8-letter strings representing ARC perf events are stores in two
32-bit registers as ASCII characters like that: "IJMP", "IALL", "IJMPTAK" etc.
And the same order of bytes in the word is used regardless CPU endianness.
Which means in case of big-endian CPU core we need to swap bytes to get
the same order as if it was on little-endian CPU.
Otherwise we're seeing the following error message on boot:
------------------------->8----------------------
ARC perf : 8 counters (32 bits), 40 conditions, [overflow IRQ support]
sysfs: cannot create duplicate filename '/devices/arc_pct/events/pmji'
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.2.18 #3
Stack Trace:
arc_unwind_core+0xd4/0xfc
dump_stack+0x64/0x80
sysfs_warn_dup+0x46/0x58
sysfs_add_file_mode_ns+0xb2/0x168
create_files+0x70/0x2a0
------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at kernel/events/core.c:12144 perf_event_sysfs_init+0x70/0xa0
Failed to register pmu: arc_pct, reason -17
Modules linked in:
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.2.18 #3
Stack Trace:
arc_unwind_core+0xd4/0xfc
dump_stack+0x64/0x80
__warn+0x9c/0xd4
warn_slowpath_fmt+0x22/0x2c
perf_event_sysfs_init+0x70/0xa0
---[ end trace a75fb9a9837bd1ec ]---
------------------------->8----------------------
What happens here we're trying to register more than one raw perf event
with the same name "PMJI". Why? Because ARC perf events are 4 to 8 letters
and encoded into two 32-bit words. In this particular case we deal with 2
events:
* "IJMP____" which counts all jump & branch instructions
* "IJMPC___" which counts only conditional jumps & branches
Those strings are split in two 32-bit words this way "IJMP" + "____" &
"IJMP" + "C___" correspondingly. Now if we read them swapped due to CPU core
being big-endian then we read "PMJI" + "____" & "PMJI" + "___C".
And since we interpret read array of ASCII letters as a null-terminated string
on big-endian CPU we end up with 2 events of the same name "PMJI".
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: stable@vger.kernel.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Move HW events mapping to separate function to make code more readable.
Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Export all available ARC architected hardware events as
kernel PMU events to make non-generic events accessible.
ARC PMU HW allow us to read the list of all available
events names. So we generate kernel PMU event list
dynamically in arc_pmu_device_probe() using
human-readable events names we got from HW instead of
using pre-defined events list.
-------------------------->8--------------------------
$ perf list
[snip]
arc_pmu/bdata64/ [Kernel PMU event]
arc_pmu/bdcstall/ [Kernel PMU event]
arc_pmu/bdslot/ [Kernel PMU event]
arc_pmu/bfbmp/ [Kernel PMU event]
arc_pmu/bfirqex/ [Kernel PMU event]
arc_pmu/bflgstal/ [Kernel PMU event]
arc_pmu/bflush/ [Kernel PMU event]
-------------------------->8--------------------------
Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
* Use BIT(), lower_32_bits(), upper_32_bits() macroses,
fix code style violations.
* Use u32, u64, s64 instead of uint32_t, uint64_t, int64_t
* Fix description comment as this code doesn't belong only to
ARC700 anymore.
* Use SPDX License Identifier.
* Remove useless ifdefs. ifdef around 'arc_pmu_match' structure
declaration is useless as we refer to 'arc_pmu_match' in
several places which aren't guarded with ifdef. Nevertheless
'ARC' option selects 'OF' unconditionally so we can simply
get rid of this ifdef.
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
use ffz primitive which maps to ARCv2 instruction, vs. non atomic
__test_and_set_bit
It is unlikely if we will even have more than 32 counters, but still add
a BUILD_BUG to catch that
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Current perf ISR loops thru all 32 counters, checking for each if it
caused the interrupt. Instead only loop thru counters which actually
interrupted (typically 1).
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
We used to live with PERF_COUNT_HW_CACHE_REFERENCES and
PERF_COUNT_HW_CACHE_REFERENCES not specified on ARC.
Those events are actually aliases to 2 cache events that we do support
and so this change sets "cache-reference" and "cache-misses" events
in the same way as "L1-dcache-loads" and L1-dcache-load-misses.
And while at it adding debug info for cache events as well as doing a
subtle fix in HW events debug info - config value is much better
represented by hex so we may see not only event index but as well other
control bits set (if they exist).
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: Vineet Gupta <vgupta@synopsys.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-snps-arc@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This makes perf_callchain_{user,kernel}() receive the max stack
as context for the perf_callchain_entry, instead of accessing
the global sysctl_perf_event_max_stack.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: He Kuang <hekuang@huawei.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: Zefan Li <lizefan@huawei.com>
Link: http://lkml.kernel.org/n/tip-kolmn1yo40p7jhswxwrc7rrd@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This was the second perf intr issue
perf sampling on multicore requires intr to be enabled on all cores.
ARC perf probe code used helper arc_request_percpu_irq() which calls
- request_percpu_irq() on core0
- enable_percpu_irq() on all all cores (including core0)
genirq requires that request be made ahead of enable call.
However if perf probe happened on non core0 (observed on a 3.18 kernel),
enable would get called ahead of request, failing obviously and
rendering perf intr disabled on all such cores
[ 11.120000] 1 ARC perf : 8 counters (48 bits), 113 conditions, [overflow IRQ support]
[ 11.130000] 1 -----> enable_percpu_irq() IRQ 20 failed
[ 11.140000] 3 -----> enable_percpu_irq() IRQ 20 failed
[ 11.140000] 2 -----> enable_percpu_irq() IRQ 20 failed
[ 11.140000] 0 =====> request_percpu_irq() IRQ 20
[ 11.140000] 0 -----> enable_percpu_irq() IRQ 20
Fix this fragility, by calling request_percpu_irq() on whatever core
calls probe (there is no requirement on which core calls this anyways)
and then calling enable on each cores.
