Merge branch 'bpf-perf'
Kaixu Xia says:
====================
bpf: Introduce the new ability of eBPF programs to access hardware PMU counter
This patchset is base on the net-next:
git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git
commit 9dc20a649609c95ce7c5ac4282656ba627b67d49.
Previous patch v6 url:
https://lkml.org/lkml/2015/8/4/188
changes in V7:
- rebase the whole patch set to net-next tree(9dc20a64);
- split out the core perf APIs into Patch 1/5;
- change the return value of function perf_event_attrs()
from struct perf_event * to const struct perf_event * in
Patch 1/5;
- rename the function perf_event_read_internal() to perf_event_
read_local() and rewrite it in Patch 1/5;
- rename the function check_func_limit() to check_map_func
_compatibility() and remove the unnecessary pass pointer to
a pointer in Patch 4/5;
changes in V6:
- make the Patch 1/4 commit message more meaning and readable;
- remove the unnecessary comment in Patch 2/4 and make it clean;
- declare the function perf_event_release_kernel() in include/
linux/perf_event.h to fix the build error when CONFIG_PERF_EVENTS
isn't configured in Patch 2/4;
- add function perf_event_attrs() to get the struct perf_event_attr
in Patch 2/4.
- move the related code from kernel/trace/bpf_trace.c to kernel/
events/core.c and add function perf_event_read_internal() to
avoid poking inside of the event outside of perf code in Patch 3/4;
- generial the func & map match-pair with an array in Patch 3/4;
changes in V5:
- move struct fd_array_map_ops* fd_ops to bpf_map;
- move array perf event decrement refcnt function to
map_free;
- fix the NULL ptr of perf_event_get();
- move bpf_perf_event_read() to kernel/bpf/bpf_trace.c;
- get rid of the remaining struct bpf_prog;
- move the unnecessay cast on void *;
changes in V4:
- make the bpf_prog_array_map more generic;
- fix the bug of event refcnt leak;
- use more useful errno in bpf_perf_event_read();
changes in V3:
- collapse V2 patches 1-3 into one;
- drop the function map->ops->map_traverse_elem() and release
the struct perf_event in map_free;
- only allow to access bpf_perf_event_read() from programs;
- update the perf_event_array_map elem via xchg();
- pass index directly to bpf_perf_event_read() instead of
MAP_KEY;
changes in V2:
- put atomic_long_inc_not_zero() between fdget() and fdput();
- limit the event type to PERF_TYPE_RAW and PERF_TYPE_HARDWARE;
- Only read the event counter on current CPU or on current
process;
- add new map type BPF_MAP_TYPE_PERF_EVENT_ARRAY to store the
pointer to the struct perf_event;
- according to the perf_event_map_fd and key, the function
bpf_perf_event_read() can get the Hardware PMU counter value;
Patch 5/5 is a simple example and shows how to use this new eBPF
programs ability. The PMU counter data can be found in
/sys/kernel/debug/tracing/trace(trace_pipe).(the cycles PMU
value when 'kprobe/sys_write' sampling)
$ cat /sys/kernel/debug/tracing/trace_pipe
$ ./tracex6
...
syslog-ng-548 [000] d..1 76.905673: : CPU-0 681765271
syslog-ng-548 [000] d..1 76.905690: : CPU-0 681787855
syslog-ng-548 [000] d..1 76.905707: : CPU-0 681810504
syslog-ng-548 [000] d..1 76.905725: : CPU-0 681834771
syslog-ng-548 [000] d..1 76.905745: : CPU-0 681859519
syslog-ng-548 [000] d..1 76.905766: : CPU-0 681890419
syslog-ng-548 [000] d..1 76.905783: : CPU-0 681914045
syslog-ng-548 [000] d..1 76.905800: : CPU-0 681935950
syslog-ng-548 [000] d..1 76.905816: : CPU-0 681958299
ls-690 [005] d..1 82.241308: : CPU-5 3138451
sh-691 [004] d..1 82.244570: : CPU-4 7324988
<...>-699 [007] d..1 99.961387: : CPU-7 3194027
<...>-695 [003] d..1 99.961474: : CPU-3 288901
<...>-695 [003] d..1 99.961541: : CPU-3 383145
<...>-695 [003] d..1 99.961591: : CPU-3 450365
<...>-695 [003] d..1 99.961639: : CPU-3 515751
<...>-695 [003] d..1 99.961686: : CPU-3 579047
...
The detail of patches is as follow:
Patch 1/5 add the necessary core perf APIs perf_event_attrs(),
perf_event_get(),perf_event_read_local() when accessing events
counters in eBPF programs
Patch 2/5 rewrites part of the bpf_prog_array map code and make it
more generic;
Patch 3/5 introduces a new bpf map type. This map only stores the
pointer to struct perf_event;
Patch 4/5 implements function bpf_perf_event_read() that get the
selected hardware PMU conuter;
Patch 5/5 gives a simple example.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>