There are some updates for the Icelake model specific uncore performance
monitors. (The update can be found at 10th generation intel core
processors families specification update Revision 004, ICL068)
1) Counter 0 of ARB uncore unit is not available for software use
2) The global 'enable bit' (bit 29) and 'freeze bit' (bit 31) of
MSR_UNC_PERF_GLOBAL_CTRL cannot be used to control counter behavior.
Needs to use local enable in event select MSR.
Accessing the modified bit/registers will be ignored by HW. Users may
observe inaccurate results with the current code.
The changes of the MSR_UNC_PERF_GLOBAL_CTRL imply that groups cannot be
read atomically anymore. Although the error of the result for a group
becomes a bit bigger, it still far lower than not using a group. The
group support is still kept. Only Remove the *_box() related
implementation.
Since the counter 0 of ARB uncore unit is not available, update the MSR
address for the ARB uncore unit.
There is no change for IMC uncore unit, which only include free-running
counters.
Fixes: 6e394376ee ("perf/x86/intel/uncore: Add Intel Icelake uncore support")
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200925134905.8839-2-kan.liang@linux.intel.com
Previously, the MSR uncore for the Ice Lake and Tiger Lake are
identical. The code path is shared. However, with recent update, the
global MSR_UNC_PERF_GLOBAL_CTRL register and ARB uncore unit are changed
for the Ice Lake. Split the Ice Lake and Tiger Lake MSR uncore support.
The changes only impact the MSR ops() and the ARB uncore unit. Other
codes can still be shared between the Ice Lake and the Tiger Lake.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200925134905.8839-1-kan.liang@linux.intel.com
The Snow Ridge integrated PCIe3 uncore unit can be used to collect
performance data, e.g. utilization, between PCIe devices, plugged into
the PCIe port, and the components (in M2IOSF) responsible for
translating and managing requests to/from the device. The performance
data is very useful for analyzing the performance of PCIe devices.
The device with the PCIe3 uncore PMON units is owned by the portdrv_pci
driver. Create a PCI sub driver for the PCIe3 uncore PMON units.
Here are some difference between PCIe3 uncore unit and other uncore
pci units.
- There may be several Root Ports on a system. But the uncore counters
only exist in the Root Port A. A user can configure the channel mask
to collect the data from other Root Ports.
- The event format of the PCIe3 uncore unit is the same as IIO unit of
SKX.
- The Control Register of PCIe3 uncore unit is 64 bits.
- The offset of each counters is 8, which is the same as M2M unit of
SNR.
- New MSR addresses for unit control, counter and counter config.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-7-git-send-email-kan.liang@linux.intel.com
Some uncore counters may be located in the configuration space of a PCI
device, which already has a bonded driver. Currently, the uncore driver
cannot register a PCI uncore PMU for these counters, because, to
register a PCI uncore PMU, the uncore driver must be bond to the device.
However, one device can only have one bonded driver.
Add an uncore PCI sub driver to support such kind of devices.
The sub driver doesn't own the device. In initialization, the sub
driver searches the device via pci_get_device(), and register the
corresponding PMU for the device. In the meantime, the sub driver
registers a PCI bus notifier, which is used to notify the sub driver
once the device is removed. The sub driver can unregister the PMU
accordingly.
The sub driver only searches the devices defined in its id table. The
id table varies on different platforms, which will be implemented in the
following platform-specific patch.
Suggested-by: Bjorn Helgaas <helgaas@kernel.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-6-git-send-email-kan.liang@linux.intel.com
The PMU unregistration in the uncore PCI sub driver is similar as the
normal PMU unregistration for a PCI device. The codes to unregister a
PCI PMU can be shared.
Factor out uncore_pci_pmu_unregister(), which will be used later.
Use uncore_pci_get_dev_die_info() to replace the codes which retrieve
the socket and die informaion.
The pci_set_drvdata() is not included in uncore_pci_pmu_unregister() as
well, because the uncore PCI sub driver will not touch the private
driver data pointer of the device.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-5-git-send-email-kan.liang@linux.intel.com
The PMU registration in the uncore PCI sub driver is similar as the
normal PMU registration for a PCI device. The codes to register a PCI
PMU can be shared.
Factor out uncore_pci_pmu_register(), which will be used later.
The pci_set_drvdata() is not included in uncore_pci_pmu_register(). The
uncore PCI sub driver doesn't own the PCI device. It will not touch the
private driver data pointer for the device.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-4-git-send-email-kan.liang@linux.intel.com
When an uncore PCI sub driver gets a remove notification, the
corresponding PMU has to be retrieved and unregistered. The codes, which
find the corresponding PMU by comparing the pci_device_id table, can be
shared.
Factor out uncore_pci_find_dev_pmu(), which will be used later.
There is no functional change.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-3-git-send-email-kan.liang@linux.intel.com
The socket and die information is required to register/unregister a PMU
in the uncore PCI sub driver. The codes, which get the socket and die
information from a BUS number, can be shared.