Interestingly this started as invesigation of STAR 9000838902:
"sporadically IRQs enabled on perf prob"
which was about occassional boot spew as request_percpu_irq got called
non-locally (from an IPI), and re-enabled interrupts in following path
proc_mkdir -> spin_unlock_irq()
which the irq work code didn't like.
| ARC perf : 8 counters (48 bits), 113 conditions, [overflow IRQ support]
|
| BUG: failure at ../kernel/irq_work.c:135/irq_work_run_list()!
| CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.18.10-01127-g285efb8e66d1 #2
|
| Stack Trace:
| arc_unwind_core.constprop.1+0x94/0x104
| dump_stack+0x62/0x98
| irq_work_run_list+0xb0/0xb4
| irq_work_run+0x22/0x3c
| do_IPI+0x74/0x9c
| handle_irq_event_percpu+0x34/0x164
| handle_percpu_irq+0x58/0x78
| generic_handle_irq+0x1e/0x2c
| arch_do_IRQ+0x3c/0x60
| ret_from_exception+0x0/0x8
Cc: Marc Zyngier <marc.zyngier@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-snps-arc@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Cc: Alexey Brodkin <abrodkin@synopsys.com>
Cc: <stable@vger.kernel.org> #4.2+
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
With all features in place, the ARC HS pct block can now be effectively
allowed to be probed/used
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
* split off pmu info into singleton and per-cpu bits
* setup PMU on all cores
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
In times of ARC 700 performance counters didn't have support of
interrupt an so for ARC we only had support of non-sampling events.
Put simply only "perf stat" was functional.
Now with ARC HS we have support of interrupts in performance counters
which this change introduces support of.
ARC performance counters act in the following way in regard of
interrupts generation.
[1] A counter counts starting from value set in PCT_COUNT register pair
[2] Once counter reaches value set in PCT_INT_CNT interrupt is raised
Basic setup look like this:
[1] PCT_COUNT = 0;
[2] PCT_INT_CNT = __limit_value__;
[3] Enable interrupts for that counter and let it run
[4] Let counter reach its limit
[5] Handle interrupt when it happens
Note that PCT HW block is build in CPU core and so ints interrupt
line (which is basically OR of all counters IRQs) is wired directly to
top-level IRQC. That means do de-assert PCT interrupt it's required to
reset IRQs from all counters that have reached their limit values.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This generalization prepares for support of overflow interrupts.
Hardware event counters on ARC work that way:
Each counter counts from programmed start value (set in
ARC_REG_PCT_COUNT) to a limit value (set in ARC_REG_PCT_INT_CNT) and
once limit value is reached this timer generates an interrupt.
Even though this hardware implementation allows for more flexibility,
in Linux kernel we decided to mimic behavior of other architectures
this way:
[1] Set limit value as half of counter's max value (to allow counter to
run after reaching it limit, see below for more explanation):
---------->8-----------
arc_pmu->max_period = (1ULL << counter_size) / 2 - 1ULL;
---------->8-----------
[2] Set start value as "arc_pmu->max_period - sample_period" and then
count up to the limit
Our event counters don't stop on reaching max value (the one we set in
ARC_REG_PCT_INT_CNT) but continue to count until kernel explicitly
stops each of them.
And setting a limit as half of counter capacity is done to allow
capturing of additional events in between moment when interrupt was
triggered until we're actually processing PMU interrupts. That way
we're trying to be more precise.
For example if we count CPU cycles we keep track of cycles while
running through generic IRQ handling code:
[1] We set counter period as say 100_000 events of type "crun"
[2] Counter reaches that limit and raises its interrupt
[3] Once we get in PMU IRQ handler we read current counter value from
ARC_REG_PCT_SNAP ans see there something like 105_000.
If counters stop on reaching a limit value then we would miss
additional 5000 cycles.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The number of counters in PCT can never be more than 32 (while
countable conditions could be 100+) for both ARCompact and ARCv2
And while at it update copyright dates.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Directly return the result of perf_pmu_register() in
arc_pmu_device_probe() instead of assigning and returning variable ret.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
static arc_pmu in the arch/arc/kernel/perf_event.c is not initialized as
it's shadowed by a local variable of the same name in the
arc_pmu_device_probe.
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Fixes: 03c94fcf95 "ARC: perf: make @arc_pmu static global"
CC: <stable@vger.kernel.org> # 4.1
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Transition to using the new generic PERF_PMU_CAP_NO_INTERRUPT method for
failing a sampling event when no PMU interrupt is available.
Signed-off-by: Vince Weaver <vincent.weaver@maine.edu>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Grant Likely <grant.likely@linaro.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Rob Herring <robh+dt@kernel.org>
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1406150159280.16738@vincent-weaver-1.umelst.maine.edu
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The ARC 700 does not have an interrupt associated with it, and as
such it cannot trigger when a counter overflows. As the counters are
48 bit, it will usually take at least 100 days before a counter
overflows, so for mere counting of events, there is no problem.
Sampling is not supported though.
Signed-off-by: Mischa Jonker <mjonker@synopsys.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This adds basic perf support for ARC700 cores. Most PERF_COUNT_HW* events
are supported now.
Signed-off-by: Mischa Jonker <mjonker@synopsys.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Alexey Brodkin
Eugeniy Paltsev
Vineet Gupta
Alexey Brodkin
Andrea Gelmini
Arnaldo Carvalho de Melo
Tobias Klauser
Max Filippov
Vince Weaver
Mischa Jonker