Pull perf tools updates from Arnaldo Carvalho de Melo:
- Introduce 'perf lock contention' subtool, using new lock contention
tracepoints and using BPF for in kernel aggregation and then
userspace processing using the perf tooling infrastructure for
resolving symbols, target specification, etc.
Since the new lock contention tracepoints don't provide lock names,
get up to 8 stack traces and display the first non-lock function
symbol name as a caller:
$ perf lock report -F acquired,contended,avg_wait,wait_total
Name acquired contended avg wait total wait
update_blocked_a... 40 40 3.61 us 144.45 us
kernfs_fop_open+... 5 5 3.64 us 18.18 us
_nohz_idle_balance 3 3 2.65 us 7.95 us
tick_do_update_j... 1 1 6.04 us 6.04 us
ep_scan_ready_list 1 1 3.93 us 3.93 us
Supports the usual 'perf record' + 'perf report' workflow as well as
a BCC/bpftrace like mode where you start the tool and then press
control+C to get results:
$ sudo perf lock contention -b
^C
contended total wait max wait avg wait type caller
42 192.67 us 13.64 us 4.59 us spinlock queue_work_on+0x20
23 85.54 us 10.28 us 3.72 us spinlock worker_thread+0x14a
6 13.92 us 6.51 us 2.32 us mutex kernfs_iop_permission+0x30
3 11.59 us 10.04 us 3.86 us mutex kernfs_dop_revalidate+0x3c
1 7.52 us 7.52 us 7.52 us spinlock kthread+0x115
1 7.24 us 7.24 us 7.24 us rwlock:W sys_epoll_wait+0x148
2 7.08 us 3.99 us 3.54 us spinlock delayed_work_timer_fn+0x1b
1 6.41 us 6.41 us 6.41 us spinlock idle_balance+0xa06
2 2.50 us 1.83 us 1.25 us mutex kernfs_iop_lookup+0x2f
1 1.71 us 1.71 us 1.71 us mutex kernfs_iop_getattr+0x2c
...
- Add new 'perf kwork' tool to trace time properties of kernel work
(such as softirq, and workqueue), uses eBPF skeletons to collect info
in kernel space, aggregating data that then gets processed by the
userspace tool, e.g.:
# perf kwork report
Kwork Name | Cpu | Total Runtime | Count | Max runtime | Max runtime start | Max runtime end |
----------------------------------------------------------------------------------------------------
nvme0q5:130 | 004 | 1.101 ms | 49 | 0.051 ms | 26035.056403 s | 26035.056455 s |
amdgpu:162 | 002 | 0.176 ms | 9 | 0.046 ms | 26035.268020 s | 26035.268066 s |
nvme0q24:149 | 023 | 0.161 ms | 55 | 0.009 ms | 26035.655280 s | 26035.655288 s |
nvme0q20:145 | 019 | 0.090 ms | 33 | 0.014 ms | 26035.939018 s | 26035.939032 s |
nvme0q31:156 | 030 | 0.075 ms | 21 | 0.010 ms | 26035.052237 s | 26035.052247 s |
nvme0q8:133 | 007 | 0.062 ms | 12 | 0.021 ms | 26035.416840 s | 26035.416861 s |
nvme0q6:131 | 005 | 0.054 ms | 22 | 0.010 ms | 26035.199919 s | 26035.199929 s |
nvme0q19:144 | 018 | 0.052 ms | 14 | 0.010 ms | 26035.110615 s | 26035.110625 s |
nvme0q7:132 | 006 | 0.049 ms | 13 | 0.007 ms | 26035.125180 s | 26035.125187 s |
nvme0q18:143 | 017 | 0.033 ms | 14 | 0.007 ms | 26035.169698 s | 26035.169705 s |
nvme0q17:142 | 016 | 0.013 ms | 1 | 0.013 ms | 26035.565147 s | 26035.565160 s |
enp5s0-rx-0:164 | 006 | 0.004 ms | 4 | 0.002 ms | 26035.928882 s | 26035.928884 s |
enp5s0-tx-0:166 | 008 | 0.003 ms | 3 | 0.002 ms | 26035.870923 s | 26035.870925 s |
--------------------------------------------------------------------------------------------------------
See commit log messages for more examples with extra options to limit
the events time window, etc.
- Add support for new AMD IBS (Instruction Based Sampling) features:
With the DataSrc extensions, the source of data can be decoded among:
- Local L3 or other L1/L2 in CCX.
- A peer cache in a near CCX.
- Data returned from DRAM.
- A peer cache in a far CCX.
