This patch is brought to you by the letter 'H'.
Commit 20b279 breaks compatiblity with older perf binaries when run with
precise modifier (:p or :pp) by requiring the exclude_guest attribute to be
set. Older binaries default exclude_guest to 0 (ie., wanting guest-based
samples) unless host only profiling is requested (:H modifier). The workaround
for older binaries is to add H to the modifier list (e.g., -e cycles:ppH -
toggles exclude_guest to 1). This was deemed unacceptable by Linus:
https://lkml.org/lkml/2012/12/12/570
Between family in town and the fresh snow in Breckenridge there is no time left
to be working on the proper fix for this over the holidays. In the New Year I
have more pressing problems to resolve -- like some memory leaks in perf which
are proving to be elusive -- although the aforementioned snow is probably why
they are proving to be elusive. Either way I do not have any spare time to work
on this and from the time I have managed to spend on it the solution is more
difficult than just moving to a new exclude_guest flag (does not work) or
flipping the logic to include_guest (which is not as trivial as one would
think).
So, two options: silently force exclude_guest on as suggested by Gleb which
means no impact to older perf binaries or revert the original patch which
caused the breakage.
This patch does the latter -- reverts the original patch that introduced the
regression. The problem can be revisited in the future as time allows.
Signed-off-by: David Ahern <dsahern@gmail.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Richter <robert.richter@amd.com>
Link: http://lkml.kernel.org/r/1356749767-17322-1-git-send-email-dsahern@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support to display hardware events translations available
through the sysfs. Add 'events' group attribute under the sysfs
x86 PMU record with attribute/file for each hardware event.
This patch adds only backbone for PMUs to display config under
'events' directory. The specific PMU support itself will come
in next patches, however this is how the sysfs group will look
like:
# ls /sys/devices/cpu/events/
branch-instructions
branch-misses
bus-cycles
cache-misses
cache-references
cpu-cycles
instructions
ref-cycles
stalled-cycles-backend
stalled-cycles-frontend
The file - hw event ID mapping is:
file hw event ID
---------------------------------------------------------------
cpu-cycles PERF_COUNT_HW_CPU_CYCLES
instructions PERF_COUNT_HW_INSTRUCTIONS
cache-references PERF_COUNT_HW_CACHE_REFERENCES
cache-misses PERF_COUNT_HW_CACHE_MISSES
branch-instructions PERF_COUNT_HW_BRANCH_INSTRUCTIONS
branch-misses PERF_COUNT_HW_BRANCH_MISSES
bus-cycles PERF_COUNT_HW_BUS_CYCLES
stalled-cycles-frontend PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
stalled-cycles-backend PERF_COUNT_HW_STALLED_CYCLES_BACKEND
ref-cycles PERF_COUNT_HW_REF_CPU_CYCLES
Each file in the 'events' directory contains the term translation
for the symbolic hw event for the currently running cpu model.
# cat /sys/devices/cpu/events/stalled-cycles-backend
event=0xb1,umask=0x01,inv,cmask=0x01
Suggested-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1349873598-12583-2-git-send-email-jolsa@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In check_hw_exists() we try to detect non-emulated MSR accesses
by writing an arbitrary value into one of the PMU registers
and check if it's value after a readout is still the same.
This algorithm silently assumes that the register does not contain
the magic value already, which is wrong in at least one situation.
Fix the algorithm to really do a read-modify-write cycle. This fixes
a warning under Xen under some circumstances on AMD family 10h CPUs.
The reasons in more details actually sound like a story from
Believe It or Not!:
First you need an AMD family 10h/12h CPU. These do not reset the
PERF_CTR registers on a reboot.
Now you boot bare metal Linux, which goes successfully through this
check, but leaves the magic value of 0xabcd in the register. You
don't use the performance counters, but do a reboot (warm reset).
