Re: [PATCH v3 16/16] fpga: dfl: fme: add performance reporting support

From: Wu Hao
Date: Sat Jun 01 2019 - 05:31:53 EST


On Thu, May 30, 2019 at 12:03:05PM -0700, Greg KH wrote:
> On Mon, May 27, 2019 at 01:22:26PM +0800, Wu Hao wrote:
> > --- /dev/null
> > +++ b/drivers/fpga/dfl-fme-perf.c
> > @@ -0,0 +1,962 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * Driver for FPGA Management Engine (FME) Global Performance Reporting
> > + *
> > + * Copyright 2019 Intel Corporation, Inc.
> > + *
> > + * Authors:
> > + * Kang Luwei <luwei.kang@xxxxxxxxx>
> > + * Xiao Guangrong <guangrong.xiao@xxxxxxxxxxxxxxx>
> > + * Wu Hao <hao.wu@xxxxxxxxx>
> > + * Joseph Grecco <joe.grecco@xxxxxxxxx>
> > + * Enno Luebbers <enno.luebbers@xxxxxxxxx>
> > + * Tim Whisonant <tim.whisonant@xxxxxxxxx>
> > + * Ananda Ravuri <ananda.ravuri@xxxxxxxxx>
> > + * Mitchel, Henry <henry.mitchel@xxxxxxxxx>
> > + */
> > +
> > +#include "dfl.h"
> > +#include "dfl-fme.h"
> > +
> > +/*
> > + * Performance Counter Registers for Cache.
> > + *
> > + * Cache Events are listed below as CACHE_EVNT_*.
> > + */
> > +#define CACHE_CTRL 0x8
> > +#define CACHE_RESET_CNTR BIT_ULL(0)
> > +#define CACHE_FREEZE_CNTR BIT_ULL(8)
> > +#define CACHE_CTRL_EVNT GENMASK_ULL(19, 16)
> > +#define CACHE_EVNT_RD_HIT 0x0
> > +#define CACHE_EVNT_WR_HIT 0x1
> > +#define CACHE_EVNT_RD_MISS 0x2
> > +#define CACHE_EVNT_WR_MISS 0x3
> > +#define CACHE_EVNT_RSVD 0x4
> > +#define CACHE_EVNT_HOLD_REQ 0x5
> > +#define CACHE_EVNT_DATA_WR_PORT_CONTEN 0x6
> > +#define CACHE_EVNT_TAG_WR_PORT_CONTEN 0x7
> > +#define CACHE_EVNT_TX_REQ_STALL 0x8
> > +#define CACHE_EVNT_RX_REQ_STALL 0x9
> > +#define CACHE_EVNT_EVICTIONS 0xa
> > +#define CACHE_EVNT_MAX CACHE_EVNT_EVICTIONS
> > +#define CACHE_CHANNEL_SEL BIT_ULL(20)
> > +#define CACHE_CHANNEL_RD 0
> > +#define CACHE_CHANNEL_WR 1
> > +#define CACHE_CHANNEL_MAX 2
> > +#define CACHE_CNTR0 0x10
> > +#define CACHE_CNTR1 0x18
> > +#define CACHE_CNTR_EVNT_CNTR GENMASK_ULL(47, 0)
> > +#define CACHE_CNTR_EVNT GENMASK_ULL(63, 60)
> > +
> > +/*
> > + * Performance Counter Registers for Fabric.
> > + *
> > + * Fabric Events are listed below as FAB_EVNT_*
> > + */
> > +#define FAB_CTRL 0x20
> > +#define FAB_RESET_CNTR BIT_ULL(0)
> > +#define FAB_FREEZE_CNTR BIT_ULL(8)
> > +#define FAB_CTRL_EVNT GENMASK_ULL(19, 16)
> > +#define FAB_EVNT_PCIE0_RD 0x0
> > +#define FAB_EVNT_PCIE0_WR 0x1
> > +#define FAB_EVNT_PCIE1_RD 0x2
> > +#define FAB_EVNT_PCIE1_WR 0x3
> > +#define FAB_EVNT_UPI_RD 0x4
> > +#define FAB_EVNT_UPI_WR 0x5
> > +#define FAB_EVNT_MMIO_RD 0x6
> > +#define FAB_EVNT_MMIO_WR 0x7
> > +#define FAB_EVNT_MAX FAB_EVNT_MMIO_WR
> > +#define FAB_PORT_ID GENMASK_ULL(21, 20)
> > +#define FAB_PORT_FILTER BIT_ULL(23)
> > +#define FAB_PORT_FILTER_DISABLE 0
> > +#define FAB_PORT_FILTER_ENABLE 1
> > +#define FAB_CNTR 0x28
> > +#define FAB_CNTR_EVNT_CNTR GENMASK_ULL(59, 0)
> > +#define FAB_CNTR_EVNT GENMASK_ULL(63, 60)
> > +
> > +/*
