The patch-set is consists of the following parts:
- A refactoring of the libbpf's hashmap interface to use `long`
instead of `void*` for keys and values. The reasoning behind the
refactoring is that integer keys / values are used in libbpf more
often then pointer keys / values. Thus the refactoring reduces the
number of awkward looking casts like `(void *)(long)off`.
The `long` is used with an implicit assumption that
`sizeof(long) == sizeof(void *)` on all supported platforms.
This includes changes in `libbpf` and `perf` source code.
- A change to `lib/bpf/btf.c:btf__dedup` that adds a new pass named
"Resolve unambiguous forward declaration". This pass builds a
hashmap `name_off -> uniquely named struct or union` and uses it to
replace FWD types by structs or unions. This is necessary for corner
cases when FWD is not used as a part of some struct or union
definition de-duplicated by `btf_dedup_struct_types`.
The goal of the patch-set is to resolve forward declarations that
don't take part in type graphs comparisons if declaration name is
unambiguous.
Example:
CU #1:
struct foo; // standalone forward declaration
struct foo *some_global;
CU #2:
struct foo { int x; };
struct foo *another_global;
Currently the de-duplicated BTF for this example looks as follows:
[1] STRUCT 'foo' size=4 vlen=1 ...
[2] INT 'int' size=4 ...
[3] PTR '(anon)' type_id=1
[4] FWD 'foo' fwd_kind=struct
[5] PTR '(anon)' type_id=4
The goal of this patch-set is to simplify it as follows:
[1] STRUCT 'foo' size=4 vlen=1
'x' type_id=2 bits_offset=0
[2] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED
[3] PTR '(anon)' type_id=1
For defconfig kernel with BTF enabled this removes 63 forward
declarations.
For allmodconfig kernel with BTF enabled this removes ~5K out of ~21K
forward declarations in ko objects. This unlocks some additional
de-duplication in ko objects, but impact is tiny: ~13K less BTF ids
out of ~2M.
Changelog:
v1 -> v2
- Style fixes in btf_dedup_resolve_fwd and btf_dedup_resolve_fwds as
suggested by Alan.
v2 -> v3
Changes suggested by Andrii:
- perf's util/hashtable.{c,h} are synchronized with libbpf
implementation, perf's source code updated accordingly;
- changes to libbpf, bpf selftests and perf are combined in a single
patch to simplify bisecting;
- hashtable interface updated to be long -> long instead of
uintptr_t -> uintptr_t;
- btf_dedup_resolve_fwds updated to correctly use IS_ERR / PTR_ERR
macro;
- test cases for btf_dedup_resolve_fwds are updated for better
clarity.
[v1] https://lore.kernel.org/bpf/20221102110905.2433622-1-eddyz87@xxxxxxxxx/
[v2] https://lore.kernel.org/bpf/20221103033430.2611623-1-eddyz87@xxxxxxxxx/
Eduard Zingerman (3):
libbpf: hashmap interface update to long -> long
libbpf: Resolve unambigous forward declarations
selftests/bpf: Tests for btf_dedup_resolve_fwds
tools/bpf/bpftool/btf.c | 25 +--
tools/bpf/bpftool/common.c | 10 +-
tools/bpf/bpftool/gen.c | 19 +-
tools/bpf/bpftool/link.c | 8 +-
tools/bpf/bpftool/main.h | 14 +-
tools/bpf/bpftool/map.c | 8 +-
tools/bpf/bpftool/pids.c | 16 +-
tools/bpf/bpftool/prog.c | 8 +-
tools/lib/bpf/btf.c | 184 +++++++++++++++---
tools/lib/bpf/btf_dump.c | 16 +-
tools/lib/bpf/hashmap.c | 16 +-
tools/lib/bpf/hashmap.h | 34 ++--
tools/lib/bpf/libbpf.c | 18 +-
tools/lib/bpf/strset.c | 18 +-
tools/lib/bpf/usdt.c | 31 ++-
tools/perf/tests/expr.c | 40 ++--
tools/perf/tests/pmu-events.c | 6 +-
tools/perf/util/bpf-loader.c | 23 ++-
tools/perf/util/expr.c | 32 ++-
tools/perf/util/hashmap.c | 16 +-
tools/perf/util/hashmap.h | 34 ++--
tools/perf/util/metricgroup.c | 12 +-
tools/perf/util/stat.c | 9 +-
tools/testing/selftests/bpf/prog_tests/btf.c | 176 +++++++++++++++++
.../bpf/prog_tests/btf_dedup_split.c | 45 +++--
.../selftests/bpf/prog_tests/hashmap.c | 102 +++++-----
.../bpf/prog_tests/kprobe_multi_test.c | 6 +-
27 files changed, 602 insertions(+), 324 deletions(-)
--
2.34.1
