[v4,bpf-next,0/4] bpf: introduce bpf_get_task_stack()

This set introduces a new helper bpf_get_task_stack(). The primary use case
is to dump all /proc/*/stack to seq_file via bpf_iter__task.

A few different approaches have been explored and compared:

  1. A simple wrapper around stack_trace_save_tsk(), as v1 [1].

     This approach introduces new syntax, which is different to existing
     helper bpf_get_stack(). Therefore, this is not ideal.

  2. Extend get_perf_callchain() to support "task" as argument.

     This approach reuses most of bpf_get_stack(). However, extending
     get_perf_callchain() requires non-trivial changes to architecture
     specific code. Which is error prone.

  3. Current (v2) approach, leverages most of existing bpf_get_stack(), and
     uses stack_trace_save_tsk() to handle architecture specific logic.

[1] https://lore.kernel.org/netdev/20200623070802.2310018-1-songliubraving@fb.com/

Changes v3 => v4:
1. Simplify the selftests with bpf_iter.h. (Yonghong)
2. Add example output to commit log of 4/4. (Yonghong)

Changes v2 => v3:
1. Rebase on top of bpf-next. (Yonghong)
2. Sanitize get_callchain_entry(). (Peter)
3. Use has_callchain_buf for bpf_get_task_stack. (Andrii)
4. Other small clean up. (Yonghong, Andrii).

Changes v1 => v2:
1. Reuse most of bpf_get_stack() logic. (Andrii)
2. Fix unsigned long vs. u64 mismatch for 32-bit systems. (Yonghong)
3. Add %pB support in bpf_trace_printk(). (Daniel)
4. Fix buffer size to bytes.

Song Liu (4):
  perf: expose get/put_callchain_entry()
  bpf: introduce helper bpf_get_task_stack()
  bpf: allow %pB in bpf_seq_printf() and bpf_trace_printk()
  selftests/bpf: add bpf_iter test with bpf_get_task_stack()

 include/linux/bpf.h                           |  1 +
 include/linux/perf_event.h                    |  2 +
 include/uapi/linux/bpf.h                      | 36 ++++++++-
 kernel/bpf/stackmap.c                         | 75 ++++++++++++++++++-
 kernel/bpf/verifier.c                         |  4 +-
 kernel/events/callchain.c                     | 13 ++--
 kernel/trace/bpf_trace.c                      | 12 ++-
 scripts/bpf_helpers_doc.py                    |  2 +
 tools/include/uapi/linux/bpf.h                | 36 ++++++++-
 .../selftests/bpf/prog_tests/bpf_iter.c       | 17 +++++
 .../selftests/bpf/progs/bpf_iter_task_stack.c | 37 +++++++++
 11 files changed, 220 insertions(+), 15 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_iter_task_stack.c

--
2.24.1

On Mon, Jun 29, 2020 at 11:54 AM Song Liu <songliubraving@fb.com> wrote:
>
> This set introduces a new helper bpf_get_task_stack(). The primary use case
> is to dump all /proc/*/stack to seq_file via bpf_iter__task.
>
> A few different approaches have been explored and compared:
>
>   1. A simple wrapper around stack_trace_save_tsk(), as v1 [1].
>
>      This approach introduces new syntax, which is different to existing
>      helper bpf_get_stack(). Therefore, this is not ideal.
>
>   2. Extend get_perf_callchain() to support "task" as argument.
>
>      This approach reuses most of bpf_get_stack(). However, extending
>      get_perf_callchain() requires non-trivial changes to architecture
>      specific code. Which is error prone.
>
>   3. Current (v2) approach, leverages most of existing bpf_get_stack(), and
>      uses stack_trace_save_tsk() to handle architecture specific logic.
>
> [1] https://lore.kernel.org/netdev/20200623070802.2310018-1-songliubraving@fb.com/
>
> Changes v3 => v4:
> 1. Simplify the selftests with bpf_iter.h. (Yonghong)
> 2. Add example output to commit log of 4/4. (Yonghong)
>
> Changes v2 => v3:
> 1. Rebase on top of bpf-next. (Yonghong)
> 2. Sanitize get_callchain_entry(). (Peter)
> 3. Use has_callchain_buf for bpf_get_task_stack. (Andrii)
> 4. Other small clean up. (Yonghong, Andrii).
>
> Changes v1 => v2:
> 1. Reuse most of bpf_get_stack() logic. (Andrii)
> 2. Fix unsigned long vs. u64 mismatch for 32-bit systems. (Yonghong)
> 3. Add %pB support in bpf_trace_printk(). (Daniel)
> 4. Fix buffer size to bytes.
>
> Song Liu (4):
>   perf: expose get/put_callchain_entry()
>   bpf: introduce helper bpf_get_task_stack()
>   bpf: allow %pB in bpf_seq_printf() and bpf_trace_printk()
>   selftests/bpf: add bpf_iter test with bpf_get_task_stack()
>
>  include/linux/bpf.h                           |  1 +
>  include/linux/perf_event.h                    |  2 +
>  include/uapi/linux/bpf.h                      | 36 ++++++++-
>  kernel/bpf/stackmap.c                         | 75 ++++++++++++++++++-
>  kernel/bpf/verifier.c                         |  4 +-
>  kernel/events/callchain.c                     | 13 ++--
>  kernel/trace/bpf_trace.c                      | 12 ++-
>  scripts/bpf_helpers_doc.py                    |  2 +
>  tools/include/uapi/linux/bpf.h                | 36 ++++++++-
>  .../selftests/bpf/prog_tests/bpf_iter.c       | 17 +++++
>  .../selftests/bpf/progs/bpf_iter_task_stack.c | 37 +++++++++
>  11 files changed, 220 insertions(+), 15 deletions(-)
>  create mode 100644 tools/testing/selftests/bpf/progs/bpf_iter_task_stack.c
>
> --
> 2.24.1

Thanks for working on this! This will enable a whole new set of tools
and applications.

Acked-by: Andrii Nakryiko <andriin@fb.com>