Date:   Wed, 7 Mar 2018 12:39:20 +0800
From:   Wu Hao <hao.wu@...el.com>
To:     Alan Tull <atull@...nel.org>
Cc:     Moritz Fischer <mdf@...nel.org>, linux-fpga@...r.kernel.org,
        linux-kernel <linux-kernel@...r.kernel.org>,
        linux-api@...r.kernel.org, "Kang, Luwei" <luwei.kang@...el.com>,
        "Zhang, Yi Z" <yi.z.zhang@...el.com>,
        Tim Whisonant <tim.whisonant@...el.com>,
        Enno Luebbers <enno.luebbers@...el.com>,
        Shiva Rao <shiva.rao@...el.com>,
        Christopher Rauer <christopher.rauer@...el.com>,
        Xiao Guangrong <guangrong.xiao@...ux.intel.com>
Subject: Re: [PATCH v4 14/24] fpga: dfl: fme: add partial reconfiguration sub
 feature support

On Tue, Mar 06, 2018 at 12:29:35PM -0600, Alan Tull wrote:
>   On Mon, Mar 5, 2018 at 8:08 PM, Wu Hao <hao.wu@...el.com> wrote:
> > On Mon, Mar 05, 2018 at 04:46:02PM -0600, Alan Tull wrote:
> >> On Tue, Feb 13, 2018 at 3:24 AM, Wu Hao <hao.wu@...el.com> wrote:
> >>
> >> Hi Hao,
> >
> > Hi Alan,
> >
> > Thanks for the comments.
> >
> >>
> >> We are going to want to be able use different FPGA managers with this
> >> framework.  The different manager may be part of a different FME in
> >> fabric or it may be a hardware FPGA manager.  Fortunately, at this
> >> point now the changes, noted below, to get there are pretty small.
> >
> > Yes, understand these could be some cases that FME having different PR
> > hardware.
> >
> 
> Or supporting reduced FME plus hardware-based FPGA manager.
> 
> Just to re-emphasize, the basic intent of the FPGA manager subsystem
> in the first place is to have FPGA managers separate from higher level
> frameworks so that the higher level frameworks will be able to able to
> use different FPGAs.
> 
> >> > +/**
> >> > + * fpga_fme_create_mgr - create fpga mgr platform device as child device
> >> > + *
> >> > + * @pdata: fme platform_device's pdata
> >> > + *
> >> > + * Return: mgr platform device if successful, and error code otherwise.
> >> > + */
> >> > +static struct platform_device *
> >> > +fpga_fme_create_mgr(struct feature_platform_data *pdata)
> >> > +{
> >> > +       struct platform_device *mgr, *fme = pdata->dev;
> >> > +       struct feature *feature;
> >> > +       struct resource res;
> >> > +       struct resource *pres;
> >> > +       int ret = -ENOMEM;
> >> > +
> >> > +       feature = get_feature_by_id(&pdata->dev->dev, FME_FEATURE_ID_PR_MGMT);
> >> > +       if (!feature)
> >> > +               return ERR_PTR(-ENODEV);
> >> > +
> >> > +       /*
> >> > +        * Each FME has only one fpga-mgr, so allocate platform device using
> >> > +        * the same FME platform device id.
> >> > +        */
> >> > +       mgr = platform_device_alloc(FPGA_DFL_FME_MGR, fme->id);
> >>
> >> At this point, the framework is assuming all FME's include the same
> >> FPGA manager device which would use the driver in dfl-fme-mgr.c.
> >>
> >> I'm thinking of two cases where the manager isn't the same as a
> >> dfl-fme-mgr.c manager are a bit different:
> >>
> >> (1) a FME-based FPGA manager, but different implementation, different
> >> registers.  The constraint is that the port implementation has to be
> >> similar enough to use the rest of the base FME code.   I am wondering
> >> if the FPGA manager can be added to the DFL.  At this point, the DFL
> >> would drive which FPGA manager is alloc'd.   That way the user gets to
> >> use all this code in dfl-fme-pr.c but with their FPGA manager.
> >
> > Actually I'm not sure how this will be implemented in the hardware in the
> > future, but from my understanding, there may be several methods to add this
> > support (a different PR hardware) to FME.
> >
> > 1) introduce a new PR sub feature to FME.
> >    driver knows it by different feature id, and create different fpga
> >    manager platform device, but may not be able to reuse dfl-fme-pr.c.
> 
> What would prevent reusing dfl-fme-pr.c?  It looks like this is 98% of
> the way there and only needs a way of knowing which FPGA manager
> driver to alloc here. I am hoping that a new PR sub feature could be
> added and that dfl-fme-pr.c can be reused.