- DRAM address map with "long latency" bit set.
- Data returned from MMIO/Config/PCI/APIC.
- Extension Memory (S-Link, GenZ, etc - identified by the CS target
and/or address map at DF's choice).
- Peer Agent Memory.
- Support hardware tracing with Intel PT on guest machines, combining
the traces with the ones in the host machine.
- Add a "-m" option to 'perf buildid-list' to show kernel and modules
build-ids, to display all of the information needed to do external
symbolization of kernel stack traces, such as those collected by
bpf_get_stackid().
- Add arch TSC frequency information to perf.data file headers.
- Handle changes in the binutils disassembler function signatures in
perf, bpftool and bpf_jit_disasm (Acked by the bpftool maintainer).
- Fix building the perf perl binding with the newest gcc in distros
such as fedora rawhide, where some new warnings were breaking the
build as perf uses -Werror.
- Add 'perf test' entry for branch stack sampling.
- Add ARM SPE system wide 'perf test' entry.
- Add user space counter reading tests to 'perf test'.
- Build with python3 by default, if available.
- Add python converter script for the vendor JSON event files.
- Update vendor JSON files for most Intel cores.
- Add vendor JSON File for Intel meteorlake.
- Add Arm Cortex-A78C and X1C JSON vendor event files.
- Add workaround to symbol address reading from ELF files without phdr,
falling back to the previoous equation.
- Convert legacy map definition to BTF-defined in the perf BPF script
test.
- Rework prologue generation code to stop using libbpf deprecated APIs.
- Add default hybrid events for 'perf stat' on x86.
- Add topdown metrics in the default 'perf stat' on the hybrid machines
(big/little cores).
- Prefer sampled CPU when exporting JSON in 'perf data convert'
- Fix ('perf stat CSV output linter') and ("Check branch stack
sampling") 'perf test' entries on s390.
* tag 'perf-tools-for-v6.0-2022-08-04' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux: (169 commits)
perf stat: Refactor __run_perf_stat() common code
perf lock: Print the number of lost entries for BPF
perf lock: Add --map-nr-entries option
perf lock: Introduce struct lock_contention
perf scripting python: Do not build fail on deprecation warnings
genelf: Use HAVE_LIBCRYPTO_SUPPORT, not the never defined HAVE_LIBCRYPTO
perf build: Suppress openssl v3 deprecation warnings in libcrypto feature test
perf parse-events: Break out tracepoint and printing
perf parse-events: Don't #define YY_EXTRA_TYPE
tools bpftool: Don't display disassembler-four-args feature test
tools bpftool: Fix compilation error with new binutils
tools bpf_jit_disasm: Don't display disassembler-four-args feature test
tools bpf_jit_disasm: Fix compilation error with new binutils
tools perf: Fix compilation error with new binutils
tools include: add dis-asm-compat.h to handle version differences
tools build: Don't display disassembler-four-args feature test
tools build: Add feature test for init_disassemble_info API changes
perf test: Add ARM SPE system wide test
perf tools: Rework prologue generation code
perf bpf: Convert legacy map definition to BTF-defined
...
The feature test to detect the availability of zlib in bpftool's
Makefile does not bring much. The library is not optional: it may or may
not be required along libbfd for disassembling instructions, but in any
case it is necessary to build feature.o or even libbpf, on which bpftool
depends.
If we remove the feature test, we lose the nicely formatted error
message, but we get a compiler error about "zlib.h: No such file or
directory", which is equally informative. Let's get rid of the test.
Suggested-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20220705200456.285943-1-quentin@isovalent.com
bpftool needs to know about the newly introduced BPF_CORE_TYPE_MATCHES
relocation for its 'gen min_core_btf' command to work properly in the
present of this relocation.
Specifically, we need to make sure to mark types and fields so that they
are present in the minimized BTF for "type match" checks to work out.
However, contrary to the existing btfgen_record_field_relo, we need to
rely on the BTF -- and not the spec -- to find fields. With this change
we handle this new variant correctly. The functionality will be tested
with follow on changes to BPF selftests, which already run against a
minimized BTF created with bpftool.
Signed-off-by: Daniel Müller <deso@posteo.net>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Quentin Monnet <quentin@isovalent.com>
Link: https://lore.kernel.org/bpf/20220628160127.607834-3-deso@posteo.net
To make it more explicit that the features listed with "bpftool feature
list" are known to bpftool, but not necessary available on the system
(as opposed to the probed features), rename the "feature list" command
into "feature list_builtins".
