mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
synced 2026-02-13 16:33:57 -05:00
feeaf1346f
When the BPF ring buffer is full, a new event cannot be recorded until one
or more old events are consumed to make enough space for it. In cases such
as fault diagnostics, where recent events are more useful than older ones,
this mechanism may lead to critical events being lost.
So add overwrite mode for BPF ring buffer to address it. In this mode, the
new event overwrites the oldest event when the buffer is full.
The basic idea is as follows:
1. producer_pos tracks the next position to record new event. When there
is enough free space, producer_pos is simply advanced by producer to
make space for the new event.
2. To avoid waiting for consumer when the buffer is full, a new variable,
overwrite_pos, is introduced for producer. It points to the oldest event
committed in the buffer. It is advanced by producer to discard one or more
oldest events to make space for the new event when the buffer is full.
3. pending_pos tracks the oldest event to be committed. pending_pos is never
passed by producer_pos, so multiple producers never write to the same
position at the same time.
The following example diagrams show how it works in a 4096-byte ring buffer.
1. At first, {producer,overwrite,pending,consumer}_pos are all set to 0.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| |
| |
| |
+-----------------------------------------------------------------------+
^
|
|
producer_pos = 0
overwrite_pos = 0
pending_pos = 0
consumer_pos = 0
2. Now reserve a 512-byte event A.
There is enough free space, so A is allocated at offset 0. And producer_pos
is advanced to 512, the end of A. Since A is not submitted, the BUSY bit is
set.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | |
| A | |
| [BUSY] | |
+-----------------------------------------------------------------------+
^ ^
| |
| |
| producer_pos = 512
|
overwrite_pos = 0
pending_pos = 0
consumer_pos = 0
3. Reserve event B, size 1024.
B is allocated at offset 512 with BUSY bit set, and producer_pos is advanced
to the end of B.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | |
| A | B | |
| [BUSY] | [BUSY] | |
+-----------------------------------------------------------------------+
^ ^
| |
| |
| producer_pos = 1536
|
overwrite_pos = 0
pending_pos = 0
consumer_pos = 0
4. Reserve event C, size 2048.
C is allocated at offset 1536, and producer_pos is advanced to 3584.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | | |
| A | B | C | |
| [BUSY] | [BUSY] | [BUSY] | |
+-----------------------------------------------------------------------+
^ ^
| |
| |
| producer_pos = 3584
|
overwrite_pos = 0
pending_pos = 0
consumer_pos = 0
5. Submit event A.
The BUSY bit of A is cleared. B becomes the oldest event to be committed, so
pending_pos is advanced to 512, the start of B.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | | |
| A | B | C | |
| | [BUSY] | [BUSY] | |
+-----------------------------------------------------------------------+
^ ^ ^
| | |
| | |
| pending_pos = 512 producer_pos = 3584
|
overwrite_pos = 0
consumer_pos = 0
6. Submit event B.
The BUSY bit of B is cleared, and pending_pos is advanced to the start of C,
which is now the oldest event to be committed.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | | |
| A | B | C | |
| | | [BUSY] | |
+-----------------------------------------------------------------------+
^ ^ ^
| | |
| | |
| pending_pos = 1536 producer_pos = 3584
|
overwrite_pos = 0
consumer_pos = 0
7. Reserve event D, size 1536 (3 * 512).
There are 2048 bytes not being written between producer_pos (currently 3584)
and pending_pos, so D is allocated at offset 3584, and producer_pos is advanced
by 1536 (from 3584 to 5120).
Since event D will overwrite all bytes of event A and the first 512 bytes of
event B, overwrite_pos is advanced to the start of event C, the oldest event
that is not overwritten.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | | |
| D End | | C | D Begin|
| [BUSY] | | [BUSY] | [BUSY] |
+-----------------------------------------------------------------------+
^ ^ ^
| | |
| | pending_pos = 1536
| | overwrite_pos = 1536
| |
| producer_pos=5120
|
consumer_pos = 0
8. Reserve event E, size 1024.
Although there are 512 bytes not being written between producer_pos and
pending_pos, E cannot be reserved, as it would overwrite the first 512
bytes of event C, which is still being written.
9. Submit event C and D.
pending_pos is advanced to the end of D.
0 512 1024 1536 2048 2560 3072 3584 4096
+-----------------------------------------------------------------------+
| | | | |
| D End | | C | D Begin|
| | | | |
+-----------------------------------------------------------------------+
^ ^ ^
| | |
| | overwrite_pos = 1536
| |
| producer_pos=5120
| pending_pos=5120
|
consumer_pos = 0
The performance data for overwrite mode will be provided in a follow-up
patch that adds overwrite-mode benchmarks.
A sample of performance data for non-overwrite mode, collected on an x86_64
CPU and an arm64 CPU, before and after this patch, is shown below. As we can
see, no obvious performance regression occurs.