Factor out uncore_pci_get_dev_die_info(), which will be used later.
There is no functional change.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1600094060-82746-2-git-send-email-kan.liang@linux.intel.com
Previously, the uncore driver would say "NB counters detected" on F17h
machines, which don't have NorthBridge (NB) counters. They have Data
Fabric (DF) counters. Just use the pmu.name to inform users which pmu
to use and its associated counter count.
F17h dmesg BEFORE:
amd_uncore: AMD NB counters detected
amd_uncore: AMD LLC counters detected
F17h dmesg AFTER:
amd_uncore: 4 amd_df counters detected
amd_uncore: 6 amd_l3 counters detected
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200921144330.6331-5-kim.phillips@amd.com
On Family 19h, the driver checks for a populated 2-bit threadmask in
order to establish that the user wants to measure individual slices,
individual cores (only one can be measured at a time), and lets
the user also directly specify enallcores and/or enallslices if
desired.
Example F19h invocation to measure L3 accesses (event 4, umask 0xff)
by the first thread (id 0 -> mask 0x1) of the first core (id 0) on the
first slice (id 0):
perf stat -a -e instructions,amd_l3/umask=0xff,event=0x4,coreid=0,threadmask=1,sliceid=0,enallcores=0,enallslices=0/ <workload>
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200921144330.6331-4-kim.phillips@amd.com
Continue to fully populate either one of threadmask or slicemask if the
user doesn't.
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200921144330.6331-3-kim.phillips@amd.com
Replace AMD_FORMAT_ATTR with the more apropos DEFINE_UNCORE_FORMAT_ATTR
stolen from arch/x86/events/intel/uncore.h. This way we can clearly
see the bit-variants of each of the attributes that want to have
the same name across families.
Also unroll AMD_ATTRIBUTE because we are going to separately add
new attributes that differ between DF and L3.
Also clean up the if-Family 17h-else logic in amd_uncore_init.
This is basically a rewrite of commit da6adaea2b
("perf/x86/amd/uncore: Update sysfs attributes for Family17h processors").
No functional changes.
Tested F17h+ /sys/bus/event_source/devices/amd_{l3,df}/format/*
content remains unchanged:
/sys/bus/event_source/devices/amd_l3/format/event:config:0-7
/sys/bus/event_source/devices/amd_l3/format/umask:config:8-15
/sys/bus/event_source/devices/amd_df/format/event:config:0-7,32-35,59-60
/sys/bus/event_source/devices/amd_df/format/umask:config:8-15
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200921144330.6331-2-kim.phillips@amd.com
Stephane Eranian found a bug in that IBS' current Fetch counter was not
being reset when the driver would write the new value to clear it along
with the enable bit set, and found that adding an MSR write that would
first disable IBS Fetch would make IBS Fetch reset its current count.
Indeed, the PPR for AMD Family 17h Model 31h B0 55803 Rev 0.54 - Sep 12,
2019 states "The periodic fetch counter is set to IbsFetchCnt [...] when
IbsFetchEn is changed from 0 to 1."
Explicitly set IbsFetchEn to 0 and then to 1 when re-enabling IBS Fetch,
so the driver properly resets the internal counter to 0 and IBS
Fetch starts counting again.
A family 15h machine tested does not have this problem, and the extra
wrmsr is also not needed on Family 19h, so only do the extra wrmsr on
families 16h through 18h.
Reported-by: Stephane Eranian <stephane.eranian@google.com>
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
[peterz: optimized]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://bugzilla.kernel.org/show_bug.cgi?id=206537
Family 19h RAPL support did not change from Family 17h; extend
the existing Fam17h support to work on Family 19h too.
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200908214740.18097-8-kim.phillips@amd.com
IBS hardware with the OpCntExt feature gets a 7-bit wider internal
counter. Both the maximum and current count bitfields in the
IBS_OP_CTL register are extended to support reading and writing it.
No changes are necessary to the driver for handling the extra
contiguous current count bits (IbsOpCurCnt), as the driver already
passes through 32 bits of that field. However, the driver has to do
some extra bit manipulation when converting from a period to the
non-contiguous (although conveniently aligned) extra bits in the
IbsOpMaxCnt bitfield.
This decreases IBS Op interrupt overhead when the period is over
1,048,560 (0xffff0), which would previously activate the driver's
software counter. That threshold is now 134,217,712 (0x7fffff0).
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200908214740.18097-7-kim.phillips@amd.com
Neither IbsBrTarget nor OPDATA4 are populated in IBS Fetch mode.
Don't accumulate them into raw sample user data in that case.
Also, in Fetch mode, add saving the IBS Fetch Control Extended MSR.