Note that "bpftool feature list" still works as before given that we
recognise arguments from their prefixes; but the real name of the
subcommand, in particular as displayed in the man page or the
interactive help, will now include "_builtins".
Since we update the bash completion accordingly, let's also take this
chance to redirect error output to /dev/null in the completion script,
to avoid displaying unexpected error messages when users attempt to
tab-complete.
Suggested-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20220701093805.16920-1-quentin@isovalent.com
Now that bpftool is able to produce a list of known program, map, attach
types, let's use as much of this as we can in the bash completion file,
so that we don't have to expand the list each time a new type is added
to the kernel.
Also update the relevant test script to remove some checks that are no
longer needed.
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Müller <deso@posteo.net>
Link: https://lore.kernel.org/bpf/20220629203637.138944-3-quentin@isovalent.com
Add a "bpftool feature list" subcommand to list BPF "features".
Contrarily to "bpftool feature probe", this is not about the features
available on the system. Instead, it lists all features known to bpftool
from compilation time; in other words, all program, map, attach, link
types known to the libbpf version in use, and all helpers found in the
UAPI BPF header.
The first use case for this feature is bash completion: running the
command provides a list of types that can be used to produce the list of
candidate map types, for example.
Now that bpftool uses "standard" names provided by libbpf for the
program, map, link, and attach types, having the ability to list these
types and helpers could also be useful in scripts to loop over existing
items.
Sample output:
# bpftool feature list prog_types | grep -vw unspec | head -n 6
socket_filter
kprobe
sched_cls
sched_act
tracepoint
xdp
# bpftool -p feature list map_types | jq '.[1]'
"hash"
# bpftool feature list attach_types | grep '^cgroup_'
cgroup_inet_ingress
cgroup_inet_egress
[...]
cgroup_inet_sock_release
# bpftool feature list helpers | grep -vw bpf_unspec | wc -l
207
The "unspec" types and helpers are not filtered out by bpftool, so as to
remain closer to the enums, and to preserve the indices in the JSON
arrays (e.g. "hash" at index 1 == BPF_MAP_TYPE_HASH in map types list).
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Müller <deso@posteo.net>
Link: https://lore.kernel.org/bpf/20220629203637.138944-2-quentin@isovalent.com
Bpftool used to bump the memlock rlimit to make sure to be able to load
BPF objects. After the kernel has switched to memcg-based memory
accounting [0] in 5.11, bpftool has relied on libbpf to probe the system
for memcg-based accounting support and for raising the rlimit if
necessary [1]. But this was later reverted, because the probe would
sometimes fail, resulting in bpftool not being able to load all required
objects [2].
Here we add a more efficient probe, in bpftool itself. We first lower
the rlimit to 0, then we attempt to load a BPF object (and finally reset
the rlimit): if the load succeeds, then memcg-based memory accounting is
supported.
This approach was earlier proposed for the probe in libbpf itself [3],
but given that the library may be used in multithreaded applications,
the probe could have undesirable consequences if one thread attempts to
lock kernel memory while memlock rlimit is at 0. Since bpftool is
single-threaded and the rlimit is process-based, this is fine to do in
bpftool itself.
This probe was inspired by the similar one from the cilium/ebpf Go
library [4].
[0] commit 97306be45f ("Merge branch 'switch to memcg-based memory accounting'")
[1] commit a777e18f1b ("bpftool: Use libbpf 1.0 API mode instead of RLIMIT_MEMLOCK")
[2] commit 6b4384ff10 ("Revert "bpftool: Use libbpf 1.0 API mode instead of RLIMIT_MEMLOCK"")
[3] https://lore.kernel.org/bpf/20220609143614.97837-1-quentin@isovalent.com/t/#u
[4] https://github.com/cilium/ebpf/blob/v0.9.0/rlimit/rlimit.go#L39
Suggested-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Link: https://lore.kernel.org/bpf/20220629111351.47699-1-quentin@isovalent.com
This reverts commit a777e18f1b.
In commit a777e18f1b ("bpftool: Use libbpf 1.0 API mode instead of
RLIMIT_MEMLOCK"), we removed the rlimit bump in bpftool, because the
kernel has switched to memcg-based memory accounting. Thanks to the
LIBBPF_STRICT_AUTO_RLIMIT_MEMLOCK, we attempted to keep compatibility
with other systems and ask libbpf to raise the limit for us if
necessary.
