Skip to contentSkip to conversation

add performance reporting support to FPGA DFL drivers

PATCH Superseded Wu Hao 3 participants 11 messages Feb 10, 2020 – Feb 13, 2020

Subsystems (4)

v7 of 10 Superseded Latest is v10 →

lore
Share & tools
lore

Wu Hao

author
#0 lore raw

v7 cover add performance reporting support to FPGA DFL drivers

Hi all,

This patchset adds performance reporting support for FPGA DFL drivers. It introduces one pmu to expose userspace interfaces via standard perf API. User could use standard perf tool to access perf events exposed via pmu.

This patchset is generated based on 5.6-rc1.

Main changes from v6:
- add a new ABI/testing/ sysfs documentation in patch #2.
- fix a warning reported by kbuild in patch #2.

Main changes from v5:
- use dev_ext_attribute instead of fme_perf_event_attr.
- use is_visible function to decide which events to expose per hardware capability, and add event_init checking for all events.

Main changes from v4:
- rebase and clean up.
- update Kconfig for PERF_EVENTS dependency.

Main changes from v3:
- add more descriptions in doc, including how to use perf tool for these hardware counters. (patch #1)
- use standard perf API instead of sysfs entries. (patch #2)

Wu Hao (1):
fpga: dfl: fme: add performance reporting support

Xu Yilun (1):
Documentation: fpga: dfl: add description for performance reporting
support

 .../ABI/testing/sysfs-bus-event_source-devices-fme | 105 +++
 Documentation/fpga/dfl.rst                         |  83 ++
 drivers/fpga/Kconfig                               |   2 +-
 drivers/fpga/Makefile                              |   1 +
 drivers/fpga/dfl-fme-main.c                        |   4 +
 drivers/fpga/dfl-fme-perf.c                        | 943 +++++++++++++++++++++
 drivers/fpga/dfl-fme.h                             |   2 +
 drivers/fpga/dfl.h                                 |   2 +
 8 files changed, 1141 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/ABI/testing/sysfs-bus-event_source-devices-fme
 create mode 100644 drivers/fpga/dfl-fme-perf.c
Signature · 1 line
1.8.3.1
author
#1 lore raw

PATCH v7 1/2 Documentation: fpga: dfl: add description for performance reporting support

From: Xu Yilun <[email protected]>

This patch adds description for performance reporting support for Device Feature List (DFL) based FPGA.

---
v5: rebase due to format change (txt->rst).

---
 Documentation/fpga/dfl.rst | 83 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 83 insertions(+)
Documentation/fpga/dfl.rst +83 -0
index 094fc8a..85b137b 100644
@@ -118,6 +118,11 @@ More functions are exposed through sysfs
      management information (current temperature, thresholds, threshold status,
      etc.).
 
+ Performance reporting
+     performance counters are exposed through perf PMU APIs. Standard perf tool
+     can be used to monitor all available perf events. Please see performance
+     counter section below for more detailed information.
+
 
 FIU - PORT
 ==========
@@ -378,6 +383,84 @@ 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
+implementation. e.g. some discrete FPGA cards don't have any cache. User could
+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/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
+the perf event belongs to. "event" is the identity of the event within its
+category. "portid" is introduced to decide counters set to monitor on FPGA
+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 fme
+
+  fme0/fab_mmio_read/                              [Kernel PMU event]
+  <...>
+  fme0/fab_port_mmio_read,portid=?/                [Kernel PMU event]
+  <...>
+
+  $# perf stat -a -e fme0/fab_mmio_read/ <command>
+  or
+  $# perf stat -a -e fme0/event=0x06,evtype=0x02,portid=0xff/ <command>
+  or
+  $# perf stat -a -e 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 fme0/fab_port_mmio_read,portid=0x0/ <command>
+  or
+  $# perf stat -a -e fme0/event=0x06,evtype=0x02,portid=0x0/ <command>
+  or
+  $# perf stat -a -e 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 fme0/fab_mmio_read/,fme0/fab_port_mmio_write,\
+                                                    portid=0/ sleep 1
+
+  Performance counter stats for 'system wide':
+
+                 3      fme0/fab_mmio_read/
+   <not supported>      fme0/fab_port_mmio_write,portid=0x0/
+
+       1.001750904 seconds time elapsed
+
+The driver also provides a "cpumask" sysfs attribute, which always shows fixed
+value cpu0 as all perf events are from 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
Signature · 1 line
1.8.3.1
author
#2 lore raw

