Re: [PATCH v9 1/2] Documentation: fpga: dfl: add description for performance reporting support
From: Randy Dunlap <hidden>
Date: 2020-04-18 00:07:56
Also in:
linux-fpga, lkml
Hi-- Here are a few comments for you: On 4/16/20 10:39 PM, Wu Hao wrote:
quoted hunk ↗ jump to hunk
From: Xu Yilun <yilun.xu@intel.com> This patch adds description for performance reporting support for Device Feature List (DFL) based FPGA. Signed-off-by: Xu Yilun <yilun.xu@intel.com> Signed-off-by: Wu Hao <redacted> --- Documentation/fpga/dfl.rst | 84 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 84 insertions(+)diff --git a/Documentation/fpga/dfl.rst b/Documentation/fpga/dfl.rst index 094fc8a..6ccaf77 100644 --- a/Documentation/fpga/dfl.rst +++ b/Documentation/fpga/dfl.rst
quoted hunk ↗ jump to hunk
@@ -378,6 +383,85 @@ The device nodes used for ioctl() or mmap() can be referenced through:: /sys/class/fpga_region/<regionX>/<dfl-port.n>/dev +Performance Counters +==================== +Performance reporting is one private feature implemented in FME. It could +supports several independent, system-wide, device counter sets in hardware to +monitor and count for performance events, including "basic", "cache", "fabric", +"vtd" and "vtd_sip" counters. Users could use standard perf tool to monitor +FPGA cache hit/miss rate, transaction number, interface clock counter of AFU +and other FPGA performance events. + +Different FPGA devices may have different counter sets, it depends on hardware
sets, depending on hardware
+implementation. e.g. some discrete FPGA cards don't have any cache. User could
E.g., some
+use "perf list" to check which perf events are supported by target hardware. + +In order to allow user to use standard perf API to access these performance +counters, driver creates a perf PMU, and related sysfs interfaces in +/sys/bus/event_source/devices/dfl_fme* to describe available perf events and +configuration options. + +The "format" directory describes the format of the config field of struct +perf_event_attr. There are 3 bitfields for config, "evtype" defines which type
config: "evtype"
+the perf event belongs to. "event" is the identity of the event within its
belongs to; "event" is
+category. "portid" is introduced to decide counters set to monitor on FPGA
category; "portid" is
+overall data or a specific port. + +The "events" directory describes the configuration templates for all available +events which can be used with perf tool directly. For example, fab_mmio_read +has the configuration "event=0x06,evtype=0x02,portid=0xff", which shows this +event belongs to fabric type (0x02), the local event id is 0x06 and it is for +overall monitoring (portid=0xff). + +Example usage of perf:: + + $# perf list |grep dfl_fme + + dfl_fme0/fab_mmio_read/ [Kernel PMU event] + <...> + dfl_fme0/fab_port_mmio_read,portid=?/ [Kernel PMU event] + <...> + + $# perf stat -a -e dfl_fme0/fab_mmio_read/ <command> + or + $# perf stat -a -e dfl_fme0/event=0x06,evtype=0x02,portid=0xff/ <command> + or + $# perf stat -a -e dfl_fme0/config=0xff2006/ <command> + +Another example, fab_port_mmio_read monitors mmio read of a specific port. So +its configuration template is "event=0x06,evtype=0x01,portid=?". The portid +should be explicitly set. + +Its usage of perf:: + + $# perf stat -a -e dfl_fme0/fab_port_mmio_read,portid=0x0/ <command> + or + $# perf stat -a -e dfl_fme0/event=0x06,evtype=0x02,portid=0x0/ <command> + or + $# perf stat -a -e dfl_fme0/config=0x2006/ <command> + +Please note for fabric counters, overall perf events (fab_*) and port perf +events (fab_port_*) actually share one set of counters in hardware, so it can't +monitor both at the same time. If this set of counters is configured to monitor +overall data, then per port perf data is not supported. See below example:: + + $# perf stat -e dfl_fme0/fab_mmio_read/,dfl_fme0/fab_port_mmio_write,\ + portid=0/ sleep 1 + + Performance counter stats for 'system wide': + + 3 dfl_fme0/fab_mmio_read/ + <not supported> dfl_fme0/fab_port_mmio_write,portid=0x0/ + + 1.001750904 seconds time elapsed + +The driver also provides a "cpumask" sysfs attribute, which contains only one +cpu id used to access these perf events. Counting on multiple CPUs is not
CPU
+allowed since they are system-wide counters on FPGA device. + +The current driver does not support sampling. So "perf record" is unsupported. + + Add new FIUs support ==================== It's possible that developers made some new function blocks (FIUs) under this
thanks. -- ~Randy