- x86_64 (AMD EPYC 9654)
Before:
Ringbuf, multi-producer contention
==================================
rb-libbpf nr_prod 1 11.623 ± 0.027M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 2 15.812 ± 0.014M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 3 7.871 ± 0.003M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 4 6.703 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 8 2.896 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 12 2.054 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 16 1.864 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 20 1.580 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 24 1.484 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 28 1.369 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 32 1.316 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 36 1.272 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 40 1.239 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 44 1.226 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 48 1.213 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 52 1.193 ± 0.001M/s (drops 0.000 ± 0.000M/s)
After:
Ringbuf, multi-producer contention
==================================
rb-libbpf nr_prod 1 11.845 ± 0.036M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 2 15.889 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 3 8.155 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 4 6.708 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 8 2.918 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 12 2.065 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 16 1.870 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 20 1.582 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 24 1.482 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 28 1.372 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 32 1.323 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 36 1.264 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 40 1.236 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 44 1.209 ± 0.002M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 48 1.189 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 52 1.165 ± 0.002M/s (drops 0.000 ± 0.000M/s)
- arm64 (HiSilicon Kunpeng 920)
Before:
Ringbuf, multi-producer contention
==================================
rb-libbpf nr_prod 1 11.310 ± 0.623M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 2 9.947 ± 0.004M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 3 6.634 ± 0.011M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 4 4.502 ± 0.003M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 8 3.888 ± 0.003M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 12 3.372 ± 0.005M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 16 3.189 ± 0.010M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 20 2.998 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 24 3.086 ± 0.018M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 28 2.845 ± 0.004M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 32 2.815 ± 0.008M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 36 2.771 ± 0.009M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 40 2.814 ± 0.011M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 44 2.752 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 48 2.695 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 52 2.710 ± 0.006M/s (drops 0.000 ± 0.000M/s)
After:
Ringbuf, multi-producer contention
==================================
rb-libbpf nr_prod 1 11.283 ± 0.550M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 2 9.993 ± 0.003M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 3 6.898 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 4 5.257 ± 0.001M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 8 3.830 ± 0.005M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 12 3.528 ± 0.013M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 16 3.265 ± 0.018M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 20 2.990 ± 0.007M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 24 2.929 ± 0.014M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 28 2.898 ± 0.010M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 32 2.818 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 36 2.789 ± 0.012M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 40 2.770 ± 0.006M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 44 2.651 ± 0.007M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 48 2.669 ± 0.005M/s (drops 0.000 ± 0.000M/s)
rb-libbpf nr_prod 52 2.695 ± 0.009M/s (drops 0.000 ± 0.000M/s)
Signed-off-by: Xu Kuohai <xukuohai@huawei.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20251018035738.4039621-2-xukuohai@huaweicloud.com
Why we want a copy of kernel headers in tools?
==============================================
There used to be no copies, with tools/ code using kernel headers
directly. From time to time tools/perf/ broke due to legitimate kernel
hacking. At some point Linus complained about such direct usage. Then we
adopted the current model.
The way these headers are used in perf are not restricted to just
including them to compile something.
There are sometimes used in scripts that convert defines into string
tables, etc, so some change may break one of these scripts, or new MSRs
may use some different #define pattern, etc.
E.g.:
$ ls -1 tools/perf/trace/beauty/*.sh | head -5
tools/perf/trace/beauty/arch_errno_names.sh
tools/perf/trace/beauty/drm_ioctl.sh
tools/perf/trace/beauty/fadvise.sh
tools/perf/trace/beauty/fsconfig.sh
tools/perf/trace/beauty/fsmount.sh
$
$ tools/perf/trace/beauty/fadvise.sh
static const char *fadvise_advices[] = {
[0] = "NORMAL",
[1] = "RANDOM",
[2] = "SEQUENTIAL",
[3] = "WILLNEED",
[4] = "DONTNEED",
[5] = "NOREUSE",
};
$
The tools/perf/check-headers.sh script, part of the tools/ build
process, points out changes in the original files.
So its important not to touch the copies in tools/ when doing changes in
the original kernel headers, that will be done later, when
check-headers.sh inform about the change to the perf tools hackers.
Another explanation from Ingo Molnar:
It's better than all the alternatives we tried so far:
- Symbolic links and direct #includes: this was the original approach but
was pushed back on from the kernel side, when tooling modified the
headers and broke them accidentally for kernel builds.
- Duplicate self-defined ABI headers like glibc: double the maintenance
burden, double the chance for mistakes, plus there's no tech-driven
notification mechanism to look at new kernel side changes.
What we are doing now is a third option:
- A software-enforced copy-on-write mechanism of kernel headers to
tooling, driven by non-fatal warnings on the tooling side build when
kernel headers get modified:
Warning: Kernel ABI header differences:
diff -u tools/include/uapi/drm/i915_drm.h include/uapi/drm/i915_drm.h
diff -u tools/include/uapi/linux/fs.h include/uapi/linux/fs.h
diff -u tools/include/uapi/linux/kvm.h include/uapi/linux/kvm.h
...
The tooling policy is to always pick up the kernel side headers as-is,
and integate them into the tooling build. The warnings above serve as a
notification to tooling maintainers that there's changes on the kernel
side.
We've been using this for many years now, and it might seem hacky, but
works surprisingly well.