Technically, there is an ABI change here with respect to the IBS raw
sample data format, but I don't see any perf driver version information
being included in perf.data file headers, but, existing users can detect
whether the size of the sample record has reduced by 8 bytes to
determine whether the IBS driver has this fix.
Fixes: 904cb3677f ("perf/x86/amd/ibs: Update IBS MSRs and feature definitions")
Reported-by: Stephane Eranian <stephane.eranian@google.com>
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/20200908214740.18097-6-kim.phillips@amd.com
get_ibs_op_count() adds hardware's current count (IbsOpCurCnt) bits
to its count regardless of hardware's valid status.
According to the PPR for AMD Family 17h Model 31h B0 55803 Rev 0.54,
if the counter rolls over, valid status is set, and the lower 7 bits
of IbsOpCurCnt are randomized by hardware.
Don't include those bits in the driver's event count.
Fixes: 8b1e13638d ("perf/x86-ibs: Fix usage of IBS op current count")
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://bugzilla.kernel.org/show_bug.cgi?id=206537
Commit 5738891229 ("perf/x86/amd: Add support for Large Increment
per Cycle Events") mistakenly zeroes the upper 16 bits of the count
in set_period(). That's fine for counting with perf stat, but not
sampling with perf record when only Large Increment events are being
sampled. To enable sampling, we sign extend the upper 16 bits of the
merged counter pair as described in the Family 17h PPRs:
"Software wanting to preload a value to a merged counter pair writes the
high-order 16-bit value to the low-order 16 bits of the odd counter and
then writes the low-order 48-bit value to the even counter. Reading the
even counter of the merged counter pair returns the full 64-bit value."
Fixes: 5738891229 ("perf/x86/amd: Add support for Large Increment per Cycle Events")
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://bugzilla.kernel.org/show_bug.cgi?id=206537
Commit 2f217d58a8 ("perf/x86/amd/uncore: Set the thread mask for
F17h L3 PMCs") inadvertently changed the uncore driver's behaviour
wrt perf tool invocations with or without a CPU list, specified with
-C / --cpu=.
Change the behaviour of the driver to assume the former all-cpu (-a)
case, which is the more commonly desired default. This fixes
'-a -A' invocations without explicit cpu lists (-C) to not count
L3 events only on behalf of the first thread of the first core
in the L3 domain.
BEFORE:
Activity performed by the first thread of the last core (CPU#43) in
CPU#40's L3 domain is not reported by CPU#40:
sudo perf stat -a -A -e l3_request_g1.caching_l3_cache_accesses taskset -c 43 perf bench mem memcpy -s 32mb -l 100 -f default
...
CPU36 21,835 l3_request_g1.caching_l3_cache_accesses
CPU40 87,066 l3_request_g1.caching_l3_cache_accesses
CPU44 17,360 l3_request_g1.caching_l3_cache_accesses
...
AFTER:
The L3 domain activity is now reported by CPU#40:
sudo perf stat -a -A -e l3_request_g1.caching_l3_cache_accesses taskset -c 43 perf bench mem memcpy -s 32mb -l 100 -f default
...
CPU36 354,891 l3_request_g1.caching_l3_cache_accesses
CPU40 1,780,870 l3_request_g1.caching_l3_cache_accesses
CPU44 315,062 l3_request_g1.caching_l3_cache_accesses
...
Fixes: 2f217d58a8 ("perf/x86/amd/uncore: Set the thread mask for F17h L3 PMCs")
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/20200908214740.18097-2-kim.phillips@amd.com
A warning as below may be triggered when sampling with large PEBS.
[ 410.411250] perf: interrupt took too long (72145 > 71975), lowering
kernel.perf_event_max_sample_rate to 2000
[ 410.724923] ------------[ cut here ]------------
[ 410.729822] WARNING: CPU: 0 PID: 16397 at arch/x86/events/core.c:1422
x86_pmu_stop+0x95/0xa0
[ 410.933811] x86_pmu_del+0x50/0x150
[ 410.937304] event_sched_out.isra.0+0xbc/0x210
[ 410.941751] group_sched_out.part.0+0x53/0xd0
[ 410.946111] ctx_sched_out+0x193/0x270
[ 410.949862] __perf_event_task_sched_out+0x32c/0x890
[ 410.954827] ? set_next_entity+0x98/0x2d0
[ 410.958841] __schedule+0x592/0x9c0
[ 410.962332] schedule+0x5f/0xd0
[ 410.965477] exit_to_usermode_loop+0x73/0x120
[ 410.969837] prepare_exit_to_usermode+0xcd/0xf0
[ 410.974369] ret_from_intr+0x2a/0x3a
[ 410.977946] RIP: 0033:0x40123c
[ 411.079661] ---[ end trace bc83adaea7bb664a ]---
In the non-overflow context, e.g., context switch, with large PEBS, perf
may stop an event twice. An example is below.