PATCH v7 2/2 fpga: dfl: fme: add performance reporting support

In reply to Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

e.g. List all events by "perf list"

perf list | grep fme

fme0/cache_read_hit/ [Kernel PMU event]
fme0/cache_read_miss/ [Kernel PMU event]
...

fme0/fab_mmio_read/ [Kernel PMU event]
fme0/fab_mmio_write/ [Kernel PMU event]
...

fme0/fab_port_mmio_read,portid=?/ [Kernel PMU event]
fme0/fab_port_mmio_write,portid=?/ [Kernel PMU event]
...

fme0/vtd_port_devtlb_1g_fill,portid=?/ [Kernel PMU event]
fme0/vtd_port_devtlb_2m_fill,portid=?/ [Kernel PMU event]
...

fme0/vtd_sip_iotlb_1g_hit/ [Kernel PMU event]
fme0/vtd_sip_iotlb_1g_miss/ [Kernel PMU event]
...

fme0/clock [Kernel PMU event]
...

e.g. check increased counter value after run one application using "perf stat" command.

perf stat -e fme0/fab_mmio_read/,fme0/fab_mmio_write/ ./test

Performance counter stats for './test':

                 1      fme0/fab_mmio_read/
                 2      fme0/fab_mmio_write/

       1.009496520 seconds time elapsed

Please note that fabric counters support both fab_* and fab_port_*, but actually they are sharing one set of performance counters in hardware. If user wants to monitor overall data events on fab_* then fab_port_* can't be supported at the same time, see example below:

perf stat -e fme0/fab_mmio_read/,fme0/fab_port_mmio_write,portid=0/

Performance counter stats for 'system wide':

0 fme0/fab_mmio_read/
<not supported> fme0/fab_port_mmio_write,portid=0/

       2.141064085 seconds time elapsed

---
v3: replace scnprintf with sprintf in sysfs interfaces.
update sysfs doc kernel version and date.
fix sysfs doc issue for fabric counter.
refine PERF_OBJ_ATTR_* macro, doesn't count on __ATTR anymore. introduce PERF_OBJ_ATTR_F_* macro, as it needs to use different filenames for some of the sysfs attributes.
remove kobject_del when destroy kobject, kobject_put is enough. do sysfs_remove_groups first when destroying perf_obj. WARN_ON_ONCE in case internal parms are wrong in read_*_count().
v4: rework this patch to use standard perf API as user interfaces.
v5: rebase and clean up.
v6: use dev_ext_attribute instead of creating new fme_perf_attribute
use is_visible function to decide which events to expose per hardware capability, and add event_init checking for all events.
v7: fix a kbuild warning and add a sysfs documentation.

