Fix the data source encodings to represent L2.1/L3.1(another core's
L2/L3 on the same node) accesses properly for power10 and older
plaforms.
Add new macros(LEVEL/REM) which can be used to add mem_lvl_num and remote
field data inside perf_mem_data_src structure.
Result in power9 system with patch changes:
localhost:~/linux/tools/perf # ./perf mem report | grep Remote
0.01% 1 252 Remote core, same node L3 or L3 hit [.] 0x0000000000002dd0 producer_consumer [.] 0x00007fff7f25eb90
anon HitM N/A No N/A 0 0
0.01% 1 220 Remote core, same node L3 or L3 hit [.] 0x0000000000002dd0 producer_consumer [.] 0x00007fff77776d90
anon HitM N/A No N/A 0 0
0.01% 1 220 Remote core, same node L3 or L3 hit [.] 0x0000000000002dd0 producer_consumer [.] 0x00007fff817d9410
anon HitM N/A No N/A 0 0
Fixes: 79e96f8f93 ("powerpc/perf: Export memory hierarchy info to user space")
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20211006140654.298352-5-kjain@linux.ibm.com
Sampled Instruction Event Register (SIER) field [46:48] identifies the
sampled instruction type. ISA v3.1 says value of 0b111 for this field as
reserved, but in POWER10 it denotes LARX/STCX type which will hopefully
be fixed in ISA v3.1 update.
Patch fixes the functions to handle type value 7 for CPU_FTR_ARCH_31.
Fixes: a64e697cef ("powerpc/perf: power10 Performance Monitoring support")
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com>
[mpe: Avoid reading mmcra until necessary, use early return to deindent if block]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1614858937-1485-1-git-send-email-atrajeev@linux.vnet.ibm.com
Performance Monitoring Unit (PMU) registers in powerpc provides
information on cycles elapsed between different stages in the
pipeline. This can be used for application tuning. On ISA v3.1
platform, this information is exposed by sampling registers.
Patch adds kernel support to capture two of the cycle counters
as part of perf sample using the sample type:
PERF_SAMPLE_WEIGHT_STRUCT.
The power PMU function 'get_mem_weight' currently uses 64 bit weight
field of perf_sample_data to capture memory latency. But following the
introduction of PERF_SAMPLE_WEIGHT_TYPE, weight field could contain
64-bit or 32-bit value depending on the architexture support for
PERF_SAMPLE_WEIGHT_STRUCT. Patches uses WEIGHT_STRUCT to expose the
pipeline stage cycles info. Hence update the ppmu functions to work for
64-bit and 32-bit weight values.
If the sample type is PERF_SAMPLE_WEIGHT, use the 64-bit weight field.
if the sample type is PERF_SAMPLE_WEIGHT_STRUCT, memory subsystem
latency is stored in the low 32bits of perf_sample_weight structure.
Also for CPU_FTR_ARCH_31, capture the two cycle counter information in
two 16 bit fields of perf_sample_weight structure.
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1616425047-1666-2-git-send-email-atrajeev@linux.vnet.ibm.com
Add platform specific attr.config value checks. Patch
includes checks for both power9 and power10.
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210408074504.248211-2-maddy@linux.ibm.com
Thresholding, a performance monitoring unit feature, can be
used to identify marked instructions which take more than
expected cycles between start event and end event.
Threshold compare (thresh_cmp) bits are programmed in MMCRA
register. In Power9, thresh_cmp bits were part of the
event code. But in case of P10, thresh_cmp are not part of
event code due to inclusion of MMCR3 bits.
Patch here adds an option to use attr.config1 variable
to be used to pass thresh_cmp value to be programmed in
MMCRA register. A new ppmu flag called PPMU_HAS_ATTR_CONFIG1
has been added and this flag is used to notify the use of
attr.config1 variable.
Patch has extended the parameter list of 'compute_mmcr',
to include power_pmu's 'flags' element and parameter list of
get_constraint to include attr.config1 value. It also extend
parameter list of power_check_constraints inorder to pass
perf_event list.