//max_samples_per_tick is adjusted to 2
//NMI is triggered
intel_pmu_handle_irq()
handle_pmi_common()
drain_pebs()
__intel_pmu_pebs_event()
perf_event_overflow()
__perf_event_account_interrupt()
hwc->interrupts = 1
return 0
//A context switch happens right after the NMI.
//In the same tick, the perf_throttled_seq is not changed.
perf_event_task_sched_out()
perf_pmu_sched_task()
intel_pmu_drain_pebs_buffer()
__intel_pmu_pebs_event()
perf_event_overflow()
__perf_event_account_interrupt()
++hwc->interrupts >= max_samples_per_tick
return 1
x86_pmu_stop(); # First stop
perf_event_context_sched_out()
task_ctx_sched_out()
ctx_sched_out()
event_sched_out()
x86_pmu_del()
x86_pmu_stop(); # Second stop and trigger the warning
Perf should only invoke the perf_event_overflow() in the overflow
context.
Current drain_pebs() is called from:
- handle_pmi_common() -- overflow context
- intel_pmu_pebs_sched_task() -- non-overflow context
- intel_pmu_pebs_disable() -- non-overflow context
- intel_pmu_auto_reload_read() -- possible overflow context
With PERF_SAMPLE_READ + PERF_FORMAT_GROUP, the function may be
invoked in the NMI handler. But, before calling the function, the
PEBS buffer has already been drained. The __intel_pmu_pebs_event()
will not be called in the possible overflow context.
To fix the issue, an indicator is required to distinguish between the
overflow context aka handle_pmi_common() and other cases.
The dummy regs pointer can be used as the indicator.
In the non-overflow context, perf should treat the last record the same
as other PEBS records, and doesn't invoke the generic overflow handler.
Fixes: 21509084f9 ("perf/x86/intel: Handle multiple records in the PEBS buffer")
Reported-by: Like Xu <like.xu@linux.intel.com>
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Like Xu <like.xu@linux.intel.com>
Link: https://lkml.kernel.org/r/20200902210649.2743-1-kan.liang@linux.intel.com
Starts from Ice Lake, the TopDown metrics are directly available as
fixed counters and do not require generic counters. Also, the TopDown
metrics can be collected per thread. Extend the RDPMC usage to support
per-thread TopDown metrics.
The RDPMC index of the PERF_METRICS will be output if RDPMC users ask
for the RDPMC index of the metrics events.
To support per thread RDPMC TopDown, the metrics and slots counters have
to be saved/restored during the context switching.
The last_period and period_left are not used in the counting mode. Use
the fields for saved_metric and saved_slots.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-12-kan.liang@linux.intel.com
Ice Lake supports the hardware TopDown metrics feature, which can free
up the scarce GP counters.
Update the event constraints for the metrics events. The metric counters
do not exist, which are mapped to a dummy offset. The sharing between
multiple users of the same metric without multiplexing is not allowed.
Implement set_topdown_event_period for Ice Lake. The values in
PERF_METRICS MSR are derived from the fixed counter 3. Both registers
should start from zero.
Implement update_topdown_event for Ice Lake. The metric is reported by
multiplying the metric (fraction) with slots. To maintain accurate
measurements, both registers are cleared for each update. The fixed
counter 3 should always be cleared before the PERF_METRICS.
Implement td_attr for the new metrics events and the new slots fixed
counter. Make them visible to the perf user tools.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-11-kan.liang@linux.intel.com
The RDPMC base offset of fixed counters is hard-code. Use a meaningful
name to replace the magic number to improve the readability of the code.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-10-kan.liang@linux.intel.com
Intro
=====
The TopDown Microarchitecture Analysis (TMA) Method is a structured
analysis methodology to identify critical performance bottlenecks in
out-of-order processors. Current perf has supported the method.
The method works well, but there is one problem. To collect the TopDown
events, several GP counters have to be used. If a user wants to collect
other events at the same time, the multiplexing probably be triggered,
which impacts the accuracy.
To free up the scarce GP counters, the hardware TopDown metrics feature
is introduced from Ice Lake. The hardware implements an additional
"metrics" register and a new Fixed Counter 3 that measures pipeline
"slots". The TopDown events can be calculated from them instead.
Events
======
The level 1 TopDown has four metrics. There is no event-code assigned to
the TopDown metrics. Four metric events are exported as separate perf
events, which map to the internal "metrics" counter register. Those
events do not exist in hardware, but can be allocated by the scheduler.
For the event mapping, a special 0x00 event code is used, which is
reserved for fake events. The metric events start from umask 0x10.
When setting up the metric events, they point to the Fixed Counter 3.
They have to be specially handled.
- Add the update_topdown_event() callback to read the additional metrics
MSR and generate the metrics.
- Add the set_topdown_event_period() callback to initialize metrics MSR
and the fixed counter 3.
- Add a variable n_metric_event to track the number of the accepted
metrics events. The sharing between multiple users of the same metric
without multiplexing is not allowed.
