Re: [PATCH v3 bpf-next 0/8] bpf: Introduce BPF timers.

Alexei Starovoitov [off-list ref] writes:

From: Alexei Starovoitov <ast@kernel.org>

The first request to support timers in bpf was made in 2013 before sys_bpf syscall
was added. That use case was periodic sampling. It was address with attaching
bpf programs to perf_events. Then during XDP development the timers were requested
to do garbage collection and health checks. They were worked around by implementing
timers in user space and triggering progs with BPF_PROG_RUN command.
The user space timers and perf_event+bpf timers are not armed by the bpf program.
They're done asynchronously vs program execution. The XDP program cannot send a
packet and arm the timer at the same time. The tracing prog cannot record an
event and arm the timer right away. This large class of use cases remained
unaddressed. The jiffy based and hrtimer based timers are essential part of the
kernel development and with this patch set the hrtimer based timers will be
available to bpf programs.

TLDR: bpf timers is a wrapper of hrtimers with all the extra safety added
to make sure bpf progs cannot crash the kernel.

v2->v3:
The v2 approach attempted to bump bpf_prog refcnt when bpf_timer_start is
called to make sure callback code doesn't disappear when timer is active and
drop refcnt when timer cb is done. That led to a ton of race conditions between
callback running and concurrent bpf_timer_init/start/cancel on another cpu,
and concurrent bpf_map_update/delete_elem, and map destroy.

Then v2.5 approach skipped prog refcnt altogether. Instead it remembered all
timers that bpf prog armed in a link list and canceled them when prog refcnt
went to zero. The race conditions disappeared, but timers in map-in-map could
not be supported cleanly, since timers in inner maps have inner map's life time
and don't match prog's life time.

This v3 approach makes timers to be owned by maps. It allows timers in inner
maps to be supported from the start. This apporach relies on "user refcnt"
scheme used in prog_array that stores bpf programs for bpf_tail_call. The
bpf_timer_start() increments prog refcnt, but unlike 1st approach the timer
callback does decrement the refcnt. The ops->map_release_uref is
responsible for cancelling the timers and dropping prog refcnt when user space
reference to a map is dropped. That addressed all the races and simplified
locking.

Great to see this! I missed v2, but the "owned by map + uref" approach
makes sense.

For the series:

Acked-by: Toke Høiland-Jørgensen <redacted>