Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
dmaengine: idxd: Add IDXD performance monitor support
Implement the IDXD performance monitor capability (named 'perfmon' in
the DSA (Data Streaming Accelerator) spec [1]), which supports the
collection of information about key events occurring during DSA and
IAX (Intel Analytics Accelerator) device execution, to assist in
performance tuning and debugging.
The idxd perfmon support is implemented as part of the IDXD driver and
interfaces with the Linux perf framework. It has several features in
common with the existing uncore pmu support:
- it does not support sampling
- does not support per-thread counting
However it also has some unique features not present in the core and
uncore support:
- all general-purpose counters are identical, thus no event constraints
- operation is always system-wide
While the core perf subsystem assumes that all counters are by default
per-cpu, the uncore pmus are socket-scoped and use a cpu mask to
restrict counting to one cpu from each socket. IDXD counters use a
similar strategy but expand the scope even further; since IDXD
counters are system-wide and can be read from any cpu, the IDXD perf
driver picks a single cpu to do the work (with cpu hotplug notifiers
to choose a different cpu if the chosen one is taken off-line).
More specifically, the perf userspace tool by default opens a counter
for each cpu for an event. However, if it finds a cpumask file
associated with the pmu under sysfs, as is the case with the uncore
pmus, it will open counters only on the cpus specified by the cpumask.
Since perfmon only needs to open a single counter per event for a
given IDXD device, the perfmon driver will create a sysfs cpumask file
for the device and insert the first cpu of the system into it. When a
user uses perf to open an event, perf will open a single counter on
the cpu specified by the cpu mask. This amounts to the default
system-wide rather than per-cpu counting mentioned previously for
perfmon pmu events. In order to keep the cpu mask up-to-date, the
driver implements cpu hotplug support for multiple devices, as IDXD
usually enumerates and registers more than one idxd device.
The perfmon driver implements basic perfmon hardware capability
discovery and configuration, and is initialized by the IDXD driver's
probe function. During initialization, the driver retrieves the total
number of supported performance counters, the pmu ID, and the device
type from idxd device, and registers itself under the Linux perf
framework.
The perf userspace tool can be used to monitor single or multiple
events depending on the given configuration, as well as event groups,
which are also supported by the perfmon driver. The user configures
events using the perf tool command-line interface by specifying the
event and corresponding event category, along with an optional set of
filters that can be used to restrict counting to specific work queues,
traffic classes, page and transfer sizes, and engines (See [1] for
specifics).
With the configuration specified by the user, the perf tool issues a
system call passing that information to the kernel, which uses it to
initialize the specified event(s). The event(s) are opened and
started, and following termination of the perf command, they're
stopped. At that point, the perfmon driver will read the latest count
for the event(s), calculate the difference between the latest counter
values and previously tracked counter values, and display the final
incremental count as the event count for the cycle. An overflow
handler registered on the IDXD irq path is used to account for counter
overflows, which are signaled by an overflow interrupt.
Below are a couple of examples of perf usage for monitoring DSA events.
The following monitors all events in the 'engine' category. Becuuse
no filters are specified, this captures all engine events for the
workload, which in this case is 19 iterations of the work generated by
the kernel dmatest module.
Details describing the events can be found in Appendix D of [1],
Performance Monitoring Events, but briefly they are:
event 0x1: total input data processed, in 32-byte units
event 0x2: total data written, in 32-byte units
event 0x4: number of work descriptors that read the source
event 0x8: number of work descriptors that write the destination
event 0x10: number of work descriptors dispatched from batch descriptors
event 0x20: number of work descriptors dispatched from work queues
# perf stat -e dsa0/event=0x1,event_category=0x1/,
dsa0/event=0x2,event_category=0x1/,
dsa0/event=0x4,event_category=0x1/,
dsa0/event=0x8,event_category=0x1/,
dsa0/event=0x10,event_category=0x1/,
dsa0/event=0x20,event_category=0x1/
modprobe dmatest channel=dma0chan0 timeout=2000
iterations=19 run=1 wait=1
Performance counter stats for 'system wide':
5,332 dsa0/event=0x1,event_category=0x1/
5,327 dsa0/event=0x2,event_category=0x1/
19 dsa0/event=0x4,event_category=0x1/
19 dsa0/event=0x8,event_category=0x1/
0 dsa0/event=0x10,event_category=0x1/
19 dsa0/event=0x20,event_category=0x1/
21.977436186 seconds time elapsed
The command below illustrates filter usage with a simple example. It
specifies that MEM_MOVE operations should be counted for the DSA
device dsa0 (event 0x8 corresponds to the EV_MEM_MOVE event - Number
of Memory Move Descriptors, which is part of event category 0x3 -
Operations. The detailed category and event IDs are available in
Appendix D, Performance Monitoring Events, of [1]). In addition to
the event and event category, a number of filters are also specified
(the detailed filter values are available in Chapter 6.4 (Filter
Support) of [1]), which will restrict counting to only those events
that meet all of the filter criteria. In this case, the filters
specify that only MEM_MOVE operations that are serviced by work queue
wq0 and specifically engine number engine0 and traffic class tc0
having sizes between 0 and 4k and page size of between 0 and 1G result
in a counter hit; anything else will be filtered out and not appear in
the final count. Note that filters are optional - any filter not
specified is assumed to be all ones and will pass anything.
# perf stat -e dsa0/filter_wq=0x1,filter_tc=0x1,filter_sz=0x7,
filter_eng=0x1,event=0x8,event_category=0x3/
modprobe dmatest channel=dma0chan0 timeout=2000
iterations=19 run=1 wait=1
Performance counter stats for 'system wide':
19 dsa0/filter_wq=0x1,filter_tc=0x1,filter_sz=0x7,
filter_eng=0x1,event=0x8,event_category=0x3/
21.865914091 seconds time elapsed
The output above reflects that the unspecified workload resulted in
the counting of 19 MEM_MOVE operation events that met the filter
criteria.
[1]: https://software.intel.com/content/www/us/en/develop/download/intel-data-streaming-accelerator-preliminary-architecture-specification.html
[ Based on work originally by Jing Lin. ]
Reviewed-by: Dave Jiang <[email protected]>
Reviewed-by: Kan Liang <[email protected]>
Signed-off-by: Tom Zanussi <[email protected]>
Link: https://lore.kernel.org/r/0c5080a7d541904c4ad42b848c76a1ce056ddac7.1619276133.git.zanussi@kernel.org
Signed-off-by: Vinod Koul <[email protected]>- Loading branch information
