Commit e679654a704e ("bpf: Fix a rcu_sched stall issue with
bpf task/task_file iterator") introduced rate limiting in
bpf_seq_read() to fix a case where traversing too many tasks
and files (tens of millions of files) may cause kernel rcu stall.
But rate limiting won't reduce the amount of work to traverse
all these files.
In practice, for a user process, typically all threads belongs
to that process share the same file table and there is no need
to visit every thread for its files.
This patch added additional logic for task_file iterator to
skip tasks if those tasks are not group_leaders and their files
are the same as those of group_leaders.
Such reduction of unnecessary work will make iterator runtime
much faster if there are a lot of non-main threads and open
files for the process.
Patch #1 is the kernel implementation and Patch #2 is the
selftest.
Changelogs:
v2 -> v3:
- add put_task_struct(task) for those skipped tasks
to avoid leaking tasks. (Josef)
v1 -> v2:
- for task_file, no need for additional user parameter,
kernel can just skip those files already visited, and
this should not impact user space. (Andrii)
- to add group_leader-only customization for task will
be considered later.
- remove Patch #1 and sent it separately as this patch set
won't depend on it any more.
Yonghong Song (2):
bpf: avoid iterating duplicated files for task_file iterator
selftests/bpf: test task_file iterator without visiting pthreads
kernel/bpf/task_iter.c | 15 +++++++++----
.../selftests/bpf/prog_tests/bpf_iter.c | 21 +++++++++++++++++++
.../selftests/bpf/progs/bpf_iter_task_file.c | 10 ++++++++-
3 files changed, 41 insertions(+), 5 deletions(-)
--
2.24.1
