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

From: Wu Hao
Date: Tue Jun 04 2019 - 11:07:48 EST


On Sat, Jun 01, 2019 at 02:42:19AM -0700, Greg KH wrote:
> On Sat, Jun 01, 2019 at 05:11:47PM +0800, Wu Hao wrote:
> > 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'?
>
> What about using the attribute group name? That gives you a subdir for
> free. Doing anything "deeper" than one level means that you really have
> a child device, and yes, you need to use a 'struct device'. Make it
> part of your bus and just have it be a different "type" and all should
> be good.
>
> Again, NEVER use a raw kobject as a child of a 'struct device', that
> will break things.
>
> And please cc: me on this series from now on, as you are obviously
> trying to do complex things with the driver model and sysfs and it is
> easy to get very wrong.
>
> But wait, step back, why does this one driver have such a "special"
> user/kernel api that unique to it and nothing else? That's also a big
> red flag, why not just use the normal perf api that everyone else uses?
>

Hi Greg,

Actually we just tried to add some interfaces for users to read hardware
counters, Yes, I fully agree that we should use the existing apis whenever
possible. I will look into perf api to see if we can use them instead.

Many thanks for the review and comments!

Hao

> thanks,
>
> greg k-h