[v4,net-next,00/15] locking: Introduce nested-BH locking.

Message

Sebastian Andrzej Siewior June 4, 2024, 3:24 p.m. UTC

Disabling bottoms halves acts as per-CPU BKL. On PREEMPT_RT code within
local_bh_disable() section remains preemtible. As a result high prior
tasks (or threaded interrupts) will be blocked by lower-prio task (or
threaded interrupts) which are long running which includes softirq
sections.

The proposed way out is to introduce explicit per-CPU locks for
resources which are protected by local_bh_disable() and use those only
on PREEMPT_RT so there is no additional overhead for !PREEMPT_RT builds.

The series introduces the infrastructure and converts large parts of
networking which is largest stake holder here. Once this done the
per-CPU lock from local_bh_disable() on PREEMPT_RT can be lifted.

Performance testing. Baseline is net-next as of commit 93bda33046e7a
("Merge branch'net-constify-ctl_table-arguments-of-utility-functions'")
plus v6.10-rc1. A 10GiG link is used between two hosts. The command
   xdp-bench redirect-cpu --cpu 3 --remote-action drop eth1 -e

was invoked on the receiving side with a ixgbe. The sending side uses
pktgen_sample03_burst_single_flow.sh on i40e.

Baseline:
| eth1->?                 9,018,604 rx/s                  0 err,drop/s
|   receive total         9,018,604 pkt/s                 0 drop/s                0 error/s
|     cpu:7               9,018,604 pkt/s                 0 drop/s                0 error/s
|   enqueue to cpu 3      9,018,602 pkt/s                 0 drop/s             7.00 bulk-avg
|     cpu:7->3            9,018,602 pkt/s                 0 drop/s             7.00 bulk-avg
|   kthread total         9,018,606 pkt/s                 0 drop/s          214,698 sched
|     cpu:3               9,018,606 pkt/s                 0 drop/s          214,698 sched
|     xdp_stats                   0 pass/s        9,018,606 drop/s                0 redir/s
|       cpu:3                     0 pass/s        9,018,606 drop/s                0 redir/s
|   redirect_err                  0 error/s
|   xdp_exception                 0 hit/s

perf top --sort cpu,symbol --no-children:
|   18.14%  007  [k] bpf_prog_4f0ffbb35139c187_cpumap_l4_hash
|   13.29%  007  [k] ixgbe_poll
|   12.66%  003  [k] cpu_map_kthread_run
|    7.23%  003  [k] page_frag_free
|    6.76%  007  [k] xdp_do_redirect
|    3.76%  007  [k] cpu_map_redirect
|    3.13%  007  [k] bq_flush_to_queue
|    2.51%  003  [k] xdp_return_frame
|    1.93%  007  [k] try_to_wake_up
|    1.78%  007  [k] _raw_spin_lock
|    1.74%  007  [k] cpu_map_enqueue
|    1.56%  003  [k] bpf_prog_57cd311f2e27366b_cpumap_drop

With this series applied:
| eth1->?                10,329,340 rx/s                  0 err,drop/s
|   receive total        10,329,340 pkt/s                 0 drop/s                0 error/s
|     cpu:6              10,329,340 pkt/s                 0 drop/s                0 error/s
|   enqueue to cpu 3     10,329,338 pkt/s                 0 drop/s             8.00 bulk-avg
|     cpu:6->3           10,329,338 pkt/s                 0 drop/s             8.00 bulk-avg
|   kthread total        10,329,321 pkt/s                 0 drop/s           96,297 sched
|     cpu:3              10,329,321 pkt/s                 0 drop/s           96,297 sched
|     xdp_stats                   0 pass/s       10,329,321 drop/s                0 redir/s
|       cpu:3                     0 pass/s       10,329,321 drop/s                0 redir/s
|   redirect_err                  0 error/s
|   xdp_exception                 0 hit/s

perf top --sort cpu,symbol --no-children:
|   20.90%  006  [k] bpf_prog_4f0ffbb35139c187_cpumap_l4_hash
|   12.62%  006  [k] ixgbe_poll
|    9.82%  003  [k] page_frag_free
|    8.73%  003  [k] cpu_map_bpf_prog_run_xdp
|    6.63%  006  [k] xdp_do_redirect
|    4.94%  003  [k] cpu_map_kthread_run
|    4.28%  006  [k] cpu_map_redirect
|    4.03%  006  [k] bq_flush_to_queue
|    3.01%  003  [k] xdp_return_frame
|    1.95%  006  [k] _raw_spin_lock
|    1.94%  003  [k] bpf_prog_57cd311f2e27366b_cpumap_drop

This diff appears to be noise.

v3…v4 https://lore.kernel.org/all/[e-mail]/:
- Removed bpf_clear_redirect_map(), moved the comment to the caller.
  Suggested by Toke.

- The bpf_redirect_info structure is memset() each time it is assigned.
  Suggested by Toke.

- The bpf_net_ctx_set() in __napi_busy_loop() has been moved from the
  top of the function to begin/ end of the BH-disabled section. This has
  been done to remain in sync with other call sites.
  After adding the memset() I've been looking at the perf-numbers in my
  test-case and I haven't noticed an impact, the numbers are in the same
  range with and without the change. Therefore I kept the numbers from
  previous posting.

- Collected Alexei's Acked-by.

v2…v3 https://lore.kernel.org/all/[e-mail]/:
- WARN checks checks for bpf_net_ctx_get() have been dropped and all
  NULL checks around it. This means bpf_net_ctx_get_ri() assumes the
  context has been set and will segfault if it is not the case.
  Suggested by Alexei and Jesper. This should always work or always
  segfault.

- It has been suggested by Toke to embed struct bpf_net_context into
  task_struct instead just a pointer to it. This would increase the size
  of task_struct by 112 bytes instead just eight and Alexei didn't like
  it due to the size impact with 1m threads. It is a pointer again.

v1…v2 https://lore.kernel.org/all/[e-mail]/:
- Jakub complained about touching networking drivers to make the
  additional locking work. Alexei complained about the additional
  locking within the XDP/eBFP case.
  This led to a change in how the per-CPU variables are accessed for the
  XDP/eBPF case. On PREEMPT_RT the variables are now stored on stack and
  the task pointer to the structure is saved in the task_struct while
  keeping every for !RT unchanged. This was proposed as a RFC in
  	v1: https://lore.kernel.org/all/[e-mail]/

  and then updated

        v2: https://lore.kernel.org/all/[e-mail]/
	  - Renamed the container struct from xdp_storage to bpf_net_context.
            Suggested by Toke Høiland-Jørgensen.
	  - Use the container struct also on !PREEMPT_RT builds. Store the
	    pointer to the on-stack struct in a per-CPU variable. Suggested by
            Toke Høiland-Jørgensen.

  This reduces the initial queue from 24 to 15 patches.

- There were complains about the scoped_guard() which shifts the whole
  block and makes it harder to review because the whole gets removed and
  added again. The usage has been replaced with local_lock_nested_bh()+
  its unlock counterpart.

Sebastian

Comments

Jakub Kicinski June 6, 2024, 2:52 a.m. UTC | #1

On Tue,  4 Jun 2024 17:24:07 +0200 Sebastian Andrzej Siewior wrote:
> Disabling bottoms halves acts as per-CPU BKL. On PREEMPT_RT code within
> local_bh_disable() section remains preemtible. As a result high prior
> tasks (or threaded interrupts) will be blocked by lower-prio task (or
> threaded interrupts) which are long running which includes softirq
> sections.

Appears to not apply cleanly, please rebase