Anton noted that for inherited counters the counter-id as provided by
PERF_SAMPLE_ID isn't mappable to the id found through PERF_RECORD_ID
because each inherited counter gets its own id.
His suggestion was to always return the parent counter id, since that
is the primary counter id as exposed. However, these inherited
counters have a unique identifier so that events like
PERF_EVENT_PERIOD and PERF_EVENT_THROTTLE can be specific about which
counter gets modified, which is important when trying to normalize the
sample streams.
This patch removes PERF_EVENT_PERIOD in favour of PERF_SAMPLE_PERIOD,
which is more useful anyway, since changing periods became a lot more
common than initially thought -- rendering PERF_EVENT_PERIOD the less
useful solution (also, PERF_SAMPLE_PERIOD reports the more accurate
value, since it reports the value used to trigger the overflow,
whereas PERF_EVENT_PERIOD simply reports the requested period changed,
which might only take effect on the next cycle).
This still leaves us PERF_EVENT_THROTTLE to consider, but since that
_should_ be a rare occurrence, and linking it to a primary id is the
most useful bit to diagnose the problem, we introduce a
PERF_SAMPLE_STREAM_ID, for those few cases where the full
reconstruction is important.
[Does change the ABI a little, but I see no other way out]
Suggested-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1248095846.15751.8781.camel@twins>
the "reserved" field was not initialized to zero, resulting in 4 bytes
of stack data leaking to userspace....
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
perf record uses -g for logging call graph data but perf report
uses -c to print call graph data. Be consistent and use -g
everywhere for call graph data.
Also update the help text to reflect the current default -
fractal,0.5
Signed-off-by: Anton Blanchard <anton@samba.org>
Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090716104817.803604373@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Right now we don't output vfork events. Even though we should
always see an exec after a vfork, we may get perfcounter
samples between the vfork and exec. These samples can lead to
some confusion when parsing perfcounter data.
To keep things consistent we should always log a fork event. It
will result in a little more log data, but is less confusing to
trace parsing tools.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090716104817.589309391@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
perf record synthesizes mmap events for the running process.
Right now it just catches file mappings, but we can check for
the vdso symbol and add that too.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090716104817.517264409@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There are a few places we are leaking tiny amounts of kernel
memory to userspace. This happens when writing out strings
because we always align the end to 64 bits.
To avoid this we should always use an appropriately sized
temporary buffer and ensure it is zeroed.
Since d_path assembles the string from the end of the buffer
backwards, we need to add 64 bits after the buffer to allow for
alignment.
We also need to copy arch_vma_name to the temporary buffer,
because if we use it directly we may end up copying to
userspace a number of bytes after the end of the string
constant.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090716104817.273972048@samba.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Fix a missed rename in EVENT_PROFILE support so that it gets
built and allows tracepoint tracing from the 'perf' tool.
Fix a typo in the (never before built & enabled) portion in
perf_counter.c as well, and update that code to the
attr.config changes as well.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ben Gamari <bgamari.foss@gmail.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <1246869094-21237-1-git-send-email-chris@chris-wilson.co.uk>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
I've attached a patch to remove the Pentium M special casing of
EMON and as noticed at least with my Pentium M the hardware PMU
now works:
Performance counter stats for '/bin/ls /var/tmp':
1.809988 task-clock-msecs # 0.125 CPUs
1 context-switches # 0.001 M/sec
0 CPU-migrations # 0.000 M/sec
224 page-faults # 0.124 M/sec
1425648 cycles # 787.656 M/sec
912755 instructions # 0.640 IPC
Vince suggested that this code was trying to address erratum
Y17 in Pentium-M's:
http://download.intel.com/support/processors/mobile/pm/sb/25266532.pdf
But that erratum (related to IA32_MISC_ENABLES.7) does not
affect perfcounters as we dont use this toggle to disable RDPMC
and WRMSR/RDMSR access to performance counters. We keep cr4's
bit 8 (X86_CR4_PCE) clear so unprivileged RDPMC access is not
allowed anyway.
Cc: Vince Weaver <vince@deater.net>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@googlemail.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When we filter by column content we may end up with a column
that has the same value for all the lines. So remove that
column and tell its unique value on the top, as a comment.