As stated by commit ef0e3b650f ("powerpc/perf: Fix Threshold
Event Counter Multiplier width for P10"), constraint bits for
thresh_cmp is also needed to be increased to 11 bits, which is
handled as part of this patch. We added bit number 53 as part
of constraint bits of thresh_cmp for power10 to make it an
11 bit field.
Updated layout for p10:
/*
* Layout of constraint bits:
*
* 60 56 52 48 44 40 36 32
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
* [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ]
* | |
* [ thresh_cmp bits for p10] thresh_sel -*
*
* 28 24 20 16 12 8 4 0
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
* [ ] | [ ] | [ sample ] [ ] [6] [5] [4] [3] [2] [1]
* | | | | |
* BHRB IFM -* | | |*radix_scope | Count of events for each PMC.
* EBB -* | | p1, p2, p3, p4, p5, p6.
* L1 I/D qualifier -* |
* nc - number of counters -*
*
* The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints
* we want the low bit of each field to be added to any existing value.
*
* Everything else is a value field.
*/
Result:
command#: cat /sys/devices/cpu/format/thresh_cmp
config1:0-17
ex. usage:
command#: perf record -I --weight -d -e
cpu/event=0x67340101EC,thresh_cmp=500/ ./ebizzy -S 2 -t 1 -s 4096
1826636 records/s
real 2.00 s
user 2.00 s
sys 0.00 s
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.038 MB perf.data (61 samples) ]
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210209095234.837356-1-kjain@linux.ibm.com
Threshold Event Counter Multiplier (TECM) is part of Monitor Mode
Control Register A (MMCRA). This field along with Threshold Event
Counter Exponent (TECE) is used to get threshould counter value.
In Power10, this is a 8bit field, so patch fixes the
current code to modify the MMCRA[TECM] extraction macro to
handle this change. ISA v3.1 says this is a 7 bit field but
POWER10 it's actually 8 bits which will hopefully be fixed
in ISA v3.1 update.
Fixes: 170a315f41 ("powerpc/perf: Support to export MMCRA[TEC*] field to userspace")
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1608022578-1532-1-git-send-email-atrajeev@linux.vnet.ibm.com
The PMU group constraints mask for threshold events covers
all thresholding bits which includes threshold control value
(start/stop), select value as well as thresh_cmp value (MMCRA[9:18].
In power9, thresh_cmp bits were part of the event code. But in case
of power10, thresh_cmp bits are not part of event code due to
inclusion of MMCR3 bits. Hence thresh_cmp is not valid for
group constraints for power10.
Fix the PMU group constraints checking for threshold events in
power10 by using constraint mask and value for only threshold control
and select bits.
Fixes: a64e697cef ("powerpc/perf: power10 Performance Monitoring support")
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1606409684-1589-4-git-send-email-atrajeev@linux.vnet.ibm.com
power10 uses bit 9 of the raw event code as RADIX_SCOPE_QUAL.
This bit is used for enabling the radix process events.
Patch fixes the PMU counter support functions to program bit
18 of MMCR1 ( Monitor Mode Control Register1 ) with the
RADIX_SCOPE_QUAL bit value. Since this field is not per-pmc,
add this to PMU group constraints to make sure events in a
group will have same bit value for this field. Use bit 21 as
constraint bit field for radix_scope_qual. Patch also updates
the power10 raw event encoding layout information, format field
and constraints bit layout to include the radix_scope_qual bit.
Fixes: a64e697cef ("powerpc/perf: power10 Performance Monitoring support")
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1606409684-1589-2-git-send-email-atrajeev@linux.vnet.ibm.com
Sparse warns about all the init functions:
symbol init_ppc970_pmu was not declared. Should it be static?
symbol init_power5p_pmu was not declared. Should it be static?
symbol init_power5_pmu was not declared. Should it be static?
symbol init_power6_pmu was not declared. Should it be static?
symbol init_power7_pmu was not declared. Should it be static?
symbol init_power9_pmu was not declared. Should it be static?
symbol init_power8_pmu was not declared. Should it be static?
symbol init_generic_compat_pmu was not declared. Should it be static?