It depeneds on how hardware implements it. : )
I agree that if it follows the same way of current PR sub feature, then we
could reuse the dfl-fme-pr.c for sure.

> 
> >
> > 2) add a different PR hardware support to current PR sub feature.
> >    It requires hardware to add registers to indicate this is a different
> >    PR hardware than dfl-fme-mgr.c, and its register region information
> >    (e.g location and size). Then this dfl-fme-pr.c driver could read
> >    these information from hardware and create a different fpga manager
> >    platform device with correct resources.
> 
> So this dfl-fme-pr.c would have to know where some ID registers are
> and the enumeration gets messier.  Some of the enumeration would be
> DFL and some would be from information that is not in the DFL headers.
> The DFL should contain the knowledge of which mgr to use.

Actually I don't know how hardware will implement this in the future, but I
just listed my ideas here. Per my understanding, driver (reuse dfl-fme-pr.c)
needs some more information to decide which platform device to create (for
fpga manager).

1) introduce a new PR sub feature. Then it has a different private feature
id in DFH (Device Feature Header). driver could use this id to create a
different platform device.

2) introduce some registers inside the current PR sub feature. Then driver
could read these registers to know which platform device to create for fpga
manager.

I think either 1) or 2) will only require small changes to current driver
code, I don't have any concern on supporting different PR hardware. :)

I understand your concern on case 2), everybody who wants to reuse dfl-fme-pr.c
needs to implement ID register which is not defined by DFL, so I guess we
should suggest 1).

> 
> >
> > I think current driver framework could support both cases above for sure.
> >
> >>
> >> (2) a FPGA manager that can be added by device tree in the case of a
> >> platform that is using device tree.  I think this will be pretty
> >> simple and can be done later when someone is actually bringing this
> >> framework up on a FPGA running under device tree.  I'm thinking that
> >> the base DFL device that reads the dfl data from hardware can have a
> >> DT property that points to the FPGA manager.  That manager can be
> >> saved somewhere handy like the pdata and passed down to this code,
> >> which realizes it can use that existing device and doesn't need to
> >> alloc a platform device.  But again, that's probably best done later.
> >
> > Sure, we can discuss further when we really need it. Actually per my
> > understanding, if hardware describes itself clearly, we may not have to
> > use DT for fpga manager, as driver could create it automatically based
> > on information read from hardware. :)
> 
> DT exists for busses that don't have that kind of discovery.  For a
> concrete example, consider how the Arria10 driver (socfpga-a10.c)
> probe function is getting its two mmio spaces and clock.

I see. If we have to create fpga mgr using DT some cases, it makes sense
that we just link the existing one instead.