---
 .../ABI/testing/sysfs-bus-event_source-devices-fme | 105 +++
 drivers/fpga/Kconfig                               |   2 +-
 drivers/fpga/Makefile                              |   1 +
 drivers/fpga/dfl-fme-main.c                        |   4 +
 drivers/fpga/dfl-fme-perf.c                        | 943 +++++++++++++++++++++
 drivers/fpga/dfl-fme.h                             |   2 +
 drivers/fpga/dfl.h                                 |   2 +
 7 files changed, 1058 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/ABI/testing/sysfs-bus-event_source-devices-fme
 create mode 100644 drivers/fpga/dfl-fme-perf.c
Documentation/ABI/testing/sysfs-bus-event_source-devices-fme new file +105 -0
new file mode 100644
index 0000000..71bcd98
@@ -0,0 +1,105 @@
+What:		/sys/bus/event_source/devices/fmeX/format
+Date:		February 2020
+KernelVersion:  5.7
+Contact:	Wu Hao <[email protected]>
+Description:	Read-only. Attribute group to describe the magic bits
+		that go into perf_event_attr.config for a particular pmu.
+		(See ABI/testing/sysfs-bus-event_source-devices-format).
+
+		Each attribute under this group defines a bit range of the
+		perf_event_attr.config. All supported attributes are listed
+		below.
+
+		  event  = "config:0-11"  - event ID
+		  evtype = "config:12-15" - event type
+		  portid = "config:16-23" - event source
+
+		For example,
+
+		  fab_mmio_read = "event=0x06,evtype=0x02,portid=0xff"
+
+		It shows this fab_mmio_read is a fabric type (0x02) event with
+		0x06 local event id for overall monitoring (portid=0xff).
+
+What:		/sys/bus/event_source/devices/fmeX/cpumask
+Date:		February 2020
+KernelVersion:  5.7
+Contact:	Wu Hao <[email protected]>
+Description:	Read-only. This file always returns 0 as all fme pmu performance
+		monitoring events are not specific to any processor, just from
+		fme device.
+
+What:		/sys/bus/event_source/devices/fmeX/events
+Date:		February 2020
+KernelVersion:  5.7
+Contact:	Wu Hao <[email protected]>
+Description:	Read-only. Attribute group to describe performance monitoring
+		events specific to fme. Each attribute in this group describes
+		a single performance monitoring event supported by this fme pmu.
+		The name of the file is the name of the event.
+		(See ABI/testing/sysfs-bus-event_source-devices-events).
+
+		All supported performance monitoring events are listed below.
+
+		Basic events (evtype=0x00)
+
+		  clock = "event=0x00,evtype=0x00,portid=0xff"
+
+		Cache events (evtype=0x01)
+
+		  cache_read_hit      = "event=0x00,evtype=0x01,portid=0xff"
+		  cache_read_miss     = "event=0x01,evtype=0x01,portid=0xff"
+		  cache_write_hit     = "event=0x02,evtype=0x01,portid=0xff"
+		  cache_write_miss    = "event=0x03,evtype=0x01,portid=0xff"
+		  cache_hold_request  = "event=0x05,evtype=0x01,portid=0xff"
+		  cache_data_write_port_contention =
+		                        "event=0x06,evtype=0x01,portid=0xff"
+		  cache_tag_write_port_contention =
+		                        "event=0x07,evtype=0x01,portid=0xff"
+		  cache_tx_req_stall  = "event=0x08,evtype=0x01,portid=0xff"
+		  cache_rx_req_stall  = "event=0x09,evtype=0x01,portid=0xff"
+		  cache_eviction      = "event=0x0a,evtype=0x01,portid=0xff"
+
+		Fabric events (evtype=0x02)
+
+		  fab_pcie0_read       = "event=0x00,evtype=0x02,portid=0xff"
+		  fab_pcie0_write      = "event=0x01,evtype=0x02,portid=0xff"
+		  fab_pcie1_read       = "event=0x02,evtype=0x02,portid=0xff"
+		  fab_pcie1_write      = "event=0x03,evtype=0x02,portid=0xff"
+		  fab_upi_read         = "event=0x04,evtype=0x02,portid=0xff"
+		  fab_upi_write        = "event=0x05,evtype=0x02,portid=0xff"
+		  fab_mmio_read        = "event=0x06,evtype=0x02,portid=0xff"
+		  fab_mmio_write       = "event=0x07,evtype=0x02,portid=0xff"
+		  fab_port_pcie0_read  = "event=0x00,evtype=0x02,portid=?"
+		  fab_port_pcie0_write = "event=0x01,evtype=0x02,portid=?"
+		  fab_port_pcie1_read  = "event=0x02,evtype=0x02,portid=?"
+		  fab_port_pcie1_write = "event=0x03,evtype=0x02,portid=?"
+		  fab_port_upi_read    = "event=0x04,evtype=0x02,portid=?"
+		  fab_port_upi_write   = "event=0x05,evtype=0x02,portid=?"
+		  fab_port_mmio_read   = "event=0x06,evtype=0x02,portid=?"
+		  fab_port_mmio_write  = "event=0x07,evtype=0x02,portid=?"
+
+		VTD events (evtype=0x03)
+
+		  vtd_port_read_transaction  = "event=0x00,evtype=0x03,portid=?"
+		  vtd_port_write_transaction = "event=0x01,evtype=0x03,portid=?"
+		  vtd_port_devtlb_read_hit   = "event=0x02,evtype=0x03,portid=?"
+		  vtd_port_devtlb_write_hit  = "event=0x03,evtype=0x03,portid=?"
+		  vtd_port_devtlb_4k_fill    = "event=0x04,evtype=0x03,portid=?"
+		  vtd_port_devtlb_2m_fill    = "event=0x05,evtype=0x03,portid=?"
+		  vtd_port_devtlb_1g_fill    = "event=0x06,evtype=0x03,portid=?"
+
+		VTD SIP events (evtype=0x04)
+
+		  vtd_sip_iotlb_4k_hit  = "event=0x00,evtype=0x04,portid=0xff"
+		  vtd_sip_iotlb_2m_hit  = "event=0x01,evtype=0x04,portid=0xff"
+		  vtd_sip_iotlb_1g_hit  = "event=0x02,evtype=0x04,portid=0xff"
+		  vtd_sip_slpwc_l3_hit  = "event=0x03,evtype=0x04,portid=0xff"
+		  vtd_sip_slpwc_l4_hit  = "event=0x04,evtype=0x04,portid=0xff"
+		  vtd_sip_rcc_hit       = "event=0x05,evtype=0x04,portid=0xff"
+		  vtd_sip_iotlb_4k_miss = "event=0x06,evtype=0x04,portid=0xff"
+		  vtd_sip_iotlb_2m_miss = "event=0x07,evtype=0x04,portid=0xff"
+		  vtd_sip_iotlb_1g_miss = "event=0x08,evtype=0x04,portid=0xff"
+		  vtd_sip_slpwc_l3_miss = "event=0x09,evtype=0x04,portid=0xff"
+		  vtd_sip_slpwc_l4_miss = "event=0x0a,evtype=0x04,portid=0xff"
+		  vtd_sip_rcc_miss      = "event=0x0b,evtype=0x04,portid=0xff"
drivers/fpga/Kconfig +1 -1
index 72380e1..b2408a7 100644
@@ -156,7 +156,7 @@ config FPGA_DFL
 
