linuxdebug/tools/perf/Documentation/perf-arm-spe.txt

219 lines
8.1 KiB
Plaintext

perf-arm-spe(1)
================
NAME
----
perf-arm-spe - Support for Arm Statistical Profiling Extension within Perf tools
SYNOPSIS
--------
[verse]
'perf record' -e arm_spe//
DESCRIPTION
-----------
The SPE (Statistical Profiling Extension) feature provides accurate attribution of latencies and
events down to individual instructions. Rather than being interrupt-driven, it picks an
instruction to sample and then captures data for it during execution. Data includes execution time
in cycles. For loads and stores it also includes data address, cache miss events, and data origin.
The sampling has 5 stages:
1. Choose an operation
2. Collect data about the operation
3. Optionally discard the record based on a filter
4. Write the record to memory
5. Interrupt when the buffer is full
Choose an operation
~~~~~~~~~~~~~~~~~~~
This is chosen from a sample population, for SPE this is an IMPLEMENTATION DEFINED choice of all
architectural instructions or all micro-ops. Sampling happens at a programmable interval. The
architecture provides a mechanism for the SPE driver to infer the minimum interval at which it should
sample. This minimum interval is used by the driver if no interval is specified. A pseudo-random
perturbation is also added to the sampling interval by default.
Collect data about the operation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Program counter, PMU events, timings and data addresses related to the operation are recorded.
Sampling ensures there is only one sampled operation is in flight.
Optionally discard the record based on a filter
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Based on programmable criteria, choose whether to keep the record or discard it. If the record is
discarded then the flow stops here for this sample.
Write the record to memory
~~~~~~~~~~~~~~~~~~~~~~~~~~
The record is appended to a memory buffer
Interrupt when the buffer is full
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When the buffer fills, an interrupt is sent and the driver signals Perf to collect the records.
Perf saves the raw data in the perf.data file.
Opening the file
----------------
Up until this point no decoding of the SPE data was done by either the kernel or Perf. Only when the
recorded file is opened with 'perf report' or 'perf script' does the decoding happen. When decoding
the data, Perf generates "synthetic samples" as if these were generated at the time of the
recording. These samples are the same as if normal sampling was done by Perf without using SPE,
although they may have more attributes associated with them. For example a normal sample may have
just the instruction pointer, but an SPE sample can have data addresses and latency attributes.
Why Sampling?
-------------
- Sampling, rather than tracing, cuts down the profiling problem to something more manageable for
hardware. Only one sampled operation is in flight at a time.
- Allows precise attribution data, including: Full PC of instruction, data virtual and physical
addresses.
- Allows correlation between an instruction and events, such as TLB and cache miss. (Data source
indicates which particular cache was hit, but the meaning is implementation defined because
different implementations can have different cache configurations.)
However, SPE does not provide any call-graph information, and relies on statistical methods.
Collisions
----------
When an operation is sampled while a previous sampled operation has not finished, a collision
occurs. The new sample is dropped. Collisions affect the integrity of the data, so the sample rate
should be set to avoid collisions.
The 'sample_collision' PMU event can be used to determine the number of lost samples. Although this
count is based on collisions _before_ filtering occurs. Therefore this can not be used as an exact
number for samples dropped that would have made it through the filter, but can be a rough
guide.
The effect of microarchitectural sampling
-----------------------------------------
If an implementation samples micro-operations instead of instructions, the results of sampling must
be weighted accordingly.
For example, if a given instruction A is always converted into two micro-operations, A0 and A1, it
becomes twice as likely to appear in the sample population.
The coarse effect of conversions, and, if applicable, sampling of speculative operations, can be
estimated from the 'sample_pop' and 'inst_retired' PMU events.
Kernel Requirements
-------------------
The ARM_SPE_PMU config must be set to build as either a module or statically.
Depending on CPU model, the kernel may need to be booted with page table isolation disabled
(kpti=off). If KPTI needs to be disabled, this will fail with a console message "profiling buffer
inaccessible. Try passing 'kpti=off' on the kernel command line".
Capturing SPE with perf command-line tools
------------------------------------------
You can record a session with SPE samples:
perf record -e arm_spe// -- ./mybench
The sample period is set from the -c option, and because the minimum interval is used by default
it's recommended to set this to a higher value. The value is written to PMSIRR.INTERVAL.
Config parameters
~~~~~~~~~~~~~~~~~
These are placed between the // in the event and comma separated. For example '-e
arm_spe/load_filter=1,min_latency=10/'
branch_filter=1 - collect branches only (PMSFCR.B)
event_filter=<mask> - filter on specific events (PMSEVFR) - see bitfield description below
jitter=1 - use jitter to avoid resonance when sampling (PMSIRR.RND)
load_filter=1 - collect loads only (PMSFCR.LD)
min_latency=<n> - collect only samples with this latency or higher* (PMSLATFR)
pa_enable=1 - collect physical address (as well as VA) of loads/stores (PMSCR.PA) - requires privilege
pct_enable=1 - collect physical timestamp instead of virtual timestamp (PMSCR.PCT) - requires privilege
store_filter=1 - collect stores only (PMSFCR.ST)
ts_enable=1 - enable timestamping with value of generic timer (PMSCR.TS)
+++*+++ Latency is the total latency from the point at which sampling started on that instruction, rather
than only the execution latency.
Only some events can be filtered on; these include:
bit 1 - instruction retired (i.e. omit speculative instructions)
bit 3 - L1D refill
bit 5 - TLB refill
bit 7 - mispredict
bit 11 - misaligned access
So to sample just retired instructions:
perf record -e arm_spe/event_filter=2/ -- ./mybench
or just mispredicted branches:
perf record -e arm_spe/event_filter=0x80/ -- ./mybench
Viewing the data
~~~~~~~~~~~~~~~~~
By default perf report and perf script will assign samples to separate groups depending on the
attributes/events of the SPE record. Because instructions can have multiple events associated with
them, the samples in these groups are not necessarily unique. For example perf report shows these
groups:
Available samples
0 arm_spe//
0 dummy:u
21 l1d-miss
897 l1d-access
5 llc-miss
7 llc-access
2 tlb-miss
1K tlb-access
36 branch-miss
0 remote-access
900 memory
The arm_spe// and dummy:u events are implementation details and are expected to be empty.
To get a full list of unique samples that are not sorted into groups, set the itrace option to
generate 'instruction' samples. The period option is also taken into account, so set it to 1
instruction unless you want to further downsample the already sampled SPE data:
perf report --itrace=i1i
Memory access details are also stored on the samples and this can be viewed with:
perf report --mem-mode
Common errors
~~~~~~~~~~~~~
- "Cannot find PMU `arm_spe'. Missing kernel support?"
Module not built or loaded, KPTI not disabled (see above), or running on a VM
- "Arm SPE CONTEXT packets not found in the traces."
Root privilege is required to collect context packets. But these only increase the accuracy of
assigning PIDs to kernel samples. For userspace sampling this can be ignored.
- Excessively large perf.data file size
Increase sampling interval (see above)
SEE ALSO
--------
linkperf:perf-record[1], linkperf:perf-script[1], linkperf:perf-report[1],
linkperf:perf-inject[1]