Then you choose to boot Xen. The check will be triggered with a
recent Linux kernel as Dom0 again, trying to write 0xabcd into the
MSR. Xen silently drops the write (expected), but the subsequent read
will return the value in the register, which just happens to be the
expected magic value. Thus the test misleadingly succeeds, leaving
the kernel in the belief that the PMU is available. This will trigger
the following message:
[ 0.020294] ------------[ cut here ]------------
[ 0.020311] WARNING: at arch/x86/xen/enlighten.c:730 xen_apic_write+0x15/0x17()
[ 0.020318] Hardware name: empty
[ 0.020323] Modules linked in:
[ 0.020334] Pid: 1, comm: swapper/0 Not tainted 3.3.8 #7
[ 0.020340] Call Trace:
[ 0.020354] [<ffffffff81050379>] warn_slowpath_common+0x80/0x98
[ 0.020369] [<ffffffff810503a6>] warn_slowpath_null+0x15/0x17
[ 0.020378] [<ffffffff810034df>] xen_apic_write+0x15/0x17
[ 0.020392] [<ffffffff8101cb2b>] perf_events_lapic_init+0x2e/0x30
[ 0.020410] [<ffffffff81ee4dd0>] init_hw_perf_events+0x250/0x407
[ 0.020419] [<ffffffff81ee4b80>] ? check_bugs+0x2d/0x2d
[ 0.020430] [<ffffffff81002181>] do_one_initcall+0x7a/0x131
[ 0.020444] [<ffffffff81edbbf9>] kernel_init+0x91/0x15d
[ 0.020456] [<ffffffff817caaa4>] kernel_thread_helper+0x4/0x10
[ 0.020471] [<ffffffff817c347c>] ? retint_restore_args+0x5/0x6
[ 0.020481] [<ffffffff817caaa0>] ? gs_change+0x13/0x13
[ 0.020500] ---[ end trace a7919e7f17c0a725 ]---
The new code will change every of the 16 low bits read from the
register and tries to write and read-back that modified number
from the MSR.
Signed-off-by: Andre Przywara <andre.przywara@amd.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Avi Kivity <avi@redhat.com>
Link: http://lkml.kernel.org/r/1349797115-28346-2-git-send-email-andre.przywara@amd.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Some PMUs don't provide a full register set for their sample,
specifically 'advanced' PMUs like AMD IBS and Intel PEBS which provide
'better' than regular interrupt accuracy.
In this case we use the interrupt regs as basis and over-write some
fields (typically IP) with different information.
The perf core however uses user_mode() to distinguish user/kernel
samples, user_mode() relies on regs->cs. If the interrupt skid pushed
us over a boundary the new IP might not be in the same domain as the
interrupt.
Commit ce5c1fe9a9 ("perf/x86: Fix USER/KERNEL tagging of samples")
tried to fix this by making the perf core use kernel_ip(). This
however is wrong (TM), as pointed out by Linus, since it doesn't allow
for VM86 and non-zero based segments in IA32 mode.
Therefore, provide a new helper to set the regs->ip field,
set_linear_ip(), which massages the regs into a suitable state
assuming the provided IP is in fact a linear address.
Also modify perf_instruction_pointer() and perf_callchain_user() to
deal with segments base offsets.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1341910954.3462.102.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
There are macros that are Intel specific and not x86 generic. Rename
them into INTEL_*.
This patch removes X86_PMC_IDX_GENERIC and does:
$ sed -i -e 's/X86_PMC_MAX_/INTEL_PMC_MAX_/g' \
arch/x86/include/asm/kvm_host.h \
arch/x86/include/asm/perf_event.h \
arch/x86/kernel/cpu/perf_event.c \
arch/x86/kernel/cpu/perf_event_p4.c \
arch/x86/kvm/pmu.c
$ sed -i -e 's/X86_PMC_IDX_FIXED/INTEL_PMC_IDX_FIXED/g' \
arch/x86/include/asm/perf_event.h \
arch/x86/kernel/cpu/perf_event.c \
arch/x86/kernel/cpu/perf_event_intel.c \
arch/x86/kernel/cpu/perf_event_intel_ds.c \
arch/x86/kvm/pmu.c
$ sed -i -e 's/X86_PMC_MSK_/INTEL_PMC_MSK_/g' \
arch/x86/include/asm/perf_event.h \
arch/x86/kernel/cpu/perf_event.c
Signed-off-by: Robert Richter <robert.richter@amd.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1340217996-2254-2-git-send-email-robert.richter@amd.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The RDPMC index calculation is wrong for AMD family 15h
(X86_FEATURE_ PERFCTR_CORE set). This leads to a #GP when
accessing the counter:
Pid: 2237, comm: syslog-ng Not tainted 3.5.0-rc1-perf-x86_64-standard-g130ff90 #135 AMD Pike/Pike
RIP: 0010:[<ffffffff8100dc33>] [<ffffffff8100dc33>] x86_perf_event_update+0x27/0x66
While the msr address offset is (index << 1) we must use index to
select the correct rdpmc.