 config FPGA_DFL_FME
 	tristate "FPGA DFL FME Driver"
-	depends on FPGA_DFL && HWMON
+	depends on FPGA_DFL && HWMON && PERF_EVENTS
 	help
 	  The FPGA Management Engine (FME) is a feature device implemented
 	  under Device Feature List (DFL) framework. Select this option to
drivers/fpga/Makefile +1 -0
index 4865b74..d8e21df 100644
@@ -40,6 +40,7 @@ obj-$(CONFIG_FPGA_DFL_FME_REGION)	+= dfl-fme-region.o
 obj-$(CONFIG_FPGA_DFL_AFU)		+= dfl-afu.o
 
 dfl-fme-objs := dfl-fme-main.o dfl-fme-pr.o dfl-fme-error.o
+dfl-fme-objs += dfl-fme-perf.o
 dfl-afu-objs := dfl-afu-main.o dfl-afu-region.o dfl-afu-dma-region.o
 dfl-afu-objs += dfl-afu-error.o
 
drivers/fpga/dfl-fme-main.c +4 -0
index 1d4690c..3533056 100644
@@ -580,6 +580,10 @@ static int fme_power_mgmt_init(struct platform_device *pdev,
 		.ops = &fme_power_mgmt_ops,
 	},
 	{
+		.id_table = fme_perf_id_table,
+		.ops = &fme_perf_ops,
+	},
+	{
 		.ops = NULL,
 	},
 };
drivers/fpga/dfl-fme-perf.c new file +943 -0

943 lines not on this page · raw message

drivers/fpga/dfl-fme.h +2 -0
index 6685c8e..4195dd6 100644
@@ -38,5 +38,7 @@ struct dfl_fme {
 extern const struct dfl_feature_ops fme_global_err_ops;
 extern const struct dfl_feature_id fme_global_err_id_table[];
 extern const struct attribute_group fme_global_err_group;
+extern const struct dfl_feature_ops fme_perf_ops;
+extern const struct dfl_feature_id fme_perf_id_table[];
 