- Only enable/disable the fixed counter 3 when there are no other active
TopDown events, which avoid the unnecessary writing of the fixed
control register.
- Disable the PMU when reading the metrics event. The metrics MSR and
the fixed counter 3 are read separately. The values may be modified by
an NMI.
All four metric events don't support sampling. Since they will be
handled specially for event update, a flag PERF_X86_EVENT_TOPDOWN is
introduced to indicate this case.
The slots event can support both sampling and counting.
For counting, the flag is also applied.
For sampling, it will be handled normally as other normal events.
Groups
======
The slots event is required in a Topdown group.
To avoid reading the METRICS register multiple times, the metrics and
slots value can only be updated by slots event in a group.
All active slots and metrics events will be updated one time.
Therefore, the slots event must be before any metric events in a Topdown
group.
NMI
======
The METRICS related register may be overflow. The bit 48 of the STATUS
register will be set. If so, PERF_METRICS and Fixed counter 3 are
required to be reset. The patch also update all active slots and
metrics events in the NMI handler.
The update_topdown_event() has to read two registers separately. The
values may be modified by an NMI. PMU has to be disabled before calling
the function.
RDPMC
======
RDPMC is temporarily disabled. A later patch will enable it.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-9-kan.liang@linux.intel.com
Currently, the if-else is used in the intel_pmu_disable/enable_event to
check the type of an event. It works well, but with more and more types
added later, e.g., perf metrics, compared to the switch statement, the
if-else may impair the readability of the code.
There is no harm to use the switch statement to replace the if-else
here. Also, some optimizing compilers may compile a switch statement
into a jump-table which is more efficient than if-else for a large
number of cases. The performance gain may not be observed for now,
because the number of cases is only 5, but the benefits may be observed
with more and more types added in the future.
Use switch to replace the if-else in the intel_pmu_disable/enable_event.
If the idx is invalid, print a warning.
For the case INTEL_PMC_IDX_FIXED_BTS in intel_pmu_disable_event, don't
need to check the event->attr.precise_ip. Use return for the case.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-7-kan.liang@linux.intel.com
Bit 15 of the PERF_CAPABILITIES MSR indicates that the perf METRICS
feature is supported. The perf METRICS is not a PEBS feature.
Rename pebs_metrics_available perf_metrics.
The bit is not used in the current code. It will be used in a later
patch.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-6-kan.liang@linux.intel.com
Magic numbers are used in the current NMI handler for the global status
bit. Use a meaningful name to replace the magic numbers to improve the
readability of the code.
Remove a Tab for all GLOBAL_STATUS_* and INTEL_PMC_IDX_FIXED_BTS macros
to reduce the length of the line.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-3-kan.liang@linux.intel.com
The RDPMC index is always re-calculated for the RDPMC userspace support,
which is unnecessary.
The RDPMC index value is stored in the variable event_base_rdpmc for
the kernel usage, which can be used for RDPMC userspace support as well.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200723171117.9918-2-kan.liang@linux.intel.com
Intel SPR platform uses fixed 16 bit energy unit for DRAM RAPL domain,
and fixed 0 bit energy unit for Psys RAPL domain.
After this, on SPR platform the energy counters appear in perf list.
Signed-off-by: Zhang Rui <rui.zhang@intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Acked-by: Len Brown <len.brown@intel.com>
Link: https://lore.kernel.org/r/20200811153149.12242-4-rui.zhang@intel.com
There will be more platforms with different fixed energy units.
Enhance the code to support different RAPL unit quirks for different
platforms.
Signed-off-by: Zhang Rui <rui.zhang@intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Len Brown <len.brown@intel.com>
Link: https://lore.kernel.org/r/20200811153149.12242-3-rui.zhang@intel.com
This fixes a problem introduced by commit:
5fb5273a90 ("perf/x86/rapl: Use new MSR detection interface")
that perf event sysfs attributes for psys RAPL domain are missing.
Fixes: 5fb5273a90 ("perf/x86/rapl: Use new MSR detection interface")
Signed-off-by: Zhang Rui <rui.zhang@intel.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Len Brown <len.brown@intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20200811153149.12242-2-rui.zhang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
-----BEGIN PGP SIGNATURE-----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=X7o0
-----END PGP SIGNATURE-----
Merge tag 'x86-cleanups-2020-08-03' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull x86 cleanups from Ingo Molnar:
"Misc cleanups all around the place"
* tag 'x86-cleanups-2020-08-03' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/ioperm: Initialize pointer bitmap with NULL rather than 0
x86: uv: uv_hub.h: Delete duplicated word
x86: cmpxchg_32.h: Delete duplicated word
x86: bootparam.h: Delete duplicated word
x86/mm: Remove the unused mk_kernel_pgd() #define
x86/tsc: Remove unused "US_SCALE" and "NS_SCALE" leftover macros
x86/ioapic: Remove unused "IOAPIC_AUTO" define
x86/mm: Drop unused MAX_PHYSADDR_BITS
x86/msr: Move the F15h MSRs where they belong
x86/idt: Make idt_descr static
initrd: Remove erroneous comment
x86/mm/32: Fix -Wmissing prototypes warnings for init.c
cpu/speculation: Add prototype for cpu_show_srbds()
x86/mm: Fix -Wmissing-prototypes warnings for arch/x86/mm/init.c
x86/asm: Unify __ASSEMBLY__ blocks
x86/cpufeatures: Mark two free bits in word 3
x86/msr: Lift AMD family 0x15 power-specific MSRs
Hygon Family 18h(Dhyana) support RAPL in bit 14 of CPUID 0x80000007 EDX,
and has MSRs RAPL_PWR_UNIT/CORE_ENERGY_STAT/PKG_ENERGY_STAT. So add Hygon
Dhyana Family 18h support for RAPL.