Example:
[acme@doppio pahole]$ perf report --sort comm,dso,symbol -d ./build/libdwarves.so.1.0.0 -C pahole | head -15
# dso: ./build/libdwarves.so.1.0.0
# comm: pahole
# Samples: 58409
#
# Overhead Symbol
# ........ ......
#
20.93% [.] tag__recode_dwarf_type
14.94% [.] namespace__recode_dwarf_types
10.38% [.] cu__table_add_tag
6.69% [.] __die__process_tag
5.05% [.] die__process_function
4.70% [.] list__for_all_tags
3.68% [.] tag__init
3.48% [.] die__create_new_parameter
[acme@doppio pahole]$
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1247325517-12272-3-git-send-email-acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The strlist__entry method allows accessing strlists like an
array, will be used in the 'perf report' to access the first
entry.
We now keep the nr_entries so that we can check if we have just
one entry, will be used in 'perf report' to improve the output
by showing just at the top when we have just, say, one DSO.
While at it use nr_entries to optimize strlist__is_empty by not
using the far more costly rb_first based implementation.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1247325517-12272-2-git-send-email-acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Always printing the level info about if it is in the kernel,
hypervisor or userspace as that is in the hist_entry.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1247325517-12272-1-git-send-email-acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Auto-adjust column width of perf report output to the
longest occuring string length.
Example:
[acme@doppio pahole]$ perf report --sort comm,dso,symbol | head -13
12.79% pahole /usr/lib64/libdw-0.141.so [.] __libdw_find_attr
8.90% pahole /lib64/libc-2.10.1.so [.] _int_malloc
8.68% pahole /usr/lib64/libdw-0.141.so [.] __libdw_form_val_len
8.15% pahole /lib64/libc-2.10.1.so [.] __GI_strcmp
6.80% pahole /lib64/libc-2.10.1.so [.] __tsearch
5.54% pahole ./build/libdwarves.so.1.0.0 [.] tag__recode_dwarf_type
[acme@doppio pahole]$
[acme@doppio pahole]$ perf report --sort comm,dso,symbol -d /lib64/libc-2.10.1.so | head -10
21.92% pahole /lib64/libc-2.10.1.so [.] _int_malloc
20.08% pahole /lib64/libc-2.10.1.so [.] __GI_strcmp
16.75% pahole /lib64/libc-2.10.1.so [.] __tsearch
[acme@doppio pahole]$
Also add these extra options to control the new behaviour:
-w, --field-width
Force each column width to the provided list, for large terminal
readability.
-t, --field-separator:
Use a special separator character and don't pad with spaces, replacing
all occurances of this separator in symbol names (and other output) with
a '.' character, that thus it's the only non valid separator.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <20090711014728.GH3452@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Instead of open coding the unclone context thingy, put it in
a common function.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ingo noticed that both AMD and P6 call
x86_pmu_disable_counter() on *_pmu_enable_counter(). This is
because we rely on the side effect of that call to program
the event config but not touch the EN bit.
We change that for AMD by having enable_all() simply write
the full config in, and for P6 by explicitly coding it.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The P6 doesn't seem to support cache ref/hit/miss counts, so
we extend the generic hardware event codes to have 0 and -1
mean the same thing as for the generic cache events.
Furthermore, it turns out the 0 event does not count
(that is, its reported that on PPro it actually does count
something), therefore use a event configuration that's
specified not to count to disable the counters.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add basic P6 PMU support. The P6 uses the EVNTSEL0 EN bit to
enable/disable both its counters. We use this for the
global enable/disable, and clear all config bits (except EN)
to disable individual counters.
Actual ia32 hardware doesn't support lfence, so use a locked
op without side-effect to implement a full barrier.
perf stat and perf record seem to function correctly.
[a.p.zijlstra@chello.nl: cleanups and complete the enable/disable code]
Signed-off-by: Vince Weaver <vince@deater.net>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <Pine.LNX.4.64.0907081718450.2715@pianoman.cluster.toy>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The cache events contain '$' which will hit shell variable
expansion. To avoid confusion change this to 'cache', ie
L1-d$-loads becomes L1-dcache-loads.
Signed-off-by: Anton Blanchard <anton@samba.org>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: Jaswinder Singh Rajput <jaswinder@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
LKML-Reference: <20090706120131.GB4391@kryten>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current callchain displays the overhead rates as absolute:
relative to the total overhead.
This patch provides relative overhead percentage, in which each
branch of the callchain tree is a independant instrumentated object.
This provides a 'fractal' view of the call-chain profile: each
sub-graph looks like a profile in itself - relative to its parent.
You can produce such output by using the "fractal" mode
that you can abbreviate via f, fr, fra, frac, etc...