They're already declared in internal.h, so just make sure all the C
files include that directly or indirectly.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://lore.kernel.org/r/20200916115637.3100484-2-mpe@ellerman.id.au
Base enablement patch to register performance monitoring hardware
support for power10. Patch introduce the raw event encoding format,
defines the supported list of events, config fields for the event
attributes and their corresponding bit values which are exported via
sysfs.
Patch also enhances the support function in isa207_common.c to include
power10 pmu hardware.
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1594996707-3727-9-git-send-email-atrajeev@linux.vnet.ibm.com
core-book3s currently uses array to store the MMCR registers as part
of per-cpu `cpu_hw_events`. This patch does a clean up to use `struct`
to store mmcr regs instead of array. This will make code easier to read
and reduces chance of any subtle bug that may come in the future, say
when new registers are added. Patch updates all relevant code that was
using MMCR array ( cpuhw->mmcr[x]) to use newly introduced `struct`.
This includes the PMU driver code for supported platforms (power5
to power9) and ISA macros for counter support functions.
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1594996707-3727-2-git-send-email-atrajeev@linux.vnet.ibm.com
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or any
later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 1 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Richard Fontana <rfontana@redhat.com>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190524100844.067492367@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In previous generation processors, both bus events and direct
events of performance monitoring unit can be individually
programmabled and monitored in PMCs.
But in Power9, L2/L3 bus events are always available as a
"bank" of 4 events. To obtain the counts for any of the
l2/l3 bus events in a given bank, the user will have to
program PMC4 with corresponding l2/l3 bus event for that
bank.
Patch enforce two contraints incase of L2/L3 bus events.
1)Any L2/L3 event when programmed is also expected to program corresponding
PMC4 event from that group.
2)PMC4 event should always been programmed first due to group constraint
logic limitation
For ex. consider these L3 bus events
PM_L3_PF_ON_CHIP_MEM (0x460A0),
PM_L3_PF_MISS_L3 (0x160A0),
PM_L3_CO_MEM (0x260A0),
PM_L3_PF_ON_CHIP_CACHE (0x360A0),
1) This is an INVALID group for L3 Bus event monitoring,
since it is missing PMC4 event.
perf stat -e "{r160A0,r260A0,r360A0}" < >
And this is a VALID group for L3 Bus events:
perf stat -e "{r460A0,r160A0,r260A0,r360A0}" < >
2) This is an INVALID group for L3 Bus event monitoring,
since it is missing PMC4 event.
perf stat -e "{r260A0,r360A0}" < >
And this is a VALID group for L3 Bus events:
perf stat -e "{r460A0,r260A0,r360A0}" < >
3) This is an INVALID group for L3 Bus event monitoring,
since it is missing PMC4 event.
perf stat -e "{r360A0}" < >
And this is a VALID group for L3 Bus events:
perf stat -e "{r460A0,r360A0}" < >
Patch here implements group constraint logic suggested by Michael Ellerman.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Raw event code has couple of fields "unit" and "cache" in it, to capture
the "unit" to monitor for a given pmcxsel and cache reload qualifier to
program in MMCR1.
isa207_get_constraint() refers "unit" field to update the MMCRC (L2/L3)
Event bus control fields with "cache" bits of the raw event code.
These are power8 specific and not supported by PowerISA v3.0 pmu. So wrap
the checks to be power8 specific. Also, "cache" bit field is referred to
update MMCR1[16:17] and this check can be power8 specific.
Fixes: 7ffd948fae ('powerpc/perf: factor out power8 pmu functions')
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
POWER9 DD1 was never a product. It is no longer supported by upstream
firmware, and it is not effectively supported in Linux due to lack of
testing.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Michael Ellerman <mpe@ellerman.id.au>
[mpe: Remove arch_make_huge_pte() entirely]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Comment explanning the raw event code encoding for Power8 was
moved to isa207_common.h file when re-factoring the code to
support power9. But then Power9 pmu branched out due to changes
specific to power9. So move the encoding comment back to power8-pmu.c
Just comment movement and no logic change.
