Re: [RFC bpf-next 1/3] bpf: add module support to btf display helpers

On Sat, 14 Nov 2020, Yonghong Song wrote:

On 11/14/20 8:04 AM, Alexei Starovoitov wrote:

quoted

On Fri, Nov 13, 2020 at 10:59 PM Andrii Nakryiko
[off-list ref] wrote:

quoted

On Fri, Nov 13, 2020 at 10:11 AM Alan Maguire [off-list ref]
wrote:

quoted

bpf_snprintf_btf and bpf_seq_printf_btf use a "struct btf_ptr *"
argument that specifies type information about the type to
be displayed.  Augment this information to include a module
name, allowing such display to support module types.

Signed-off-by: Alan Maguire <redacted>
---
  include/linux/btf.h            |  8 ++++++++
  include/uapi/linux/bpf.h       |  5 ++++-
  kernel/bpf/btf.c               | 18 ++++++++++++++++++
  kernel/trace/bpf_trace.c       | 42
  ++++++++++++++++++++++++++++++++----------
  tools/include/uapi/linux/bpf.h |  5 ++++-
  5 files changed, 66 insertions(+), 12 deletions(-)

diff --git a/include/linux/btf.h b/include/linux/btf.h
index 2bf6418..d55ca00 100644
--- a/include/linux/btf.h
+++ b/include/linux/btf.h

@@ -209,6 +209,14 @@ static inline const struct btf_var_secinfo

*btf_type_var_secinfo(
  const struct btf_type *btf_type_by_id(const struct btf *btf, u32
  type_id);
  const char *btf_name_by_offset(const struct btf *btf, u32 offset);
  struct btf *btf_parse_vmlinux(void);
+#ifdef CONFIG_DEBUG_INFO_BTF_MODULES
+struct btf *bpf_get_btf_module(const char *name);
+#else
+static inline struct btf *bpf_get_btf_module(const char *name)
+{
+       return ERR_PTR(-ENOTSUPP);
+}
+#endif
  struct btf *bpf_prog_get_target_btf(const struct bpf_prog *prog);
  #else
  static inline const struct btf_type *btf_type_by_id(const struct btf
  *btf,

diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 162999b..26978be 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h

@@ -3636,7 +3636,8 @@ struct bpf_stack_build_id {
   *             the pointer data is carried out to avoid kernel crashes
   during
   *             operation.  Smaller types can use string space on the
   stack;
   *             larger programs can use map data to store the string
- *             representation.
+ *             representation.  Module-specific data structures can be
+ *             displayed if the module name is supplied.
   *
   *             The string can be subsequently shared with userspace via
   *             bpf_perf_event_output() or ring buffer interfaces.

@@ -5076,11 +5077,13 @@ struct bpf_sk_lookup {
   * potentially to specify additional details about the BTF pointer
   * (rather than its mode of display) - is included for future use.
   * Display flags - BTF_F_* - are passed to bpf_snprintf_btf separately.
+ * A module name can be specified for module-specific data.
  */
  struct btf_ptr {
         void *ptr;
         __u32 type_id;
         __u32 flags;            /* BTF ptr flags; unused at present. */
+       const char *module;     /* optional module name. */

I think module name is a wrong API here, similarly how type name was
wrong API for specifying the type (and thus we use type_id here).
Using the module's BTF ID seems like a more suitable interface. That's
what I'm going to use for all kinds of existing BPF APIs that expect
BTF type to attach BPF programs.

Right now, we use only type_id and implicitly know that it's in
vmlinux BTF. With module BTFs, we now need a pair of BTF object ID +
BTF type ID to uniquely identify the type. vmlinux BTF now can be
specified in two different ways: either leaving BTF object ID as zero
(for simplicity and backwards compatibility) or specifying it's actual
BTF obj ID (which pretty much always should be 1, btw). This feels
like a natural extension, WDYT?

And similar to type_id, no one should expect users to specify these
IDs by hand, Clang built-in and libbpf should work together to figure
this out for the kernel to use.

BTW, with module names there is an extra problem for end users. Some
types could be either built-in or built as a module (e.g., XFS data
structures). Why would we require BPF users to care which is the case
on any given host?

+1.
As much as possible libbpf should try to hide the difference between
type in a module vs type in the vmlinux, since that difference most of the
time is irrelevant from bpf prog pov.

All sounds good to me - I've split out the libbpf fix and 
put together an updated patchset for the helpers/test which
converts the currently unused __u32 "flags" field in
struct btf_ptr to an "obj_id" field.  If obj_id is > 1 it
is presumed to be a module ID.  I'd suggest we could move
ahead with those changes, using the more clunky methods
to retrieve the module-specific BTF id, and later fix up
the test to use Yonghong's __builtin_btf_type_id()
modification.  Does that sound reasonable?

In connection to this, I wonder if libbpf could
benefit from a simple helper btf__id() (similar
to btf__fd()), allowing easy retrieval of the
object ID associated with module BTF?  I suspect
we will always have cases in general-purpose
tracers where we need to look up BTF ids of
objects dynamically, so such a function would
help in that case.

I just crafted a llvm patch where for __builtin_btf_type_id(), a 64bit value
is returned instead of a 32bit value. libbpf can use the lower
32bit as the btf_type_id and upper 32bit as the kernel module btf id.

   https://reviews.llvm.org/D91489

feel free to experiment with it to see whether it helps.

Great! I'll give it a try, thanks!

Alan