Signed-off-by: Robert Richter <robert.richter@amd.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vince Weaver <vweaver1@eecs.utk.edu>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Rename checking_wrmsrl() to wrmsrl_safe(), to match the naming
convention used by all the other MSR access functions/macros.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
The rdpmc instruction is faster than the equivelant rdmsr call,
so use it when possible in the kernel.
The perfctr kernel patches did this, after extensive testing showed
rdpmc to always be faster (One can look in etc/costs in the perfctr-2.6
package to see a historical list of the overhead).
I have done some tests on a 3.2 kernel, the kernel module I used
was included in the first posting of this patch:
rdmsr rdpmc
Core2 T9900: 203.9 cycles 30.9 cycles
AMD fam0fh: 56.2 cycles 9.8 cycles
Atom 6/28/2: 129.7 cycles 50.6 cycles
The speedup of using rdpmc is large.
[ It's probably possible (and desirable) to do this without
requiring a new field in the hw_perf_event structure, but
the fixed events make this tricky. ]
Signed-off-by: Vince Weaver <vweaver1@eecs.utk.edu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/alpine.DEB.2.00.1203011724030.26934@cl320.eecs.utk.edu
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Zheng Yan reported that event group validation can wreck event state
when Intel extra_reg allocation changes event state.
Validation shouldn't change any persistent state. Cloning events in
validate_{event,group}() isn't really pretty either, so add a few
special cases to avoid modifying the event state.
The code is restructured to minimize the special case impact.
Reported-by: Zheng Yan <zheng.z.yan@linux.intel.com>
Acked-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1338903031.28282.175.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We always need to pass the last sample period to
perf_sample_data_init(), otherwise the event distribution will be
wrong. Thus, modifiyng the function interface with the required period
as argument. So basically a pattern like this:
perf_sample_data_init(&data, ~0ULL);
data.period = event->hw.last_period;
will now be like that:
perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
Avoids unininitialized data.period and simplifies code.
Signed-off-by: Robert Richter <robert.richter@amd.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1333390758-10893-3-git-send-email-robert.richter@amd.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf updates and fixes from Ingo Molnar:
"It's mostly fixes, but there's also two late items:
- preliminary GTK GUI support for perf report
- PMU raw event format descriptors in sysfs, to be parsed by tooling
The raw event format in sysfs is a new ABI. For example for the 'CPU'
PMU we have:
aldebaran:~> ll /sys/bus/event_source/devices/cpu/format/*
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/any
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/cmask
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/edge
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/event
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/inv
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/offcore_rsp
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/pc
-r--r--r--. 1 root root 4096 Mar 31 10:29 /sys/bus/event_source/devices/cpu/format/umask
those lists of fields contain a specific format:
aldebaran:~> cat /sys/bus/event_source/devices/cpu/format/offcore_rsp
config1:0-63
So, those who wish to specify raw events can now use the following
event format:
-e cpu/cmask=1,event=2,umask=3
Most people will not want to specify any events (let alone raw
events), they'll just use whatever default event the tools use.
But for more obscure PMU events that have no cross-architecture
generic events the above syntax is more usable and a bit more
structured than specifying hex numbers."
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (41 commits)
perf tools: Remove auto-generated bison/flex files
perf annotate: Fix off by one symbol hist size allocation and hit accounting
perf tools: Add missing ref-cycles event back to event parser
perf annotate: addr2line wants addresses in same format as objdump
perf probe: Finder fails to resolve function name to address
tracing: Fix ent_size in trace output
perf symbols: Handle NULL dso in dso__name_len
perf symbols: Do not include libgen.h
perf tools: Fix bug in raw sample parsing
perf tools: Fix display of first level of callchains
perf tools: Switch module.h into export.h
perf: Move mmap page data_head offset assertion out of header
perf: Fix mmap_page capabilities and docs
perf diff: Fix to work with new hists design
perf tools: Fix modifier to be applied on correct events
perf tools: Fix various casting issues for 32 bits
perf tools: Simplify event_read_id exit path
tracing: Fix ftrace stack trace entries
tracing: Move the tracing_on/off() declarations into CONFIG_TRACING
perf report: Add a simple GTK2-based 'perf report' browser
...