./perf report -s sym -c fractal
Example:
8.46% [k] copy_user_generic_string
|
|--52.01%-- generic_file_aio_read
| do_sync_read
| vfs_read
| |
| |--97.20%-- sys_pread64
| | system_call_fastpath
| | pread64
| |
| --2.81%-- sys_read
| system_call_fastpath
| __read
|
|--39.85%-- generic_file_buffered_write
| __generic_file_aio_write_nolock
| generic_file_aio_write
| do_sync_write
| reiserfs_file_write
| vfs_write
| |
| |--97.05%-- sys_pwrite64
| | system_call_fastpath
| | __pwrite64
| |
| --2.95%-- sys_write
| system_call_fastpath
| __write_nocancel
[...]
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246772361-9960-5-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The cumul hits are the number of hits of every childs of a node
plus the hits of the current nodes, required for percentage
computing of a branch.
Theses numbers are calculated during the sorting of the branches of
the callchain tree using a depth first postfix traversal, so that
cumulative hits are propagated in the right order.
But if we plan to implement percentages relative to the parent and not
absolute percentages (relative to the whole overhead), we need to know
the cumulative hits of the parent before computing the children
because the relative minimum acceptable number of entries (ie: minimum
rate against the cumulative hits from the parent) is the basis to
filter the children against a given rate.
Then we need to handle the cumul hits on the fly to prepare the
implementation of relative overhead rates.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246772361-9960-4-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The default callchain parameters are set to use the flat mode and never
filter any overhead threshold of backtrace.
But flat mode is boring compared to graph mode.
Also the number of callchains may be very high if none is
filtered.
Let's change this to set the graph view and a minimum overhead of 0.5%
as default parameters.
Reported-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246772361-9960-3-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
If the user doesn't provide options to tune his callchain output
(ie: if he uses -c without arguments) then the default value passed
in the OPT_CALLBACK_DEFAULT() macro is used.
But it's parsed later by strtok() which will replace comma separators
to a zero. This may segfault as we are using a read-only string.
Use a modifiable one instead, and also fix the "100%" default
minimum threshold value by turning it into a 0 (output every callchains)
as it was intended in the origin.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246772361-9960-2-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
perf report segfaults while trying to handle callchains from a non
callchain data file.
Instead of a segfault, print a useful message to the user.
Reported-by: Jens Axboe <jens.axboe@oracle.com>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246772361-9960-1-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Now atomic64_read() is light weight (no register pressure and
small icache), we can inline it again.
Also use "=&A" constraint instead of "+A" to avoid warning
about unitialized 'res' variable. (gcc had to force 0 in eax/edx)
$ size vmlinux.prev vmlinux.after
text data bss dec hex filename
4908667 451676 1684868 7045211 6b805b vmlinux.prev
4908651 451676 1684868 7045195 6b804b vmlinux.after
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <4A4E1AA2.30002@gmail.com>
[ Also fix typo in atomic64_set() export ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noticed that the variable name 'old_val' is
confusingly named in these functions - the correct
naming is 'new_val'.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907030942260.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Remove the read-first logic from atomic64_xchg() and simplify
the loop.
This function was the last user of __atomic64_read() - remove it.
Also, change the 'real_val' assumption from the somewhat quirky
1ULL << 32 value to the (just as arbitrary, but simpler) value
of 0.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <tip-05118ab8859492ac9ddda0154cf90e37b0a4a0b0@git.kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
atomic64_t primitives are used by a handful of drivers,
so export the APIs consistently. These were inlined
before.
Also mark atomic64_32.o a core object, so that the symbols
are available even if not linked to core kernel pieces.
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <tip-05118ab8859492ac9ddda0154cf90e37b0a4a0b0@git.kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Occasionally we get bugs where atomic_read or atomic_set are
used on atomic64_t variables or vice versa. These bugs don't
generate warnings on x86 because atomic_read and atomic_set are
coded as macros rather than C functions, so we don't get any
type-checking on their arguments; similarly for atomic64_read
and atomic64_set in 64-bit kernels.
This converts them to C functions so that the arguments are
type-checked and bugs like this will get caught more easily. It
also converts atomic_cmpxchg and atomic_xchg, and
atomic64_cmpxchg and atomic64_xchg on 64-bit, so we get
type-checking on their arguments too.
Compiling a typical 64-bit x86 config, this generates no new
warnings, and the vmlinux text is 86 bytes smaller.
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noticed that atomic64_xchg() uses atomic_read(), which
happens to work because atomic_read() is a macro so the
.counter value gets u64-read on 32-bit too - but this is really
bogus and serious bugs are waiting to happen.
Fix atomic64_xchg() to use __atomic64_read() instead.