 #endif /* __DFL_FME_H */
drivers/fpga/dfl.h +2 -0
index 9f0e656..d312678 100644
@@ -197,12 +197,14 @@ struct dfl_feature_driver {
  *		    feature dev (platform device)'s reources.
  * @ioaddr: mapped mmio resource address.
  * @ops: ops of this sub feature.
+ * @priv: priv data of this feature.
  */
 struct dfl_feature {
 	u64 id;
 	int resource_index;
 	void __iomem *ioaddr;
 	const struct dfl_feature_ops *ops;
+	void *priv;
 };
 
 #define DEV_STATUS_IN_USE	0
Signature · 1 line
1.8.3.1
#3 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

- I would still prefer for the PMU drivers to live under drivers/perf/

- You should probably give the PMU a better name than "fme%d", for example

    "intel_fpga_dfl_fme%d".

- CPU0 can be hotplugged off on non-x86 systems. How do you cope with that?

- readq() will emit 2x32-bit reads on some architectures. What happens in this case?

Will

author
#4 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 11 lines · Will Deacon

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

Hi Will

Thanks a lot for the review! :)

Quoted 2 lines

- I would still prefer for the PMU drivers to live under drivers/perf/

Hm.. one possible way is to create a platform device, and introduce a new platform device driver under drivers/perf/.

Quoted 3 lines

- You should probably give the PMU a better name than "fme%d", for example

    "intel_fpga_dfl_fme%d".

Sure. as it's not intel only, so will use "fpga_dfl_fme" or "dfl_fme" instead.

Quoted 3 lines

- CPU0 can be hotplugged off on non-x86 systems. How do you cope with that?

Oh.. then we can't just always return CPU0 for cpumask, have to switch to another available cpu. Will add code to handle this case in the next version. Thanks!

Quoted 3 lines

- readq() will emit 2x32-bit reads on some architectures. What happens in this case?

64bit counter may increase and carry out of bit31 between 2-32bit reads, we probably need some more code (add extra reads?) to handle this case.

Thanks again for these valuable comments, will try to fix all of them in the next version. : )

Hao

Quoted 2 lines

Will

#5 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 15 lines · Wu Hao
Quoted 11 lines · Will Deacon

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

Hi Will

Quoted 7 lines

Thanks a lot for the review! :)

Quoted 2 lines

- I would still prefer for the PMU drivers to live under drivers/perf/

Hm.. one possible way is to create a platform device, and introduce a new platform device driver under drivers/perf/.

No, do not abuse platform drivers, you have a real device, use it.

author
#6 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 19 lines · Greg KH
Quoted 18 lines · Wu Hao
Quoted 11 lines · Will Deacon

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

Hi Will

Thanks a lot for the review! :)

Quoted 1 line
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

- I would still prefer for the PMU drivers to live under drivers/perf/

Hm.. one possible way is to create a platform device, and introduce a new platform device driver under drivers/perf/.

No, do not abuse platform drivers, you have a real device, use it.

Sure, thanks for the comments. Then I don't have any other idea to move code to drivers/perf/ directory, so probably only can live with current code.

Thanks
Hao

#7 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 22 lines · Wu Hao
Quoted 21 lines · Greg KH
Quoted 20 lines · Wu Hao
Quoted 11 lines · Will Deacon

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

Hi Will

Thanks a lot for the review! :)

Quoted 2 lines

- I would still prefer for the PMU drivers to live under drivers/perf/

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Hm.. one possible way is to create a platform device, and introduce a new platform device driver under drivers/perf/.

No, do not abuse platform drivers, you have a real device, use it.

Sure, thanks for the comments. Then I don't have any other idea to move code to drivers/perf/ directory, so probably only can live with current code.

The location of the file in the kernel tree has no bearing on if you use a platform device, a USB device, or a PCI device. It is just a location of a file.