> > + * Performance Counter Registers for Clock.
> > + *
> > + * Clock Counter can't be reset or frozen by SW.
> > + */
> > +#define CLK_CNTR 0x30
> > +
> > +/*
> > + * Performance Counter Registers for IOMMU / VT-D.
> > + *
> > + * VT-D Events are listed below as VTD_EVNT_* and VTD_SIP_EVNT_*
> > + */
> > +#define VTD_CTRL 0x38
> > +#define VTD_RESET_CNTR BIT_ULL(0)
> > +#define VTD_FREEZE_CNTR BIT_ULL(8)
> > +#define VTD_CTRL_EVNT GENMASK_ULL(19, 16)
> > +#define VTD_EVNT_AFU_MEM_RD_TRANS 0x0
> > +#define VTD_EVNT_AFU_MEM_WR_TRANS 0x1
> > +#define VTD_EVNT_AFU_DEVTLB_RD_HIT 0x2
> > +#define VTD_EVNT_AFU_DEVTLB_WR_HIT 0x3
> > +#define VTD_EVNT_DEVTLB_4K_FILL 0x4
> > +#define VTD_EVNT_DEVTLB_2M_FILL 0x5
> > +#define VTD_EVNT_DEVTLB_1G_FILL 0x6
> > +#define VTD_EVNT_MAX VTD_EVNT_DEVTLB_1G_FILL
> > +#define VTD_CNTR 0x40
> > +#define VTD_CNTR_EVNT GENMASK_ULL(63, 60)
> > +#define VTD_CNTR_EVNT_CNTR GENMASK_ULL(47, 0)
> > +#define VTD_SIP_CTRL 0x48
> > +#define VTD_SIP_RESET_CNTR BIT_ULL(0)
> > +#define VTD_SIP_FREEZE_CNTR BIT_ULL(8)
> > +#define VTD_SIP_CTRL_EVNT GENMASK_ULL(19, 16)
> > +#define VTD_SIP_EVNT_IOTLB_4K_HIT 0x0
> > +#define VTD_SIP_EVNT_IOTLB_2M_HIT 0x1
> > +#define VTD_SIP_EVNT_IOTLB_1G_HIT 0x2
> > +#define VTD_SIP_EVNT_SLPWC_L3_HIT 0x3
> > +#define VTD_SIP_EVNT_SLPWC_L4_HIT 0x4
> > +#define VTD_SIP_EVNT_RCC_HIT 0x5
> > +#define VTD_SIP_EVNT_IOTLB_4K_MISS 0x6
> > +#define VTD_SIP_EVNT_IOTLB_2M_MISS 0x7
> > +#define VTD_SIP_EVNT_IOTLB_1G_MISS 0x8
> > +#define VTD_SIP_EVNT_SLPWC_L3_MISS 0x9
> > +#define VTD_SIP_EVNT_SLPWC_L4_MISS 0xa
> > +#define VTD_SIP_EVNT_RCC_MISS 0xb
> > +#define VTD_SIP_EVNT_MAX VTD_SIP_EVNT_RCC_MISS
> > +#define VTD_SIP_CNTR 0X50
> > +#define VTD_SIP_CNTR_EVNT GENMASK_ULL(63, 60)
> > +#define VTD_SIP_CNTR_EVNT_CNTR GENMASK_ULL(47, 0)
> > +
> > +#define PERF_OBJ_ROOT_ID (~0)
> > +
> > +#define PERF_TIMEOUT 30
> > +
> > +/**
> > + * struct perf_object - object of performance counter
> > + *
> > + * @id: instance id. PERF_OBJ_ROOT_ID indicates it is a parent object which
> > + * counts performance counters for all instances.
> > + * @attr_groups: the sysfs files are associated with this object.
> > + * @feature: pointer to related private feature.
> > + * @node: used to link itself to parent's children list.
> > + * @children: used to link its children objects together.
> > + * @kobj: generic kobject interface.
> > + *
> > + * 'node' and 'children' are used to construct parent-children hierarchy.
> > + */
> > +struct perf_object {
> > + int id;
> > + const struct attribute_group **attr_groups;
> > + struct dfl_feature *feature;
> > +
> > + struct list_head node;
> > + struct list_head children;
> > + struct kobject kobj;
>
> Woah, why are you using a "raw" kobject and not a 'struct device' here?
> You just broke userspace and no libraries will see your kobject's
> properties as the "chain" of struct devices is not happening anymore.
>
> Why can this not just be a 'struct device'?