No code changed:
arch/x86/lib/atomic64_32.o:
text data bss dec hex filename
435 0 0 435 1b3 atomic64_32.o.before
435 0 0 435 1b3 atomic64_32.o.after
md5:
bd8ab95e69c93518578bfaf0ea3be4d9 atomic64_32.o.before.asm
bd8ab95e69c93518578bfaf0ea3be4d9 atomic64_32.o.after.asm
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noticed that atomic64_xchg() uses atomic_read(), which
happens to work because atomic_read() is a macro so the
.counter value gets u64-read on 32-bit too - but this is really
bogus and serious bugs are waiting to happen.
Change atomic_read() to be a type-safe inline, and this exposes
the atomic64 bogosity as well:
arch/x86/lib/atomic64_32.c: In function ‘atomic64_xchg’:
arch/x86/lib/atomic64_32.c:39: warning: passing argument 1 of ‘atomic_read’ from incompatible pointer type
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
cmpxchg8b is a huge instruction in terms of register footprint,
we almost never want to inline it, not even within the same
code module.
GCC 4.3 still messes up for two functions, under-judging the
true cost of this instruction - so annotate two key functions
to reduce the bloat:
arch/x86/lib/atomic64_32.o:
text data bss dec hex filename
1763 0 0 1763 6e3 atomic64_32.o.before
435 0 0 435 1b3 atomic64_32.o.after
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noted (based on Eric Dumazet's numbers) that we would
probably be better off not trying an atomic_read() in
atomic64_add_return() but intead intentionally let the first
cmpxchg8b fail - to get a cache-friendly 'give me ownership
of this cacheline' transaction. That can then be followed
by the real cmpxchg8b which sets the value local to the CPU.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Rewrite cmpxchg8b() to not use %edi register but a generic "+m"
constraint, to increase compiler freedom in code generation and
possibly better code.
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noticed that the 32-bit version of atomic64_read() was
being overly complex with re-reading the value and doing a
retry loop over that.
Instead we can just rely on cmpxchg8b returning either the new
value or returning the current value.
We can use any 'old' value, which will be faster as it can be
loaded via immediates. Using some value that is not equal to
the real value in memory the instruction gets faster.
This also has the advantage that the CPU could avoid dirtying
the cacheline.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Linus noted that the atomic64_t primitives are all inlines
currently which is crazy because these functions have a large
register footprint anyway.
Move them to a separate file: arch/x86/lib/atomic64_32.c
Also, while at it, rename all uses of 'unsigned long long' to
the much shorter u64.
This makes the appearance of the prototypes a lot nicer - and
it also uncovered a few bugs where (yet unused) API variants
had 'long' as their return type instead of u64.
[ More intrusive changes are not yet done in this patch. ]
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Locked instructions on two cache lines at once are painful. If
atomic64_t uses two cache lines, my test program is 10x slower.
The chance for that is significant: 4/32 or 12.5%.
Make sure an atomic64_t is 8 bytes aligned.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
LKML-Reference: <alpine.LFD.2.01.0907021653030.3210@localhost.localdomain>
[ changed it to __aligned(8) as per Andrew's suggestion ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Certain versions of GCC dont see the initialization that is done here:
builtin-report.c: In function ‘__cmd_report’:
builtin-report.c:1038: warning: ‘syms’ may be used uninitialized in this function
So annotate it with a NULL initialization.
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ingo Molnar wrote:
> i just bisected a 'perf report' bug that would cause us to not
> resolve all user-space symbols in a 'git gc' run to:
>
> f5812a7a33 is first bad commit
> commit f5812a7a33
> Author: Arnaldo Carvalho de Melo <acme@redhat.com>
> Date: Tue Jun 30 11:43:17 2009 -0300
>
> perf_counter tools: Adjust only prelinked symbol's addresses
Rename ->prelinked to ->adjust_symbols and making what was done
only for prelinked libraries also to ET_EXEC binaries, such as
/usr/bin/git:
[acme@doppio pahole]$ readelf -h /usr/bin/git | grep Type
Type: EXEC (Executable file)
[acme@doppio pahole]$
And after installing the 'git-debuginfo' package, I get correct results:
[acme@doppio linux-2.6-tip]$ perf report --sort comm,dso,symbol -d /usr/bin/git | head -20
#
# (1139614 samples)
#
# Overhead Command Shared Object Symbol
# ........ ................ ......................... ......
#
34.98% git /usr/bin/git [.] send_sideband
33.39% git /usr/bin/git [.] enter_repo
6.81% git /usr/bin/git [.] diff_opt_parse
4.95% git /usr/bin/git [.] is_repository_shallow
3.24% git /usr/bin/git [.] odb_mkstemp
1.39% git /usr/bin/git [.] output
1.34% git /usr/bin/git [.] xmmap
1.25% git /usr/bin/git [.] receive_pack_config
1.16% git /usr/bin/git [.] git_pathdup
0.90% git /usr/bin/git [.] read_object_with_reference
0.86% git /usr/bin/git [.] show_patch_diff
0.85% git /usr/bin/git 0x00000000095e2e
0.69% git /usr/bin/git [.] display
[acme@doppio linux-2.6-tip]$
I'll check what are the last cases where we can't resolve symbols, like
this 0x00000000095e2e later.