You are interacting with the perf api as the driver's primary userspace api, so put the driver into the drivers/perf/ directory. That's all that Will is asking you to do here.

greg k-h

author
#8 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 31 lines · Greg KH
Quoted 22 lines · Wu Hao
Quoted 21 lines · Greg KH
Quoted 20 lines · Wu Hao
Quoted 11 lines · Will Deacon

Hi,

Quoted 5 lines · Wu Hao

This patch adds support for performance reporting private feature for FPGA Management Engine (FME). Now it supports several different performance counters, including 'basic', 'cache', 'fabric', 'vtd' and 'vtd_sip'. It allows user to use standard linux tools to access these performance counters.

I had a quick look at this, and it mostly looks alright to me. Just a few high-level comments/questions:

Hi Will

Thanks a lot for the review! :)

Quoted 2 lines

- I would still prefer for the PMU drivers to live under drivers/perf/

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Hm.. one possible way is to create a platform device, and introduce a new platform device driver under drivers/perf/.

No, do not abuse platform drivers, you have a real device, use it.

Sure, thanks for the comments. Then I don't have any other idea to move code to drivers/perf/ directory, so probably only can live with current code.

Quoted 7 lines

The location of the file in the kernel tree has no bearing on if you use a platform device, a USB device, or a PCI device. It is just a location of a file.

You are interacting with the perf api as the driver's primary userspace api, so put the driver into the drivers/perf/ directory. That's all that Will is asking you to do here.

Thanks a lot for the patient explanation. : )

Actually this patch only adds a new file to existing fme platform driver, fme platform driver already has several different types userspace interfaces, including hwmon, sysfs and etc, so the new perf api is only one of them, as we can't move the whole fme platform driver into drivers/perf/, we feel that maybe we should keep that code together with other fme files in drivers/fpga/.

Thanks
Hao

Quoted 2 lines

greg k-h

#9 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

In reply to Wu Hao
Quoted 9 lines · Wu Hao
Documentation/ABI/testing/sysfs-bus-event_source-devices-fme new file +9 -0
+What:		/sys/bus/event_source/devices/fmeX/format
+Date:		February 2020
+KernelVersion:  5.7
+Contact:	Wu Hao <[email protected]>
+Description:	Read-only. Attribute group to describe the magic bits
+		that go into perf_event_attr.config for a particular pmu.
+		(See ABI/testing/sysfs-bus-event_source-devices-format).
+
+		Each attribute under this group defines a bit range of the

Documentation/ABI/testing/sysfs-bus-event_source-devices-fme line 19

Quoted 10 lines
Documentation/ABI/testing/sysfs-bus-event_source-devices-fme new file +10 -0
+		perf_event_attr.config. All supported attributes are listed
+		below.
+
+		  event  = "config:0-11"  - event ID
+		  evtype = "config:12-15" - event type
+		  portid = "config:16-23" - event source
+
+		For example,
+
+		  fab_mmio_read = "event=0x06,evtype=0x02,portid=0xff"

Are perf sysfs files always this bad "multiple values per file"? Or is that unique to this driver? If not unique, do you have specific examples in the kernel that currently do this today?

Review

Quoted 2 lines
(quoted diff) +2 -0
+static struct attribute *fme_perf_events_attrs_empty[] = {
+	NULL,

Huh?

drivers/fpga/dfl-fme-perf.c line 229

Quoted 5 lines
drivers/fpga/dfl-fme-perf.c new file +5 -0
+};
+
+static struct attribute_group fme_perf_events_group = {
+	.name = "events",
+	.attrs = fme_perf_events_attrs_empty,

You create an empty directory? Why? What goes in here?

very odd...

greg k-h

author
#10 lore raw

Re: [PATCH v7 2/2] fpga: dfl: fme: add performance reporting support

Quoted 16 lines · Greg KH
Quoted 15 lines · Wu Hao
Documentation/ABI/testing/sysfs-bus-event_source-devices-fme new file +15 -0
+What:		/sys/bus/event_source/devices/fmeX/format
+Date:		February 2020
+KernelVersion:  5.7
+Contact:	Wu Hao <[email protected]>
+Description:	Read-only. Attribute group to describe the magic bits
+		that go into perf_event_attr.config for a particular pmu.
+		(See ABI/testing/sysfs-bus-event_source-devices-format).
+
+		Each attribute under this group defines a bit range of the
+		perf_event_attr.config. All supported attributes are listed
+		below.
+
+		  event  = "config:0-11"  - event ID
+		  evtype = "config:12-15" - event type
+		  portid = "config:16-23" - event source
Quoted 8 lines

