. We have discussed that.
>
>> + return 0;
>> +}
>
>
