On 8/18/20 3:23 PM, Yonghong Song wrote:
Patch #1 fixed a rcu stall warning when traversing large number
of tasks/files without overflowing seq_file buffer. The method
is to control the number of visited objects in bpf_seq_read()
so all bpf iterators will benefit.
Patch #2 calculated tid properly in a namespace in order to avoid
visiting the name task multiple times.
Patch #3 handled read() error code EAGAIN properly in bpftool
bpf_iter userspace code to collect pids. The change is needed
due to Patch #1.
Sorry. Missed the changelog below.
Changelogs:
- v1 -> v2:
. do ratelimiting in bpf_seq_read() with a counter.
cond_resched() doesn't work as some iterators (e.g., tcp,
netlink, etc.) has rcu read critical section across
consecutive seq_ops->next() functions.
Yonghong Song (3):
bpf: fix a rcu_sched stall issue with bpf task/task_file iterator
bpf: avoid visit same object multiple times
bpftool: handle EAGAIN error code properly in pids collection
kernel/bpf/bpf_iter.c | 15 ++++++++++++++-
kernel/bpf/task_iter.c | 3 ++-
tools/bpf/bpftool/pids.c | 2 ++
3 files changed, 18 insertions(+), 2 deletions(-)