The output is available via the energy-pkg pseudo event:
$ perf stat -a -I 1000 --per-socket -e power/energy-pkg/
[ mingo: Tidied up the initializers. ]
Signed-off-by: Pu Wen <puwen@hygon.cn>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20200720082205.1307-1-puwen@hygon.cn
Reading LBR registers in a perf NMI handler for a non-PEBS event
causes a high overhead because the number of LBR registers is huge.
To reduce the overhead, the XSAVES instruction should be used to replace
the LBR registers' reading method.
The XSAVES buffer used for LBR read has to be per-CPU because the NMI
handler invoked the lbr_read(). The existing task_ctx_data buffer
cannot be used which is per-task and only be allocated for the LBR call
stack mode. A new lbr_xsave pointer is introduced in the cpu_hw_events
as an XSAVES buffer for LBR read.
The XSAVES buffer should be allocated only when LBR is used by a
non-PEBS event on the CPU because the total size of the lbr_xsave is
not small (~1.4KB).
The XSAVES buffer is allocated when a non-PEBS event is added, but it
is lazily released in x86_release_hardware() when perf releases the
entire PMU hardware resource, because perf may frequently schedule the
event, e.g. high context switch. The lazy release method reduces the
overhead of frequently allocate/free the buffer.
If the lbr_xsave fails to be allocated, roll back to normal Arch LBR
lbr_read().
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dave Hansen <dave.hansen@intel.com>
Link: https://lkml.kernel.org/r/1593780569-62993-24-git-send-email-kan.liang@linux.intel.com
In the LBR call stack mode, LBR information is used to reconstruct a
call stack. To get the complete call stack, perf has to save/restore
all LBR registers during a context switch. Due to a large number of the
LBR registers, this process causes a high CPU overhead. To reduce the
CPU overhead during a context switch, use the XSAVES/XRSTORS
instructions.
Every XSAVE area must follow a canonical format: the legacy region, an
XSAVE header and the extended region. Although the LBR information is
only kept in the extended region, a space for the legacy region and
XSAVE header is still required. Add a new dedicated structure for LBR
XSAVES support.
Before enabling XSAVES support, the size of the LBR state has to be
sanity checked, because:
- the size of the software structure is calculated from the max number
of the LBR depth, which is enumerated by the CPUID leaf for Arch LBR.
The size of the LBR state is enumerated by the CPUID leaf for XSAVE
support of Arch LBR. If the values from the two CPUID leaves are not
consistent, it may trigger a buffer overflow. For example, a hypervisor
may unconsciously set inconsistent values for the two emulated CPUID.
- unlike other state components, the size of an LBR state depends on the
max number of LBRs, which may vary from generation to generation.
Expose the function xfeature_size() for the sanity check.
The LBR XSAVES support will be disabled if the size of the LBR state
enumerated by CPUID doesn't match with the size of the software
structure.
The XSAVE instruction requires 64-byte alignment for state buffers. A
new macro is added to reflect the alignment requirement. A 64-byte
aligned kmem_cache is created for architecture LBR.
Currently, the structure for each state component is maintained in
fpu/types.h. The structure for the new LBR state component should be
maintained in the same place. Move structure lbr_entry to fpu/types.h as
well for broader sharing.
Add dedicated lbr_save/lbr_restore functions for LBR XSAVES support,
which invokes the corresponding xstate helpers to XSAVES/XRSTORS LBR
information at the context switch when the call stack mode is enabled.
Since the XSAVES/XRSTORS instructions will be eventually invoked, the
dedicated functions is named with '_xsaves'/'_xrstors' postfix.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dave Hansen <dave.hansen@intel.com>
Link: https://lkml.kernel.org/r/1593780569-62993-23-git-send-email-kan.liang@linux.intel.com
A new kmem_cache method is introduced to allocate the PMU specific data
task_ctx_data, which requires the PMU specific code to create a
kmem_cache.