Pull x32 support for x86-64 from Ingo Molnar:
"This tree introduces the X32 binary format and execution mode for x86:
32-bit data space binaries using 64-bit instructions and 64-bit kernel
syscalls.
This allows applications whose working set fits into a 32 bits address
space to make use of 64-bit instructions while using a 32-bit address
space with shorter pointers, more compressed data structures, etc."
Fix up trivial context conflicts in arch/x86/{Kconfig,vdso/vma.c}
* 'x86-x32-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (71 commits)
x32: Fix alignment fail in struct compat_siginfo
x32: Fix stupid ia32/x32 inversion in the siginfo format
x32: Add ptrace for x32
x32: Switch to a 64-bit clock_t
x32: Provide separate is_ia32_task() and is_x32_task() predicates
x86, mtrr: Use explicit sizing and padding for the 64-bit ioctls
x86/x32: Fix the binutils auto-detect
x32: Warn and disable rather than error if binutils too old
x32: Only clear TIF_X32 flag once
x32: Make sure TS_COMPAT is cleared for x32 tasks
fs: Remove missed ->fds_bits from cessation use of fd_set structs internally
fs: Fix close_on_exec pointer in alloc_fdtable
x32: Drop non-__vdso weak symbols from the x32 VDSO
x32: Fix coding style violations in the x32 VDSO code
x32: Add x32 VDSO support
x32: Allow x32 to be configured
x32: If configured, add x32 system calls to system call tables
x32: Handle process creation
x32: Signal-related system calls
x86: Add #ifdef CONFIG_COMPAT to <asm/sys_ia32.h>
...
Complete the syscall-less self-profiling feature and address
all complaints, namely:
- capabilities, so we can detect what is actually available at runtime
Add a capabilities field to perf_event_mmap_page to indicate
what is actually available for use.
- on x86: RDPMC weirdness due to being 40/48 bits and not sign-extending
properly.
- ABI documentation as to how all this stuff works.
Also improve the documentation for the new features.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vince Weaver <vweaver1@eecs.utk.edu>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/1332433596.2487.33.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Adding sysfs group 'format' attribute for pmu device that
contains a syntax description on how to construct raw events.
The event configuration is described in following
struct pefr_event_attr attributes:
config
config1
config2
Each sysfs attribute within the format attribute group,
describes mapping of name and bitfield definition within
one of above attributes.
eg:
"/sys/...<dev>/format/event" contains "config:0-7"
"/sys/...<dev>/format/umask" contains "config:8-15"
"/sys/...<dev>/format/usr" contains "config:16"
the attribute value syntax is:
line: config ':' bits
config: 'config' | 'config1' | 'config2"
bits: bits ',' bit_term | bit_term
bit_term: VALUE '-' VALUE | VALUE
Adding format attribute definitions for x86 cpu pmus.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/n/tip-vhdk5y2hyype9j63prymty36@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With branch stack sampling, it is possible to filter by priv levels.
In system-wide mode, that means it is possible to capture only user
level branches. The builtin SW LBR filter needs to disassemble code
based on LBR captured addresses. For that, it needs to know the task
the addresses are associated with. Because of context switches, the
content of the branch stack buffer may contain addresses from
different tasks.
We need a callback on context switch to either flush the branch stack
or save it. This patch adds a new callback in struct pmu which is called
during context switches. The callback is called only when necessary.
That is when a system-wide context has, at least, one event which
uses PERF_SAMPLE_BRANCH_STACK. The callback is never called for
per-thread context.
In this version, the Intel x86 code simply flushes (resets) the LBR
on context switches (fills it with zeroes). Those zeroed branches are
then filtered out by the SW filter.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1328826068-11713-11-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
If precise sampling is enabled on Intel x86 then perf_event uses PEBS.
To correct for the off-by-one error of PEBS, perf_event uses LBR when
precise_sample > 1.
On Intel x86 PERF_SAMPLE_BRANCH_STACK is implemented using LBR,
therefore both features must be coordinated as they may not
configure LBR the same way.
For PEBS, LBR needs to capture all branches at the priv level of
the associated event.
This patch checks that the branch type and priv level of BRANCH_STACK
is compatible with that of the PEBS LBR requirement, thereby allowing:
$ perf record -b any,u -e instructions:upp ....
But:
$ perf record -b any_call,u -e instructions:upp
Is not possible.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1328826068-11713-5-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The Intel LBR on some recent processor is capable
of filtering branches by type. The filter is configurable
via the LBR_SELECT MSR register.
