At Meta we have a profiling daemon which periodically collects
information on many hosts. This collection usually involves grabbing
stacks (user and kernel) using perf_event BPF progs and later symbolicating
them. For user stacks we try to use BPF_F_USER_BUILD_ID and rely on
remote symbolication, but BPF_F_USER_BUILD_ID doesn't always succeed. In
those cases we must fall back to digging around in /proc/PID/maps to map
virtual address to (binary, offset). The /proc/PID/maps digging does not
occur synchronously with stack collection, so the process might already
be gone, in which case it won't have /proc/PID/maps and we will fail to
symbolicate.
This 'exited process problem' doesn't occur very often as
most of the prod services we care to profile are long-lived daemons, but
there are enough usecases to warrant a workaround: a BPF program which
can be optionally loaded at data collection time and essentially walks
/proc/PID/maps. Currently this is done by walking the vma list:
struct vm_area_struct* mmap = BPF_CORE_READ(mm, mmap);
mmap_next = BPF_CORE_READ(rmap, vm_next); /* in a loop */
Since commit 763ecb035029 ("mm: remove the vma linked list") there's no
longer a vma linked list to walk. Walking the vma maple tree is not as
simple as hopping struct vm_area_struct->vm_next. Luckily,
commit f39af05949a4 ("mm: add VMA iterator"), another commit in that series,
added struct vma_iterator and for_each_vma macro for easy vma iteration. If
similar functionality was exposed to BPF programs, it would be perfect for our
usecase.
This series adds such functionality, specifically a BPF equivalent of
for_each_vma using the open-coded iterator style.
Notes:
* This approach was chosen after discussion on a previous series [0] which
attempted to solve the same problem by adding a BPF_F_VMA_NEXT flag to
bpf_find_vma.
* Unlike the task_vma bpf_iter, the open-coded iterator kfuncs here do not
drop the vma read lock between iterations. See Alexei's response in [0].
* The [vsyscall] page isn't really part of task->mm's vmas, but
/proc/PID/maps returns information about it anyways. The vma iter added
here does not do the same. See comment on selftest in patch 3.
* bpf_iter_task_vma allocates a _data struct which contains - among other
things - struct vma_iterator, using BPF allocator and keeps a pointer to
the bpf_iter_task_vma_data. This is done in order to prevent changes to
struct ma_state - which is wrapped by struct vma_iterator - from
necessitating changes to uapi struct bpf_iter_task_vma.
Changelog:
v3 -> v4: https://lore.kernel.org/bpf/20230822050558.2937659-1-davemarchevsky@xxxxxx/
Patch 1 ("bpf: Don't explicitly emit BTF for struct btf_iter_num")
* Add Andrii ack
Patch 2 ("bpf: Introduce task_vma open-coded iterator kfuncs")
* Mark bpf_iter_task_vma_new args KF_RCU and remove now-unnecessary !task
check (Yonghong)
* Although KF_RCU is a function-level flag, in reality it only applies to
the task_struct *task parameter, as the other two params are a scalar int
and a specially-handled KF_ARG_PTR_TO_ITER
* Remove struct bpf_iter_task_vma definition from uapi headers, define in
kernel/bpf/task_iter.c instead (Andrii)
Patch 3 ("selftests/bpf: Add tests for open-coded task_vma iter")
* Use a local var when looping over vmas to track map idx. Update vmas_seen
global after done iterating. Don't start iterating or update vmas_seen if
vmas_seen global is nonzero. (Andrii)
* Move getpgid() call to correct spot - above skel detach. (Andrii)
v2 -> v3: https://lore.kernel.org/bpf/20230821173415.1970776-1-davemarchevsky@xxxxxx/
Patch 1 ("bpf: Don't explicitly emit BTF for struct btf_iter_num")
* Add Yonghong ack
Patch 2 ("bpf: Introduce task_vma open-coded iterator kfuncs")
* UAPI bpf header and tools/ version should match
* Add bpf_iter_task_vma_kern_data which bpf_iter_task_vma_kern points to,
bpf_mem_alloc/free it instead of just vma_iterator. (Alexei)
* Inner data ptr == NULL implies initialization failed
v1 -> v2: https://lore.kernel.org/bpf/20230810183513.684836-1-davemarchevsky@xxxxxx/
* Patch 1
* Now removes the unnecessary BTF_TYPE_EMIT instead of changing the
type (Yonghong)
* Patch 2
* Don't do unnecessary BTF_TYPE_EMIT (Yonghong)
* Bump task refcount to prevent ->mm reuse (Yonghong)
* Keep a pointer to vma_iterator in bpf_iter_task_vma, alloc/free
via BPF mem allocator (Yonghong, Stanislav)
* Patch 3
Patch summary:
* Patch 1 is a tiny fix I ran into while implementing the vma iter in this
series. It can be applied independently.
* Patch 2 is the meat of the implementation
* Patch 3 adds tests for the new functionality
* Existing iter tests exercise failure cases (e.g. prog that doesn't call
_destroy()). I didn't replicate them in this series, but am happy to add
them in v2 if folks feel that it would be worthwhile.
[0]: https://lore.kernel.org/bpf/20230801145414.418145-1-davemarchevsky@xxxxxx/
Dave Marchevsky (3):
bpf: Don't explicitly emit BTF for struct btf_iter_num
bpf: Introduce task_vma open-coded iterator kfuncs
selftests/bpf: Add tests for open-coded task_vma iter
kernel/bpf/bpf_iter.c | 2 -
kernel/bpf/helpers.c | 3 +
kernel/bpf/task_iter.c | 85 +++++++++++++++++++
tools/lib/bpf/bpf_helpers.h | 8 ++
.../selftests/bpf/prog_tests/bpf_iter.c | 26 +++---
.../testing/selftests/bpf/prog_tests/iters.c | 71 ++++++++++++++++
...f_iter_task_vma.c => bpf_iter_task_vmas.c} | 0
.../selftests/bpf/progs/iters_task_vma.c | 63 ++++++++++++++
8 files changed, 243 insertions(+), 15 deletions(-)
rename tools/testing/selftests/bpf/progs/{bpf_iter_task_vma.c => bpf_iter_task_vmas.c} (100%)
create mode 100644 tools/testing/selftests/bpf/progs/iters_task_vma.c
--
2.34.1