And I guess this will fix the problems Mike were seeing too:
[acme@doppio linux-2.6-tip]$ readelf -h ../build/perf/vmlinux | grep Type
Type: EXEC (Executable file)
[acme@doppio linux-2.6-tip]$
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This adds the use of colors to signal at a glance the important
overhead thresholds in callchains hit rates.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246558475-10624-3-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Among perf annotate, perf report and perf top, we can find the
common colored printing of percents according to the following
rules:
High overhead = > 5%, colored in red
Mid overhead = > 0.5%, colored in green
Low overhead = < 0.5%, default color
Factorize these multiple checks in a single function named
percent_color_fprintf() and also provide a get_percent_color()
for sites which print percentages and other things at the same
time.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246558475-10624-2-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Callchains output may become a burden on a trace because even
rarely hit site are exposed. This can be too much information.
Let the user set a threshold as a minimum percent of hits using
the new pattern for the -c option:
-c mode,min_percent
Example:
$ perf report -s sym -c flat,4
8.25% [k] copy_user_generic_string
4.19%
copy_user_generic_string
generic_file_aio_read
do_sync_read
vfs_read
sys_pread64
system_call_fastpath
pread64
5.39% [k] search_by_key
4.63% 0x00000000009e0a
2.36% [k] memcpy_c
[...]
$ perf report -s sym -c graph,2
8.25% [k] copy_user_generic_string
|
|--4.31%-- generic_file_aio_read
| do_sync_read
| vfs_read
| |
| --4.19%-- sys_pread64
| system_call_fastpath
| pread64
|
--3.24%-- generic_file_buffered_write
__generic_file_aio_write_nolock
generic_file_aio_write
do_sync_write
reiserfs_file_write
vfs_write
|
--3.14%-- sys_pwrite64
system_call_fastpath
__pwrite64
5.39% [k] search_by_key
|
--2.23%-- reiserfs_update_sd_size
4.63% 0x00000000009e0a
2.36% [k] memcpy_c
[...]
You can also omit it and it will default to 0.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246558475-10624-1-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently, the printing of callchains is done in a single
vertical level, this is the "flat" mode:
8.25% [k] copy_user_generic_string
4.19%
copy_user_generic_string
generic_file_aio_read
do_sync_read
vfs_read
sys_pread64
system_call_fastpath
pread64
This patch introduces a new "graph" mode which provides a
hierarchical output of factorized paths recursively sorted:
8.25% [k] copy_user_generic_string
|
|--4.31%-- generic_file_aio_read
| do_sync_read
| vfs_read
| |
| |--4.19%-- sys_pread64
| | system_call_fastpath
| | pread64
| |
| --0.12%-- sys_read
| system_call_fastpath
| __read
|
|--3.24%-- generic_file_buffered_write
| __generic_file_aio_write_nolock
| generic_file_aio_write
| do_sync_write
| reiserfs_file_write
| vfs_write
| |
| |--3.14%-- sys_pwrite64
| | system_call_fastpath
| | __pwrite64
| |
| --0.10%-- sys_write
[...]
The command line has then changed.
By providing the -c option, the callchain will output in the
flat mode by default.
But you can override it:
perf report -c graph
or
perf report -c flat
You can also pass the abreviated mode:
perf report -c g
or
perf report -c gra
will both make use of the graph mode.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246550301-8954-3-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There is no predefined macro to create an option that can have
a custom value or a default one if none is given.
This patch provides a new helper OPT_CALLBACK_DEFAULT() which
defines such kind of option.
For example, considering an option -c, we want to get the
default value in the following cases:
perf command -c -d
perf command -d -c
And the foo value when it's given:
perf command -c foo -d
perf command -d -c foo
That's also why PARSE_OPT_LASTARG_DEFAULT is extended here to
support default values whatever the position of the option, not
only in the end.
Should it now be renamed to PARSE_OPT_ARG_DEFAULT ?
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: git@vger.kernel.org
LKML-Reference: <1246550301-8954-2-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