Hi Greg,

Many thanks for the review and comments.

Actually we are just trying to create sysfs hierarchy for performance
counters using these data structures.

If we use 'struct device' instead of kobject, then we have to let userspace
code to deal with device's sysfs (e.g. ignore 'uevent' below). This is the
only concern from my side now, as I know that using 'struct device'
saves code as we don't need to introduce a new perf_obj_attribute then.

dfl-fme.0/perf/
âââ iommu
â âââ afu0
â â âââ devtlb_1g_fill
â â âââ devtlb_2m_fill
â â âââ devtlb_4k_fill
â â âââ devtlb_read_hit
â â âââ devtlb_write_hit
â â âââ read_transaction
â â âââ uevent
â â âââ write_transaction
â âââ freeze
â âââ iotlb_1g_hit
â âââ iotlb_1g_miss
...
ÂÂ âââ uevent
...

Do you think if we could keep it or it's better to use 'struct device'?

>
>
> > +};
> > +
> > +/**
> > + * struct perf_obj_attribute - attribute of perf object
> > + *
> > + * @attr: attribute of this perf object.
> > + * @show: show callback for sysfs attribute.
> > + * @store: store callback for sysfs attribute.
> > + */
> > +struct perf_obj_attribute {
> > + struct attribute attr;
> > + ssize_t (*show)(struct perf_object *pobj, char *buf);
> > + ssize_t (*store)(struct perf_object *pobj,
> > + const char *buf, size_t n);
> > +};
> > +
> > +#define to_perf_obj_attr(_attr) \
> > + container_of(_attr, struct perf_obj_attribute, attr)
> > +#define to_perf_obj(_kobj) \
> > + container_of(_kobj, struct perf_object, kobj)
> > +
> > +#define __POBJ_ATTR(_name, _mode, _show, _store) { \
> > + .attr = {.name = __stringify(_name), \
> > + .mode = VERIFY_OCTAL_PERMISSIONS(_mode) }, \
> > + .show = _show, \
> > + .store = _store, \
> > +}
> > +
> > +#define PERF_OBJ_ATTR_F_RO(_name, _filename) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_filename, 0444, _name##_show, NULL)
> > +
> > +#define PERF_OBJ_ATTR_F_WO(_name, _filename) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_filename, 0200, NULL, _name##_store)
> > +
> > +#define PERF_OBJ_ATTR_F_RW(_name, _filename) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_filename, 0644, _name##_show, _name##_store)
> > +
> > +#define PERF_OBJ_ATTR_RO(_name) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_name, 0444, _name##_show, NULL)
> > +
> > +#define PERF_OBJ_ATTR_WO(_name) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_name, 0200, NULL, _name##_store)
> > +
> > +#define PERF_OBJ_ATTR_RW(_name) \
> > +struct perf_obj_attribute perf_obj_attr_##_name = \
> > + __POBJ_ATTR(_name, 0644, _name##_show, _name##_store)
>
> When you have to roll your own sysfs attributes for a single driver,
> that is a HUGE hint you are doing something wrong. No driver for an
> individual device should EVER have to do this.
>
> Please use the driver core properly and do not route around it.

Sure. Thanks!

Hao

>
> thanks,
>
> greg k-h