On Tue, Apr 2, 2024 at 9:00 PM Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote: > > On Tue, Apr 2, 2024 at 5:52 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > > > > On Wed, 3 Apr 2024 09:40:48 +0900 > > Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote: > > > > > OK, for me, this last sentence is preferred for the help message. That explains > > > what this is for. > > > > > > All callbacks that attach to the function tracing have some sort > > > of protection against recursion. This option is only to verify that > > > ftrace (and other users of ftrace_test_recursion_trylock()) are not > > > called outside of RCU, as if they are, it can cause a race. > > > But it also has a noticeable overhead when enabled. > > Sounds good to me, I can add this to the description of the Kconfig option. > > > > > > > BTW, how much overhead does this introduce? and the race case a kernel crash? > > I just checked our fleet-wide production data for the last 24 hours. > Within the kprobe/kretprobe code path (ftrace_trampoline and > everything called from it), rcu_is_watching (both calls, see below) > cause 0.484% CPU cycles usage, which isn't nothing. So definitely we'd > prefer to be able to avoid that in production use cases. > I just ran synthetic microbenchmark testing multi-kretprobe throughput. We get (in millions of BPF kretprobe-multi program invocations per second): - 5.568M/s as baseline; - 5.679M/s with changes in this patch (+2% throughput improvement); - 5.808M/s with disabling rcu_is_watching in rethook_try_get() (+2.3% more vs just one of rcu_is_watching, and +4.3% vs baseline). It's definitely noticeable. > > > or just messed up the ftrace buffer? > > > > There's a hypothetical race where it can cause a use after free. > > > > That is, after you shutdown ftrace, you need to call synchronize_rcu_tasks(), > > which requires RCU to be watching. There's a theoretical case where that > > task calls the trampoline and misses the synchronization. Note, these > > locations are with preemption disabled, as rcu is always watching when > > preemption is enabled. Thus it would be extremely fast where as the > > synchronize_rcu_tasks() is rather slow. > > > > We also have synchronize_rcu_tasks_rude() which would actually keep the > > trace from happening, as it would schedule on each CPU forcing all CPUs to > > have RCU watching. > > > > I have never heard of this race being hit. I guess it could happen on a VM > > where a vCPU gets preempted at the right moment for a long time and the > > other CPUs synchronize. > > > > But like lockdep, where deadlocks can crash the kernel, we don't enable it > > for production. > > > > The overhead is another function call within the function tracer. I had > > numbers before, but I guess I could run tests again and get new numbers. > > > > I just noticed another rcu_is_watching() call, in rethook_try_get(), > which seems to be a similar and complementary validation check to the > one we are putting under CONFIG_FTRACE_VALIDATE_RCU_IS_WATCHING option > in this patch. It feels like both of them should be controlled by the > same settings. WDYT? Can I add CONFIG_FTRACE_VALIDATE_RCU_IS_WATCHING > guard around rcu_is_watching() check in rethook_try_get() as well? > > > > Thanks, > > > > -- Steve