> 
> >
> >>
> >> > +       if (!mgr)
> >> > +               return ERR_PTR(ret);
> >> > +
> >> > +       mgr->dev.parent = &fme->dev;
> >> > +
> >> > +       pres = platform_get_resource(fme, IORESOURCE_MEM,
> >> > +                                    feature->resource_index);
> >> > +       if (!pres) {
> >> > +               ret = -ENODEV;
> >> > +               goto create_mgr_err;
> >> > +       }
> >> > +
> >> > +       memset(&res, 0, sizeof(struct resource));
> >> > +
> >> > +       res.start = pres->start;
> >> > +       res.end = pres->end;
> >> > +       res.name = pres->name;
> >> > +       res.flags = IORESOURCE_MEM;
> >> > +
> >> > +       ret = platform_device_add_resources(mgr, &res, 1);
> >> > +       if (ret)
> >> > +               goto create_mgr_err;
> >> > +
> >> > +       ret = platform_device_add(mgr);
> >> > +       if (ret)
> >> > +               goto create_mgr_err;
> >> > +
> >> > +       return mgr;
> >> > +
> >> > +create_mgr_err:
> >> > +       platform_device_put(mgr);
> >> > +       return ERR_PTR(ret);
> >> > +}
> 
> >> > diff --git a/drivers/fpga/dfl-fme-pr.h b/drivers/fpga/dfl-fme-pr.h
> >> > new file mode 100644
> >> > index 0000000..11bd001
> >> > --- /dev/null
> >> > +++ b/drivers/fpga/dfl-fme-pr.h
> >> > @@ -0,0 +1,113 @@
> >> > +/* SPDX-License-Identifier: GPL-2.0 */
> >> > +/*
> >> > + * Header file for FPGA Management Engine (FME) Partial Reconfiguration Driver
> >> > + *
> >> > + * Copyright (C) 2017 Intel Corporation, Inc.
> >> > + *
> >> > + * Authors:
> >> > + *   Kang Luwei <luwei.kang@...el.com>
> >> > + *   Xiao Guangrong <guangrong.xiao@...ux.intel.com>
> >> > + *   Wu Hao <hao.wu@...el.com>
> >> > + *   Joseph Grecco <joe.grecco@...el.com>
> >> > + *   Enno Luebbers <enno.luebbers@...el.com>
> >> > + *   Tim Whisonant <tim.whisonant@...el.com>
> >> > + *   Ananda Ravuri <ananda.ravuri@...el.com>
> >> > + *   Henry Mitchel <henry.mitchel@...el.com>
> >> > + */
> >> > +
> >> > +#ifndef __DFL_FME_PR_H
> >> > +#define __DFL_FME_PR_H
> >> > +
> >> > +#include <linux/platform_device.h>
> >> > +
> >> > +/**
> >> > + * struct fme_region - FME fpga region data structure
> >> > + *
> >> > + * @region: platform device of the FPGA region.
> >> > + * @node: used to link fme_region to a list.
> >> > + * @port_id: indicate which port this region connected to.
> >> > + */
> >> > +struct fme_region {
> >> > +       struct platform_device *region;
> >> > +       struct list_head node;
> >> > +       int port_id;
> >> > +};
> >> > +
> >> > +/**
> >> > + * struct fme_region_pdata - platform data for FME region platform device.
> >> > + *
> >> > + * @mgr: platform device of the FPGA manager.
> >> > + * @br: platform device of the FPGA bridge.
> >> > + * @region_id: region id (same as port_id).
> >> > + */
> >> > +struct fme_region_pdata {
> >> > +       struct platform_device *mgr;
> >> > +       struct platform_device *br;
> >> > +       int region_id;
> >> > +};
> >> > +
> >> > +/**
> >> > + * struct fme_bridge - FME fpga bridge data structure
> >> > + *
> >> > + * @br: platform device of the FPGA bridge.
> >> > + * @node: used to link fme_bridge to a list.
> >> > + */
> >> > +struct fme_bridge {
> >> > +       struct platform_device *br;
> >> > +       struct list_head node;
> >> > +};
> >> > +
> >> > +/**
> >> > + * struct fme_bridge_pdata - platform data for FME bridge platform device.
> >> > + *
> >> > + * @port: platform device of the port feature dev.
> >> > + */
> >> > +struct fme_br_pdata {
> >> > +       struct platform_device *port;
> >> > +};
> >> > +
> >> > +#define FPGA_DFL_FME_MGR       "dfl-fme-mgr"
> >> > +#define FPGA_DFL_FME_BRIDGE    "dfl-fme-bridge"
> >> > +#define FPGA_DFL_FME_REGION    "dfl-fme-region"
> >> > +
> >>
> >> Everything in dfl-fme-pr.h up to this point is good and general and
> >> should remain in dfl-fme-pr.h.  The register #defines for this FME's
> >> FPGA manager device below should be associated with the FPGA manager
> >> driver.  Sorry if the way I stated that in the v3 review wasn't clear.
> >
> > Actually I put the PR sub feature register set definitions in this header
> > file (dfl-fme-pr.h), because it's possible the driver (dfl-fme-pr.c) of
> > this PR sub feature access some of the registers in the future. e.g read
> > some PR sub feature registers to create different fpga manager platform
> > devices as I mentioned above.
> 
> That sounds like a workaround.  Since you're adding a new method of
> enumeration, you should use that new method of enumeration to choose
> which FPGA manager is being used.  Otherwise we are ending up with
> multi-level enumeration, i.e. look at the DFL and then look at a
> specific register location in the device.
> 
> >
> > I have to say this is only future consideration, and in this dfl-fme-pr.c
> > driver there is no code to access below registers currently. I can move all
> > of them to dfl-fme-mgr.h or dfl-fme-mgr.c in the next version if this is
> > preferred. : )
> 
> That sounds good.  That makes the mgr driver its own separate thing
> which is what is supposed to happen in this framework.