How do we know if memcg-based accounting is supported? There is a probe
in libbpf to check this. But this probe currently relies on the
availability of a given BPF helper, bpf_ktime_get_coarse_ns(), which
landed in the same kernel version as the memory accounting change. This
works in the generic case, but it may fail, for example, if the helper
function has been backported to an older kernel. This has been observed
for Google Cloud's Container-Optimized OS (COS), where the helper is
available but rlimit is still in use. The probe succeeds, the rlimit is
not raised, and probing features with bpftool, for example, fails.
A patch was submitted [0] to update this probe in libbpf, based on what
the cilium/ebpf Go library does [1]. It would lower the soft rlimit to
0, attempt to load a BPF object, and reset the rlimit. But it may induce
some hard-to-debug flakiness if another process starts, or the current
application is killed, while the rlimit is reduced, and the approach was
discarded.
As a workaround to ensure that the rlimit bump does not depend on the
availability of a given helper, we restore the unconditional rlimit bump
in bpftool for now.
[0] https://lore.kernel.org/bpf/20220609143614.97837-1-quentin@isovalent.com/
[1] https://github.com/cilium/ebpf/blob/v0.9.0/rlimit/rlimit.go#L39
Signed-off-by: Quentin Monnet <quentin@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Yafang Shao <laoar.shao@gmail.com>
Cc: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/bpf/20220610112648.29695-2-quentin@isovalent.com
This change adjusts the Makefile to use "HOSTAR" as the archive tool
to keep the sanity of the build process for the bootstrap part in
check. For the rationale, please continue reading.
When cross compiling bpftool with buildroot, it leads to an invocation
like:
$ AR="/path/to/buildroot/host/bin/arc-linux-gcc-ar" \
CC="/path/to/buildroot/host/bin/arc-linux-gcc" \
...
make
Which in return fails while building the bootstrap section:
----------------------------------8<----------------------------------
make: Entering directory '/src/bpftool-v6.7.0/src'
... libbfd: [ on ]
... disassembler-four-args: [ on ]
... zlib: [ on ]
... libcap: [ OFF ]
... clang-bpf-co-re: [ on ] <-- triggers bootstrap
.
.
.
LINK /src/bpftool-v6.7.0/src/bootstrap/bpftool
/usr/bin/ld: /src/bpftool-v6.7.0/src/bootstrap/libbpf/libbpf.a:
error adding symbols: archive has no index; run ranlib
to add one
collect2: error: ld returned 1 exit status
make: *** [Makefile:211: /src/bpftool-v6.7.0/src/bootstrap/bpftool]
Error 1
make: *** Waiting for unfinished jobs....
AR /src/bpftool-v6.7.0/src/libbpf/libbpf.a
make[1]: Leaving directory '/src/bpftool-v6.7.0/libbpf/src'
make: Leaving directory '/src/bpftool-v6.7.0/src'
---------------------------------->8----------------------------------
This occurs because setting "AR" confuses the build process for the
bootstrap section and it calls "arc-linux-gcc-ar" to create and index
"libbpf.a" instead of the host "ar".
Signed-off-by: Shahab Vahedi <shahab@synopsys.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Quentin Monnet <quentin@isovalent.com>
Cc: Jean-Philippe Brucker <jean-philippe@linaro.org>
Link: https://lore.kernel.org/bpf/8d297f0c-cfd0-ef6f-3970-6dddb3d9a87a@synopsys.com
Add BTF_KIND_ENUM64 support.
For example, the following enum is defined in uapi bpf.h.
$ cat core.c
enum A {
BPF_F_INDEX_MASK = 0xffffffffULL,
BPF_F_CURRENT_CPU = BPF_F_INDEX_MASK,
BPF_F_CTXLEN_MASK = (0xfffffULL << 32),
} g;
Compiled with
clang -target bpf -O2 -g -c core.c
Using bpftool to dump types and generate format C file:
$ bpftool btf dump file core.o
...
[1] ENUM64 'A' encoding=UNSIGNED size=8 vlen=3
'BPF_F_INDEX_MASK' val=4294967295ULL
'BPF_F_CURRENT_CPU' val=4294967295ULL
'BPF_F_CTXLEN_MASK' val=4503595332403200ULL
$ bpftool btf dump file core.o format c
...
enum A {
BPF_F_INDEX_MASK = 4294967295ULL,
BPF_F_CURRENT_CPU = 4294967295ULL,
BPF_F_CTXLEN_MASK = 4503595332403200ULL,
};
...
Note that for raw btf output, the encoding (UNSIGNED or SIGNED)
is printed out as well. The 64bit value is also represented properly
in BTF and C dump.
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20220607062652.3722649-1-yhs@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