Fixes: 4d3576b207 ('powerpc/perf: factor out power8 pmu macros and defines')
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
There are some hardware events on Power systems which only count when
the processor is not idle, and there are some fixed-function counters
which count such events. For example, the "run cycles" event counts
cycles when the processor is not idle. If the user asks to count
cycles, we can use "run cycles" if this is a per-task event, since the
processor is running when the task is running, by definition. We can't
use "run cycles" if the user asks for "cycles" on a system-wide
counter.
Currently in power8 this check is done using PPMU_ONLY_COUNT_RUN flag
in power8_get_alternatives() function. Based on the flag, events are
switched if needed. This function should also be enabled in power9, so
factor out the code to isa207_get_alternatives().
Fixes: efe881afdd ('powerpc/perf: Factor out event_alternative function')
Reported-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Commit 20dd4c624d ('powerpc/perf: Fix SDAR_MODE value for continous
sampling on Power9') set the default sdar_mode value in MMCRA[SDAR_MODE]
to be used as 0b01 (Update on TLB miss). And this value is set if sdar_mode
from event is zero, or we are in continous sampling mode in power9 dd1.
But it is preferred to have the sdar_mode value for power9 as
0b10 (Update on dcache miss) for better sampling updates instead
of 0b01 (Update on TLB miss).
From Anton:
Using a bandwidth test case with a 1MB footprint, I profiled cycles and
chose TLB updates of the SDAR:
$ perf record -d -e r000400000000001E:u ./bw2001 1M
^
SDAR TLB
$ perf report -D | grep PERF_RECORD_SAMPLE | sed 's/.*addr: //' | sort -u | wc -l
4
I get 4 unique addresses. If I ran with dcache misses:
$ perf record -d -e r000800000000001E:u ./bw2001 1M
^
SDAR dcache miss
$ perf report -D|grep PERF_RECORD_SAMPLE| sed 's/.*addr: //'|sort -u | wc -l
5217
I get 5217 unique addresses. No surprises here, but it does show why
TLB misses is the wrong event to default to - we get very little useful
information out of it.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Acked-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Threshold feature when used with MMCRA [Threshold Event Counter Event],
MMCRA[Threshold Start event] and MMCRA[Threshold End event] will update
MMCRA[Threashold Event Counter Exponent] and MMCRA[Threshold Event
Counter Multiplier] with the corresponding threshold event count values.
Patch to export MMCRA[TECX/TECM] to userspace in 'weight' field of
struct perf_sample_data.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The LDST field and DATA_SRC in SIER identifies the memory hierarchy level
(eg: L1, L2 etc), from which a data-cache miss for a marked instruction
was satisfied. Use the 'perf_mem_data_src' object to export this
hierarchy level to user space.
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
MMCRA[SDAR_MODE] specifices how the SDAR should be updated in
continous sampling mode. On P9 it must be set to 0b00 when
MMCRA[63] is set.
Fixes: c7c3f568be ('powerpc/perf: macros for power9 format encoding')
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
PMC5 on POWER9 DD1 may not provide right counts in all
sampling scenarios, hence use PM_INST_DISP event instead
in PMC2 or PMC3 in preference.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Factor out the power8 event_alternative function to share
the code with power9.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Patch to add macros and contants to support the power9 raw
event encoding format. Couple of functions added since some of the
bits fields like PMCxCOMB and THRESH_CMP has different width and location
within MMCR* in power9.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Factor out some of the power8 pmu functions
to new file "isa207-common.c" to share with
power9 pmu code. Only code movement and no
logic change
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Factor out some of the power8 pmu macros to
new a header file to share with power9 pmu code.
Just code movement and no logic change.
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Kajol Jain
Athira Rajeev
Madhavan Srinivasan
Michael Ellerman
Thomas Gleixner
Madhavan Srinivasan
Nicholas Piggin