On 2024-10-22 15:53, Andrii Nakryiko wrote:
On Tue, Oct 22, 2024 at 10:55 AM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
On 2024-10-22 12:14, Jordan Rife wrote:
I assume this patch isn't meant to fix the related issues with freeing
BPF programs/links with call_rcu?
No, indeed. I notice that bpf_link_free() uses a prog->sleepable flag to
choose between:
if (sleepable)
call_rcu_tasks_trace(&link->rcu, bpf_link_defer_dealloc_mult_rcu_gp);
else
call_rcu(&link->rcu, bpf_link_defer_dealloc_rcu_gp);
But the faultable syscall tracepoint series does not require syscall programs
to be sleepable. So some changes may be needed on the ebpf side there.
Your fix now adds a chain of call_rcu -> call_rcu_tasks_trace ->
kfree, which should work regardless of sleepable/non-sleepable. For
the BPF-side, yes, we do different things depending on prog->sleepable
(adding extra call_rcu_tasks_trace for sleepable, while still keeping
call_rcu in the chain), so the BPF side should be good, I think.
On the BPF side I think there needs to be some smarter handling of
when to use call_rcu or call_rcu_tasks_trace to free links/programs
based on whether or not the program type can be executed in this
context. Right now call_rcu_tasks_trace is used if the program is
sleepable, but that isn't necessarily the case here. Off the top of my
head this would be BPF_PROG_TYPE_RAW_TRACEPOINT and
BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE, but may extend to
BPF_PROG_TYPE_TRACEPOINT? I'll let some of the BPF folks chime in
here, as I'm not entirely sure.
From the BPF standpoint, as of right now, neither of RAW_TRACEPOINT or
TRACEPOINT programs are sleepable. So a single RCU grace period is
fine. But even if they were (and we'll allow that later on), we handle
sleepable programs with the same call_rcu_tasks_trace -> call_rcu
chain.
Good points, in this commit:
commit 4aadde89d8 ("tracing/bpf: disable preemption in syscall probe")
I took care to disable preemption around use of the bpf program attached
to a syscall tracepoint, which makes this change a no-op from the
tracers' perspective.
It's only when you'll decide to remove this preempt-off and allow
syscall tracepoints to sleep in bpf that you'll need to tweak that.
That's just to say that I don't think that we need any BPF-specific
fix beyond what Mathieu is doing in this patch, so:
Acked-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
Thanks!
Mathieu
A big hammer solution would be to make all grace periods waited for after
a bpf tracepoint probe unregister chain call_rcu and call_rcu_tasks_trace.
Else, if we properly tag all programs attached to syscall tracepoints as
sleepable, then keeping the call_rcu_tasks_trace() only for those would
work.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