Currently, the task_ctx_data is only used by the Intel LBR call stack
feature, which is introduced since Haswell. The kmem_cache should be
only created for Haswell and later platforms. There is no alignment
requirement for the existing platforms.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-18-git-send-email-kan.liang@linux.intel.com
Last Branch Records (LBR) enables recording of software path history by
logging taken branches and other control flows within architectural
registers now. Intel CPUs have had model-specific LBR for quite some
time, but this evolves them into an architectural feature now.
The main improvements of Architectural LBR implemented includes:
- Linux kernel can support the LBR features without knowing the model
number of the current CPU.
- Architectural LBR capabilities can be enumerated by CPUID. The
lbr_ctl_map is based on the CPUID Enumeration.
- The possible LBR depth can be retrieved from CPUID enumeration. The
max value is written to the new MSR_ARCH_LBR_DEPTH as the number of
LBR entries.
- A new IA32_LBR_CTL MSR is introduced to enable and configure LBRs,
which replaces the IA32_DEBUGCTL[bit 0] and the LBR_SELECT MSR.
- Each LBR record or entry is still comprised of three MSRs,
IA32_LBR_x_FROM_IP, IA32_LBR_x_TO_IP and IA32_LBR_x_TO_IP.
But they become the architectural MSRs.
- Architectural LBR is stack-like now. Entry 0 is always the youngest
branch, entry 1 the next youngest... The TOS MSR has been removed.
The way to enable/disable Architectural LBR is similar to the previous
model-specific LBR. __intel_pmu_lbr_enable/disable() can be reused, but
some modifications are required, which include:
- MSR_ARCH_LBR_CTL is used to enable and configure the Architectural
LBR.
- When checking the value of the IA32_DEBUGCTL MSR, ignoring the
DEBUGCTLMSR_LBR (bit 0) for Architectural LBR, which has no meaning
and always return 0.
- The FREEZE_LBRS_ON_PMI has to be explicitly set/clear, because
MSR_IA32_DEBUGCTLMSR is not touched in __intel_pmu_lbr_disable() for
Architectural LBR.
- Only MSR_ARCH_LBR_CTL is cleared in __intel_pmu_lbr_disable() for
Architectural LBR.
Some Architectural LBR dedicated functions are implemented to
reset/read/save/restore LBR.
- For reset, writing to the ARCH_LBR_DEPTH MSR clears all Arch LBR
entries, which is a lot faster and can improve the context switch
latency.
- For read, the branch type information can be retrieved from
the MSR_ARCH_LBR_INFO_*. But it's not fully compatible due to
OTHER_BRANCH type. The software decoding is still required for the
OTHER_BRANCH case.
LBR records are stored in the age order as well. Reuse
intel_pmu_store_lbr(). Check the CPUID enumeration before accessing
the corresponding bits in LBR_INFO.
- For save/restore, applying the fast reset (writing ARCH_LBR_DEPTH).
Reading 'lbr_from' of entry 0 instead of the TOS MSR to check if the
LBR registers are reset in the deep C-state. If 'the deep C-state
reset' bit is not set in CPUID enumeration, ignoring the check.
XSAVE support for Architectural LBR will be implemented later.
The number of LBR entries cannot be hardcoded anymore, which should be
retrieved from CPUID enumeration. A new structure
x86_perf_task_context_arch_lbr is introduced for Architectural LBR.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-15-git-send-email-kan.liang@linux.intel.com
The way to store the LBR information from a PEBS LBR record can be
reused in Architecture LBR, because
- The LBR information is stored like a stack. Entry 0 is always the
youngest branch.
- The layout of the LBR INFO MSR is similar.
The LBR information may be retrieved from either the LBR registers
(non-PEBS event) or a buffer (PEBS event). Extend rdlbr_*() to support
both methods.
Explicitly check the invalid entry (0s), which can avoid unnecessary MSR
access if using a non-PEBS event. For a PEBS event, the check should
slightly improve the performance as well. The invalid entries are cut.
The intel_pmu_lbr_filter() doesn't need to check and filter them out.
Cannot share the function with current model-specific LBR read, because
the direction of the LBR growth is opposite.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-14-git-send-email-kan.liang@linux.intel.com
The previous model-specific LBR and Architecture LBR (legacy way) use a
similar method to save/restore the LBR information, which directly
accesses the LBR registers. The codes which read/write a set of LBR
registers can be shared between them.
Factor out two functions which are used to read/write a set of LBR
registers.
Add lbr_info into structure x86_pmu, and use it to replace the hardcoded
LBR INFO MSR, because the LBR INFO MSR address of the previous
model-specific LBR is different from Architecture LBR. The MSR address
should be assigned at boot time. For now, only Sky Lake and later
platforms have the LBR INFO MSR.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-13-git-send-email-kan.liang@linux.intel.com
The {rd,wr}lbr_{to,from} wrappers are invoked in hot paths, e.g. context
switch and NMI handler. They should be always inline to achieve better
performance. However, the CONFIG_OPTIMIZE_INLINING allows the compiler
to uninline functions marked 'inline'.
