Re: [PATCH bpf-next v4 17/27] bpf: Add multi trampoline attach support

On Tue, Aug 31, 2021 at 4:36 PM Andrii Nakryiko
[off-list ref] wrote:

On Thu, Aug 26, 2021 at 12:41 PM Jiri Olsa [off-list ref] wrote:

quoted

Adding new multi trampoline link (BPF_LINK_TYPE_TRACING_MULTI)
as an interface to attach program to multiple functions.

The link_create bpf_attr interface already has 'bpf_prog' file
descriptor, that defines the program to be attached. It must be
loaded with BPF_F_MULTI_FUNC flag.

Adding new multi_btf_ids/multi_btf_ids_cnt link_create bpf_attr
fields that provides BTF ids.

The new link gets multi trampoline (via bpf_trampoline_multi_get)
and links the provided program with embedded trampolines and the
'main' trampoline with new multi link/unlink functions:

  int bpf_trampoline_multi_link_prog(struct bpf_prog *prog,
                                     struct bpf_trampoline_multi *tr);
  int bpf_trampoline_multi_unlink_prog(struct bpf_prog *prog,
                                       struct bpf_trampoline_multi *tr);

If embedded trampoline contains fexit programs, we need to switch
its model to the multi trampoline model (because of the final 'ret'
argument). We keep the count of attached multi func programs for each
trampoline, so we can tell when to switch the model.

Related to my comments on the next patch, if we switch the order of
return value and always reserve 6 slots for input args, regardless of
the actual number of function input args, that should make this
upgrade logic unnecessary, right?

quoted

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 include/linux/bpf.h            |   5 ++
 include/uapi/linux/bpf.h       |   5 ++
 kernel/bpf/core.c              |   1 +
 kernel/bpf/syscall.c           | 120 +++++++++++++++++++++++++++++++++
 kernel/bpf/trampoline.c        |  87 ++++++++++++++++++++++--
 tools/include/uapi/linux/bpf.h |   5 ++
 6 files changed, 219 insertions(+), 4 deletions(-)

[...]

quoted

diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 1f9d336861f0..9533200ffadf 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h

@@ -1008,6 +1008,7 @@ enum bpf_link_type {
        BPF_LINK_TYPE_NETNS = 5,
        BPF_LINK_TYPE_XDP = 6,
        BPF_LINK_TYPE_PERF_EVENT = 7,
+       BPF_LINK_TYPE_TRACING_MULTI = 8,

        MAX_BPF_LINK_TYPE,
 };

@@ -1462,6 +1463,10 @@ union bpf_attr {
                                 */
                                __u64           bpf_cookie;
                        } perf_event;
+                       struct {
+                               __aligned_u64   multi_btf_ids;          /* addresses to attach */
+                               __u32           multi_btf_ids_cnt;      /* addresses count */
+                       };

Please follow the pattern of perf_event, name this struct "multi".

quoted

                };
        } link_create;

diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index bad03dde97a2..6c16ac43dd91 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c

[...]

quoted

+
+       bpf_link_init(&link->link, BPF_LINK_TYPE_TRACING_MULTI,
+                     &bpf_tracing_multi_link_lops, prog);
+       link->attach_type = prog->expected_attach_type;
+       link->multi = multi;
+
+       err = bpf_link_prime(&link->link, &link_primer);
+       if (err)
+               goto out_free;
+       err = bpf_trampoline_multi_link_prog(prog, multi);
+       if (err)
+               goto out_free;

bpf_link_cleanup(), can't free link after priming. Look at other
places using bpf_link.

quoted

+       return bpf_link_settle(&link_primer);
+
+out_free:
+       bpf_trampoline_multi_put(multi);
+       kfree(link);
+out_free_ids:
+       kfree(btf_ids);
+       return err;
+}
+

[...]