There are limitation on how this register can be used.
On Nehalem/Westmere, the LBR_SELECT is shared by the two HT threads
when HT is on. It is private to each core when HT is off.
On SandyBridge, the LBR_SELECT register is private to each thread
when HT is on. It is private to each core when HT is off.
The kernel must manage the sharing of LBR_SELECT. It allows
multiple users on the same logical CPU to use LBR_SELECT as
long as they program it with the same value. Across sibling
CPUs (HT threads), the same restriction applies on NHM/WSM.
This patch implements this sharing logic by leveraging the
mechanism put in place for managing the offcore_response
shared MSR.
We modify __intel_shared_reg_get_constraints() to cause
x86_get_event_constraint() to be called because LBR may
be associated with events that may be counter constrained.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1328826068-11713-4-git-send-email-eranian@google.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
So that we can get the perf bench exec stack fixes and then apply the
remaining fix for the files added after what is in perf/urgent.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With the new throttling/unthrottling code introduced with
commit:
e050e3f0a7 ("perf: Fix broken interrupt rate throttling")
we occasionally hit two WARN_ON_ONCE() checks in:
- intel_pmu_pebs_enable()
- intel_pmu_lbr_enable()
- x86_pmu_start()
The assertions are no longer problematic. There is a valid
path where they can trigger but it is harmless.
The assertion can be triggered with:
$ perf record -e instructions:pp ....
Leading to paths:
intel_pmu_pebs_enable
intel_pmu_enable_event
x86_perf_event_set_period
x86_pmu_start
perf_adjust_freq_unthr_context
perf_event_task_tick
scheduler_tick
And:
intel_pmu_lbr_enable
intel_pmu_enable_event
x86_perf_event_set_period
x86_pmu_start
perf_adjust_freq_unthr_context.
perf_event_task_tick
scheduler_tick
cpuc->enabled is always on because when we get to
perf_adjust_freq_unthr_context() the PMU is not totally
disabled. Furthermore when we need to adjust a period,
we only stop the event we need to change and not the
entire PMU. Thus, when we re-enable, cpuc->enabled is
already set. Note that when we stop the event, both
pebs and lbr are stopped if necessary (and possible).
Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: peterz@infradead.org
Link: http://lkml.kernel.org/r/20120202110401.GA30911@quad
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current x86 event scheduler fails to resolve scheduling problems
of certain combinations of events and constraints. This happens if the
counter mask of such an event is not a subset of any other counter
mask of a constraint with an equal or higher weight, e.g. constraints
of the AMD family 15h pmu:
counter mask weight
amd_f15_PMC30 0x09 2 <--- overlapping counters
amd_f15_PMC20 0x07 3
amd_f15_PMC53 0x38 3
The scheduler does not find then an existing solution. Here is an
example:
event code counter failure possible solution
0x02E PMC[3,0] 0 3
0x043 PMC[2:0] 1 0
0x045 PMC[2:0] 2 1
0x046 PMC[2:0] FAIL 2
The event scheduler may not select the correct counter in the first
cycle because it needs to know which subsequent events will be
scheduled. It may fail to schedule the events then.
To solve this, we now save the scheduler state of events with
overlapping counter counstraints. If we fail to schedule the events
we rollback to those states and try to use another free counter.
Constraints with overlapping counters are marked with a new introduced
overlap flag. We set the overlap flag for such constraints to give the
scheduler a hint which events to select for counter rescheduling. The
EVENT_CONSTRAINT_OVERLAP() macro can be used for this.
Care must be taken as the rescheduling algorithm is O(n!) which will
increase scheduling cycles for an over-commited system dramatically.
The number of such EVENT_CONSTRAINT_OVERLAP() macros and its counter
masks must be kept at a minimum. Thus, the current stack is limited to
2 states to limit the number of loops the algorithm takes in the worst
case.
On systems with no overlapping-counter constraints, this
implementation does not increase the loop count compared to the
previous algorithm.
V2:
* Renamed redo -> overlap.
* Reimplementation using perf scheduling helper functions.
V3:
* Added WARN_ON_ONCE() if out of save states.
* Changed function interface of perf_sched_restore_state() to use bool
as return value.
Signed-off-by: Robert Richter <robert.richter@amd.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1321616122-1533-3-git-send-email-robert.richter@amd.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