Mark the {rd,wr}lbr_{to,from} wrappers as __always_inline to force
inline the wrappers.
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-12-git-send-email-kan.liang@linux.intel.com
Current LBR information in the structure x86_perf_task_context is stored
in a different format from the PEBS LBR record and Architecture LBR,
which prevents the sharing of the common codes.
Use the format of the PEBS LBR record as a unified format. Use a generic
name lbr_entry to replace pebs_lbr_entry.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-11-git-send-email-kan.liang@linux.intel.com
An IA32_LBR_CTL is introduced for Architecture LBR to enable and config
LBR registers to replace the previous LBR_SELECT.
All the related members in struct cpu_hw_events and struct x86_pmu
have to be renamed.
Some new macros are added to reflect the layout of LBR_CTL.
The mapping from PERF_SAMPLE_BRANCH_* to the corresponding bits in
LBR_CTL MSR is saved in lbr_ctl_map now, which is not a const value.
The value relies on the CPUID enumeration.
For the previous model-specific LBR, most of the bits in LBR_SELECT
operate in the suppressed mode. For the bits in LBR_CTL, the polarity is
inverted.
For the previous model-specific LBR format 5 (LBR_FORMAT_INFO), if the
NO_CYCLES and NO_FLAGS type are set, the flag LBR_NO_INFO will be set to
avoid the unnecessary LBR_INFO MSR read. Although Architecture LBR also
has a dedicated LBR_INFO MSR, perf doesn't need to check and set the
flag LBR_NO_INFO. For Architecture LBR, XSAVES instruction will be used
as the default way to read the LBR MSRs all together. The overhead which
the flag tries to avoid doesn't exist anymore. Dropping the flag can
save the extra check for the flag in the lbr_read() later, and make the
code cleaner.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-10-git-send-email-kan.liang@linux.intel.com
The LBR capabilities of Architecture LBR are retrieved from the CPUID
enumeration once at boot time. The capabilities have to be saved for
future usage.
Several new fields are added into structure x86_pmu to indicate the
capabilities. The fields will be used in the following patches.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-9-git-send-email-kan.liang@linux.intel.com
The type of task_ctx is hardcoded as struct x86_perf_task_context,
which doesn't apply for Architecture LBR. For example, Architecture LBR
doesn't have the TOS MSR. The number of LBR entries is variable. A new
struct will be introduced for Architecture LBR. Perf has to determine
the type of task_ctx at run time.
The type of task_ctx pointer is changed to 'void *', which will be
determined at run time.
The generic LBR optimization can be shared between Architecture LBR and
model-specific LBR. Both need to access the structure for the generic
LBR optimization. A helper task_context_opt() is introduced to retrieve
the pointer of the structure at run time.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-7-git-send-email-kan.liang@linux.intel.com
To reduce the overhead of a context switch with LBR enabled, some
generic optimizations were introduced, e.g. avoiding restore LBR if no
one else touched them. The generic optimizations can also be used by
Architecture LBR later. Currently, the fields for the generic
optimizations are part of structure x86_perf_task_context, which will be
deprecated by Architecture LBR. A new structure should be introduced
for the common fields of generic optimization, which can be shared
between Architecture LBR and model-specific LBR.
Both 'valid_lbrs' and 'tos' are also used by the generic optimizations,
but they are not moved into the new structure, because Architecture LBR
is stack-like. The 'valid_lbrs' which records the index of the valid LBR
is not required anymore. The TOS MSR will be removed.
LBR registers may be cleared in the deep Cstate. If so, the generic
optimizations should not be applied. Perf has to unconditionally
restore the LBR registers. A generic function is required to detect the
reset due to the deep Cstate. lbr_is_reset_in_cstate() is introduced.
Currently, for the model-specific LBR, the TOS MSR is used to detect the
reset. There will be another method introduced for Architecture LBR
later.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-6-git-send-email-kan.liang@linux.intel.com
The MSRs of Architectural LBR are different from previous model-specific
LBR. Perf has to implement different functions to save and restore them.
The function pointers for LBR save and restore are introduced. Perf
should initialize the corresponding functions at boot time.
The generic optimizations, e.g. avoiding restore LBR if no one else
touched them, still apply for Architectural LBRs. The related codes are
not moved to model-specific functions.
Current model-specific LBR functions are set as default.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-5-git-send-email-kan.liang@linux.intel.com
The method to read Architectural LBRs is different from previous
model-specific LBR. Perf has to implement a different function.
A function pointer for LBR read is introduced. Perf should initialize
the corresponding function at boot time, and avoid checking lbr_format
at run time.
The current 64-bit LBR read function is set as default.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1593780569-62993-4-git-send-email-kan.liang@linux.intel.com