Sure, will fix this.

Thanks
Hao

> 
> >
> >>
> >> > +/* FME Partial Reconfiguration Sub Feature Register Set */
> >> > +#define FME_PR_DFH             0x0
> >> > +#define FME_PR_CTRL            0x8
> >> > +#define FME_PR_STS             0x10
> >> > +#define FME_PR_DATA            0x18
> >> > +#define FME_PR_ERR             0x20
> >> > +#define FME_PR_INTFC_ID_H      0xA8
> >> > +#define FME_PR_INTFC_ID_L      0xB0
> >> > +
> >> > +/* FME PR Control Register Bitfield */
> >> > +#define FME_PR_CTRL_PR_RST     BIT(0)  /* Reset PR engine */
> >> > +#define FME_PR_CTRL_PR_RSTACK  BIT(4)  /* Ack for PR engine reset */
> >> > +#define FME_PR_CTRL_PR_RGN_ID  GENMASK_ULL(9, 7)       /* PR Region ID */
> >> > +#define FME_PR_CTRL_PR_START   BIT(12) /* Start to request for PR service */
> >> > +#define FME_PR_CTRL_PR_COMPLETE        BIT(13) /* PR data push complete notification */
> >> > +
> >> > +/* FME PR Status Register Bitfield */
> >> > +/* Number of available entries in HW queue inside the PR engine. */
> >> > +#define FME_PR_STS_PR_CREDIT   GENMASK_ULL(8, 0)
> >> > +#define FME_PR_STS_PR_STS      BIT(16) /* PR operation status */
> >> > +#define FME_PR_STS_PR_STS_IDLE 0
> >> > +#define FME_PR_STS_PR_CTRLR_STS        GENMASK_ULL(22, 20)     /* Controller status */
> >> > +#define FME_PR_STS_PR_HOST_STS GENMASK_ULL(27, 24)     /* PR host status */
> >> > +
> >> > +/* FME PR Data Register Bitfield */
> >> > +/* PR data from the raw-binary file. */
> >> > +#define FME_PR_DATA_PR_DATA_RAW        GENMASK_ULL(32, 0)
> >> > +
> >> > +/* FME PR Error Register */
> >> > +/* PR Operation errors detected. */
> >> > +#define FME_PR_ERR_OPERATION_ERR       BIT(0)
> >> > +/* CRC error detected. */
> >> > +#define FME_PR_ERR_CRC_ERR             BIT(1)
> >> > +/* Incompatible PR bitstream detected. */
> >> > +#define FME_PR_ERR_INCOMPATIBLE_BS     BIT(2)
> >> > +/* PR data push protocol violated. */
> >> > +#define FME_PR_ERR_PROTOCOL_ERR                BIT(3)
> >> > +/* PR data fifo overflow error detected */
> >> > +#define FME_PR_ERR_FIFO_OVERFLOW       BIT(4)
> >>
> >> This stuff is specific to this FPGA manager device, so it should
> >> either be in the dfl-fme-mgr.c or in a dfl-fme-mgr.h
> >
> > same as above, I can fix this in the next version.
> >
> > Thanks
> > Hao
> 
> Thanks,
> Alan

Powered by blists - more mailing lists