Documentation/ABI/testing/sysfs-bus-event_source-devices-fme line 19

Quoted 4 lines
Documentation/ABI/testing/sysfs-bus-event_source-devices-fme new file +4 -0
+
+		For example,
+
+		  fab_mmio_read = "event=0x06,evtype=0x02,portid=0xff"

Are perf sysfs files always this bad "multiple values per file"? Or is that unique to this driver? If not unique, do you have specific examples in the kernel that currently do this today?

Hi Greg,

Thanks a lot for the review. : )

Perf sysfs files allow this kind of output, so some perf drivers are using the similar format for their jobs.

Examples from my machine.

 # cat /sys/bus/event_source/devices/cpu/events/cycles-ct
 event=0x3c,in_tx=1,in_tx_cp=1
 # cat /sys/bus/event_source/devices/cpu/events/el-start
 event=0xc8,umask=0x1
 # cat /sys/bus/event_source/devices/cpu/events/instructions
 event=0xc0
 # cat /sys/bus/event_source/devices/cpu/events/branch-instructions
 event=0xc4

See arch/x86/events/intel/core.c

EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1");
...

And descriptions/examples from ABI/testing/sysfs-bus-event_source-devices-events

What: /sys/bus/event_source/devices/<pmu>/events/<event>
Date: 2014/02/24
Contact: Linux kernel mailing list <[email protected]>
Description: Per-pmu performance monitoring events specific to the running system

Each file (except for some of those with a '.' in them, '.unit' and '.scale') in the 'events' directory describes a single performance monitoring event supported by the <pmu>. The name of the file is the name of the event.

File contents:

<term>[=<value>][,<term>[=<value>]]...

Where <term> is one of the terms listed under /sys/bus/event_source/devices/<pmu>/format/ and <value> is a number is base-16 format with a '0x' prefix (lowercase only). If a <term> is specified alone (without an assigned value), it is implied that 0x1 is assigned to that <term>.

Examples (each of these lines would be in a seperate file):

event=0x2abc event=0x423,inv,cmask=0x3 domain=0x1,offset=0x8,starting_index=0xffff domain=0x1,offset=0x8,core=?

Each of the assignments indicates a value to be assigned to a particular set of bits (as defined by the format file corresponding to the <term>) in the perf_event structure passed to the perf_open syscall.

In the case of the last example, a value replacing "?" would need to be provided by the user selecting the particular event. This is referred to as "event parameterization". Event parameters have the format 'param=?'.

So this is not something new introduced by this patch.

Quoted 15 lines

Review

Quoted 4 lines
(quoted diff) +2 -0
+static struct attribute *fme_perf_events_attrs_empty[] = {
+	NULL,

Huh?

drivers/fpga/dfl-fme-perf.c line 229

Quoted 5 lines
drivers/fpga/dfl-fme-perf.c new file +5 -0
+};
+
+static struct attribute_group fme_perf_events_group = {
+	.name = "events",
+	.attrs = fme_perf_events_attrs_empty,

You create an empty directory? Why? What goes in here?

Quoted 1 line

very odd...

Actually events are filled into this "events" from several different groups via pmu->attr_update[1].

	pmu->attr_update =      fme_perf_events_groups;

pmu->attr_update allows us to update "events" directories with attributes that depend on various HW conditions. In our case, several different groups with different is_visible functions are filled into "events" using this method. And several existing pmu drivers (e.g. arch/x86/events/intel/cstate.c) are using the same way (having an empty directory first and update it using pmu->attr_update).

But I have to admit that I should add some comments there to avoid confusion, sorry, will do that in the next version.

[1] https://lkml.org/lkml/2019/5/4/188

Thanks
Hao

Quoted 2 lines

greg k-h