-
Notifications
You must be signed in to change notification settings - Fork 51
/
TODO
127 lines (76 loc) · 3.3 KB
/
TODO
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
+ Add KVM tests, specifically
PERF_RECORD_MISC_GUEST_KERNEL and PERF_RECORD_MISC_GUEST_USER events.
+ Test the new cycles field in LBR entries on Sandybridge
+ Test for new CONTEXT_SWITCH records
+ In sampling, be sure to handle the various sample_header values
+ Test for PERF_FLAG_FD_CLOEXEC
+ Fix up the test for PERF_IOC_PERIOD and if it takes place right away.
+ Add test for Fixed Counter 2 (retired instructions) being different.
Basically run 5 instructions events, compare results
+ Add test for ring-buffer size of 0 or 1
(possibly causes kernel bug?)
+ Add test for IVB constraints
+ Wakeup events not equal to 1
+ single-shot RT counts
+ the bug with overflow even after event closed
PERF_FLAG_FD_OUTPUT
PERF_FLAG_FD_NO_GROUP
get perf cache-detection code going
inherit?
inherit and PERF_FORMAT_GROUP?
document:
pinned flag?
perf tests:
+ All generalized events
+ More multi-threaded test cases
+ More assembler code in the cache load/stores tests
PAPI tests:
+ Invert flag
+ Max number of SW events
From Phil:
Do you guys have a test for:
a) grouped, non-multiplexed counters (where you can read/start/stop on one FD)
b) multiplexed, grouped counters
c) the portion of perf_events that uses counter indexing (returns a [5] array).
(I'm not sure what this is useful for yet...sampling?)
Will Deacon:
If I try using perf with event groups on a mainline kernel, only the group
leader counts:
linaro@mashed-potato:~$ perf stat --no-scale -g -e task-clock -e cpu-clock -e cs
+-- ls
hwb linux
Performance counter stats for 'ls':
17.545541 task-clock # 0.634 CPUs utilized
0.000000 cpu-clock
0 cs # 0.000 M/sec
0.027684984 seconds time elapsed
This is fixed by b79387ef ("perf: Fix enable_on_exec for sibling events")
in -next (20111222) but I can't see this queued anywhere for 3.2 (I thought
I might see it in -rc7).
Is this going in for the final cut, or will groups just be borked until 3.3?
How can this be a regression? Its always been like that. At worst its a
broken new feature.
Also, you really shouldn't set enable_on_exec on non group leaders,
that's just daft. But yes that patch sorts it. Something like the below
should too.
+ test guest exclude support
attr.exclude_{guest,host}
commit a4f144ebbdf6f7807c477bce8e136047ed27321f
Author: David Carrillo-Cisneros <[email protected]>
Date: Wed Jun 1 12:33:05 2016 -0700
perf/core: Fix crash due to account/unaccount_sb_event() inconsistency
unaccount_pmu_sb_event() did not check for attributes in event->attr
before calling detach_sb_event(), while account_pmu_event() did.
This caused NULL pointer reference in cgroup events that did not
have any of the attributes checked by account_pmu_event().
To trigger the bug just wait for a cgroup event to terminate, e.g.:
$ mkdir /dev/cgroup/devices/test
$ perf stat -e cycles -a -G test sleep 0
... see crash ...
Date: Wed, 14 Sep 2016 22:38:45 -0700
From: Stephane Eranian
Subject: slow perf ioctl()
Here is another thing I've run into lately having to do with
ioctl(fd, PERF_IOC_REFRESH, 1) on self-monitoring thread.
The ioctl() takes seconds to execute. I have not yet had the time
to track down why. Any of your tests exhibiting this problem?